Introduction: When Infrastructure Begins to Understand Itself
On the morning of October 5, 2026, Schneider Electric, the French multinational whose business spans energy management, electrical distribution, industrial automation, and the software that governs much of the world’s critical infrastructure, announced a definitive agreement to acquire PTC, the Boston-based industrial engineering and product-lifecycle software company, in an all-cash transaction at $205 per share. The price valued PTC’s equity at approximately $22.6 billion, implied an enterprise value of roughly $23.7 billion, and represented a 42.3 percent premium to PTC’s last closing price before the announcement. The transaction, which both boards approved unanimously, is expected to close by the third quarter of 2027, subject to the approval of a majority of PTC’s shareholders and the receipt of regulatory clearances in the relevant jurisdictions. Schneider stated that it would fund the acquisition through a combination of an equity issuance of roughly €5 to €6 billion and new debt of roughly €16 to €17 billion, and it identified approximately €250 million of annual run-rate cost synergies by the third year and approximately €800 million of expected revenue synergies. [1][2]
The financial terms are large but, taken on their own, they would not justify a research paper. What makes the transaction consequential is the logic the two companies offered for it. PTC builds the software that manufacturers use to design, engineer, and manage complex physical products across computer-aided design and product-lifecycle management, serving more than 30,000 customers. Schneider already owns AVEVA, whose software describes how plants and energy systems operate, and in June 2026 it agreed to acquire Cognite, a Norwegian company specializing in the contextualization of industrial operational data, for approximately $3.1 billion in cash. The stated ambition of bringing these capabilities together is to connect the knowledge used to design industrial machinery with the information generated by that machinery throughout its operating life, so that engineering drawings, equipment configurations, maintenance histories, electricity consumption, manufacturing processes, and operational performance can become parts of one connected intelligence environment capable of supporting industrial AI agents. [1][3]
“The acquisition of PTC represents an important step forward in our ambition to lead the new era of Energy and Industrial Intelligence.”
— Olivier Blum, Chief Executive Officer, Schneider Electric [1]
Seventeen days later, on October 22, 2026, at NVIDIA GTC Berlin, Schneider Electric announced a second development, this time in collaboration with NVIDIA and Hewlett Packard Enterprise. The three companies introduced what they described as the industry’s first Industrial AI Agents at the Edge: software agents deployed on HPE Private Cloud AI infrastructure, built on NVIDIA Nemotron open models, and governed through Schneider’s EcoStruxure Industrial Insights applications, intended to monitor operational technology, correlate cybersecurity signals with process context, and identify energy-optimization opportunities inside process and heavy industries such as pulp and paper, chemicals, and energy. The announcement was explicit that the opportunity it was pursuing differed from the robotics and humanoid narratives that dominate most discussion of “physical AI,” because in a process plant the intelligence does not inhabit a machine that walks; it inhabits the plant itself. [4][5]
“Industrial AI Agents can monitor, reason and optimize operations in real time.”
— Sri Kodali, Global Head of Developer Relations, Energy and Utilities, NVIDIA [4]
These two Schneider announcements did not occur in isolation. On October 12, Microsoft said it would scale its Azure Maia accelerator family to multi-gigawatt deployments beginning in 2027 and introduced Maia 300, claiming 40 percent more tokens per megawatt than leading GPU-based systems, a framing that treats electricity rather than silicon as the binding unit of account. On October 14, NVIDIA announced that Applied Materials, Cadence, KLA, Lam Research, Siemens, Synopsys, and others were adopting its CUDA-X libraries to build what it called AI-defined fabs, and on the same day Honeywell introduced Forge Cognition, an agentic AI and digital-twin platform for buildings, campuses, and industrial facilities developed with NVIDIA and Dell Technologies, whose chief technology officer told Fortune that the platform’s autonomous operating mode had been deliberately left switched off for now. Two days after that, Schneider reported record third-quarter revenues of €11.46 billion, with triple-digit demand growth from datacenters, and raised its full-year outlook. [6][7][8][9][10]
Taken together, these developments point to something more specific than the familiar story that AI is spreading into every industry. They point to a change in what AI is being asked to do with respect to the physical foundation of the economy. For most of the past three years, the central question of the AI investment cycle has been how much energy, how many chips, how many datacenters, and how much capital would be required to train and serve increasingly capable models. That question has not gone away; the four largest hyperscalers guided in their July 2026 earnings to roughly $720 to $745 billion of combined capital expenditure for calendar 2026, and NVIDIA reported $89.0 billion of datacenter revenue in a single quarter. But a second question is now visible alongside the first: whether the intelligence being produced at the top of the stack can be turned back on the physical systems at the bottom of the stack, so that datacenters, grids, fabs, and factories become capable of understanding and improving their own operations. [11][12]
Why the Title “Infrastructure Cognition”?
I have chosen the title Infrastructure Cognition because it describes a transformation in the relationship between artificial intelligence and physical infrastructure that neither “digitalization” nor “automation” adequately captures. Traditionally, datacenters, power grids, manufacturing facilities, industrial machinery, and energy systems have been understood as productive physical assets whose performance depends on human engineers, operating procedures, monitoring equipment, and periodic maintenance. Digitalization added records and dashboards to those assets. Automation added the execution of predefined procedures. Cognition introduces a more demanding possibility: infrastructure equipped with AI systems that can interpret operating conditions, understand relationships among interconnected components, anticipate emerging problems, simulate potential interventions, express uncertainty, and recommend or undertake carefully authorized actions whose consequences can be verified.
In this paper, Infrastructure Cognition therefore refers to the continuously evolving body of machine-readable engineering, operational, environmental, commercial, and maintenance information — together with the AI models, simulations, knowledge structures, and decision systems capable of interpreting it — that enables intelligent understanding of industrial and infrastructure assets throughout their lifecycles. The definition deliberately places information before models, because the research reviewed in this paper, from Tao Fei’s 2024 assessment of industrial digital twins in Nature Computational Science to the May 2026 systematic survey of foundation-model agents in industrial automation, consistently finds that the limiting factor in industrial AI is not the sophistication of the model but the fidelity, provenance, and semantic coherence of the information the model is given about the physical world. [13][14]
The title is also appropriate because the next generation of investment in the Five-Layer AI Economy may extend well beyond purchasing GPUs, constructing hyperscale datacenters, and securing electricity supplies. Companies such as NVIDIA, Schneider Electric, Siemens, Honeywell, Microsoft, and the major hyperscalers are developing or acquiring technologies that connect physical infrastructure with engineering data, digital twins, industrial software, and autonomous agents. Strategic acquisitions could accelerate this integration by combining previously separated bodies of industrial knowledge. The central economic opportunity is not simply to create more sophisticated digital representations of factories, power plants, and datacenters. It is to convert connected industrial knowledge into a strategic productive asset whose value may become comparable to that of the equipment and facilities themselves — and, as this paper will argue at length, to do so without confusing the ability to explain a physical system with the authority to control it.
The Central Research Question
Modern industrial enterprises possess enormous quantities of information, yet much of that information remains isolated across technologies, departments, vendors, and operational systems that were never designed to interpret one another. A manufacturer may maintain engineering drawings within one platform, equipment telemetry within another, energy consumption records in a separate monitoring environment, maintenance schedules in enterprise software, and manufacturing performance data within specialized industrial control systems. Each system may function effectively within its own responsibilities. None necessarily understands how changes in one system affect the others, and the people who could reconstruct those relationships are often the scarcest resource in the plant.
Consider an industrial facility experiencing an unexpected increase in electricity consumption. Its energy-management software identifies the anomaly. Maintenance software shows that a particular motor is approaching its recommended replacement interval. Engineering records reveal that the motor’s original specifications assumed different operating conditions. Production scheduling software shows that output recently increased, while environmental monitoring equipment records an unusual temperature pattern in the hall where the motor is installed. Individually, these observations offer partial explanations, and each department will tend to favor the explanation that its own data supports. When connected through a reliable, contextualized information environment, however, they could provide a much more complete understanding of what is actually happening. An AI system with access to all of them could potentially distinguish among excessive mechanical wear, inefficient operating schedules, abnormal cooling requirements, a defective sensor, or a mismatch between equipment configuration and current production demands. It could then evaluate alternatives, simulate their effects, recommend a maintenance intervention, and help determine afterwards whether the resulting action improved efficiency without compromising safety or production reliability.
Traditional industrial monitoring tells an operator what is happening. Infrastructure Cognition seeks to understand why it is happening, what could happen next, and which interventions would produce better outcomes. The important distinction — and it will recur throughout this paper — is that understanding a physical system is not the same as having unrestricted authority to control it. A credible cognitive infrastructure architecture must preserve engineering limits, independent safety protection, human accountability, and an auditable boundary between observation, recommendation, approval, and actuation. Daniela Rus, who directs MIT’s Computer Science and Artificial Intelligence Laboratory and whose definition of physical AI explicitly includes power grids, sensors, and satellites alongside robots, has put the engineering constraint more bluntly than most industry marketing will admit. [15][16]
“In the physical world, we cannot have mistakes and hallucinations.”
— Daniela Rus, Director, MIT Computer Science and Artificial Intelligence Laboratory [15]
This paper examines five related questions. First, why does industrial information remain fragmented despite decades of digital investment, and what would it take to make it coherent? Second, why is industrial knowledge becoming a strategic acquisition target, and how should investors evaluate the premiums being paid for it? Third, when can AI agents responsibly participate in industrial operations, and what architecture separates legitimate cognition from unaccountable automation? Fourth, how might these changes affect regional and geopolitical competition, including the state-level politics of datacenter growth in the United States? Fifth, how can the success of intelligent infrastructure be measured by 2030 in a way that distinguishes verified productivity from announced capacity?
The underlying analytical framework is the Five-Layer AI Economy: Layer 1 Energy, Layer 2 Chips, Layer 3 Datacenters, Layer 4 Models, and Layer 5 Applications and Agents. Infrastructure Cognition introduces a feedback relationship into that framework, in which intelligence produced at the upper layers improves the systems supporting all five. The remainder of the paper proceeds from the information problem (Section 1), to the acquisition strategy and the hyperscaler capital cycle (Section 2), to the architecture of bounded autonomy (Section 3), to the geopolitical and regional consequences (Section 4), to measurement and capital allocation for 2030 (Section 5), and finally to seven pillars that summarize what has been learned (Section 6).
Section 1: The Industrial Information Problem: Why Physical Assets Remain Digitally Fragmented
1.1. The Difference Between Data Collection and Industrial Understanding
Over several decades, industrial enterprises invested heavily in computerized maintenance management, supervisory control and data acquisition, equipment condition monitoring, engineering design tools, enterprise resource planning, manufacturing execution systems, and industrial networks. The results have been impressive within particular functions. Operators can observe alarms at remarkable granularity. Engineers can test components in simulation before manufacturing them. Finance departments can allocate equipment expenditures to cost centers. Maintenance crews can trace spare parts, repair orders, and service contracts across years of operating history. By any reasonable standard, the modern industrial facility is drowning in data rather than starving for it.
Yet an abundance of digital information does not guarantee a coherent explanation of how a facility works. Industrial data is typically organized according to the software application, contractor, asset owner, or department that created it, not according to the physical object it describes. An engineering model may identify an asset by a design-stage tag assigned by the original equipment manufacturer. The facility’s control system may use a different identifier assigned during commissioning. A maintenance database may record a replacement part under a third label chosen by the procurement system. Financial records may group several physical machines into a single capitalized asset for depreciation purposes. These naming differences are not merely an administrative nuisance to be resolved by a data-cleansing project. They create an analytical failure: the organization cannot reliably determine which observations belong to the same asset at a specific moment in its operating life. Even when all records are technically accessible, they may not be mutually interpretable.
The scale of this problem is visible in the operational-technology security literature, which has long treated the asset inventory as the foundation of every subsequent control. A 2026 industry survey cited by SecurityWeek found that only 21 percent of industrial security leaders reported possessing a complete OT asset inventory, even as 88 percent described their programs as mature — a gap that says as much about industrial information as it does about cybersecurity. If an organization does not reliably know what equipment it operates, it cannot expect an AI system to reason correctly about that equipment’s behavior. [17]
The relevant question is therefore not how many terabytes of data the industrial enterprise possesses. It is whether its systems can establish trustworthy relationships among physical objects, engineering specifications, operating events, decisions, and consequences. That requires a common — or at least interoperable — semantic context: what a component is, where it is situated, what it connects to, how it is supposed to behave, which records describe it, and how confident the organization is in each of those answers. An industrial AI model trained on generic technical language cannot automatically infer these site-specific facts. Its useful intelligence depends on access to validated engineering context and sufficiently current operating conditions. Without both, seemingly persuasive recommendations may be mechanically impossible, economically misleading, or unsafe, and the persuasiveness itself becomes a hazard because it invites trust the recommendation has not earned.
1.2. Four Information Environments That Rarely Form a Single System
Industrial fragmentation is easiest to understand by separating four information environments that coexist in almost every facility but rarely form a single system. The first is the design environment, which contains computer-aided design files, engineering calculations, equipment specifications, piping and instrumentation diagrams, software requirements, and the original assumptions about performance. Design information describes the intended asset, including its operating limits and its dependencies on other equipment. This is the environment in which PTC has built its business, and it is the environment Schneider Electric is paying a 42 percent premium to acquire. [1]
The second is the operational environment, which contains sensor readings, control-system states, event logs, alarms, production rates, power consumption, temperatures, pressures, and observed physical behavior. It describes the asset under actual conditions, which may depart substantially from design assumptions after many years of use, modification, and wear. This is the environment of AVEVA, of Schneider’s EcoStruxure platform, and of the supervisory control systems that Siemens, Honeywell, Emerson, Rockwell, and others have installed in hundreds of thousands of sites.
The third is the maintenance environment, which records inspections, incidents, repairs, spare-part substitutions, technician observations, warranty claims, and remaining service-life estimates. It shows how a physical asset has changed through intervention and degradation, and it is frequently the environment with the richest tacit knowledge and the poorest data structure, because much of it lives in free-text work orders written by technicians under time pressure. The fourth is the commercial and governance environment, which records energy contracts, insurance requirements, customer service obligations, capital budgets, procurement commitments, operating permits, cybersecurity rules, and the allocation of human decision authority. It establishes which technically possible actions are economically rational or institutionally permitted.
Table 1. The four industrial information environments and what each one knows
| Environment | Representative records | What it describes | Typical custodian | Characteristic failure |
| Design | CAD, PLM, P&IDs, specifications, design calculations | The asset as intended, with limits and dependencies | Engineering, OEMs, PLM vendors (e.g., PTC) | Assumptions silently outdated by field changes |
| Operational | Telemetry, alarms, control states, production and power data | The asset as it behaves now | Operations, SCADA/DCS vendors (e.g., AVEVA, Siemens, Honeywell) | Identifiers differ from design tags; sensor drift |
| Maintenance | Work orders, inspections, repairs, spares, warranty claims | How the asset has changed through wear and intervention | Maintenance, CMMS/EAM systems | Tacit knowledge trapped in free text |
| Commercial and governance | Energy contracts, permits, SLAs, budgets, cyber rules, authority matrices | Which actions are permitted and economically rational | Finance, legal, compliance, management | Constraints invisible to engineering tools |
An intelligent system must cross all four environments. Suppose a datacenter considers reducing pump speed in a cooling loop to save electricity. A local control algorithm may know the pump’s current flow and the water temperature, but not the warranty restriction recorded in the engineering file that limits sustained operation below a certain speed. A model may estimate energy savings with great precision, but not recognize that a redundant cooling loop is scheduled for maintenance that afternoon, leaving the facility with less resilience than the model assumes. A finance system may value the savings accurately, but not account for the service-level penalties that apply if cooling resilience falls below a contractually agreed threshold. In that situation, data is abundant in every environment, but understanding is incomplete in all of them. Infrastructure Cognition requires connections that allow a proposed decision to be evaluated against physical, operational, economic, and contractual realities simultaneously, and it requires the system to know when one of those realities is unknown.
1.3. The Industrial Asset as an Evolving Body of Knowledge
A physical asset does not remain identical to its original blueprint. Motors are rewound; transformers age; firmware is updated; valves are replaced with near-equivalents from different suppliers; cooling systems are reconfigured; production lines are adjusted to new products; emergency operating procedures change after incidents. Consequently, infrastructure intelligence cannot be a static digital archive created once during commissioning and consulted thereafter. It must be a versioned, continuously maintained knowledge environment in which each critical asset has a stable identity, documented relationships to neighboring assets, design constraints, a record of changes, time-stamped observations, a maintenance history, and an explicit statement of known uncertainties. When an asset is altered, its digital description must change in a controlled manner. When a sensor fails, the system must preserve the distinction between missing information and an actual physical change, because an AI model that cannot tell those two conditions apart will eventually act on an absence as if it were a presence.
The academic concept most often invoked for this purpose is the digital twin. Michael Grieves introduced the idea in a 2003 product-lifecycle-management course at the University of Michigan and formalized it with John Vickers in 2017 as a virtual representation that spans the asset’s entire lifecycle, from pre-construction to post-disposal. Fei Tao and colleagues at Beihang University, whose 2019 state-of-the-art review in IEEE Transactions on Industrial Informatics became one of the most cited papers in the field, extended the concept into a five-dimensional model encompassing the physical entity, the virtual entity, services, data, and the connections among them. Tao’s 2024 perspective in Nature Computational Science was notably more sober than much of the commercial literature, describing digital twins as having attracted a substantial amount of attention while cautioning that their maturity for large-scale industrial application depended on avoiding a specific set of pitfalls, including models that are visually impressive but physically unvalidated, data that is collected but never contextualized, and twins that are built once and never kept current. [13][18][19]
“Highlight the main pitfalls to avoid and challenges to overcome.”
— Fei Tao, Beihang University, writing in Nature Computational Science [13]
The empirical literature supports that caution. A 2025 Winter Simulation Conference study by Wuttke, Onggo, and Rabe systematically reviewed 124 publications from 2021 to 2024 on digital twin implementation for simulation-based industrial applications and identified 33 distinct challenges across nine dimensions, concluding that despite their potential, implementation remains complex and adoption in actual industrial operations remains limited. A 2025 review in the Journal of Intelligent Construction surveying aerospace, manufacturing, construction, and power applications reached a similar conclusion, identifying data management, model complexity, cybersecurity, and standardization as the four persistent obstacles. These findings matter for the acquisition thesis discussed in Section 2, because they suggest that the value Schneider Electric is purchasing in PTC and Cognite lies less in any particular software product than in the possibility of solving the contextualization problem at scale. [20][21]
The essential discipline, which I will call configuration fidelity, is the degree to which the information used for a recommendation reflects the asset that actually exists at the moment the recommendation is made. A cognitive system whose understanding is outdated becomes a potential hazard precisely because its conclusions can appear sophisticated and authoritative. This creates a new investment category for industrial enterprises: continuously validating the meaning, provenance, and physical correspondence of operational data. The expenditure is less visually impressive than a new robot or a new datacenter hall, but it may determine whether those assets deliver the productivity their purchasers were promised.
1.4. The Five-Layer AI Economy Meets the Industrial Operating Environment
The Five-Layer AI Economy provides a structure for locating both the source and the destination of infrastructure intelligence. Layer 1, Energy, comprises power generation, transmission, substations, transformers, power quality, backup resources, electricity contracts, and grid conditions, and it establishes the physical envelope within which computational and industrial systems operate. Layer 2, Chips, comprises GPUs, CPUs, industrial controllers, network processors, accelerators, and specialized edge devices that perform the calculations needed for inference, simulation, optimization, and physical control. Layer 3, Datacenters, comprises hyperscale campuses, enterprise clusters, local inference servers, and edge installations that provide different locations for storing, processing, and securing industrial information. Layer 4, Models, comprises language models, forecasting models, physical simulators, anomaly detectors, control models, and domain-specific engineering systems that interpret information and estimate possible outcomes. Layer 5, Applications and Agents, comprises asset-health applications, energy optimizers, maintenance coordinators, security analysts, and engineering agents that translate model outputs into workflows that people can inspect, approve, or authorize.
Table 2. The Five-Layer AI Economy and the Infrastructure Cognition feedback loop
| Layer | Core assets | Upward contribution (traditional) | Downward contribution (Infrastructure Cognition) | October 2026 example |
| 1. Energy | Generation, transmission, substations, contracts | Powers everything above | Receives load forecasts, flexibility signals, power-quality diagnostics | DOE 2026 Draft Transmission Needs Study; Texas ERCOT audit |
| 2. Chips | GPUs, accelerators, controllers, edge devices | Compute for models | Fab yield improved by AI simulation and inspection | NVIDIA AI-defined fabs; Microsoft Maia 300 (40% more tokens/MW claim) |
| 3. Datacenters | Hyperscale campuses, edge sites | Hosts models and data | Cooling, power, and maintenance optimized by facility-wide models | Microsoft “grid to chip” architecture; Honeywell Forge Cognition pilots |
| 4. Models | LLMs, simulators, anomaly detectors | Interprets information | Validated against physical constraints and configuration fidelity | NVIDIA Nemotron, Omniverse, CUDA-X libraries |
| 5. Applications and Agents | Asset-health, energy, security, engineering agents | Delivers intelligence to workflows | Recommends or executes bounded actions, verifies outcomes | Schneider–NVIDIA–HPE Industrial AI Agents at the Edge |
The traditional direction of causation moves upward: electricity supports chips, chips support datacenters, datacenters support models, and models support applications. Infrastructure Cognition adds a loop back down. An application could identify deteriorating power quality in a substation feed, which changes operational decisions in the datacenter. A simulation could improve cooling efficiency, thereby reducing the electricity needed to support the very models running the simulation. An engineering agent could help redesign components in the next generation of industrial hardware. NVIDIA’s October 14 announcement that TSMC, Siemens, Lam Research, and others are using its computational lithography, molecular-dynamics, and inspection libraries inside fabs is the clearest current example of Layer 4 improving Layer 2. [7][22]
However, no layer can substitute for another. Better software cannot manufacture a missing transformer, create transmission rights, or repair a failed cooling tower without human and physical execution. The International Energy Agency’s executive director has made the point from the energy side with a phrase that has become a slogan, and it bears repeating here because the entire argument of this paper depends on it: cognition can increase the usefulness of scarce infrastructure, but it cannot abolish the material constraints of that infrastructure. [23]
“There is no AI without energy.”
— Fatih Birol, Executive Director, International Energy Agency [23]
1.5. From Datacenter Expansion to Datacenter Understanding
The urgency of the industrial information problem is intensified by the speed of AI infrastructure construction. A hyperscale AI facility contains interconnected systems for power distribution, liquid cooling, thermal control, networking, cybersecurity, access management, and computing workloads, and its operational state changes continuously as models, chips, racks, and maintenance requirements evolve. A high-density accelerator cluster cannot be evaluated solely by the price of its processors. Its performance depends on power delivery, cooling capacity, rack-level design, water strategy, network architecture, outage tolerance, and regional electricity conditions, and these dependencies have tightened with every hardware generation. Microsoft’s October 13 technical discussion of Azure infrastructure described power availability and electrical efficiency as system-level concerns extending from the grid through power conversion, racks, servers, and silicon, which is to say that the company now designs the datacenter as a single energy-to-compute chain rather than as a building that happens to contain servers. [24]
The demand side of that chain has become a matter of national planning. The Lawrence Berkeley National Laboratory’s December 2024 report for the U.S. Department of Energy found that datacenters consumed about 4.4 percent of U.S. electricity in 2023 and projected that they would consume between 6.7 and 12 percent by 2028, a range of roughly 325 to 580 terawatt-hours. The IEA’s April 2025 Energy and AI report projected that global datacenter consumption would reach approximately 945 terawatt-hours by 2030, slightly more than Japan’s total consumption, and its April 2026 follow-up, Key Questions on Energy and AI, found that datacenter electricity demand had grown 17 percent in 2025 — more than five times the 3 percent growth in overall global demand — with AI-focused facilities growing by roughly 50 percent. The IMF’s Power Hungry working paper, drawing on OPEC estimates, suggested datacenter consumption could triple to around 1,500 terawatt-hours by 2030, comparable to India’s current consumption, and warned that the uncertainty of the forecast itself could delay energy investment and raise prices. [76] Stanford’s 2026 AI Index, which added an infrastructure chapter for the first time, recorded AI datacenter power capacity of 29.6 gigawatts, a figure its authors compared to New York State’s peak demand. [25][26][27][28][29]
“AI needs more and more electricity for the data centers that make it possible.”
— International Monetary Fund, IMF Blog, May 2025 [28]
Against this backdrop, the U.S. Department of Energy’s 2026 Draft National Transmission Needs Study, released July 9 for public comment, identified load growth from datacenters, expanding domestic manufacturing, and large industrial loads as the principal force behind new transmission investment, synthesizing more than 120 industry studies across 20 regions. The assistant secretary for the Office of Electricity described demand as accelerating faster than anything seen in decades. The immediate research question for this paper is therefore whether intelligent operation can improve reliability, utilization, maintenance, and energy performance fast enough to contribute meaningfully during the period when new substations, power plants, and transmission lines are still being permitted and built. This is not a substitute for grid investment. It is a reason to evaluate the productivity of existing equipment alongside the volume of new equipment purchased. [30]
“Electricity demand is accelerating faster than anything we’ve seen in decades.”
— Catherine Jereza, Assistant Secretary, Office of Electricity, U.S. Department of Energy [30]
1.6. What Cognition Means — and What It Does Not Mean
The language of cognition invites anthropomorphic interpretation, and it is worth being precise about what is and is not being claimed. A datacenter does not become conscious because it contains advanced monitoring systems. A substation does not acquire intentions because a forecasting model analyzes its behavior. A factory does not possess independent judgment simply because software can generate a maintenance plan. In this paper, cognition is a functional and engineering description, not a biological or philosophical claim. The term identifies a specific combination of capabilities: perceiving relevant conditions, representing relationships, maintaining contextual information, evaluating alternatives, recognizing uncertainty, and contributing to decisions whose consequences can be verified. These capabilities may be implemented through many technologies — traditional statistical models, physics-based simulators, specialized neural networks, knowledge graphs, large language models, optimization algorithms, or carefully designed software agents — and the choice among them is an engineering decision, not an ideological one.
The critical distinction is between information retrieval and causal adequacy. An agent may retrieve a technical manual and correctly summarize it without knowing whether the manual applies to the currently installed equipment. It may recognize that two events occurred together without knowing whether one caused the other. It may produce a plausible explanation of a transformer alarm without identifying a measurement error in the upstream sensor. Stanford’s 2026 AI Index documented what its authors called the jagged frontier of AI capability, in which the same models that exceed human baselines on doctoral-level science questions read an analog clock correctly only about half the time; the industrial analogue of that jaggedness is a model that can summarize a thousand pages of maintenance history fluently while failing to notice that the pump it is reasoning about was replaced last Tuesday. Fei-Fei Li of Stanford has argued that the gap between language-based AI and physical reality has become the field’s most urgent problem, and that closing it requires models built around spatial and physical reasoning rather than language alone. [29][31]
“Turn seeing into reasoning, perception into action, and imagination into creation.”
— Fei-Fei Li, Professor of Computer Science, Stanford University, on spatial intelligence [31]
Industrial understanding therefore requires multiple kinds of evidence. A model must know which records are authoritative, which measurements are current, whether the relevant system is within the conditions for which its model was validated, and which physical relationships are sufficiently well established to support action. Where causal understanding is weak, the system should express uncertainty and require additional testing. This suggests a useful taxonomy that will structure the rest of the paper. Descriptive cognition identifies what exists and what is happening. Diagnostic cognition evaluates why the condition may have arisen. Predictive cognition estimates future states under specified assumptions. Prescriptive cognition compares interventions. Operational cognition connects the analysis to an authorized workflow and verifies the physical result. The categories differ sharply in their requirements and risks; a facility can have excellent descriptive cognition without being remotely ready for operational cognition, and the marketing of industrial AI frequently blurs that distinction.
Perhaps the strongest evidence of useful intelligence is not a sophisticated explanation but an ability to say that available information is insufficient. If an agent cannot identify the calibration date of a critical sensor, it should avoid treating the reading as certain. If engineering documentation conflicts with a maintenance report, it should flag the disagreement rather than silently choosing one. If the operating regime lies outside the model’s validated domain, it should not imply that a numerical prediction is reliable. The industrial enterprise of 2030 will need mechanisms that preserve these distinctions. Marketing language may describe systems as thinking or reasoning; engineering governance must translate such claims into measurable functions, bounded permissions, and known limitations. Infrastructure Cognition is credible only when its demonstrated capabilities are narrower than, or equal to, the authority entrusted to it. That definition keeps this paper distinct from a general discussion of digital transformation: industrial digitalization can succeed at automating records and dashboards without creating any integrated understanding of physical assets, and Infrastructure Cognition is the additional, harder capacity to derive and validate operational meaning from connected information.
Section 2: The Acquisition Strategy Behind Industrial Intelligence
2.1. Schneider Electric–PTC: Purchasing Access to Engineering Context
The October 5 Schneider Electric–PTC agreement offers a revealing illustration of how industrial information is being revalued, and it deserves to be read as a statement about the economics of knowledge rather than merely as the largest acquisition in Schneider’s history. Schneider already possesses a substantial position in power systems, energy management, electrical distribution, automation, and industrial operations; its 2026 half-year results showed Energy Management growing 18 percent organically and Industrial Automation accelerating to 11 percent, with AVEVA’s annualized recurring revenue up 11 percent year over year. PTC brings established expertise in product engineering, computer-aided design, product-lifecycle management, and the associated engineering records of more than 30,000 customers. Bringing these capabilities together could connect the logic of how products are designed with the operational evidence of how they perform in physical environments — the design and operational environments of Table 1, which have historically belonged to different vendors, different budgets, and different professional cultures. [1][32]
The announced transaction has an approximately $22.6 billion equity value and a $23.7 billion implied enterprise value. The distinction matters for anyone attempting to evaluate the price. Equity value measures what shareholders receive; enterprise value incorporates net debt and is the figure properly compared with operating earnings. Schneider’s strategic argument includes a larger software-and-services business, cross-selling into its installed base, and joint development around what both companies call an industrial digital thread. None of those projected benefits should be treated as realized. They depend on regulatory approval, transaction closing, technical integration, customer adoption, and execution over a period that will extend well past the anticipated third-quarter 2027 close. The deal is also a capital-allocation statement: on October 20, Schneider announced the successful completion of a €6.0 billion capital increase, conducted without shareholders’ preferential subscription rights, to partially finance the acquisition. A company that raises six billion euros of fresh equity to buy an engineering-software firm is telling the market that it believes the commercial value of engineering context will become central to the sale and operation of energy and automation infrastructure. [2][33]
The reasoning is coherent even if the outcome is uncertain. An owner of engineering information may occupy a privileged position in the future AI workflow. Before an agent can recommend changing a turbine setting, it must know which turbine exists, what equipment has been modified since commissioning, what constraints apply, and how the change affects related systems. The data required to answer those questions is distributed among precisely the software categories being consolidated. PTC’s chief executive, Neil Barua, framed the combination in terms of scale and resources to accelerate innovation and reach new geographies and end markets; the more interesting framing is the one implicit in the deal structure, which treats the design record as the missing upstream half of an industrial intelligence system whose downstream half Schneider already owns. [1]
2.2. Cognite and AVEVA: The Value of Contextualization
The proposed PTC acquisition follows Schneider’s June 30, 2026 agreement to acquire Cognite for approximately $3.1 billion in cash. Cognite specializes in connecting industrial operational information through contextualized data models, knowledge graphs, and agent-oriented software, and Schneider identified the prospective combination with AVEVA as a foundation for industrial AI systems that could connect engineering and operational information. That acquisition also remained subject to closing conditions when announced, and Schneider’s half-year report in July described both Cognite and AiDASH — a smaller acquisition focused on AI for utility vegetation management and climate resilience — as proposed rather than completed. [3][32]
The strategic sequence is important and, in my judgment, more deliberate than the market initially appreciated. AVEVA supplies industrial software used in design and operations. Cognite offers methods for linking operational information and giving it usable context — the semantic layer that Section 1 identified as the binding constraint. PTC provides deeper access to product and engineering lifecycles. Taken together, these capabilities could support a chain from initial design, to commissioning, to operation, to maintenance, to redesign, with each stage’s records interpretable by the next. This does not mean a single company would own every industrial fact in the market. Customers retain multiple platforms, local systems, vendor contracts, and regulatory obligations, and the digital-twin literature reviewed in Section 1 makes clear how far most facilities remain from any integrated representation. Nevertheless, an integrated portfolio could become a powerful intermediary through which industrial organizations discover, structure, and interpret information, and the competitive advantage might reside less in any single model than in the ongoing ability to maintain relationships among assets, records, processes, and decisions over decades.
There are two possible outcomes, and procurement officers should understand both before signing. In an open architecture, data from different vendors can be exchanged with documented meanings and customer-controlled permissions, and the consolidated platform competes on the quality of its contextualization. In a closed architecture, integration becomes a means of making competitors expensive to replace, and the engineering record — which the customer created and paid for — becomes a source of leverage against the customer. Schneider’s own language emphasizes open and diverse ecosystems, and the company’s commercial history with AVEVA suggests it understands that industrial customers punish lock-in over long horizons. But procurement decisions made between 2027 and 2030 may determine not merely software costs but the long-term bargaining power of manufacturers and utilities, and the contractual framework proposed in Section 5.2 is written with that risk in mind. [32]
Table 3. The 2026 industrial-intelligence acquisition and partnership map
| Acquirer or lead | Target or partner | Date (2026) | Announced value | What it adds to the cognitive stack | Status at October 29, 2026 |
| Schneider Electric | PTC | Oct 5 | ~$22.6B equity / $23.7B EV | Design environment: CAD, PLM, product records | Signed; close expected by Q3 2027 |
| Schneider Electric | Cognite | Jun 30 | ~$3.1B cash | Contextualization: knowledge graphs, industrial data models, agents | Signed; subject to closing |
| Schneider Electric | NVIDIA, HPE | Oct 22 | Partnership | Edge agents for OT monitoring, cybersecurity, energy optimization | Announced at GTC Berlin |
| Siemens | NVIDIA | Jan 6 | Partnership | Industrial AI operating system, Digital Twin Composer, adaptive factories | Expanding; Erlangen blueprint |
| Honeywell | NVIDIA, Dell | Oct 14 | Partnership | Forge Cognition: agents and twins for buildings and industrial sites | Launched; pilots at six sites |
| NVIDIA | Applied Materials, Cadence, KLA, Lam, Siemens, Synopsys, TSMC | Oct 14 | Ecosystem | CUDA-X libraries for AI-defined fabs; cuLitho, cuPotential | Announced; vendor-reported speedups |
| Microsoft | Siemens, NVIDIA, KUKA | Apr 16 | Partnership | Azure IoT Operations, Fabric, industrial demonstration at Hannover | Demonstrated |
2.3. Siemens and NVIDIA: Simulation Becomes an Operational Instrument
Schneider Electric is not alone in pursuing this transformation, and its principal European rival has arguably been pursuing it longer. In January 2026, at CES, Siemens and NVIDIA expanded their partnership around what Siemens describes as an industrial AI operating system and AI-driven, adaptive manufacturing. Siemens identified its own Erlangen electronics facility as an initial blueprint and introduced its Digital Twin Composer, which combines industrial engineering information with NVIDIA Omniverse simulation capabilities. Siemens also reported early performance outcomes from digital-twin work with PepsiCo, including throughput and design-validation improvements, which PepsiCo described as an industry-first AI and digital-twin collaboration. Such figures are company-reported results for defined deployments, useful as evidence of the kinds of value companies are testing but not a universal return-on-investment benchmark. [34][35]
An important distinction must be preserved here between visual representation and operational reasoning, because the two are frequently conflated in demonstrations. A visually accurate virtual factory can help a planner understand spatial relationships and communicate a design to stakeholders. A decision to alter a production process, however, requires reliable physics, valid operating assumptions, current configurations, and a means of determining whether the simulated result transfers to the real facility. Siemens’ commercial results suggest that customers are willing to pay for the second kind of capability: in its third quarter of fiscal 2026, ended June 30, the company reported record orders of €27.9 billion, up 14 percent on a comparable basis, with Digital Industries’ profit margin rising to 18.7 percent from 14.5 percent a year earlier, its software business making the largest contribution to the improvement, and its digital business growing 18 percent over nine months against a 15 percent target. Chief executive Roland Busch described it as another record third quarter. [36]
NVIDIA’s October 14 announcement concerning AI-defined fabs extends this logic into the manufacturing facilities that produce the chips used by AI systems. Its expanded CUDA-X portfolio — including cuPotential for atomistic simulation, NanoVDB for three-dimensional etch modeling, and NPP for inspection imaging — supports simulation, inspection, and engineering workflows across design, manufacturing preparation, inspection and test, and fab operations. NVIDIA reported early-access performance measurements ranging from roughly 8x at Cadence to approximately 85x at Teradyne, while describing partnerships with Cadence and Synopsys and, separately, with TSMC, which is applying cuLitho and GPU-accelerated scheduling inside its own fabs. These are vendor-reported figures for specific libraries rather than demonstrated changes in saleable wafer output, and the trade press was careful to note that no uniform operational outcome was announced. Even so, they illustrate a feedback pattern in which AI infrastructure supports new industrial engineering capability, which may eventually improve the fabrication of future AI chips — Layer 4 improving Layer 2 in the vocabulary of this paper. [7][22]
2.4. Microsoft, Honeywell, and Competing Architecture Models
Microsoft’s industrial architecture provides another approach, one that begins from the cloud and data platform rather than from the control system. Azure IoT Operations and related tools connect devices and equipment at the edge with broader data systems such as Microsoft Fabric, and the company’s April 2026 demonstration at Hannover Messe connected design, simulation, factory execution, live telemetry, and AI analysis with partners including Siemens, NVIDIA, and KUKA. Microsoft describes industrial data as a foundation for agent-assisted detection and action, which is a reasonable description of the Layer 5 ambition, though it leaves open the question — central to Section 4 — of where that data should reside and who should govern it. [37][38]
Honeywell’s October 14 launch of Forge Cognition offers a different commercial entry point: the operation of buildings, campuses, and existing industrial facilities. Developed with NVIDIA’s Nemotron models and Omniverse libraries and running on Dell’s AI Factory infrastructure, the platform combines real-time digital twins, a structured knowledge base that encodes engineering expertise alongside equipment data, and AI agents for alarm interpretation, predictive asset health, and energy management. Honeywell described manual, semi-autonomous, and autonomous operating modes with operator override, and it reported pilot results from NVIDIA’s 750,000-square-foot Voyager headquarters, the Fairmont Hotel Dubai, NEX Mall Singapore, Vanderbilt University, and Honeywell campuses in Bangalore and Richardson, including up to a 90 percent reduction in unnecessary security alerts and up to 10 percent potential energy savings at Voyager. The company is positioning the product against an installed base of more than 300,000 sites. [8][40]
“Intelligent systems that can reason, learn and act in the context of real-world operations.”
— John Roese, Global Chief Technology and AI Officer, Dell Technologies [8]
What makes Honeywell’s launch analytically significant is less the pilot figures than the governance posture. Suresh Venkatarayalu, Honeywell’s chief technology officer, told Fortune ahead of the launch that the company had not opened up autonomy and that humans would remain in control for probably the next two to three years. In a market in which the word “autonomous” is used freely, a vendor that ships an autonomous mode and then publicly declines to enable it is making a statement about the engineering safety case that Section 3 develops in detail: the commercial success of physical AI will depend on customers’ ability to set enforceable limits on decision authority, and the vendors most likely to be trusted may be those who impose such limits on themselves first. [9]
“We have not opened up an autonomy right now.”
— Suresh Venkatarayalu, Chief Technology Officer, Honeywell Technologies [9]
The competitive landscape is therefore not a simple contest between one industrial software platform and another. It includes automation vendors, energy specialists, cloud providers, chip manufacturers, equipment makers, simulation platforms, and independent industrial software developers, and each approaches the same customer from a different point of leverage. The engineering platform knows what the asset was designed to do. The automation supplier knows what the control system is doing. The cloud provider manages large-scale storage, analytics, and model services. The accelerator supplier provides computation and simulation libraries. The maintenance platform knows the history of interventions and failures. The asset owner retains responsibility for production, safety, compliance, and financial outcomes. ARC Advisory Group’s assessment of the Schneider–NVIDIA–HPE agents captured the customer’s perspective precisely: the value of agent-based AI will depend on its ability to work with contextualized operational data, fit within established governance structures, and support rather than compromise critical operational decision-making. The winning architecture may be the one capable of preserving these distinct competencies while making the information between them intelligible and actionable. [39]
2.5. The New Economic Logic of Industrial Technology Acquisitions
Traditional industrial acquisitions emphasize geographic reach, installed equipment, service revenue, factory capacity, or product synergies. Infrastructure Cognition introduces a further category: the value of combining engineering and operational information that becomes more useful when interpreted jointly than when held separately. In economic terms, the acquirer is paying for a complementarity between two bodies of knowledge, and the acquisition premium is justified only if that complementarity can be realized in customer operations rather than merely asserted in investor presentations.
Here the macroeconomic debate about AI productivity becomes directly relevant to industrial strategy, because the two most influential academic positions in that debate make opposite predictions about exactly the kind of investment Schneider and Siemens are undertaking. Erik Brynjolfsson of Stanford’s Digital Economy Lab, drawing on his 2021 work with Daniel Rock and Chad Syverson on the productivity J-curve, has argued that general-purpose technologies initially depress measured productivity because firms must invest in complementary intangible assets — redesigned processes, new data structures, retrained workers — before the gains appear, and that the United States began transitioning from the investment phase to the payoff phase in 2025. In a February 2026 Financial Times op-ed he pointed to Bureau of Labor Statistics data as evidence that the long-awaited take-off was becoming visible. Daron Acemoglu of MIT, by contrast, has argued in The Simple Macroeconomics of AI that so long as AI’s effects operate through task-level cost savings, the aggregate gains are bounded by a version of Hulten’s theorem and are likely to be nontrivial but modest — on the order of 0.66 percent of total factor productivity over ten years in his baseline, and less when hard-to-learn tasks are properly accounted for. [44][45][46]
“The fog may finally be lifting.”
— Erik Brynjolfsson, Director, Stanford Digital Economy Lab, on AI and productivity data [44]
“Nontrivial but modest.”
— Daron Acemoglu, Institute Professor, MIT, on AI’s ten-year total-factor-productivity effect [46]
The industrial acquisitions examined in this paper can be read as bets on Brynjolfsson’s mechanism over Acemoglu’s bound. The complementary intangible assets that the J-curve requires — contextualized data, reconciled asset identities, validated digital twins, redesigned maintenance workflows — are precisely what Cognite, PTC, and Forge Cognition are supposed to supply. If those investments allow AI to do new things in industrial operations, such as preventing outages that would otherwise have occurred or extending equipment life in ways that defer capital expenditure, then the gains escape the task-substitution logic that constrains Acemoglu’s estimate. If, instead, the acquired software merely automates existing reporting tasks a little faster, Acemoglu’s bound will hold and the premiums will prove expensive. The honest position for an investor is that both outcomes remain possible and that the evidence will arrive site by site, which is why Section 5 insists on measurement rather than narrative.
Integration also imposes real costs that acquisition models tend to understate. Legacy software interfaces may be incompatible. Customers may resist transferring sensitive engineering data to a vendor that also serves their competitors. Existing contracts may limit data reuse. Platform consolidation can create cybersecurity exposure by connecting systems that were previously, and sometimes deliberately, separated. Engineers may distrust an agent whose reasoning is not traceable to a validated source. Investors should consequently demand evidence on more than annual recurring revenue and software cross-selling. A credible acquisition thesis should explain the operational mechanism: which previously separated data types will be connected, which decisions will improve, what measurable benefit is expected, which risks remain, and how the transaction will preserve customer choice. The structural possibility is compelling — as industrial knowledge becomes more consequential to physical performance, the market may increasingly value firms that can establish, maintain, and govern a trustworthy representation of infrastructure across its lifecycle — but the acquisition announcement itself is not evidence that the possibility has been realized.
2.6. The Hyperscaler Investment Cycle Meets Industrial Engineering
The competitive stakes are enlarged by the extraordinary infrastructure commitments of the hyperscale cloud companies, whose July 2026 earnings reports marked a new high in an investment cycle that has already exceeded most forecasts. Amazon’s chief executive told investors on July 30 that the company now expected approximately $220 billion of cash capital expenditure in 2026, up from the roughly $200 billion guided in February, citing higher memory prices among other drivers. Alphabet raised its full-year range to $195 to $205 billion after second-quarter capital expenditure of $44.9 billion, roughly 60 percent of it in servers. Meta spent $31.08 billion in the quarter and guided to $130 to $145 billion for the year, a range it raised again to $135 to $145 billion with its third-quarter results on October 21. Microsoft’s fiscal fourth-quarter capital expenditure including finance leases reached $41 billion, and its calendar-2026 figure of roughly $175 billion is lower than its earlier guidance only because of a lease-accounting change that extended the useful lives of datacenters from 15 to 25 years. The four companies’ combined second-quarter cash capital expenditure of $165 billion was up 87 percent year over year, and the implied second-half run rate is higher still. [11][41][43]
“We now believe we will spend approximately $220 billion in cash CapEx in 2026.”
— Andy Jassy, Chief Executive Officer, Amazon [11]
Table 4. Hyperscaler capital expenditure, calendar 2026 (company guidance and reported quarters)
| Company | Q1 2026 capex | Q2 2026 capex | 2026 guidance (early year) | 2026 guidance (as of Oct 22, 2026) | Capex as share of Q2 operating cash flow |
| Amazon (cash capex) | ~$44.2B | ~$53.1–54.2B | ~$200B | ~$220B | ~119% |
| Alphabet (purchases of P&E) | $35.7B | $44.9B | $175–185B | $195–205B | ~115% |
| Microsoft (capex incl. finance leases) | $31.9B (FQ3) | $41.0B (FQ4) | ~$190B | ~$175B (after lease-accounting change) | ~65% |
| Meta (incl. finance-lease principal) | $19.84B | $31.08B | $115–135B | $135–145B (raised Oct 21) | ~95–98% |
| Four-company total | ~$131.6B | ~$171B | ~$700B | ~$725–745B | — |
Source: company earnings releases and calls as compiled by Datacentres.com, NextMSC, and Analysis Atlas; figures differ slightly across compilations because of definitional differences noted in the row labels. [11][43]
The cash-flow consequences are now visible. In the second quarter, cash capital expenditure was approximately 115 percent of operating cash flow at Alphabet and 119 percent at Amazon, Alphabet’s free cash flow turned negative by $5.9 billion, Amazon’s trailing free cash flow moved from a $1.2 billion inflow to a $7.6 billion outflow, long-term debt at both companies roughly doubled from December 2025, and Alphabet raised $49.6 billion of equity and convertible preferred stock in June. Only Microsoft, which spent about 65 percent of operating cash flow, generated substantial free cash flow in the quarter. Whatever one’s view of the ultimate returns, the hyperscalers have moved from funding AI infrastructure out of surplus to funding it out of the capital markets, and that change raises the premium on operating the resulting assets well. [43]
The supplier side of the same cycle is equally stark. NVIDIA reported revenue of $96.2 billion for its fiscal second quarter ended July 26, 2026, up 106 percent from a year earlier, with datacenter revenue of $89.0 billion, up 117 percent, representing more than 92 percent of the total; it guided to $108 billion for the following quarter while assuming no datacenter compute revenue from China, and it disclosed approximately $105 billion in supply and capacity obligations alongside infrastructure guarantees for AI cloud power and land. In August, AWS and NVIDIA announced plans involving two million additional NVIDIA GPUs across AWS infrastructure. Jensen Huang’s framing of the quarter — that compute had become revenue — is the clearest possible statement of the first phase of the AI economy, in which the physical inputs to intelligence are themselves the product. [12][42]
“Now, compute is revenue.”
— Jensen Huang, Founder and Chief Executive Officer, NVIDIA [12]
Microsoft’s October announcements are particularly relevant to the title of this paper because they describe the second phase. The company’s October 13 technical discussion of Azure infrastructure treated power availability and electrical efficiency as system-level concerns extending from the grid through power conversion, racks, servers, and silicon, and its October 12 Maia 300 announcement emphasized co-engineering across accelerator systems, networking, and datacenter deployment, with each 96-accelerator rack drawing 167 kilowatts and the company claiming 40 percent more tokens per megawatt than leading GPU systems. These are company descriptions of engineering strategy and should be evaluated against subsequent operating evidence — independent validation of the efficiency claims will not be possible before the 2027 deployments begin — but the unit of account is telling. When a hyperscaler advertises tokens per megawatt rather than teraflops per chip, it is acknowledging that the binding constraint has moved from Layer 2 to Layer 1, and that the productivity of the datacenter as an energy-conversion system now matters as much as the capability of the silicon inside it. [6][24]
This is a foundational shift in procurement logic. A conventional datacenter project could assess power, cooling, servers, and workload software through relatively separate planning processes. An advanced AI facility increasingly requires simultaneous design decisions because modifying one layer changes requirements in the others. A new accelerator architecture alters rack power density. Higher rack power changes electrical distribution requirements. A new cooling approach changes pumping energy, water strategy, heat removal, redundancy, and building design, and those changes affect the speed and cost of construction. Infrastructure Cognition can help manage these dependencies by preserving the design intent and continuously comparing it with observed operating results. A hyperscaler operating hundreds of facilities might use validated models to identify which cooling configurations perform best under particular climatic conditions, which equipment exhibits common failure patterns, and how specific hardware changes alter the balance among throughput, power draw, and reliability.
The economic advantage would emerge from learning across a fleet without ignoring differences among sites. A facility in the American Southwest faces different weather and water constraints from one in Northern Europe. A campus with dedicated on-site generation — of which Cleanview identified 46 planned in the United States by early 2026 — has different operating options from one dependent entirely on a constrained utility connection. A model trained on one region cannot be assumed to transfer without calibration and review. Hyperscalers may also become consumers and suppliers of industrial intelligence simultaneously: consumers because their own datacenters require increasingly complex engineering, suppliers because their cloud platforms host the data systems, AI services, simulation tools, and agent development environments used by manufacturing and infrastructure customers. This dual role creates new strategic relationships with Schneider Electric, Siemens, Honeywell, and the utilities, and it creates tensions — a manufacturer may hesitate to place sensitive operating data into a platform operated by a major customer or competitor, and a utility may require strict local custody of operational technology records — that the commercial architecture must reconcile rather than assume away. [47]
From an investment perspective, the relevant question is whether AI infrastructure owners can establish a continuous learning mechanism around the entire asset lifecycle. The construction phase produces engineering assumptions. Commissioning generates verification records. Operations provide performance data. Maintenance yields failure and repair histories. Retirement or equipment replacement reveals the actual lifetime economics of the original design. If this information becomes machine-readable and comparable, successive infrastructure generations can be designed with better evidence, and the lease-accounting decision that extended Microsoft’s datacenter useful lives to 25 years becomes something more than an accounting choice: it becomes a claim about how long the knowledge embedded in those facilities will remain productive. In that sense, hyperscale capital expenditure may generate a second asset beyond the datacenter itself — accumulated, validated knowledge about the engineering and operation of large-scale AI infrastructure — whose commercial value will depend on how responsibly it is governed and how reliably it improves future outcomes.
Section 3: When AI Agents Become Industrial Operators
3.1. Five Levels of Operational Intelligence
The phrase “autonomous industrial operation” obscures enormous differences in capability and authority, and a good deal of confusion in both the investment community and the engineering community arises from using one word for five distinct things. A useful framework distinguishes five levels, ordered not by technical sophistication but by the degree of authority delegated to the system. At Level 1, Observation, sensors, alarms, dashboards, and rule-based monitors report operating states; the system recognizes a condition but does not necessarily explain its cause. At Level 2, Interpretation, models combine equipment history, environmental variables, and engineering context to suggest explanations; the system can prioritize probable causes, but its conclusions remain uncertain and are offered to a human. At Level 3, Simulation, the system estimates the consequences of potential interventions, comparing operating schedules, maintenance plans, energy settings, or production configurations against physical and economic constraints before anything is changed. At Level 4, Supervised Action, an agent prepares work orders, operating recommendations, procurement requests, or configuration changes for authorized personnel to review, and certain low-risk changes may be executable within preapproved limits. At Level 5, Bounded Autonomy, an agent can initiate specifically permitted operational changes, monitor the result, and revert or escalate when a predefined limit is reached, while human supervisors remain accountable for the operating envelope and for every decision beyond delegated authority.
Table 5. Five levels of operational intelligence in physical infrastructure
| Level | Name | What the system does | Who decides | Evidence required before adoption | Representative 2026 deployments |
| 1 | Observation | Reports states, alarms, trends | Human | Sensor validity, asset inventory | Conventional SCADA/BMS; OT monitoring baseline at CMPC |
| 2 | Interpretation | Ranks probable causes with context | Human | Contextualized data, provenance | Schneider–NVIDIA–HPE OT monitoring and cybersecurity agents |
| 3 | Simulation | Tests interventions against physics and economics | Human | Validated twin, configuration fidelity | Siemens Digital Twin Composer; Honeywell operating-scenario simulation |
| 4 | Supervised action | Prepares actions; executes low-risk ones within limits | Human approves; agent executes | Approval workflow, rollback, audit trail | Forge Cognition semi-autonomous mode; Osprey plan-first orchestration |
| 5 | Bounded autonomy | Executes permitted changes, monitors, reverts, escalates | Agent within envelope; human owns envelope | Safety case, independent interlocks, change control | Forge Cognition autonomous mode (shipped but not enabled) |
The progression through these levels is not inevitable, and it should not be treated as a maturity ladder that every facility must climb. A nuclear plant, a grid substation, a semiconductor fabrication tool, a hospital facility, and an ordinary commercial office building have radically different failure consequences. A Level 4 workflow that is entirely appropriate for adjusting conference-room setpoints might be inappropriate even at Level 2 for high-voltage switching. The essential measure of sophistication is not the number of actions the agent can perform without permission; it is the reliability with which the system produces beneficial outcomes within a defensible safety boundary. Honeywell’s decision to ship Forge Cognition with a Level 5 mode and keep it disabled is, in this framework, not a limitation of the product but a correct reading of where the evidence currently stands. [9]
The academic evidence supports that reading. A systematic literature survey published in May 2026 by researchers at the University of Stuttgart and the Institut für Automation und Kommunikation, following PRISMA 2020 guidelines, screened 2,341 publications on foundation-model-based agents in industrial automation and synthesized 88 of them. It found that reported systems were predominantly at prototype and early-validation stages — 75 percent at technology-readiness levels 4 to 6 — with deployment-oriented evidence remaining rare at 9.1 percent. Operational goals were most frequently positioned in user assistance, monitoring, and process optimization (Levels 1 through 3 in the present taxonomy), while conventional production-control purposes such as planning and scheduling were less prominent. Compared with an established baseline for industrial agent systems, the foundation-model agents showed substantial gains in human interaction and in dealing with uncertainty, but a pronounced deficit in negotiation, and the most widely reported limitations were lack of generalization, hallucination and output instability, data scarcity, and inference latency. In other words, the research community’s own accounting places industrial agents exactly where Honeywell’s chief technology officer placed them: capable, promising, and not yet ready to be left alone with the plant. [14]
3.2. Predictive Maintenance Moves From Alerts to Decisions
Predictive maintenance is a useful initial application because the underlying industrial problem is concrete, the economic value is measurable, and the failure modes are well understood. Machines wear out, bearings degrade, temperature patterns change, lubrication quality deteriorates, components drift, and equipment failures impose costs that range from inconvenience to catastrophe. Many organizations already deploy condition monitoring at Level 1. The next step — and it is a larger step than the vendors’ slide decks imply — is to integrate condition information with engineering limits, service history, inventory, production commitments, and financial priorities so that the output is a decision rather than an alert.
Consider two compressors exhibiting similar vibration signatures. One serves a production line scheduled for shutdown within three days; its replacement components are in stock. The other supports continuous production during a contractual delivery period, has a history of unsuccessful repairs, and depends on a part with a twelve-week lead time. An anomaly detector will assign both compressors a comparable health score, because the physics of the vibration is the same. A contextual industrial agent should recognize that the appropriate maintenance actions differ profoundly, because the consequences, options, and constraints differ profoundly. Infrastructure Cognition would connect diagnosis to scheduling, procurement, and post-repair evaluation: the agent could prepare a maintenance recommendation with supporting evidence, request parts, estimate downtime, check whether redundancy is available, and propose a service window, while human operators review the critical decisions. After the repair, the system would compare the subsequent equipment state to the original diagnosis, allowing the organization to measure whether the recommendation was correct — and, over hundreds of such cases, whether the agent’s judgment deserves expanded authority.
The value is not avoiding all failures, which is impossible and, beyond a certain point, uneconomic. It is improving the choice among preventive maintenance, condition-based maintenance, controlled continued operation, and immediate shutdown. Good maintenance decisions balance the probability of failure, the consequence of failure, the cost of intervention, the availability of spares, and the risk introduced by the maintenance activity itself, which is frequently underestimated: a significant fraction of industrial incidents occur during or shortly after maintenance, when systems are in non-standard configurations and protective functions may be bypassed. An agent that recommends more maintenance is not necessarily an agent that improves reliability. The Schneider–NVIDIA–HPE collaboration’s reference customer, the Chilean forestry and pulp company CMPC, illustrated the appropriate posture when its digital and process manager described the existing OT monitoring deployment as a proven foundation on which agents could help specialists investigate issues faster and move from insight to action, while keeping people in control of critical decisions. [57]
“Keeping people in control of critical decisions.”
— Matías Jory, Digital and Process Manager, Industrial Operations, CMPC [57]
3.3. Power Systems and Datacenters: Cognition Under Physical Constraints
An AI datacenter makes the maintenance problem more complex because the interactions among subsystems are tighter and the consequences of error propagate faster. High-density accelerators produce heat that must be removed within seconds. Pumps, cooling loops, distribution equipment, backup generators, batteries, power-management systems, and network devices interact continuously. A facility-wide model may identify opportunities to improve cooling performance, shift certain computing activities, or prioritize maintenance before a critical component fails. Yet facilities cannot treat all equipment as economically interchangeable. A server workload might be rescheduled to another region; a transformer cannot be moved across a continent, and the IEA noted in April 2026 that transformers, along with gas turbines and advanced chips, were among the supply-chain shortages slowing datacenter development worldwide. A liquid cooling system has thermal limits. Backup generation may be constrained by emissions permits and fuel availability. Batteries have charge and discharge restrictions. Grid conditions may limit when additional load can be added. The model must therefore distinguish flexible decisions from inflexible physical constraints, and it must do so with a representation of the facility that is accurate at the level of individual components. [23]
For electricity networks there is a further boundary that datacenter operators have been slow to internalize: one operator’s local optimization can create consequences for neighboring systems. A recommendation that looks efficient for a datacenter may affect a distribution utility, another large customer, or a regional transmission operator. The technical literature on AI datacenters as grid loads has begun to document why. A 2025 review of electricity demand and grid impacts noted that AI datacenters interface with the grid primarily through power-electronic converters that exhibit low inertia, fast response dynamics, and harmonic distortion, and that their rapid, large swings in demand — synchronized across thousands of accelerators during training runs — differ fundamentally from the electromechanical loads for which grids were designed. Researchers at Georgia Tech have reviewed the strategies by which datacenters could instead provide flexibility, including rerouting workloads, smart scheduling, pre-cooling, and the use of backup systems to support the grid during emergencies, while observing that operators have historically had little incentive to respond to electricity prices because the revenue from a computing job dwarfs its energy cost. [55][56]
That incentive structure is changing under regulatory pressure. During a winter storm in early 2026, the Department of Energy directed datacenters in the PJM Interconnection, the country’s largest regional grid, to run on backup generators to free capacity for the grid, and PJM has proposed requiring large new loads to bring their own supply or accept curtailment. Google has announced contracts with several utilities to reduce consumption at certain datacenters when called upon. A useful architecture must therefore incorporate explicit grid-service agreements and verified dispatch authority rather than assuming that a datacenter agent can make power-system decisions independently, and the datacenter’s internal cognitive system must be able to represent its obligations to the grid as constraints of the same standing as its obligations to its own tenants. The Department of Energy’s September 2026 “Speed to Power” announcements, which described investments across 26 states in reconductoring and grid-enhancing technologies, demonstrate that operational improvements and physical expansion are proceeding together: smarter use of existing lines can contribute capacity, but the public investment, permitting, engineering, and safety framework remains essential. [53][54]
“Demand response has to be part of the solution.”
— Matt O’Connor, Chief Investment Officer, International Energy, Carlyle [54]
3.4. The Industrial Decision Loop: Observe, Contextualize, Simulate, Authorize, Verify
The central operating mechanism proposed in this paper is a controlled decision loop of five stages, each of which must preserve provenance and each of which can be audited independently. In the Observe stage, the system collects and validates sensor readings, event logs, operating states, engineering records, and environmental conditions, and it records the age and quality of each input. In the Contextualize stage, it identifies the affected assets, their dependencies, the relevant constraints, the maintenance history, and the uncertainty in the data, drawing on the four information environments of Section 1. In the Simulate stage, it estimates possible causes and tests proposed interventions using an appropriate physical or statistical model whose validated domain is known. In the Authorize stage, it verifies that the proposed action is permitted, that the relevant risks have been evaluated, and that any required human approval has been obtained through a channel that cannot be forged by the model itself. In the Verify stage, it measures the real-world outcome, compares it with the expected result, documents deviations, and updates the operational knowledge base so that the next cycle begins with better information than the last.
Table 6. The governed industrial decision loop and the provenance each stage must preserve
| Stage | Core question | Inputs | Provenance recorded | Failure to guard against |
| Observe | What is happening? | Telemetry, logs, environmental data, engineering records | Source, timestamp, calibration status, data quality | Treating a failed sensor as a changed asset |
| Contextualize | Which assets, constraints, and dependencies are involved? | Asset registry, twin, maintenance history, contracts | Identifier reconciliation, configuration version | Reasoning about an asset that was modified or replaced |
| Simulate | What would happen if we intervened? | Physics or statistical model, scenario parameters | Model version, validated domain, assumptions | Extrapolating outside the validated regime |
| Authorize | Are we permitted, and who approves? | Policy, authority matrix, risk classification | Approving actor, time, scope of permission | A natural-language instruction becoming physical authority |
| Verify | Did reality match the prediction? | Post-action measurements | Observed outcome, deviation, knowledge-base update | Crediting the agent for outcomes it did not cause |
This loop separates legitimate cognition from unaccountable automation. Each stage must preserve provenance: the information used, its age, its confidence or quality, the model version, the simulated assumptions, the approving actor, and the actual physical outcome. The loop also requires a safe response to uncertainty. If two sensors disagree, the system must not create false certainty by selecting whichever value supports its preferred recommendation. If a digital twin no longer matches the actual configuration, the system must limit its actions until the discrepancy is resolved. If connectivity fails, the facility should continue operating under defined local control and safety procedures that do not depend on the agent at all. The research literature on language-model agents in control systems has converged on a similar structure. A 2026 review of large language models in sensor-driven control systems concluded that in viable architectures the model functions as a semantic and orchestration layer that augments rather than replaces classical sensing, estimation, planning, and control. The Osprey framework, developed for large scientific facilities whose control systems expose tens of thousands to millions of channels, adopted what its authors called plan-first orchestration with human oversight, translating operator intent into approved, isolated execution rather than allowing the model direct access to hardware. A 2026 scoping review in the Journal of Manufacturing Systems reached the same conclusion from the human-factors side, proposing a hybrid intelligence model in which human-in-the-loop verification, validation, and uncertainty management are built into the architecture rather than added afterwards. [48][49][50]
By 2030, the maturity of industrial AI may be judged less by the fluency of agent explanations than by how consistently these five stages produce auditable, reliable decisions. The Verify stage deserves particular emphasis because it is the stage most often omitted in commercial deployments and the stage on which every claim of value ultimately rests. An agent that recommends an intervention and never learns whether the intervention worked is not learning; it is guessing with a record of its guesses.
3.5. Edge Inference, Cybersecurity, and Human Engineering Judgment
Deploying every industrial task in a distant hyperscale datacenter is neither technically necessary nor always desirable, and the October 22 Schneider–NVIDIA–HPE announcement was notable for saying so explicitly. Some monitoring and decision functions require low latency, local availability, controlled data residency, and a predictable operating cost. Others benefit from large cloud models, fleet-wide analytics, specialized simulation, and access to enormous computing resources. The collaboration expressly emphasized industrial edge agents, data control, cybersecurity, and the cost of AI consumption, and it drew an important boundary in its CMPC example by describing an existing operational-technology monitoring deployment as a foundation for future agent capabilities rather than as evidence that autonomous plant operation had been achieved. An effective architecture may combine edge-based monitoring with periodic cloud analysis and centralized engineering-model updates: local systems handle immediate alarms and permitted responses, larger models investigate complex patterns across facilities, and human experts approve policy changes and high-consequence interventions. [4][5]
“In process industries, the plant itself is the body.”
— Schneider Electric executive, quoted in Manufacturing Digital on physical AI for process industries [5]
This structure creates cybersecurity obligations that the industrial sector is only beginning to formalize for the agentic case. Industrial systems cannot automatically trust model outputs, retrieved documents, third-party plugins, external communications, or compromised sensor feeds. NIST’s initial public draft of Special Publication 800-82 Revision 4, Guide to Operational Technology Security, released on September 21, 2026 with comments due November 30, restructured its guidance around the Cybersecurity Framework 2.0, expanded its sector coverage to building automation, water and wastewater, food and agriculture, freight rail, and maritime vessels, addressed the convergence of industrial IoT and cloud, and added security-architecture guidance focused on protecting system-management functions and applying zero-trust principles. Its emphasis on asset management and network monitoring as foundational controls aligns directly with the configuration-fidelity argument of Section 1, and its zero-trust orientation implies that an AI agent, like any other principal, should be granted the minimum authority its function requires and no more. NIST’s AI Risk Management Framework offers a complementary approach to managing AI-specific risks, including the profile work for critical infrastructure. [51][52]
Human engineering expertise remains decisive because industrial autonomy is not merely a software engineering problem. It requires knowledge of local conditions, failure modes, regulations, physical safety, vendor limitations, and operational trade-offs that are rarely written down anywhere a model can read. AI may change how engineers spend their time — reducing routine diagnosis while increasing their role in validation, exception handling, and system assurance — but the economic benefit must be measured alongside the costs of workforce training and the redistribution of responsibility. Stanford’s 2026 AI Index found that 73 percent of U.S. AI experts viewed the technology’s labor-market impact positively while only 23 percent of the general public agreed; in industrial settings, where the workforce holds the tacit knowledge on which safe operation depends, that perception gap is not a public-relations problem but an operational risk. [29]
3.6. The Engineering Safety Case for Agentic Infrastructure
In traditional engineering practice, an intervention in a critical system is justified through a combination of specifications, testing, documented procedures, and professional accountability. The introduction of AI agents should not eliminate those requirements; it makes their explicit formulation more important, because the agent introduces a source of variability that conventional control logic does not have. A safety case is a structured argument, supported by evidence, that a system is acceptably safe for a defined use in a defined operating environment. A cognitive infrastructure application should have a corresponding assurance record that specifies what the agent is designed to do, which physical assets it can observe, which actions it can propose, which actions it may execute, what conditions invalidate its recommendations, and how the facility maintains control if the AI system fails or behaves unexpectedly.
Consider an energy-optimization agent assigned to a campus cooling system — the application Honeywell is piloting and the one Schneider’s energy-optimization agents target. The operator may authorize the system to adjust a noncritical setpoint within a narrow range while preserving minimum flow, equipment temperature, redundancy, and service obligations. The agent must not interpret a general instruction to save energy as permission to disable a protective alarm or reduce a safety margin below its engineering requirement. The control architecture should enforce this boundary independently of the language model. A natural-language instruction should never be the sole source of a machine’s physical authority. The agent may submit a structured proposed change to a trusted application, but a separate control layer — deterministic, tested, and outside the model’s reach — should validate identity, authorization, permissible parameters, equipment status, and required interlocks before the action can reach the physical system. This is the architectural meaning of Daniela Rus’s observation that the physical world cannot tolerate hallucinations: the system is designed so that a hallucination, when it occurs, is caught before it becomes a command. [15]
The distinction is especially important when the agent uses external information. A retrieved maintenance manual, a network message, or a vendor support note might contain erroneous or malicious instructions, and the prompt-injection literature has demonstrated repeatedly that language models cannot reliably distinguish data from commands when both arrive as text. The system must treat such material as evidence to be evaluated, not as authority to rewrite operating permissions. Industrial cybersecurity, model security, and ordinary access control converge at the point where an AI recommendation is translated into an action, and the Schneider–NVIDIA–HPE cybersecurity agents — which correlate network activity, control-system behavior, asset configurations, operational events, and process context — are themselves an acknowledgment that the agentic layer both requires and can contribute to that convergence. [39]
Safety also requires attention to degraded conditions. What should happen if an edge server fails? What if the wide-area network is disconnected? What if a model update introduces new behavior? What if an equipment identifier is duplicated or a sensor reports physically impossible values? A mature design should fall back to a verified mode of operation and preserve enough records for engineers to understand the incident afterwards. Not all errors deserve the same response. An incorrect prediction of office comfort may justify a new setpoint or an operator correction. An unreliable recommendation involving high-voltage switching, toxic materials, pressurized equipment, or nuclear safety should trigger a far more conservative response, up to and including automatic suspension of the agent’s authority. Risk classification should consider both the probability of failure and the severity of its consequences, not simply the model’s average test accuracy, and the organization should establish a change-control process for the models and data connectors used in industrial settings: model versions need documented tests; new permissions should require approval; significant data-schema changes should be validated; and system updates should be reversible where technically feasible. Historical test performance does not guarantee reliability after equipment modifications, seasonal changes, altered products, or different workload conditions.
The human operating model must be similarly deliberate. Engineers should be able to see why the agent selected a recommendation, which evidence it relied upon, what alternatives it evaluated, and where its uncertainty lies. Supervisors should receive enough information to make genuine decisions rather than being presented with a polished answer designed to elicit routine approval — the automation-complacency failure that Brynjolfsson’s work on physicians paired with AI assistants has documented in a very different domain. Over time, the strongest industrial agents may become those that are demonstrably conservative under uncertainty, transparent about limits, and reliable in routine conditions. Their competitive value could arise from reducing avoidable human workload while preserving the expertise and authority necessary when circumstances become exceptional. The conclusion is not that industrial AI should never act autonomously. It is that autonomy must be engineered as a bounded, testable, reversible, and accountable capability, and that the more consequential the physical system, the stronger the assurance required before the AI crosses from recommendation into execution.
Section 4: The Geopolitical and Regional Consequences of Intelligent Infrastructure
4.1. Industrial Information Becomes a Strategic Resource
The geopolitical importance of advanced AI is usually framed around semiconductors, computational capacity, model performance, and access to electricity. Infrastructure Cognition adds a less visible resource to that list: verified knowledge about how strategically important physical systems are designed, maintained, and operated. Industrial information is economically valuable because it makes complicated systems interpretable, and it is sensitive for exactly the same reason. Engineering records for a semiconductor fabrication plant may reveal production processes, specialized equipment, supplier relationships, control configurations, and physical dependencies. Infrastructure information for an electric grid could expose critical assets, maintenance weaknesses, system topology, and operating constraints. A complete model of a datacenter could contain commercially sensitive capacity information, security architecture, or service obligations. The very contextualization that makes an industrial agent useful makes the agent’s knowledge base a target.
This creates a tension between interoperability and national security that will shape procurement, regulation, and corporate structure through 2030. Industrial organizations benefit from combining data across vendors and jurisdictions. Yet moving detailed operational information into shared platforms may create disclosure, cyber-risk, intellectual-property, or export-control concerns. The appropriate response is not necessarily to prohibit all sharing; it is to design data-access boundaries according to sensitivity, purpose, authorization, and jurisdiction, which is precisely what the zero-trust orientation of the NIST SP 800-82 Revision 4 draft recommends for operational technology generally. The United States’ Bureau of Industry and Security has continued to administer restrictions and licensing policies involving advanced computing semiconductors and semiconductor technologies, and in January 2026 it revised its license review policy for certain accelerator exports to China under specified conditions; NVIDIA’s decision to exclude China datacenter compute revenue from its third-quarter guidance shows how directly such rules reach into Layer 2 economics. Such rules govern particular products and transactions and should not be casually generalized to all industrial data or every AI model. The broader analytical point is that technical capability and cross-border access increasingly require legal scrutiny across several layers of the AI economy, and that an engineering-data platform, once embedded in industrial operations, may remain there for decades — long after the chip generation that prompted the original export rule has been retired. [12][51][58]
4.2. The United States, Europe, and China: Different Sources of Industrial Advantage
The United States has formidable advantages in frontier AI models, hyperscale cloud platforms, advanced chip architecture, enterprise software, and AI investment. Stanford’s 2026 AI Index found that the performance gap between the best American and Chinese models had effectively closed — Anthropic’s leading model held a 2.7 percent edge on the Index’s benchmark basket as of March 2026 — but the United States retains an overwhelming lead in installed compute and in the capital that funds it. Its challenge, from the perspective of this paper, is connecting those capabilities to factories, utilities, industrial equipment, and infrastructure assets that often use heterogeneous legacy systems and operate under fragmented state and federal regulatory regimes. The federal government’s July 2025 AI Action Plan and Executive Order 14318, Accelerating Federal Permitting of Data Center Infrastructure, addressed the physical side of that challenge by streamlining permitting for facilities requiring more than 100 megawatts of new load and making federal lands available, on the explicit premise that the existing permitting system made it almost impossible to build infrastructure at the required speed. The Action Plan also called for an AI Information Sharing and Analysis Center to address critical-infrastructure cybersecurity, an acknowledgment that the federal government regards the agentic layer as a security question as well as an economic one. [29][74][75]
“Almost impossible to build this infrastructure in the United States.”
— America’s AI Action Plan, The White House, July 2025, on the permitting system [75]
Europe possesses deep strengths in industrial automation, engineering software, machinery, electrical infrastructure, and manufacturing expertise, and the two companies at the center of this paper — Schneider Electric and Siemens — are the clearest expressions of those strengths. The European Commission’s 2026 AI Gigafactories initiative, with its July 2026 call for proposals, aims to enlarge computing capacity and technological autonomy, but the European industrial AI opportunity is broader than building additional clusters. It lies in embedding AI into established industrial workflows, equipment ecosystems, and energy management, where European firms already hold the installed base and the engineering relationships. Schneider’s chief financial officer noted in February 2026 that European datacenter construction was beginning to unlock in places where permitting, electrical connection, and government support had been assembled, and the company’s third-quarter 2026 results showed European organic growth of 8.6 percent against 19.9 percent in North America and 26.1 percent in China and East Asia — a reminder that Europe’s industrial-software strength coexists with a slower physical build-out. [10][59]
“We start to see some unlock of data centers in Europe.”
— Hilary Maxson, Chief Financial Officer, Schneider Electric [78]
China is pursuing its own industrial AI integration along a path that emphasizes manufacturing breadth. On October 20, 2026, Chinese government reporting described work on next-generation smart-manufacturing policies, including industrial AI agents, intelligent inspection and control equipment, robotics, and specialized models, building on a January 2026 plan to deepen the integration of the industrial internet with artificial intelligence. These announcements are policy intentions and objectives; they do not establish uniform implementation across Chinese factories, and the digital-twin implementation literature suggests that Chinese facilities face the same contextualization obstacles as everyone else. But the strategic emphasis is unmistakable: where the United States leads with models and Europe with automation, China is attempting to lead with the sheer number of industrial sites in which agents can be deployed and improved. [60]
The strategic competition could therefore develop along multiple axes — the quality of frontier models, access to accelerators, the breadth of installed industrial equipment, the engineering data available for model contextualization, control of standards, cybersecurity and trust, and the pace at which firms demonstrate measurable physical productivity — and an economy might lead in one axis while remaining dependent on foreign suppliers in another. The United States could lead in model capability while importing certain industrial components and much of its automation software. Europe could lead in aspects of automation while relying on external accelerator infrastructure. China could move quickly to apply AI across manufacturing while facing restrictions on certain advanced technologies. The future should not be reduced to a single national leaderboard, and the central geopolitical question is which countries can link scientific innovation, software, hardware, infrastructure operations, and institutional trust into repeatable industrial improvements. The World Bank’s World Development Report 2026 added a further dimension by arguing that developing economies need not build frontier infrastructure to benefit from AI, while its president observed that the power, computing, data, skills, and capital required for the largest AI datacenters are beyond most emerging markets and will remain concentrated in the richest grids. [71][72]
“AI has thrown developing economies a lifeline, and they should seize it.”
— Indermit Gill, Senior Vice President and Chief Economist, World Bank Group [71]
4.3. Texas: Datacenter Demand Meets Grid Accountability
Texas illustrates the local consequences of a rapidly expanding AI infrastructure economy more vividly than any other jurisdiction, and the events of August and September 2026 mark a turning point that investors in Layer 3 should study carefully. The state’s advantages include substantial energy resources, an extensive industrial base, technology investment, and a regulatory environment that attracted datacenter projects at a scale that eventually overwhelmed the planning system: by December 2025, ERCOT reported 226 gigawatts of large-load projects, primarily datacenters, seeking grid connections, roughly three times the entire installed datacenter capacity of the United States. On August 3, 2026, Governor Greg Abbott directed the Public Utility Commission and ERCOT to audit every datacenter in the interconnection queue. On September 14, he directed the Texas Water Development Board to compel datacenters to comply with water-use reporting requirements and to impose legal consequences for past noncompliance. On September 21, he directed the Texas Commission on Environmental Quality to halt the issuance of all permits related to datacenter projects until ERCOT completed its audit, which law firms advising developers expected in mid-December 2026, and he announced that he would work with the Legislature to eliminate financial incentives for datacenters in the next session. His stated conditions were that datacenter projects must cover all electrical infrastructure costs, result in lower residential electricity bills, and complete the audit. [47][61][62]
“Simply put, Texans must come first.”
— Greg Abbott, Governor of Texas, directing the halt of datacenter permits [61]
The significance of the Texas sequence for this paper is twofold. First, it demonstrates that announcing a multibillion-dollar investment is no longer sufficient evidence that a facility can be reliably powered on the desired schedule, in a state that had been the most accommodating in the country. Second, it demonstrates that the information problem of Section 1 is now a public-policy problem: the audit exists because neither the grid operator nor the state knew which of the 226 gigawatts of requests were supported by committed construction, equipment, financing, and power arrangements, and which were speculative placeholders. Infrastructure Cognition has two potential roles in this setting. At the facility level, improved operational models could help developers estimate actual power requirements, cooling behavior, flexibility, and resilience with enough credibility to satisfy an auditor. At the system-planning level, better-verified operating information could help utilities and regulators distinguish credible load forecasts from unsupported requests. However, regulators should not mistake vendor models for independent evidence. A project sponsor’s AI-generated power forecast remains a forecast; it requires transparent assumptions, validation, and verification. The productive direction, as Texas confronts large-load growth, is to require increasingly auditable engineering and operating information while protecting commercially sensitive data — and the broader lesson is that intelligent infrastructure must serve both the private owner and the public system on which it depends.
4.4. Arizona and Ohio: Intelligent Manufacturing Must Be Built Alongside Physical Capacity
Arizona’s semiconductor investment is a case of physical infrastructure and technical knowledge developing together. TSMC’s Arizona operation involves advanced manufacturing capacity, specialized labor, utilities, packaging capabilities, equipment supply, and continuous engineering refinement, and the company’s 2026 disclosures describe further expansion intentions and additional advanced packaging and manufacturing activity that should be distinguished from completed operating capacity. Semiconductor production is particularly suitable for studying Infrastructure Cognition because yield improvements depend on interactions among materials, equipment conditions, process parameters, inspection data, and design changes — a system of such complexity that it has always relied on statistical process control and is now, as NVIDIA’s October announcements with TSMC make clear, adopting GPU-accelerated lithography, atomistic simulation, vision-based defect inspection, and AI-assisted fab scheduling. A simulated yield gain is not identical to a sustained increase in saleable output, and the results must be proven in fabrication conditions, but the fab is among the most structured and instrumented environments in all of manufacturing, which makes it the most plausible early site for Level 3 and Level 4 cognition to deliver verified value. [22][63]
Ohio presents another dimension. The state’s $600 million onshoring incentive agreement with Intel, finalized in 2023, demonstrates how major investments require industrial sites, roads, water, energy, suppliers, workforce development, and public infrastructure support, and the subsequent history of that project demonstrates that earlier state commitments are no guarantee of a construction timetable. For both states, a valuable 2027–2030 agenda would connect publicly supported industrial investment with measurable technical capabilities: maintenance reliability, energy efficiency, equipment utilization, workforce training, and transferable engineering skills. A new fab is a physical asset. A fab that continually learns from validated engineering and operational information has the potential to become a more productive asset, provided its learning mechanisms are tested and governed. This distinction matters for public incentives, because governments should measure the durable benefits of industrial projects — local supplier development, engineering employment, operational resilience, tax-base contributions — rather than treating the initial capital-expenditure announcement as the final economic outcome. [64]
4.5. Indiana and Michigan: The Convergence of AI, Manufacturing, and Energy
Indiana combines a traditional manufacturing base with a rapidly expanding datacenter and cloud investment environment. The state previously announced an $11 billion AWS datacenter campus in northern Indiana, supported by available tax-exemption mechanisms, and in September 2026 Governor Mike Braun announced Amazon’s plan for an advanced manufacturing operation in central Indiana with 300 high-wage jobs, describing future use of AI-powered manufacturing systems and robotics. The investments represent different facilities and purposes, but together they illustrate a broadening AI industrial footprint in a single state, and they raise the policy question that Texas has already confronted: whether power systems, workforce programs, infrastructure permitting, and local communities can support the resulting development without shifting costs onto residential customers. [65]
Michigan offers a related but distinct example in which the regulatory machinery has moved earlier. Governor Gretchen Whitmer’s administration announced a July 2026 datacenter affordability and responsible-growth action plan emphasizing protection for ratepayers and local resources, and on October 1 the Michigan Public Service Commission conditionally approved electric-service arrangements involving a Google datacenter, demonstrating the direct role of utility regulation in large AI infrastructure projects. The state is also pursuing a broader energy and industrial transition involving the Palisades nuclear restart, advanced manufacturing, and workforce development. These separate initiatives should not be conflated into a single power contract or assumed to serve particular AI projects; rather, they illustrate how electricity reliability, industrial capability, and datacenter development are becoming linked policy questions in states whose economies were built on heavy industry. Infrastructure Cognition could help these states evaluate the performance of established and newly built facilities, but public agencies would need trustworthy data, skilled technical reviewers, clearly allocated responsibilities, and transparent measures before relying on AI-generated analyses in regulatory decisions. [66]
4.6. California, Virginia, and Pennsylvania: Cognition Must Include Public Accountability
The most advanced technology cannot earn public legitimacy by ignoring the costs it imposes on others, and several states made this point in increasingly concrete terms during 2026. In September, Governor Gavin Newsom signed what his office described as the most comprehensive datacenter laws in the nation, addressing electricity, water, and land-use information, with provisions intended to increase public transparency and protect customers against inappropriate cost shifting. These laws focus on different aspects of project impacts and should be interpreted according to their individual provisions, not as a blanket approval or prohibition of datacenter construction. Virginia, which hosts the world’s largest concentration of datacenters and where roughly a quarter of electricity consumption already serves them, adopted new accountability standards under Governor Abigail Spanberger, and the State Corporation Commission ordered in July and August that certain direct-connect substations and transmission lines serving large loads be paid through mandatory contributions in aid of construction rather than socialized across all ratepayers. Pennsylvania Governor Josh Shapiro’s August executive order introduced state-level requirements for proposed datacenter projects, emphasizing local approvals, economic and environmental responsibilities, and the financial consequences for other electricity customers. [67][68][69]
Table 7. State-level datacenter accountability actions, 2026
| State | Date | Action | Mechanism | Relevance to Infrastructure Cognition |
| Texas | Aug 3 – Sep 21 | Grid audit of all queued datacenters; water-use enforcement; halt of all TCEQ permits pending audit | Executive directives to PUCT, ERCOT, TWDB, TCEQ | Forces verifiable load and water data; penalizes speculative queue entries |
| Michigan | Jul 15; Oct 1 | Affordability and responsible-growth plan; conditional approval of Google electric-service contracts | Gubernatorial plan; MPSC order | Regulatory review of large-load contracts before energization |
| Virginia | Jul 31 – Sep 18 | Direct-connect infrastructure costs assigned to large loads; accountability standards and Executive Order 22 | SCC order; executive order | Cost causation made explicit; reporting obligations |
| Pennsylvania | Aug 18 | Executive order on datacenter development | Executive order | Local approvals; ratepayer-impact disclosure |
| California | Sep 21 | Package of datacenter laws on electricity, water, land use | Legislation | Public transparency on resource use |
These initiatives are directly relevant to Infrastructure Cognition because a truly intelligent facility must be understood within a larger community and infrastructure system. An operator may optimize energy consumption inside a building while creating external costs associated with water supply, transmission upgrades, emissions, or local public infrastructure. The United Nations Secretary-General made the same point at the global scale in June 2026 when he launched an AI Environmental Transparency Initiative calling on every major AI company to measure and publicly disclose its energy, water, and land impacts and to commit to renewable power for all datacenters by 2030, citing a UN study finding that datacenters consumed more electricity than all but ten countries in 2025. The next generation of industrial intelligence should consequently be able to support independently verifiable reporting on resource use and operational commitments. Such transparency need not expose confidential model weights, proprietary manufacturing recipes, or cybersecurity-sensitive diagrams. It should instead distinguish the information legitimately needed for regulatory oversight from the information that must remain protected. The policy principle is straightforward: intelligence that optimizes only the private asset while ignoring public consequences is incomplete infrastructure intelligence. [73]
“It is time to come clean.”
— António Guterres, Secretary-General of the United Nations, on AI’s environmental footprint [73]
4.7. Sovereignty, Industrial Standards, and the New Geography of Trust
The next geopolitical competition may involve not only the location of compute and manufacturing equipment but also the legal and technical custody of industrial knowledge. A manufacturer might use a European engineering platform, American cloud models, Asian machine tools, and local plant-control software, and each relationship can raise separate questions about intellectual property, operating data, cybersecurity, and applicable law. The challenge becomes sharper when the combined system can do more than display information. A model with a complete representation of industrial process dependencies could support maintenance and optimization, but the same information could reveal sensitive production methods or vulnerabilities. An agent with the ability to issue commands across an international network creates questions about who may authorize the commands, where the execution environment is situated, and which legal jurisdiction governs an incident.
These are not reasons to assume that every cross-border platform is inherently unsafe. They are reasons to classify industrial information and delegate authority explicitly. Less sensitive performance statistics may be shared across a global fleet to improve reliability. Detailed control-system credentials and protective-system configurations may need strict local custody. Engineering models may be licensed for a particular purpose without granting permission to use the underlying records for unrelated model training — a provision that will become a standard negotiating point as vendors seek to train foundation models on customer data. Europe’s regulatory environment illustrates the complexity. The EU AI Act, through the Digital Omnibus on AI that entered into force on July 27, 2026 as Regulation (EU) 2026/1744, deferred the obligations for standalone high-risk systems under Annex III — a category that expressly includes the management and operation of critical infrastructure — from August 2, 2026 to December 2, 2027, and for product-embedded high-risk systems to August 2, 2028, while leaving the substantive requirements for risk management, data governance, technical documentation, and human oversight unchanged. Whether a specific industrial agent is legally classified as high-risk depends on its intended use and the applicable provisions; the label cannot be determined merely because software operates inside a factory, but an energy-optimization agent with authority over grid-connected equipment should be designed on the assumption that it may be. [70]
For U.S. companies operating in multiple jurisdictions, this argues for designing compliance into the architecture rather than bolting it on. A globally consistent model may be deployed under different local permissions, audit requirements, retention rules, and restrictions on autonomous action. Large firms may prefer standardized core technologies with site-specific policy layers. Startups may gain an advantage by providing well-documented compliance, security, and interoperability components. Industrial standards are equally important: an equipment manufacturer may describe a pump, transformer, or robotic cell using one data model while a plant owner records different names and attributes, and a common standard can reduce ambiguity, but forcing every asset into an oversimplified schema can erase important engineering distinctions. Useful interoperability requires both shared descriptions and the ability to preserve specialized technical meaning.
A mature system should therefore represent not only its knowledge but the limits of that knowledge. It should record which source supplied an equipment specification, which local engineer approved a configuration, which rules apply to the facility, and which data-sharing rights were granted. Where sources conflict, the system should preserve the disagreement instead of silently selecting an answer. The resulting investment opportunities extend beyond frontier AI models to industrial data validation, semantic mapping, secure edge execution, access-control systems, digital-twin verification, model testing, cryptographic provenance, and cross-border compliance infrastructure, and some of the most commercially important firms in this field may be specialized providers that remain nearly invisible to consumers. There is also a strategic argument for building domestic engineering capability: a government may subsidize a new factory or datacenter, but its resilience depends on local expertise capable of inspecting and maintaining the physical systems. Reliance on foreign software is not automatically unacceptable; reliance on inaccessible knowledge or unsupported systems can be dangerous. Policies that encourage data portability, workforce development, and independent testing can reduce vulnerability without requiring every technology component to be domestically produced.
Infrastructure Cognition thus introduces a new concept of sovereignty. It is not merely the physical location of a server. It is an organization’s effective control over the knowledge, permissions, operating decisions, and recovery capabilities essential to its assets, and this form of sovereignty becomes increasingly important as industrial intelligence crosses borders and corporate ownership boundaries. For multinational enterprises, the most durable strategy may combine local operational control with globally shared learning: a facility retains authority over safety-critical actions while contributing anonymized or appropriately protected performance evidence to models used across other facilities, so that engineering teams learn from failures and maintenance patterns across regions without exposing sensitive diagrams or permitting one remote software service to become a single point of failure.
Section 5: The Intelligent Industrial Enterprise of 2030
5.1. From Isolated AI Projects to a Governed Operating Architecture
The transition to 2030 is unlikely to occur as one simultaneous upgrade across the industrial economy, and any forecast that assumes otherwise should be discounted accordingly. Older plants will retain legacy control systems whose vendors may no longer exist. New datacenters may incorporate integrated digital engineering from the first day of construction. Highly regulated industries will impose stricter approval requirements, and the EU’s December 2027 deadline for high-risk systems in critical infrastructure will shape procurement timetables well beyond Europe. Some companies will lack reliable asset records or sufficient cybersecurity resources to deploy complex agents safely, and the survey evidence suggesting that only a fifth of industrial security leaders possess a complete OT asset inventory implies that this group is large. This creates a practical distinction between AI experimentation and industrial transformation. A successful pilot might classify alarms or summarize maintenance reports. An intelligent enterprise must be capable of repeating such functions across multiple sites, recording errors, correcting inaccurate context, changing models safely, proving economic value, and preserving human authority — and the May 2026 systematic survey’s finding that only 9.1 percent of reported industrial agent systems had deployment-oriented evidence is a measure of how far most organizations remain from the second condition. [14][17]
An industrial cognitive architecture should therefore have several separable parts: an asset-information foundation; a layer for validating incoming observations; models and simulations specific to the relevant physical processes; workflow applications; permissions and authorization; operational monitoring; and independent safety systems. These parts need not all be supplied by one vendor, and the acquisition strategies reviewed in Section 2 should not be read as implying that they must be. Indeed, separation can make it easier to validate and replace components without rewriting the entire facility, and the architectural principle that a model which can read a maintenance record does not automatically need permission to alter a programmable logic controller is easier to enforce when the two capabilities are supplied by different systems. The engineering challenge is to create useful connections without allowing every software component to become a route into the physical control system. By 2030, the most mature deployments may resemble a collection of carefully governed capabilities rather than a single omnipotent AI system, and their strength will lie in coordination, context, accountability, and resilience rather than in the breadth of any one model’s authority.
5.2. Interoperability and the Industrial Data Ownership Question
Interoperability is a critical commercial principle because industrial assets typically remain in operation far longer than software product cycles. Factories use equipment from dozens of suppliers. Grid operators depend on networks installed across many decades. Datacenters combine power equipment, servers, networking, cooling systems, and facility management from different manufacturers, and a facility whose useful life has just been extended to 25 years for accounting purposes will outlive several generations of the software that manages it. For Infrastructure Cognition to remain useful across these horizons, customers must be able to retain asset histories and interpret technical information even if a software vendor changes, is acquired, or discontinues a product. Data export alone is insufficient when exported records lose identifiers, relationships, engineering units, timestamps, or provenance, because the contextualization that gave the data its value is exactly what is lost in a flat export.
A practical contractual framework should therefore address at least six questions before a mission-critical system becomes dependent on a particular platform. Who owns the original engineering and operational information? Who may use that information to train or refine a proprietary model, and under what anonymization and consent conditions? Can customers retrieve contextual relationships and decision histories — not merely raw records — in portable, documented formats? Which party is responsible when supplier software recommends an unsafe or erroneous intervention, and how is that responsibility allocated when a human approved the recommendation? What happens to essential monitoring and control functions when network access or a commercial subscription is interrupted? Can the customer independently audit the basis of a high-consequence recommendation, including the model version and the data it relied upon? The relevant risk is not only excessive licensing fees. It is the loss of operational freedom when the customer cannot reconstruct the knowledge required to understand and maintain its own assets. Open interfaces can reduce this risk, but only if they preserve useful meaning and are implemented effectively; true interoperability is semantic, operational, and contractual, not merely technical connectivity. The consolidation underway at Schneider Electric makes these questions more urgent, not less, and the company’s stated commitment to open ecosystems will be tested precisely on whether its customers can leave with their context intact. [32]
5.3. Measuring Cognition: An Industrial Productivity Scorecard
The economic promise of Infrastructure Cognition requires measurable results, and the measurement problem is harder than it first appears. Counting agent deployments, connected sensors, or model-generated recommendations can exaggerate progress because activity is not the same as productive output, and vendor-reported pilot figures — a 90 percent reduction in unnecessary alerts, a 10 percent potential energy saving — describe defined deployments under favorable conditions rather than fleet-wide outcomes. An effective scorecard should measure at least five dimensions, and it should be applied with the same discipline to a vendor’s pilot as to an operator’s annual report.
Table 8. An industrial productivity scorecard for Infrastructure Cognition
| Dimension | Representative metrics | What it reveals | Common distortion to avoid |
| Reliability | Unplanned downtime; critical-asset availability; mean time to repair; recurrence of identical failures; share of high-consequence incidents prevented or correctly escalated | Whether the system improves physical outcomes | Crediting all avoided failures to the agent regardless of cause |
| Resource productivity | Electricity per unit of useful output; cooling efficiency; material waste; water consumption; equipment utilization; yield | Whether scarce inputs are used better | Comparing across workload types or climates without normalization |
| Decision quality | Share of recommendations validated by specialists; false or non-actionable share; detection-to-intervention time; predicted-versus-actual match rate | Whether the agent’s judgment is trustworthy | Counting recommendations rather than verified outcomes |
| Financial performance | Net operating savings; avoided losses; software and compute cost; maintenance spend; capex avoided through verified optimization; integration cost | Whether the investment pays | Double counting availability gains already captured in output |
| Governance and safety | Unauthorized interventions; policy violations; audit-trail completeness; time to rollback; cyber incidents; human-override performance | Whether authority is bounded and accountable | Treating absence of incidents in a short pilot as proof of safety |
No single index can capture every industry’s outcome. A refinery, a semiconductor fab, a hyperscale datacenter, a distribution utility, and a commercial office building face very different risks and value different things. Still, a common analytical structure makes evaluations more comparable, and it gives boards a language for distinguishing a vendor demonstration from an operating improvement. The most important financial measure is the verified net value of the intervention: demonstrated operational gains and avoided costs, less technology expenditures, integration, supervision, cybersecurity, and any additional operating costs or losses, compared with an appropriate baseline and evaluated over time. Claims about theoretical efficiency gains are not substitutes for realized benefits, and a company that cannot produce the baseline cannot credibly claim the gain.
5.4. The Capital Allocation Question: New Assets Versus Better Assets
The artificial intelligence economy is entering an era in which infrastructure capital allocation has two related purposes. The first is expansion: constructing power plants, datacenters, substations, advanced semiconductor fabs, cooling systems, and industrial facilities, at the $700-billion-plus annual pace the hyperscalers have now guided. The second is optimization: increasing the safe, productive, and reliable use of existing infrastructure. Investment in cognition should not be treated as a replacement for building new assets. A factory already operating close to physical limits cannot reason its way to unlimited output. An electric grid with insufficient transmission capacity cannot create copper, transformers, or permitted rights-of-way through software, and the IEA has estimated that meeting electricity demand through 2030 would require annual grid investment to rise by roughly 50 percent from today’s $400 billion. An AI model cannot compensate for equipment that has exceeded its safe operating life. [47]
Nevertheless, modest changes in reliability or utilization have substantial financial consequences in expensive facilities, and the more expensive the facility, the larger the consequence. Preventing an avoidable outage in a datacenter whose accelerators cost more per rack than many factories cost in total, extending the interval between planned maintenance activities on a transformer with a multi-year replacement lead time, identifying a bottleneck before construction rather than after, or making more effective use of redundant equipment can create value that would otherwise require additional capital spending at a moment when that capital is being raised in the debt and equity markets. Corporate boards should evaluate cognition investments against alternative uses of capital. If a company can obtain a specified throughput increase through validated software and operating changes, that may be economically preferable to a new production line. If the proposed optimization would compromise redundancy, safety, or contractual performance, new physical capacity may still be the better choice. Acquisitions such as the proposed Schneider Electric–PTC combination should be judged by this logic: a high purchase price may be justified if the transaction produces durable additional value for customers and shareholders, but the announcement itself is not evidence of that value. The same principle applies to hyperscalers. The most attractive 2030 infrastructure portfolio may not be the one with the most megawatts of installed capacity. It could be the portfolio with the highest reliable productive output per dollar of total lifecycle expenditure, subject to service quality, security, and community obligations — which is, not coincidentally, the quantity that Microsoft’s tokens-per-megawatt framing and the state-level cost-causation orders are both, from opposite directions, beginning to measure. [6][68]
5.5. Three Scenarios for 2027–2030
The path from the present to 2030 can be described through three scenarios that differ primarily in how far the organizations deploying industrial AI progress along the five levels of Section 3 and how much verified evidence they accumulate along the way.
Table 9. Three scenarios for Infrastructure Cognition, 2027–2030
| Scenario | Characteristic state in 2030 | Dominant level (Table 5) | Where plausible | Principal constraint |
| A. Fragmented intelligence | Pilots, monitoring tools, isolated vendor applications; engineering and operating data still poorly connected; agents improve administrative tasks but rarely influence physical decisions | Levels 1–2 | Legacy-heavy sites; organizations without asset truth or cyber resources | Integration cost, trust, data fragmentation |
| B. Supervised industrial intelligence | Better asset models; selected data sources integrated; simulation used to evaluate changes; agents prepare or execute low-risk workflows under controlled approval; measurable gains in maintenance, reliability, energy, engineering productivity | Levels 2–4 | Broad adoption path across process, discrete, buildings, datacenters | Change management, workforce trust, measurement discipline |
| C. Bounded adaptive infrastructure | Facilities continuously compare digital models with physical operations, propose improvements, verify outcomes, and make authorized changes autonomously; engineers focus on design, safety assurance, exceptions, and model governance | Levels 4–5 | Well-instrumented environments: new hyperscale campuses, advanced fabs, selected process units | Safety cases, regulation, independent interlocks, proof of transferability |
Scenario A, fragmented intelligence, is plausible wherever legacy systems, integration costs, and trust barriers remain high, and the research evidence reviewed in Sections 1 and 3 suggests it will describe a large fraction of the industrial base in 2030 regardless of how quickly the frontier moves. Scenario B, supervised industrial intelligence, is the most plausible broad adoption pathway because it respects industrial safety requirements and existing organizational structures while capturing the gains that the Brynjolfsson mechanism predicts once complementary investments in data and process have been made. Scenario C, bounded adaptive infrastructure, is plausible in selected, well-instrumented environments by 2030 — new hyperscale campuses designed with cognition from the outset, advanced fabs whose instrumentation already approaches the ideal, and process units where the safety case has been built and tested — but it should not be generalized to all critical infrastructure, and the vendors who have shipped Level 5 capability while declining to enable it are behaving rationally. The path among these scenarios depends on more than computing performance. It requires high-quality data, reliable simulation, cybersecure deployment, successful change management, acceptable regulatory treatment, workforce trust, and reproducible business results. The central forecast of this paper is therefore conditional: infrastructure intelligence will become strategically significant where it can consistently improve physical outcomes without undermining safety, accountability, or operational control, and it will remain a demonstration everywhere else.
5.6. An Implementation Agenda for Corporate and Government Leaders
An effective transformation program can be organized into four overlapping phases, each of which produces evidence that the next phase requires. In Phase 1, establishing the asset truth, the organization inventories its most important equipment, identifies authoritative engineering records, maps relationships among systems, reconciles asset identifiers across the four information environments, and evaluates the quality of its operational telemetry, determining which records are missing and which data exchanges are contractually permitted. This phase is unglamorous and frequently skipped, and it is the phase without which everything downstream rests on sand. In Phase 2, proving a limited economic outcome, the organization selects a small number of consequential but manageable applications — alarm classification, fault diagnosis, energy-optimization recommendations, maintenance scheduling — establishes the baseline before deployment, validates recommendations with responsible engineers, and measures actual operating benefits and error costs using the scorecard of Table 8. In Phase 3, introducing bounded agent workflows, it documents permitted actions, escalation thresholds, human-review requirements, rollback procedures, model-update policies, and cybersecurity controls, and it requires evidence that the system can detect uncertainty, handle missing data, and decline unsafe actions before any Level 4 authority is granted. In Phase 4, scaling with accountability, it extends successful capabilities to additional sites only after validating transferability, reports results using comparable measures, and maintains vendor exit options, independent auditability, workforce training, and appropriate public disclosures.
For governors and regional development agencies, the parallel agenda includes transparent large-load forecasting of the kind the Texas audit is now compelling, realistic incentives, local workforce investment, grid modernization, cybersecurity preparedness, and public accountability for resource costs. Governments do not need to design each industrial AI model, but they do need competent institutions able to evaluate its effects when public infrastructure and public money are involved, and the Virginia and Michigan commissions’ willingness to examine large-load contracts in detail suggests what such competence looks like in practice. For universities and engineering schools, the opportunity is to connect computer science with mechanical, electrical, industrial, civil, and systems engineering; MIT’s Energy Initiative launched a Data Center Power Forum in late 2025 to bring faculty and member companies together on exactly these questions, and the University of Southern California’s Annenberg Center devoted a 2026 season of public programming to explaining datacenters to the communities that host them. Industrial AI systems require professionals who understand both machine-learning limitations and the physical consequences of operating decisions, and the pipeline for such professionals is thin. For startups, a large opportunity may lie in the unglamorous work of industrial integration — identifying equipment accurately, converting documentation into usable context, assessing sensor quality, validating digital twins, building permissioned workflows, and measuring outcomes — which can create substantial value even when the startup never trains a frontier model. [79][80]
5.7. A Research Method for Determining Whether Infrastructure Cognition Works
The most important research challenge between 2027 and 2030 will be distinguishing the economic effects of intelligent infrastructure from the effects of ordinary modernization, because the two almost always arrive together. A company that replaces unreliable equipment, installs new sensors, changes maintenance procedures, and deploys an AI application simultaneously may observe an improvement in productivity, and it cannot automatically attribute the full improvement to AI. This is the industrial version of the attribution problem that divides Brynjolfsson and Acemoglu at the macroeconomic level, and it will be resolved, if at all, at the level of individual facilities and comparable cohorts rather than in national statistics.
A credible study should begin with a specific intervention and a clearly defined outcome. An industrial facility might evaluate whether an AI-assisted maintenance system reduces unplanned compressor downtime; the relevant output is not the number of maintenance recommendations but the change in lost production hours or related economic losses after accounting for operating conditions, equipment age, maintenance expenditures, and other concurrent changes. Where possible, studies should use comparable assets or sites, with one group adopting the AI workflow while another continues established procedures, provided the comparison is ethical and operationally acceptable. Researchers should document pre-deployment trends, seasonal demand, production volume, staff changes, and exceptional events. If random assignment is impractical, a carefully designed observational comparison can still be informative, but the uncertainty should be reported rather than hidden. Measurement should include false positives and false negatives: a system that predicts many failures may appear successful if every subsequent defect is credited to it, but excessive warnings cause unnecessary maintenance, production disruption, and operator fatigue, while a system that produces very few warnings may reduce labor costs while missing dangerous conditions. Industrial results also require time. A one-month pilot may capture faster report generation but cannot demonstrate improved equipment life or lower annual maintenance spending, so a meaningful program should specify short-term measures (diagnostic accuracy, time savings), medium-term measures (maintenance quality, downtime, resource productivity, operating costs), and long-term measures (asset lifetime, capital replacement timing, sustained reliability, transferability across sites).
The financial model should separate five elements. The first is gross operating benefit: incremental saleable output, avoided downtime, reduced energy use, lower maintenance cost, or other outcomes that can be documented and appropriately valued. The second is acquisition and integration cost: software licenses, computing equipment, sensors, interfaces, engineering services, model development, and implementation disruption. The third is recurring operating cost: inference, storage, model monitoring, network connectivity, vendor support, cybersecurity, and ongoing validation — a category that the Schneider–NVIDIA–HPE announcement’s explicit concern with “AI consumption costs” suggests vendors are beginning to take seriously. The fourth is human and organizational cost: training, supervision, procedure development, quality assurance, safety review, and responsibility for approving actions. The fifth is risk-adjusted consequence: expected losses or avoided losses associated with errors, outages, cyber incidents, and other operational failures, using defensible estimates rather than invented certainty. The resulting measure can be expressed conceptually as follows. [4]
Verified Net Value = Attributable Operating Benefits − Integration Costs − Recurring Costs − Human Assurance Costs − Risk-Adjusted Additional Losses
The formula is a decision framework, not an accounting standard, and organizations must avoid double counting: improved equipment availability may already be reflected in higher saleable output, and an avoided outage should not be counted again as a separate revenue gain unless the effects are distinct. A hypothetical example shows why the discipline matters. Suppose a facility reports an annualized $5 million benefit from reduced downtime and energy use. If $1.5 million of that improvement came from unrelated equipment upgrades installed in the same year, the gross incremental benefit attributable to the AI program would be no more than $3.5 million under that assumption. If annualized software, infrastructure, support, and assurance costs are $2 million, the preliminary net benefit becomes $1.5 million before considering remaining implementation and risk costs. If the same program generated several incorrect interventions that forced shutdowns, the economic calculation would need to include those losses, and the $1.5 million could become negative. Conversely, a reduction in severe safety incidents might produce substantial value even if ordinary operating-cost savings were modest. In either case, the study should distinguish observed benefits from modeled future benefits, and the numbers here are illustrative, not a claim about any company.
Four further measures complete the method. Decision Verification Rate is the share of consequential recommendations for which an evaluator can establish the information used, the authorization obtained, the action taken, and the outcome observed; a high rate does not prove the decisions were correct, but it demonstrates that the organization possesses a basis for evaluating them. Configuration Correspondence is the proportion of critical assets whose digital records accurately represent the current physical configuration at the level required for a given task, assessed through physical inspection, controlled sampling, engineering reviews, or verified records rather than by asking the AI system whether its own data is correct. Safe Automation Coverage is the proportion of relevant workflows for which automatic action is both technically validated and explicitly permitted within a documented envelope, excluding high-risk actions the organization has deliberately reserved for humans, because broader automation is not necessarily better. Knowledge Portability is the extent to which an asset owner can preserve essential data, relationships, and decision histories when changing suppliers — not a conventional manufacturing metric, but one that may prove decisive for enterprise resilience and long-term bargaining power as the industrial-software market consolidates. Together, these measures provide a basis for comparing different industrial intelligence investments, encourage companies to establish evidence before promising transformational returns, and give public institutions a stronger basis for evaluating claims made in connection with subsidies, permits, and infrastructure commitments.
5.8. The Four Investment Decisions That Could Define 2030
The argument of this paper leads to four capital-allocation decisions that every major industrial company, hyperscaler, and infrastructure investor should examine, and they interact in ways that make sequencing as important as selection. The first concerns the quality of the information foundation: whether capital should be directed to better sensors, engineering-document reconciliation, digital asset registries, maintenance-system integration, or contextual knowledge platforms. These investments provide little immediate publicity, but without them expensive AI models risk operating on incomplete or misleading descriptions of the physical world, and Schneider’s willingness to pay $25 billion across Cognite and PTC for exactly this layer is a signal that at least one sophisticated acquirer believes the foundation is where the value lies.
The second concerns where computation should occur. A company may place advanced reasoning and fleet analytics in cloud infrastructure while keeping time-sensitive control and sensitive operational information at the edge, and different workloads may justify different arrangements. The optimal choice depends on latency, energy use, data movement, cybersecurity, reliability, and total lifecycle cost, not simply the advertised price of inference, and the hyperscalers’ own move toward tokens-per-megawatt accounting suggests that the energy cost of cognition will become a visible line item in industrial budgets. The third concerns whether to build, partner, or acquire. Large industrial incumbents may buy engineering-data companies to extend an installed software portfolio, as Schneider has. Hyperscalers may partner with industrial firms to make AI services accessible to physical operations, as Microsoft has with Siemens. Chip and infrastructure suppliers may build ecosystems, as NVIDIA has with Honeywell, HPE, Dell, and the semiconductor-equipment industry. Startups may specialize in validated simulation, data interoperability, or governance. The best arrangement preserves expertise, customer trust, and the ability to integrate with existing equipment.
The fourth concerns the boundary of autonomous authority. Boards and operators must decide which tasks may be automated, how decisions are tested, who remains accountable, and what evidence justifies expanding permissions from Level 2 to Level 3 to Level 4. This investment is partly technical and partly institutional: approval processes, engineering training, audit systems, safety cases, and organizational responsibilities require attention alongside software licenses. These decisions interact. Investing in a large model before correcting equipment records may create impressive demonstrations but little reliable industrial value. Purchasing a software platform without guaranteeing data portability may create dependence. Pursuing automatic control before defining safe operating envelopes may increase rather than reduce risk. Building a new facility without preserving engineering and operating knowledge may repeat avoidable mistakes in the next generation. The 2030 industrial enterprise should therefore not be defined simply by the number of AI systems it possesses. It should be defined by its ability to make successive capital and operating decisions with better evidence, lower avoidable risk, and greater verified productivity. That is the investment discipline required to turn Infrastructure Cognition from a promising technology narrative into a durable economic capability.
Section 6: What Have We Learned? Seven Pillars of Infrastructure Cognition
Pillar 1 — Connected Industrial Knowledge Is a Strategic Asset
The first lesson is that industrial information becomes substantially more useful when it is connected to the physical objects and processes it describes and interpretable across the four environments in which it was created. A factory’s engineering drawings, telemetry, service records, energy use, and production commitments are not independent economic realities; they are different perspectives on the same operating system, and the value of each rises when the others are available. Organizations therefore need to treat the quality of connected industrial knowledge as part of their productive capability. Accurate context supports better diagnosis, more effective maintenance, stronger simulation, and more informed capital allocation; inaccurate context generates operational mistakes even when the underlying AI model is sophisticated. The practical implication is that data governance is no longer a back-office concern but an engineering requirement, and that asset identity, provenance, configuration management, record quality, and access permissions deserve investment and executive attention. The Schneider Electric agreements concerning PTC and Cognite illustrate the strategic importance that at least one major acquirer now assigns to this layer; their ultimate commercial success remains to be proven through completed transactions and customer outcomes. The core principle is that the value of industrial intelligence depends on the quality of its understanding of real assets, not simply the volume of information it collects. [1][3]
Pillar 2 — The AI Capital Expenditure Cycle Is Moving Toward Operational Productivity
The second lesson concerns investment. The first phase of the AI infrastructure expansion emphasized the physical inputs required to create intelligence — electricity, chips, datacenters, networking, and model training — and that phase is far from over: the hyperscalers’ 2026 guidance exceeds $700 billion, NVIDIA’s datacenter revenue has more than doubled year over year, and the IEA expects the five largest technology companies’ AI investment to rise a further 75 percent in 2026. But the second phase is now visible alongside the first. Investors have begun to distinguish between infrastructure installed and productive value delivered, free cash flow at three of the four largest spenders has come under visible pressure, and the vendors themselves have shifted their units of account toward tokens per megawatt, uptime, and yield. A large datacenter becomes economically stronger when it operates reliably and efficiently. A factory becomes more valuable when its equipment produces more saleable output without disproportionate additional cost. An electricity network becomes more useful when operational insight helps increase dependable service within engineering limits. NVIDIA’s industrial software initiatives, Siemens’ simulation work, Honeywell’s facility-focused products, and Microsoft’s grid-to-chip engineering all show companies seeking applications of AI to operational productivity rather than treating model benchmarks as the only evidence of progress. The core principle is that the next measure of AI infrastructure success is not merely how much capacity was financed but how much dependable economic output that capacity helps produce. [6][7][12][23][43]
Pillar 3 — Industrial AI Agents Must Earn Authority Through Verified Reliability
The third lesson is that autonomy must follow evidence and responsibility. Physical systems differ fundamentally from text-generation environments: a poor paragraph can be revised, while an unsafe intervention involving power distribution, pressure, heat, or rotating equipment may cause injury, damage, service interruption, or environmental harm. Consequently, industrial agents should earn incremental authority through demonstrated competence under controlled conditions, moving from observation and interpretation through simulation and supervised execution to narrowly bounded autonomous action only after appropriate engineering validation. Human authorization is not evidence that a system is primitive; in high-consequence settings, it is a defining feature of a mature and accountable architecture, as are independent safety systems, cybersecure operating boundaries, model traceability, and effective rollback mechanisms. The research literature places the overwhelming majority of industrial agent systems at prototype and early-validation stages, and the most governance-conscious vendor in the October 2026 launches shipped autonomous capability while publicly declining to enable it. The industry should judge agents by the quality and safety of verified decisions, not by promotional claims about independent activity. The core principle is that intelligence without enforceable limits is not a suitable foundation for critical infrastructure. [9][14][15]
Pillar 4 — Energy Is the Binding Layer, and Cognition Cannot Repeal It
The fourth lesson is physical. Every institution that has examined the AI build-out seriously — the IEA, the IMF, the World Bank, the U.S. Department of Energy, Lawrence Berkeley National Laboratory, Stanford’s AI Index, and the United Nations — has reached the same conclusion from a different direction: electricity, and the grid that delivers it, is the constraint against which the rest of the Five-Layer AI Economy must be planned. Datacenter electricity demand grew 17 percent in 2025, is expected to double by 2030, and in the United States may account for nearly half of all demand growth this decade. Transformers, turbines, and transmission are in short supply; interconnection queues contain more than 2,500 gigawatts of stalled projects globally; and the grid investment required through 2030 is roughly 50 percent above current levels. Infrastructure Cognition can make existing assets more productive — flattening load, extending equipment life, improving cooling efficiency, enabling demand response — and that contribution is valuable precisely because it arrives faster than new transmission can be permitted. But it cannot manufacture a transformer, and any strategy that treats intelligence as a substitute for physical capacity will fail. The core principle, in the IEA’s words, is that there is no AI without energy. [23][25][26][28][30]
Pillar 5 — Industrial Competitiveness Will Depend on Integration Across All Five Layers
The fifth lesson is geopolitical and regional. AI leadership cannot be evaluated exclusively by access to the most powerful accelerators or the largest language models, and the narrowing of the U.S.–China model gap documented by Stanford makes that point for anyone who still needed it. Industrial productivity emerges from connecting energy infrastructure, manufacturing equipment, computing resources, engineering models, skilled workers, and trustworthy operating institutions. The Five-Layer AI Economy captures these dependencies, and Infrastructure Cognition adds the possibility that Layers 4 and 5 help improve the performance of Layers 1 through 3 and the industrial systems they support, through chip manufacturing, power management, maintenance scheduling, and infrastructure design. States such as Texas, Arizona, Ohio, Indiana, Michigan, Virginia, Pennsylvania, and California provide different laboratories for this transition, with challenges ranging from energy demand and large-load forecasting to industrial workforce development, grid-cost allocation, local permitting, and public trust. Internationally, the United States, Europe, and China bring different industrial strengths and policy priorities, and the decisive advantage may arise from their capacity to integrate those strengths while maintaining security, reliability, and customer confidence. The core principle is that sustainable AI leadership requires the ability to turn advanced computing into measurable industrial capability. [29][61][68]
Pillar 6 — Trust, Interoperability, and Public Accountability Determine Scalability
The sixth lesson is institutional. Infrastructure Cognition cannot flourish if organizations cannot trust the source of operating information, understand why agents recommend particular actions, move their data between systems, or identify responsibility for failure. Similarly, communities may reasonably oppose infrastructure projects that shift significant costs to other utility customers or obscure water, energy, and environmental effects, and the 2026 record — from the Texas permit halt to Virginia’s cost-causation orders to the United Nations’ transparency initiative — shows that they are doing so with increasing effect. These concerns are not separate from AI productivity. They influence whether facilities can be built, financed, insured, and operated over time. NIST’s work on OT security and AI risk management, the EU’s deferred but unchanged high-risk requirements, and the state-level accountability measures together illustrate the need for safeguards around intelligent infrastructure that are embedded into technical architecture rather than added after deployment. Interoperable systems, independently verifiable results, transparent accountability, and strong human oversight may therefore become competitive strengths: they reduce integration risk, enable more trustworthy deployment, and help companies maintain the social and regulatory permission to operate. The core principle is that the intelligence of infrastructure must be judged by its contribution to safety, productivity, resilience, and the public systems on which it relies. [51][52][67][70][73]
Pillar 7 — The Productivity Debate Will Be Settled in Plants, Not in Papers
The seventh lesson concerns evidence. The disagreement between those who see an AI productivity take-off becoming visible and those who expect gains that are nontrivial but modest is, at bottom, a disagreement about whether AI will do new things or merely do existing tasks more cheaply. Infrastructure Cognition is the domain in which that question will be answered most concretely, because the outcomes — downtime avoided, yield improved, energy saved, outages prevented, capital deferred — are physical, measurable, and attributable at the level of individual assets if the measurement is designed properly. The World Bank’s observation that the window to get the foundations right is narrow applies to industrial enterprises as much as to developing countries. The Verified Net Value framework, the Decision Verification Rate, Configuration Correspondence, Safe Automation Coverage, and Knowledge Portability proposed in Section 5 are offered as instruments for that measurement. Their purpose is not to constrain ambition but to discipline it: to ensure that by 2030 the industrial economy can distinguish the facilities where cognition has created durable value from those where it has created impressive demonstrations. The core principle is that the burden of proof rests with the intervention, and that an industrial AI program unable to produce its own baseline has not yet earned the right to claim a gain. [44][46][77]
“The window to get this right is narrow.”
— Gaurav Nayyar, Director, World Development Report 2026, World Bank Group [77]
Conclusion: Why Infrastructure Cognition Defines the Next Industrial Phase of the Five-Layer AI Economy
The October 2026 developments surrounding Schneider Electric, PTC, NVIDIA, HPE, Honeywell, Siemens, and Microsoft offer an early view of a potentially important transformation. Companies that once focused primarily on selling electrical equipment, industrial control systems, engineering software, or computing platforms are moving toward products and partnerships that connect these capabilities, and their commercial objective is increasingly to help customers understand complex physical environments and use that understanding to improve operations. The proposed $22.6 billion Schneider Electric–PTC acquisition is particularly revealing because it assigns substantial strategic value to the connection between engineering knowledge and operational performance, and because Schneider raised €6 billion of fresh equity to pursue it. The company’s separate agreement to acquire Cognite reinforces the importance of contextualizing industrial data, while its collaboration with NVIDIA and HPE extends the discussion toward industrial edge agents that monitor, reason, and optimize under human control. Siemens and NVIDIA present another route through simulation and digitally connected manufacturing, and NVIDIA’s work with TSMC and the semiconductor-equipment industry carries that route into the fabs that produce the chips themselves. Honeywell’s October launch demonstrates how similar principles may be applied to buildings and existing facilities, and its decision to withhold autonomy demonstrates how the principles should be governed. None of these announcements, individually, proves that a fully autonomous industrial economy is imminent. Together, they show where several of the world’s most capable technology and industrial companies are investing attention, capital, and reputation. [1][3][4][7][8][34]
This paper has argued that the next phase of AI development may be defined partly by a change in the economic role of information. Industrial data has traditionally been collected to satisfy a particular operational requirement: monitor a machine, record a repair, calculate energy costs, document a design. Infrastructure Cognition proposes that these separate records can become a connected, continuously updated body of machine-readable knowledge that allows models and agents to interpret the operating life of a physical asset. Its value depends on more than sophisticated algorithms. The asset information must be trustworthy; simulations must correspond to actual conditions; decision authority must be explicit; cybersecurity must be integral; and the consequences of interventions must be measured. An intelligent system that cannot recognize its own uncertainty is not ready to be trusted with critical operations. A facility that optimizes its private efficiency while transferring costs to neighboring communities has not achieved a complete understanding of its place within the broader infrastructure system.
The Five-Layer AI Economy explains the enormous chain of investment enabling artificial intelligence: energy supplies computing facilities, semiconductor chips provide processing capacity, datacenters organize those resources, models produce machine intelligence, and applications and agents bring intelligence into economic activity. Infrastructure Cognition enriches that structure with a consequential feedback loop in which the models and agents at the top of the stack contribute to improving the design, utilization, reliability, and efficiency of the physical infrastructure below. This feedback loop changes how investors and policymakers should measure progress. The number of accelerators acquired is meaningful, but it does not alone establish economic productivity. The megawatts connected to a datacenter are important, but they do not reveal whether its cooling, computing, or maintenance operations are optimized. A factory’s construction expenditure matters, but its lasting economic contribution depends on output, reliability, workforce capability, and the intelligence with which it is operated. The hyperscalers’ own shift toward tokens per megawatt, the regulators’ shift toward cost causation and verified load, and the academic literature’s insistence on configuration fidelity and human oversight are all expressions of the same underlying realization: that the AI economy is becoming, once again, an economy of physical performance.
By 2030, some industrial assets may be surrounded by systems capable of continually comparing engineering expectations with actual conditions, diagnosing deviations, simulating alternatives, requesting approvals, executing limited authorized responses, and verifying results. Other facilities will continue relying on conventional automation and human expertise, and many will occupy the supervised middle ground that this paper regards as the most plausible broad outcome. All of these realities will coexist, and none should be forced into an exaggerated narrative of universal machine autonomy or an equally exaggerated narrative of permanent stagnation.
The title Infrastructure Cognition is therefore deliberate. Infrastructure identifies the physical foundation of the economy — datacenters, electric grids, factories, semiconductor plants, energy systems, and related industrial assets. Cognition identifies an emerging capacity to interpret the relationships among those assets, explain changes, anticipate failures, and support better decisions through connected knowledge and carefully governed AI. It is a more precise description than a general claim that infrastructure is becoming digital or automated. Digitalization records information. Automation executes defined procedures. Infrastructure Cognition seeks to connect evidence, physical context, simulation, judgment, and verified action across the asset lifecycle, within limits that engineers can defend and communities can trust. The greatest long-term opportunity may not be a future in which machines operate without people. It may be a future in which engineers, operators, companies, and public institutions understand the physical systems on which civilization depends with a level of depth, continuity, and responsiveness that was previously impractical — and in which the enormous capital now flowing into the Five-Layer AI Economy is judged, at last, by what it helps the physical world produce. That is the research premise of Infrastructure Cognition, and why it belongs at the center of the Five-Layer AI Economy’s 2027–2030 agenda.
Endnotes and Sources:
[1] Schneider Electric and PTC, “Schneider Electric to Acquire PTC, Creating the Next Level of Energy and Industrial Intelligence,” press release, October 5, 2026. https://investor.ptc.com/resources/news/news-details/2026/Schneider-Electric-to-acquire-PTC-creating-the-next-level-of-Energy-and-Industrial-Intelligence/default.aspx
[2] Reuters News via Capital.com, “Schneider Electric To Buy PTC For $205 Per Share,” October 5, 2026 (deal terms, financing structure, synergies). https://capital.com/en-int/news/schneider-electric-to-buy-ptc-for-205-per-share
[3] Cognite and Schneider Electric, “Schneider Electric Announces Agreement to Acquire Cognite,” June 30, 2026. https://resources.cognite.com/en/company/newsroom/schneider-electric-announces-agreement-to-acquire-cognite
[4] Schneider Electric, “Schneider Electric, NVIDIA & HPE Announce Strategic Collaboration to Accelerate Physical AI for Process Industries,” GlobeNewswire, Berlin, October 22, 2026. https://www.se.com/ie/en/about-us/newsroom/news/press-releases/schneider-electric-nvidia-hpe-announce-strategic-collaboration-to-accelerate-physical-ai-for-process-industries-6ad9dfc2a43b7054780ca48f/
[5] Manufacturing Digital, “Schneider, NVIDIA & HPE: Physical Edge AI For Factory Floors,” October 2026. https://manufacturingdigital.com/news/schneider-nvidia-hpe-physical-edge-ai-for-factory-floors
[6] Microsoft, “Microsoft to Expand Azure Maia System to Multi-Gigawatt Deployments,” October 12, 2026; see also Data Center Knowledge, “Microsoft Pushes Custom AI Chips with Azure Maia 300.” https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/
[7] NVIDIA, “NVIDIA and Semiconductor Leaders Advance AI Across Chip Design and Manufacturing for AI-Defined Fabs,” October 14, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-and-Semiconductor-Leaders-Advance-AI-Across-Chip-Design-and-Manufacturing-for-AI-Defined-Fabs/default.aspx
[8] Honeywell Technologies, “Honeywell Technologies Launches Forge Cognition with NVIDIA and Dell Technologies, Advancing Physical AI Across Buildings and Industrial Operations,” October 14, 2026 (includes John Roese quotation). https://www.honeywell.com/us/en/news/press-releases/2026/10/honeywell-technologies-launches-forge-cognition-with-nvidia-and-dell-technologies-advancing-physical-ai-across-buildings-and-industrial-operations
[9] Fortune, reporting on Honeywell Forge Cognition and remarks by CTO Suresh Venkatarayalu, as summarized by AI Weekly, October 14, 2026. https://aiweekly.co/alerts/honeywell-launches-forge-cognition-building-ai-platform-with-nvidia-and-dell
[10] Sharecast News via Hargreaves Lansdown, “Schneider Electric raises outlook as data centre boom drives record sales,” October 16, 2026 (Q3 2026 results). https://www.hl.co.uk/shares/stock-market-news/company–news/schneider-electric-raises-outlook-as-data-centre-boom-drives-record-sales
[11] Datacentres.com, “Hyperscaler CapEx Trends: 2026 Guidance Passes $700 Billion,” compiling Amazon, Alphabet, Microsoft, and Meta July 2026 earnings releases and calls (includes Andy Jassy quotation). https://www.datacentres.com/guides/hyperscaler-capex-trends
[12] NVIDIA, “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027,” August 26, 2026 (includes Jensen Huang quotation). https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/
[13] Fei Tao, “Advancements and challenges of digital twins in industry,” Nature Computational Science 4, 169–177, March 26, 2024. https://www.nature.com/articles/s43588-024-00603-w
[14] David Kube, Mario Thron et al., “Foundation-Model-Based Agents in Industrial Automation: Purposes, Capabilities, and Open Challenges,” arXiv:2605.02592, May 4, 2026 (Institute for Industrial Automation and Software Engineering, University of Stuttgart; ifak). https://arxiv.org/abs/2605.02592
[15] Daniela Rus, “When AI meets robotics: A conversation with Daniela Rus,” Capgemini Research Institute, 2025. https://www.capgemini.com/dk-en/insights/research-library/a-conversation-with-daniela-rus/
[16] VinFuture Prize, “Prof. Daniela Rus: A new voice in the VinFuture Prize Council – Honoring AI and Robotics revolution,” 2025 (physical AI including power grids, sensors, and satellites). https://vinfutureprize.org/?p=11934
[17] SecurityWeek, “OT Security Guidance: NIST Drafts Updated Guide, CISA/FBI Advise on ICS Integrators,” September 2026 (OT asset-inventory survey figures). https://www.securityweek.com/ot-security-guidance-nist-drafts-updated-guide-cisa-fbi-advise-on-ics-integrators/
[18] Michael Grieves and John Vickers, “Digital Twin: Mitigating Unpredictable, Undesirable Emergent Behavior in Complex Systems,” in Transdisciplinary Perspectives on Complex Systems, Springer, 2017. https://link.springer.com/chapter/10.1007/978-3-319-38756-7_4
[19] Fei Tao, He Zhang, Ang Liu, and A. Y. C. Nee, “Digital Twin in Industry: State-of-the-Art,” IEEE Transactions on Industrial Informatics 15(4), 2405–2415, 2019. https://ieeexplore.ieee.org/document/8477101
[20] Alexander Wuttke, Bhakti Stephan Onggo, and Markus Rabe, “Review and classification of challenges in digital twin implementation for simulation-based industrial applications,” Proceedings of the 2025 Winter Simulation Conference, Seattle, December 2025. https://eprints.soton.ac.uk/504059
[21] “Advances in digital twin technology in industry: A review of applications, challenges, and standardization,” Journal of Intelligent Construction 3, 2025. https://www.sciopen.com/article/10.26599/JIC.2025.9180083
[22] NVIDIA, “NVIDIA and TSMC Bring AI Into Fabs to Advance Semiconductor Design and Manufacturing,” NVIDIA GTC Taipei, 2026. https://nvidianews.nvidia.com/news/nvidia-and-tsmc-bring-ai-into-fabs-to-advance-semiconductor-design-and-manufacturing
[23] International Energy Agency, “Key Questions on Energy and AI,” April 2026, as reported by W.Media, “Data center power demand will double by 2030: IEA,” April 20, 2026 (includes Fatih Birol quotation). https://w.media/data-center-power-demand-will-double-by-2030-iea/
[24] Microsoft Azure Infrastructure Blog, “Powering Azure Infrastructure from Grid to Chip,” October 13, 2026. https://techcommunity.microsoft.com/blog/AzureInfrastructureBlog/powering-azure-infrastructure-from-grid-to-chip/4563890
[25] U.S. Department of Energy, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers,” announcing Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, December 2024. https://content.govdelivery.com/accounts/USEERE/bulletins/3c88056
[26] International Energy Agency, “Energy and AI,” April 2025. https://www.iea.org/reports/energy-and-ai
[27] Enlit World, “AI and data centre electricity use continue to surge, IEA finds,” April 2026 (485 TWh to 950 TWh projection; AI-focused demand to 465 TWh). https://www.enlit.world/library/ai-and-data-centre-electricity-use-continue-to-surge-iea-finds
[28] Christian Bogmans, Patricia Gomez-Gonzalez, Giovanni Melina, and Sneha Thube, “AI Needs More Abundant Power Supplies to Keep Driving Economic Growth,” IMF Blog, May 13, 2025, summarizing the April 2025 World Economic Outlook and the IMF working paper “Power Hungry: How AI Will Drive Energy Demand.” https://www.imf.org/en/Blogs/Articles/2025/05/13/ai-needs-more-abundant-Power-supplies-to-keep-driving-economic-growth
[29] Stanford Institute for Human-Centered Artificial Intelligence, “2026 AI Index Report,” April 13, 2026; summary of infrastructure and capability findings via Burges Salmon, “AI in 2026 – What Does Stanford’s AI Index Tell Us?” https://hai.stanford.edu/ai-index/2026-ai-index-report
[30] U.S. Department of Energy, Office of Electricity, “DOE’s Office of Electricity Publishes 2026 Draft National Transmission Needs Study,” July 9, 2026; Catherine Jereza quotation via T&D World, “DOE Study Finds Significant U.S. Transmission Expansion Needed.” https://www.energy.gov/oe/articles/does-office-electricity-publishes-2026-draft-national-transmission-needs-study
[31] Decrypt, “AI’s Progress Now Depends on ‘World Models’ That Grasp Physical Reality,” reporting Fei-Fei Li’s essay “From Words to Worlds: Spatial Intelligence Is AI’s Next Frontier,” November 2025. https://decrypt.co/348051/ai-progress-now-depends-world-models-grasp-physical-reality
[32] Schneider Electric, “2026 Half Year Results: Record H1 driven by strong execution and broad-based demand, Full Year Target upgraded,” July 30, 2026; analysis via ARC Advisory Group. https://www.arcweb.com/blog/schneider-electric-q2-growth-highlights-data-center-automation-industrial-software-momentum
[33] Schneider Electric, “Schneider Electric announces the success of its €6.0 billion capital increase without shareholders’ preferential subscription rights to partially finance the acquisition of PTC Inc.,” October 20, 2026. https://www.webdisclosure.com/press-release/schneider-electric-epa-su-schneider-electric-announces-the-success-of-its-60-billion-capital-increase-without-shareholders-preferential-subscription-rights-to-partially-finance-the-acquisition-of-ptc-inc-eERB5BBo9Lz
[34] Siemens, “Siemens Unveils Technologies to Accelerate the Industrial AI Revolution at CES 2026,” January 6, 2026. https://press.siemens.com/global/en/pressrelease/siemens-unveils-technologies-accelerate-industrial-ai-revolution-ces-2026
[35] PepsiCo, “PepsiCo Announces Industry-First AI and Digital Twin Collaboration with Siemens and NVIDIA,” January 6, 2026. https://www.pepsico.com/newsroom/press-releases/2025/pepsico-announces-industry-first-ai-and-digital-twin-collaboration-with-siemens-and-nvidia
[36] Siemens AG, “Record third quarter – Outlook raised,” Q3 fiscal 2026 press release (quarter ended June 30, 2026), including remarks by Roland Busch. https://press.siemens.com/global/uk/node/7434
[37] Microsoft, “Industrial Intelligence Unlocked: Microsoft at Hannover Messe 2026,” April 16, 2026. https://www.microsoft.com/en-us/microsoft-cloud/blog/manufacturing/2026/04/16/industrial-intelligence-unlocked-microsoft-at-hannover-messe-2026/
[38] Microsoft Learn, “What Is Azure IoT Operations?” 2026 documentation. https://learn.microsoft.com/azure/iot-operations/overview-iot-operations
[39] Craig Resnick, “Schneider Electric, NVIDIA, and HPE: Industrial AI Agents at the Edge for Process Industries,” ARC Advisory Group, October 2026. https://www.arcweb.com/node/164418
[40] ChemXplore, “Honeywell launches Forge Cognition with NVIDIA and Dell to add agentic AI to buildings and industrial operations,” October 2026 (pilot sites and reported results). https://chemxplore.com/news/honeywell-forge-cognition-physical-ai
[41] Amazon, “Amazon.com Announces Fourth Quarter Results,” February 2026 (initial 2026 capital-expenditure expectation of approximately $200 billion). https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Fourth-Quarter-Results/
[42] Amazon Web Services and NVIDIA, “AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI,” August 26, 2026. https://press.aboutamazon.com/aws/2026/8/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai
[43] NextMSC Insights, “Hyperscaler Capex Comparison 2026: Microsoft, Google, Amazon and Meta,” October 2026 (capex as share of operating cash flow; debt and equity issuance). https://www.nextmsc.com/industry-deep-dive/hyperscaler-capex-comparison-2026
[44] Fortune, “One of Stanford’s original AI gurus says productivity liftoff has begun after doubling in 2025 amid transition to ‘harvest phase’ along J-curve,” February 15, 2026, reporting Erik Brynjolfsson’s Financial Times op-ed “The AI productivity take-off is finally visible.” https://fortune.com/2026/02/15/ai-productivity-liftoff-doubling-2025-jobs-report-transition-harvest-phase-j-curve/
[45] Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics 13(1), 333–372, 2021. https://www.aeaweb.org/articles?id=10.1257/mac.20180386
[46] Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487, 2024; published in Economic Policy 40(121), 13–58, 2025. https://www.nber.org/papers/w32487
[47] Ron Bousso, “US AI boom faces electric shock,” Reuters column via New Straits Times, February 26, 2026 (ERCOT 226 GW queue; Cleanview on-site generation count; IEA grid-queue and investment figures). https://www.nst.com.my/amp/opinion/columnists/2026/02/1385522/us-ai-boom-faces-electric-shock
[48] “Osprey: Production-Ready Agentic AI for Safety-Critical Control Systems,” arXiv:2508.15066, versions August–December 2025. https://arxiv.org/abs/2508.15066
[49] “Large Language Models in Sensor-Driven Control Systems: Architectures, Challenges, and Opportunities,” peer-reviewed review, 2026 (PMC). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13417370/
[50] “Human-in-the-loop and large language models in smart manufacturing: A systematic scoping review,” Journal of Manufacturing Systems, accepted April 9, 2026. https://documentserver.uhasselt.be/bitstream/1942/49037/1/2026_04_JMS_LLM%20and%20HITL.pdf
[51] National Institute of Standards and Technology, SP 800-82 Revision 4, “Guide to Operational Technology (OT) Security,” Initial Public Draft, September 21, 2026 (comments due November 30, 2026). https://csrc.nist.gov/pubs/sp/800/82/r4/ipd
[52] National Institute of Standards and Technology, “AI Risk Management Framework,” including critical-infrastructure profile work. https://www.nist.gov/itl/ai-risk-management-framework
[53] U.S. Department of Energy, “Energy Department Announces Speed to Power Investments Across 26 States to Lower Electricity Costs and Improve Grid Reliability,” September 24, 2026. https://www.energy.gov/articles/energy-department-announces-speed-power-investments-across-26-states-lower-electricity
[54] Reuters via The Star, “Stressed US grid forcing data centers to get more flexible,” March 27, 2026 (includes Matt O’Connor quotation; DOE PJM backup-generation directive; Google utility contracts). https://www.thestar.com.my/tech/tech-news/2026/03/27/stressed-us-grid-forcing-data-centers-to-get-more-flexible-
[55] Constance Crozier and Matthew Liska, review on data-center flexibility for the electric grid, Georgia Institute of Technology EPIcenter, 2025. https://hg.gatech.edu/node/689280
[56] “Electricity Demand and Grid Impacts of AI Data Centers: Challenges and Prospects,” arXiv:2509.07218, 2025. https://arxiv.org/abs/2509.07218
[57] IT Brief UK, “Schneider Electric teams with NVIDIA, HPE on industrial AI,” October 2026 (includes Matías Jory, CMPC, quotation). https://itbrief.co.uk/story/schneider-electric-teams-with-nvidia-hpe-on-industrial-ai
[58] U.S. Department of Commerce, Bureau of Industry and Security, “Department of Commerce Revises License Review Policy for Semiconductors Exported to China,” January 13, 2026. https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china
[59] European Commission, “AI Gigafactories,” including the July 2026 call. https://commission.europa.eu/topics/competitiveness/competitiveness-coordination-tool-projects/ai-gigafactories_en
[60] State Council of the People’s Republic of China, “China Mulls Policies to Accelerate Smart Manufacturing,” October 20, 2026; “China Unveils Plan to Deepen Integration of Industrial Internet, AI,” January 7, 2026. https://english.www.gov.cn/news/202610/20/content_WS6ad76e83c6d00ca5f9a0dda6.html
[61] Office of the Texas Governor, “Governor Abbott Directs TCEQ To Halt Data Center Permits,” September 21, 2026; “Governor Abbott Directs Comprehensive Data Center Audit,” August 3, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-tceq-to-halt-data-center-permits
[62] Bracewell LLP, “Governor Abbott Orders TCEQ to Pause All Permit Issuance for Data Center Projects,” September 23, 2026 (audit timing). https://www.bracewell.com/resources/governor-abbott-orders-tceq-to-pause-all-permit-issuance-for-data-center-projects/
[63] TSMC, “TSMC Arizona,” including 2026 expansion intentions. https://www.tsmc.com/static/abouttsmcaz/index.htm
[64] Ohio Department of Development, “Ohio Announces Finalization of $600 Million Onshoring Incentive Grant Agreement with Intel,” June 23, 2023. https://development.ohio.gov/home/news-and-events/all-news/2023-0623-ohio-announces-finalization-of-600m-onshoring-incentive-grant-agreement-with-intel
[65] Indiana Economic Development Corporation, “Gov. Braun Announces 300 High-Wage Jobs at New Amazon Manufacturing Facility,” September 24, 2026; “AWS Plans $11 Billion Northern Indiana Datacenter Campus,” April 25, 2024. https://iedc.in.gov/events/news/details/2026/09/24/gov.-braun-announces-300-high-wage-jobs-at-new-amazon-manufacturing-facility
[66] Office of Michigan Governor Gretchen Whitmer, “Michigan Affordable and Responsible Growth Action Plan,” July 15, 2026; Michigan Public Service Commission, “Data Center and Electric Service,” October 1, 2026 conditional approval. https://www.michigan.gov/whitmer/news/press-releases/2026/07/15/gov-whitmer-data
[67] Office of California Governor Gavin Newsom, “Governor Newsom Signs Most Comprehensive Data Center Laws in the Nation,” September 21, 2026. https://www.gov.ca.gov/2026/09/21/governor-newsom-signs-most-comprehensive-data-center-laws-in-the-nation-providing-communities-more-control-on-water-electricity-and-land-use/
[68] Office of Virginia Governor Abigail Spanberger, “Data Center Accountability Standards and Executive Order 22,” September 18, 2026; “SCC Orders Data Centers to Cover Dedicated Transmission Infrastructure Costs,” August 5, 2026. https://www.governor.virginia.gov/newsroom/news-releases/2026/september-releases/name-1123696-en.html
[69] Commonwealth of Pennsylvania, “Governor Shapiro Signs Executive Order on Data Center Development,” August 18, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen
[70] National Law Review / K&L Gates, “EU Digital Omnibus on AI Enters Into Force,” Regulation (EU) 2026/1744, published July 24, 2026, in force July 27, 2026. https://natlawreview.com/article/eu-digital-omnibus-ai-enters-force
[71] World Bank Group, “World Development Report 2026: The Promise of Artificial Intelligence,” August 4, 2026, as reported by BusinessDay (includes Indermit Gill quotation). https://businessday.ng/?p=1171936
[72] International Business Times UK, “‘It’s Not Coming, Trust Me’: World Bank Chief Says Big AI Data Centres Will Stay in Rich Grids,” 2026 (remarks by Ajay Banga). https://www.ibtimes.co.uk/emerging-markets-ai-data-centre-infrastructure-1821787
[73] E&E News, “UN chief urges AI industry to quit fossil fuels,” June 23, 2026 (António Guterres, London Climate Action Week; AI Environmental Transparency Initiative); see also Asharq Al-Awsat, “UN Chief Urges AI Firms ‘Come Clean’ on Environmental Footprint.” https://www.eenews.net/articles/un-chief-urges-ai-industry-to-quit-fossil-fuels/
[74] Perkins Coie, “How the White House’s AI Action Plan Aims To Ensure American Leadership in AI,” July 2025 (Executive Order 14318, Accelerating Federal Permitting of Data Center Infrastructure). https://perkinscoie.com/insights/update/how-white-houses-ai-action-plan-aims-ensure-american-leadership-ai
[75] Arnold & Porter, “Data Centers and Permitting Reform: Is Artificial Intelligence Up to the Challenge?” July 2025, quoting “Winning the Race: America’s AI Action Plan,” The White House, July 23, 2025. https://www.arnoldporter.com/en/perspectives/blogs/environmental-edge/2025/07/data-centers-and-permitting-reform
[76] International Monetary Fund, “Power Hungry: How AI Will Drive Energy Demand,” IMF Working Paper, April 2025, as reported by BusinessWorld. https://www.bworldonline.com/?p=667775
[77] Prime Business Africa, “World Bank Sees AI as Growth Catalyst for Developing Economies, Warns of Digital Divide,” August 5, 2026 (includes Gaurav Nayyar quotation). https://www.primebusiness.africa/?p=253411
[78] Communications Today, “Schneider beats profit expectations as data center demand offsets weak dollar,” February 2026 (includes Hilary Maxson quotation). https://www.communicationstoday.co.in/?p=397171
[79] MIT News, “MIT Energy Initiative launches Data Center Power Forum,” November 7, 2025. https://news.mit.edu/2025/mit-energy-initiative-launches-data-center-power-forum-1107
[80] University of Southern California, Annenberg Center for Climate Journalism and Communication, “Electric Futures” podcast, Season 3, Episode 7: “Everything You Need to Know About Data Centers,” recorded April 2026. https://app.podcastguru.io/podcast/1718960690/episode/episode-7-electric-futures-live-everything-you-need-to-know-about-data-centers-056953ef35191455e061d021bf1ef5aa
