Introduction: The Day the AI Bottleneck Moved From the GPU to the Fiber
On September 9, 2026, an unusually revealing number emerged from a semiconductor company that most investors would not instinctively place at the center of the artificial-intelligence boom. Speaking at Citi’s Global TMT Conference in New York, STMicroelectronics Chief Financial Officer Lorenzo Grandi disclosed that approximately 80 percent of the more than $2 billion in AI-datacenter revenue the company expects to generate in 2027 will come from chips used in fiber-optic data links, while only about 20 percent is expected to come from chips associated with server cooling and power conversion, and even that smaller portion is not expected to become meaningful before 2028.[1] The disclosure provided the first detailed breakdown of a revenue target that the company had already raised twice during 2026, moving from an original ambition of “nicely above $500 million” to approximately $1 billion for 2026 and then to well over $2 billion for 2027, as demand for AI connectivity silicon repeatedly exceeded the company’s own planning assumptions.[2] In other words, one of Europe’s largest semiconductor companies looked at the next phase of artificial-intelligence infrastructure and saw its most immediate, most bankable opportunity not in the processors that calculate intelligence, but in the components that allow those processors to communicate.
“Short term will be mainly driven by the connectivity portion.”
— Lorenzo Grandi, Chief Financial Officer, STMicroelectronics, at Citi’s Global TMT Conference, September 9, 2026 [1]
That single sentence, delivered in the dry cadence of a finance chief walking analysts through segment guidance, deserves to be read as an industrial signal rather than a corporate footnote. Behind it sits a portfolio of silicon photonics built on the company’s PIC100 technology platform, BiCMOS electronic integrated circuits for optical transceivers, and long-term customer agreements that include capacity commitments, volume commitments, pricing structures, and cash advances that give the company revenue visibility beyond 2027—an arrangement that resembles the way memory makers and foundries contract with hyperscalers far more than the way a traditional European industrial chipmaker historically sold microcontrollers into automobiles and washing machines.[3] When Morningstar’s analysts examined the same disclosures, they concluded that even the $2 billion figure for 2027 might prove conservative, and raised their own revenue estimates for the company above prior consensus specifically because of the silicon-photonics and optical-content opportunity attached to Nvidia’s most advanced rack-scale architectures.[4]
The disclosure would have been interesting standing alone. Coming exactly one day after another major semiconductor announcement, however, it looked less like a coincidence and more like the second data point on a trend line. On September 8, 2026, Qualcomm and Amazon announced a multigenerational collaboration covering customized silicon for Amazon’s large-scale AI datacenters, with the initial technical focus on AI inference.[5] The commercial framework attached to the collaboration could involve as much as $60 billion of Qualcomm server products and services over the coming decade, structured through a warrant granting Amazon rights to acquire up to 25 million Qualcomm shares at $161.26 per share, vesting in tranches tied to binding orders and commercial milestones through September 2036.[6] The headline number dominated coverage, as headline numbers always do. But hidden inside the much larger financial architecture was an equally consequential technical detail: the two companies will also collaborate on advanced optical connectivity for AI datacenters, with links extending to 1.6 terabits per second and future generations beyond that, built on Qualcomm’s high-speed SerDes and optical digital-signal-processing technologies.[5] Analysts at Futurum observed that of Qualcomm’s four datacenter product lines, the two closest to actual revenue are precisely custom silicon and connectivity—the compute and the communication, designed together, sold together, and deployed together.[7]
Neither announcement is principally about replacing Nvidia, and neither should be read that way. What they reveal is something more fundamental about what happens when Nvidia, Amazon, Google, Microsoft, Meta, OpenAI, Anthropic, xAI, and the emerging neocloud industry attempt to assemble hundreds of thousands—and eventually millions—of accelerators into coordinated computing systems. Buying more GPUs increases theoretical computation. But theoretical computation is economically useful only when processors can continuously obtain model parameters, activations, expert outputs, memory contents, and intermediate results from other parts of the machine quickly enough to remain productive. A modern frontier-AI cluster is therefore becoming less like an enormous collection of independent computers and more like a single distributed machine whose internal distances may extend across racks, buildings, campuses, and eventually metropolitan regions. Every additional accelerator creates potential computation, but it simultaneously creates additional communication, and the larger the machine becomes, the harder it becomes to keep all of its components synchronized.
This produces a deceptively simple problem with genuinely enormous financial consequences. A modern GPU can perform an extraordinary number of mathematical operations per second, and each successive generation performs more of them. But if that GPU spends a growing percentage of those seconds waiting for data to arrive from another GPU, another rack, or another building, then its owner has purchased expensive theoretical FLOPs that cannot be fully monetized. The industry traditionally addressed short-distance communication with electrical connections, and copper remains extraordinarily effective where distances are small, because it is mature, inexpensive, and well understood. But as signaling rates climb through 100, 200, and 400 gigabits per second per lane, electrical transmission encounters increasingly severe problems involving attenuation, heat, signal integrity, power consumption, and physical reach. Moving ever larger quantities of information electrically across ever larger machines consumes increasing amounts of the same scarce electricity that hyperscalers would rather devote to useful computation—electricity that the International Energy Agency projects will push global datacenter consumption from roughly 415 terawatt-hours in 2024 to approximately 945 terawatt-hours by 2030, more than the entire present-day electricity consumption of Japan.[35]
Light changes that equation. Optical communication can transport enormous volumes of information over greater distances with fundamentally different power and signal-loss characteristics than electrical interconnects, and the academic community saw this coming long before the market did. Stanford’s David A. B. Miller, one of the founding intellectual figures of optical interconnect research, argued nearly a decade before the current buildout that the energy dissipated in computing systems was progressively shifting from logic operations to interconnect, and that this shift would eventually become the binding limit on the growth of information processing itself.[31]
“Optics may be the only scalable solution to such problems.”
— David A. B. Miller, Professor of Electrical Engineering, Stanford University [31]
Nvidia itself has now effectively institutionalized this transition. Its Vera Rubin platform, which the company announced was ramping into full production at GTC Taipei in May 2026, introduces Spectrum-X Ethernet Photonics, the world’s first co-packaged-optics-based switches with 200 gigabit-per-second SerDes, integrating optical engines directly into the switch package to serve what Nvidia explicitly describes as future million-GPU AI factories. Nvidia says the technology delivers five times better power efficiency, five times longer sustained AI uptime, and 1.3 times faster deployment than networks built from traditional pluggable transceivers, with CoreWeave, Lambda, and Oracle Cloud Infrastructure among the first ecosystem adopters.[8] The industry is simultaneously developing alternative architectures and open standards. The Ultra Accelerator Link Consortium’s UALink 200G 1.0 specification provides 200 gigabit-per-second-per-lane connectivity intended to connect as many as 1,024 accelerators inside a single AI computing pod with load/store memory semantics.[11] Broadcom is shipping its third-generation co-packaged-optics Ethernet switch. Marvell has assembled, through acquisition and internal development, an end-to-end portfolio of 1.6-terabit optical DSPs and a Photonic Fabric for multi-rack AI scale-up. Ayar Labs raised $500 million in March 2026—the largest financing any photonics company has ever completed—explicitly to move co-packaged optical I/O into high-volume production.[26] And Nvidia, in a move almost without precedent for a fabless chip designer, invested $4 billion of its own capital directly into the laser and optics manufacturers Coherent and Lumentum to secure the photonic supply chain beneath its future architectures.[23]
These developments, taken together, suggest that the next era of AI infrastructure cannot be adequately understood by counting GPUs. The relevant question becomes whether those GPUs can operate as one machine. That is the foundation of the Photon Constraint. Within the Five-Layer AI Economy—Energy, Chips, Datacenters, Models, and Applications and Agents—optical connectivity sits in an unusual position. It originates primarily inside Layer 2, the semiconductor and systems layer, but it determines the efficiency of Layer 3 datacenters, the practical scalability of Layer 4 models, and ultimately the economics of Layer 5 applications and autonomous agents, while continuously drawing down the finite electricity budget governed by Layer 1. This paper argues that between 2027 and 2030, artificial-intelligence infrastructure will increasingly encounter a Photon Constraint: a condition in which additional accelerator capacity produces diminishing economic value unless optical bandwidth, switching capacity, memory movement, and network topology scale alongside computation. The next AI shortage, therefore, may not simply be GPUs. It may be the ability to make GPUs see one another quickly enough.
Why I Chose the Title “Photon Constraint”
I chose Photon Constraint because the phrase captures, in two words, the next physical boundary of AI scaling. The first era of generative AI was dominated by GPU scarcity: whoever could obtain the largest quantity of advanced accelerators could build the greatest amount of computational capacity, and the industrial politics, capital flows, and export-control regimes of 2022 through 2026 all organized themselves around that single scarcity. But as clusters become vastly larger, the constraint increasingly migrates into the space between the processors. The economic value of a GPU depends on whether data, parameters, and intermediate calculations can reach it fast enough, and at sufficiently large scale, communication becomes computation—not metaphorically, but as a matter of accounting, because every microsecond an accelerator spends waiting for the network is a microsecond of purchased capability that produces nothing.
The word Photon deliberately shifts the center of attention from silicon processing to optical movement, from the transistor that switches to the light that travels. The word Constraint emphasizes that connectivity can become a binding limit on AI expansion in exactly the way that electricity, high-bandwidth memory, advanced packaging, and grid interconnection already have—each of which was once dismissed as a supporting detail before it became the headline shortage of its year. The title therefore fits this paper because it describes not merely a networking trend, but a structural transition in the Five-Layer AI Economy: the point at which adding more GPUs is no longer enough to guarantee proportional increases in usable intelligence, and at which the industry’s decisive question changes from “how many operations can silicon perform?” to “how much intelligence can light carry?”
Section 1: From Compute Scarcity to Communication Scarcity
Every industrial revolution eventually discovers that its founding scarcity was only the first scarcity. The railway age began with a shortage of locomotives and ended obsessed with signaling, scheduling, and gauge standardization, because it turned out that owning trains mattered less than coordinating them. The electrical age began with a race to build generators and ended by building grids, because a dynamo without transmission is a very expensive paperweight. Artificial intelligence is now living through the same transition, and this section traces how the industry’s founding scarcity—the advanced accelerator—is giving way to a subtler and more structural scarcity: the capacity to move information among accelerators fast enough, far enough, and efficiently enough for enormous machines to behave as one machine. Understanding that transition requires holding two ideas simultaneously: that GPUs remain the irreplaceable engines of machine intelligence, and that their aggregate usefulness is increasingly decided by everything around them.
1.1 The GPU Era Created the Next Bottleneck
Between 2022 and 2026, the dominant infrastructure question was disarmingly straightforward: how many GPUs can you obtain? Nvidia’s H100, H200, Blackwell, and subsequent Rubin generations became symbols of strategic computational capacity in the way that oil reserves once symbolized industrial power. Hyperscalers and frontier laboratories signed enormous multi-year supply agreements; governments debated accelerator export restrictions as instruments of statecraft; startups raised billions of dollars against nothing more than allocation letters; and datacenter developers redesigned entire buildings around increasingly dense AI racks whose power draw per cabinet rose from tens of kilowatts toward hundreds. The market capitalization of the semiconductor industry reorganized itself around this single scarcity with astonishing speed, and for four years the thesis held: whoever accumulated the most accelerators accumulated the most capability.
But every successful solution to one bottleneck exposes the next bottleneck, and it does so precisely because the first solution worked. As accelerator availability expands and individual processors become more powerful, keeping those processors continuously occupied becomes progressively more difficult, because each new generation of accelerator consumes data faster than the previous generation while the infrastructure feeding it improves on a slower curve. An accelerator waiting for information is economically similar to an idle factory machine: the asset exists, the capital has been deployed, the electricity may still be flowing, the depreciation clock is certainly still running, and yet productive output falls below theoretical capacity. When the machine in question costs tens of thousands of dollars and the cluster contains a hundred thousand of them, the difference between theoretical and realized output stops being an engineering curiosity and becomes one of the largest line items in the global economy. The central distinction of the Photon Constraint thesis is therefore between installed compute capacity and usable synchronized compute capacity, and the argument of this entire paper is that the gap between the two will become one of the defining economic variables of the 2027–2030 period.
The clearest evidence that the industry itself has internalized this distinction comes not from networking vendors talking their own book, but from the accelerator champions themselves. Broadcom, whose custom XPU business serves the very hyperscalers building the largest clusters on Earth, told investors during its second-quarter fiscal 2026 earnings call that networking is the key to building scalable accelerator clusters, describing a portfolio that spans co-packaged copper for scale-up inside racks, its 102.4-terabit Tomahawk 6 Ethernet switch for scale-out between racks, and co-packaged optics, 1.6T optical DSPs, and continuous-wave and electro-absorption-modulated lasers for extending AI clusters across entire datacenters.[16] One quarter later, with AI semiconductor revenue having more than tripled year-over-year to $16.7 billion, chief executive Hock E. Tan made the strategic implication explicit.[17]
“Our AI networking revenue is expected to grow just as fast as XPUs.”
— Hock E. Tan, President and CEO, Broadcom, Q3 FY2026 earnings call [17]
When the company that sells the accelerators tells the market that the wires and lasers connecting those accelerators will grow just as fast as the accelerators themselves, the era of counting GPUs as a sufficient measure of AI capacity is already over.
1.2 FLOPs Are Not the Same as Intelligence Production
AI infrastructure is still most frequently measured through theoretical processor performance—petaflops per chip, exaflops per cluster, and the marketing arithmetic that multiplies one by the other. Yet large models do not execute on isolated processors, and they never have. Training distributes mathematical operations across enormous numbers of accelerators, requiring gradients, weights, and optimizer states to circulate continuously among them through collective operations whose completion time is governed by the slowest participant. Inference increasingly distributes memory, attention caches, long contexts, and mixture-of-experts model shards across multiple devices, so that answering a single query can involve traffic across dozens of chips. Agentic systems compound the pattern further, generating repeated inference calls, tool invocations, retrieval operations, and model-to-model interactions whose economics depend on latency as much as throughput. The useful output of the system therefore depends upon three things simultaneously—calculation, memory movement, and communication—and improving only the first eventually produces an imbalanced machine whose most expensive component spends its time waiting for its cheapest components.
This is not a new insight to the research community, even if it is a new insight to capital markets. Columbia University’s Keren Bergman, the Charles Batchelor Professor of Electrical Engineering and director of the Lightwave Research Laboratory, has spent two decades demonstrating that high-performance systems are increasingly bottlenecked by the energy and communication costs of moving data across compute and memory resources, that off-chip communication capacity lags on-chip capacity by nearly two orders of magnitude, and that the energy per bit of data movement now dominates the energy of computation itself. As early as 2019, describing the challenge for AI and machine-learning workloads specifically, she identified data movement in unambiguous terms.[29]
“The singular roadblock to the scalability of these systems.”
— Keren Bergman, Charles Batchelor Professor of Electrical Engineering, Columbia University [29]
The result of taking this seriously is a new economics of goodput: not how much theoretical processing capacity exists on the invoice, but how much useful model work a cluster completes within a given amount of time, energy, and capital. Goodput is the quantity that determines training-run duration, tokens served per dollar, and ultimately the gross margin of every AI product built on top of the infrastructure. Photon Constraint is fundamentally a goodput problem, and everything that follows in this paper—the supply-chain analysis, the economic framework, the geopolitics, and the 2027–2030 outlook—is an examination of what happens when an industry that spent four years maximizing installed FLOPs is forced to begin maximizing realized goodput instead.
1.3 Scale-Up, Scale-Out, and Scale-Across
To reason precisely about where light replaces electricity, the paper must distinguish three increasingly distinct levels of AI connectivity, because each level has different distance requirements, different latency budgets, different economics, and therefore a different timetable for the optical transition. Scale-up connects accelerators inside tightly integrated systems or computing pods, where latency must be extremely low because the processors frequently behave like portions of a single logical machine sharing a memory domain; this is the territory of NVLink, of UALink’s 1,024-accelerator pods with load/store semantics, and of the co-packaged optical chiplets now being designed directly beside the accelerator die.[11] Scale-out connects large numbers of servers and racks across an AI datacenter, where Ethernet and InfiniBand fabrics move enormous quantities of east-west traffic among computational resources and where the pluggable optical transceiver has already become one of the highest-volume precision components in the global electronics industry. Scale-across increasingly connects separate datacenters or entire AI campuses, potentially allowing facilities separated by meaningful geographic distances to contribute to coordinated workloads through coherent optics and data-center-interconnect systems whose market, according to Dell’Oro Group, is now pulling the entire optical transport industry to revenue levels not seen since the year 2000.[32]
| Dimension | Scale-Up | Scale-Out | Scale-Across |
| Physical span | Within a rack or pod (centimeters to a few meters) | Across a datacenter hall (meters to ~2 km) | Between buildings, campuses, metros (km to ~100+ km) |
| What travels | Memory-semantic traffic: activations, KV caches, expert exchange | Collective operations, gradients, parameter shards, storage traffic | Checkpoints, replication, coordinated multi-site workloads |
| Latency tolerance | Nanoseconds to low microseconds | Microseconds | Tens of microseconds to milliseconds |
| Incumbent medium | Copper (NVLink, co-packaged copper, UALink over electrical lanes) | Pluggable optics on Ethernet / InfiniBand | Coherent DWDM optics, ZR/ZR+ pluggables |
| Optical successor | Co-packaged optics and optical I/O chiplets (Ayar Labs TeraPHY, Celestial AI Photonic Fabric) | CPO switching (Nvidia Spectrum-X Ethernet Photonics, Broadcom TH6-Davisson) | 1.6T coherent DSPs, IPoDWDM, intelligence corridors |
| Representative 2026 signal | Marvell/Celestial 16 Tb/s optical chiplet; UALink 200G specifications | CPO switches in production at Nvidia and Broadcom | Dell’Oro: optical transport past $18B on AI DCI demand |
Table 1. The three connectivity domains of the AI factory, and the optical transition underway in each. Sources: UALink Consortium, Nvidia, Broadcom, Marvell, Dell’Oro Group.[8][11][14][19][32]
Each transition outward—from pod to hall to campus—increases the importance of optical communication, because distance is precisely the variable on which electrical signaling degrades and photonic signaling does not. The farther AI computation spreads physically, the more difficult it becomes to pretend that processor performance alone determines system performance, and the more clearly the network reveals itself as the true architecture of the machine.
1.4 Why Copper Encounters a Physics Problem
Copper is not disappearing, and any honest version of the Photon Constraint thesis must begin by saying so plainly. Electrical interconnect will remain the dominant medium at the shortest distances for years because it is mature, inexpensive, and remarkably efficient when the electrical reach required is small; Broadcom itself emphasizes co-packaged copper as a leadership technology for scale-up inside the rack.[16] The problem is not copper today; the problem is copper’s trajectory. As signaling rates double from 100 to 200 gigabits per second per lane and beyond, transmission losses across circuit-board traces and cable assemblies rise steeply, driven by the skin effect and dielectric absorption that scale nonlinearly with frequency. As electrical signals travel farther from the switch ASIC or the accelerator, maintaining signal quality increasingly requires additional circuitry—retimers, equalizers, digital signal processors—each of which consumes power. That power becomes heat. Heat requires cooling. Cooling requires additional electricity. And the electrical reach over which any of this remains workable keeps shrinking with every speed generation, which is why the practical radius of copper is collapsing from meters toward fractions of a meter at exactly the moment AI machines are growing from racks toward campuses.
The communication problem therefore connects directly back to Layer 1 of the Five-Layer AI Economy: a watt consumed transporting information is a watt that cannot be used performing useful matrix multiplication, inside a global power envelope that the IEA already projects will more than double by 2030.[35] This is precisely the framing that the co-packaged-optics industry has adopted. By bringing optical engines physically onto the same package as the switching silicon, CPO shrinks the distance over which extremely high-speed electrical signals must travel from tens of centimeters of circuit board to millimeters of package substrate, eliminating entire stages of power-hungry signal conditioning. Broadcom’s third-generation CPO switch, the Tomahawk 6–Davisson, delivers 102.4 terabits per second of optically enabled switching capacity—doubling any CPO switch previously available—using sixteen 6.4-terabit optical engines built on TSMC’s COUPE compact photonic engine process, with the company claiming roughly 70 percent lower optical interconnect power than traditional pluggable approaches.[14][15]
1.5 From Electrons to Photons
The long-run transformation is not that computers cease using electrons; computation itself remains resolutely electronic, and nothing in this paper forecasts otherwise. The transformation is that a growing portion of the communication between computational components moves into the optical domain, so that the canonical data path of the AI factory becomes hybrid: electronic computation, followed by electro-optical conversion, followed by photon transmission, followed by optical-electronic conversion, followed by electronic computation again. Every stage of that sentence is an industry. The conversion stages require lasers, modulators, and photodetectors; the transmission stage requires fiber, connectors, and optical switching; the whole assembly requires photonic integrated circuits, optical DSPs, advanced packaging that can marry photonics to logic dies, test equipment that can validate optical paths at manufacturing scale, and manufacturing processes—like TSMC’s COUPE—that can produce all of it in the volumes that million-GPU factories will demand.[15]
Artificial intelligence is therefore beginning to pull the photonics industry into its gravitational center, exactly as it previously pulled in memory, advanced packaging, and power generation. An industry that spent decades serving telecommunications carriers on multi-year procurement cycles is being annexed, at extraordinary speed, into the fastest capital-deployment cycle in economic history. Section 2 maps who is being pulled in, what they disclosed during 2026, and why the revenue is arriving faster than almost anyone forecast.
Section 2: The Emerging Photonic AI Supply Chain
Investment cycles usually become visible first through supplier revenue, because suppliers report what customers are actually buying rather than what conference keynotes say they intend to buy. The GPU boom announced itself through Nvidia’s datacenter segment before most of the world had used a chatbot. The memory bottleneck elevated SK Hynix, Micron, and Samsung before “HBM” entered the financial vocabulary. Advanced packaging turned CoWoS capacity at TSMC into a strategic variable tracked as closely as OPEC quotas. The Photon Constraint is now producing the same pattern of early supplier signals, and 2026 has been the year in which those signals stopped being anecdotes and became a coordinated, capital-backed industrial buildout. This section reads that buildout company by company, because the specific commitments—who invested what, who acquired whom, who signed which capacity agreement—are the empirical spine of the entire thesis.
| Date | Company / Parties | Event | Why It Matters for the Photon Constraint |
| Dec 2, 2025 | Marvell / Celestial AI | Definitive agreement to acquire Celestial AI for $3.25B (up to $5.5B with earn-outs) | First multi-billion-dollar acquisition specifically for scale-up optical interconnect; 16 Tb/s Photonic Fabric chiplet[19] |
| Jan 9, 2026 | Marvell / XConn | Acquisition of PCIe/CXL switching silicon and UALink scale-up switch team | Consolidation of the scale-up switching stack around one connectivity vendor[21] |
| Feb 2, 2026 | Marvell | Celestial AI acquisition completed | Photonic Fabric moves inside a top-tier merchant silicon roadmap[20] |
| Feb 9, 2026 | STMicroelectronics / AWS | Multi-year, multi-billion-dollar commercial agreement | A hyperscaler contracts directly for photonics supply, with capacity and cash advances[36] |
| Mar 2, 2026 | Nvidia / Coherent, Lumentum | $2B investment in each, plus multi-billion purchase commitments and capacity rights | The GPU champion vertically de-risks lasers and optical components; Lumentum commits to a new fab[23][24] |
| Mar 3, 2026 | Ayar Labs | $500M Series E at $3.75B valuation; largest photonics financing ever | Optical I/O chiplets move from research to high-volume production funding[26] |
| May 31, 2026 | Nvidia | Vera Rubin in full production; Spectrum-X Ethernet Photonics CPO switches in production | Co-packaged optics enters the mainstream AI factory roadmap at 200G SerDes[8] |
| Jun 2, 2026 | STMicroelectronics | Datacenter revenue guidance raised to ~$1B (2026), ~2x in 2027 | Optical connectivity revenue arrives faster than the supplier itself forecast[2] |
| Sep 4, 2026 | Broadcom | Q3 FY2026: AI semiconductor revenue $16.7B, +221% YoY; Tomahawk 7 (200T) taped out | Networking guided to grow as fast as XPUs; CPO shipping at 102.4T[17][18] |
| Sep 8, 2026 | Qualcomm / Amazon | Multigenerational custom silicon + 1.6T optical connectivity; up to $60B framework | Compute and optics co-designed inside one strategic relationship[5][6] |
| Sep 9, 2026 | STMicroelectronics | CFO: ~80% of >$2B 2027 AI revenue from fiber-optic data-link chips | The revenue breakdown that anchors this paper[1] |
Table 2. The photonic capital cycle of 2025–2026: ten months in which optical interconnect moved from peripheral component to strategic infrastructure.
2.1 STMicroelectronics: The Revenue Signal
STMicroelectronics provides the cleanest single piece of evidence that the photonic transition is becoming commercially material, precisely because the company is not a storyteller of the AI boom but a components manufacturer with a conservative Franco-Italian disclosure culture. Its expectation that approximately four-fifths of its targeted 2027 AI-datacenter revenue will come from fiber-optic data-link chips—silicon photonics on its PIC100 platform, BiCMOS electronic ICs for optical transceivers, and the associated microcontrollers—demonstrates that AI infrastructure spending is spreading outward from headline processors into supporting communication silicon at a scale that now moves the composite margins of a $14-billion-revenue company.[1][3] The trajectory of the guidance itself is as informative as its level: from “nicely above $500 million” to approximately $1 billion for 2026 in a single June revision, and from “well above $1 billion” to more than $2 billion for 2027, each upgrade attributed to AI infrastructure-led demand outrunning the company’s capacity ramp.[2] Long-term customer agreements including committed capacity, volumes, pricing, and cash advances give the company visibility beyond 2027—contract structures that historically appear only where customers fear shortage.[3] The multi-year, multi-billion-dollar commercial agreement with Amazon Web Services announced on February 9, 2026 shows who those fearful customers are.[36] History suggests a simple reading: the GPU boom created Nvidia; the memory bottleneck elevated SK Hynix, Micron, and Samsung; advanced packaging made TSMC’s capacity a geopolitical variable; the Photon Constraint could similarly elevate companies whose products solve data-movement problems, and the market has begun repricing them accordingly.
2.2 Amazon and Qualcomm: Compute and Connectivity Converge
The Qualcomm–Amazon agreement of September 8, 2026 is particularly revealing because customized compute and optical connectivity appear inside the same strategic relationship, negotiated together, priced together, and milestoned together. Amazon is not merely buying an alternative inference accelerator to complement its Trainium line; the companies are jointly engineering how that silicon communicates at high speed across Amazon’s AI infrastructure, including optical connectivity reaching 1.6 terabits per second and explicitly extending to future generations, built on Qualcomm’s SerDes and optical digital-signal-processing technologies.[5] The warrant structure—up to 25 million shares vesting against as much as $60 billion in purchases through September 2036—binds the relationship across a full decade of silicon generations rather than a single product cycle.[6] Futurum’s analysis of the deal noted that connectivity was one of the two Qualcomm datacenter product lines closest to revenue, ramping even before the custom compute silicon itself.[7]
This indicates that future AI-chip competition will increasingly occur at the system level, and that a processor can no longer be evaluated independently of its memory bandwidth, its interconnect bandwidth, its network topology, its power efficiency, its software ecosystem, and its optical architecture. The processor becomes one component of a larger machine, and the buyer of that machine—in this case the largest cloud provider on Earth—now contracts for the light and the logic in a single negotiation. When procurement behavior changes at that level, it is because the customer has concluded that the two can no longer be optimized separately.
2.3 Nvidia’s Expansion From GPU Company to System Architect
Nvidia’s own behavior during 2026 constitutes perhaps the strongest single endorsement of the Photon Constraint thesis, because no company has more to lose from a world in which the GPU alone stops being the unit of value. Nvidia increasingly competes not simply through GPU performance but through full-stack system design: CPUs, GPUs, NVLink, networking, DPUs, switches, software, and now silicon photonics woven into the fabric of the platform itself. The Vera Rubin platform, ramping into full production as of GTC Taipei in May 2026, delivers what Nvidia describes as ten times the agent throughput at scale of the prior Grace Blackwell generation, and it introduces Spectrum-X Ethernet Photonics—the world’s first co-packaged-optics-based switches with 200G SerDes—as the foundational fabric for million-GPU AI factories, claiming five times better power efficiency, five times longer sustained AI uptime, and 1.3 times faster deployment than transceiver-based networks.[8] The underlying silicon-photonics program, first unveiled at GTC in March 2025 with joint inventions and supply-chain collaborations spanning TSMC, Coherent, Corning, Foxconn, Lumentum, and SENKO, integrated optics innovations using four times fewer lasers to deliver 3.5 times more power efficiency and 63 times greater signal integrity than conventional approaches.[9] TSMC’s chairman and chief executive, whose COUPE process and SoIC three-dimensional stacking make the co-packaging physically possible, framed the ambition in terms that would have sounded like science fiction three years earlier.[9]
“Scale to a million GPUs and beyond, pushing the boundaries of AI.”
— C. C. Wei, Chairman and CEO, TSMC [9]
Then, in March 2026, Nvidia did something that fabless semiconductor companies almost never do: it committed $4 billion of its own balance sheet—$2 billion each into Coherent and Lumentum—paired with multi-billion-dollar purchase commitments and future capacity rights for advanced laser and optical networking components, with Lumentum building an entirely new fabrication facility on the strength of the agreement.[23][24] Lumentum’s chief executive confirmed the capacity dimension directly.[24]
“We are also investing in a new fabrication facility to increase capacity and accelerate innovation.”
— Michael Hurlston, CEO, Lumentum [24]
Coherent’s chief executive, whose company has supplied Nvidia for two decades, described the arrangement in terms that place optics squarely at the center of the buildout.[25]
“A key enabler of next-generation AI data center infrastructure.”
— Jim Anderson, CEO, Coherent [25]
The implication is profound: Nvidia’s competitive moat may increasingly depend as much on moving bits between GPUs as on calculating operations inside GPUs, and the company is spending shareholder capital to make sure the photonic layer of that moat cannot be constrained by anyone else’s fab schedule. A company does not pre-purchase laser capacity years in advance to solve a problem it considers peripheral.
2.4 Broadcom, Marvell, Ayar Labs, and the Photonic Challengers
Photon Constraint also expands the universe of strategically important AI companies well beyond the accelerator vendors, and the expansion is proceeding along two distinct fronts: incumbent connectivity giants racing to assemble complete optical portfolios, and specialized photonic companies racing to production before the incumbents close the gap. Broadcom anchors the first front. The company has spent years developing co-packaged optics alongside its dominant switching franchise, and in October 2025 it began shipping Tomahawk 6–Davisson, its third-generation CPO Ethernet switch, delivering 102.4 terabits per second of optically enabled capacity—double any CPO switch then available—while its Bailly predecessor accumulated millions of field-test hours that hyperscalers regard as the real currency of adoption.[14] By the third quarter of fiscal 2026, Broadcom’s AI semiconductor revenue had reached $16.7 billion in a single quarter, up 221 percent year-over-year, its Tomahawk 7 200-terabit switch had taped out, and the company described itself as the leader in optical DSPs while rapidly expanding capacity in EMLs, VCSELs, and continuous-wave lasers—the laser diet of the optical AI factory.[17][18]
Marvell has pursued the same destination through the boldest acquisition strategy in the sector. Its December 2025 agreement to acquire Celestial AI for $3.25 billion—rising to as much as $5.5 billion against revenue milestones—brought in the Photonic Fabric platform, whose first-generation chiplet integrates all required electrical and optical components and delivers 16 terabits per second of bandwidth in a single chiplet, ten times the capacity of today’s state-of-the-art 1.6T scale-out ports, designed to be co-packaged directly with XPUs and scale-up switches.[19] The acquisition closed on February 2, 2026, and Marvell projects the technology will reach a $500 million annualized revenue run rate during fiscal 2028 and $1 billion by the end of fiscal 2029.[19][20] Chairman and chief executive Matt Murphy articulated the strategic logic in a single phrase that could serve as the epigraph of this entire section.[19]
“Scale-up becomes the next frontier in AI infrastructure.”
— Matt Murphy, Chairman and CEO, Marvell Technology [19]
Around the Celestial acquisition, Marvell has layered the January 2026 purchase of XConn Technologies—adding PCIe and CXL switching silicon and a UALink scale-up switch engineering team—and an April 2026 expansion of its 1.6T optical DSP portfolio spanning the Ara, Alaska, and Nova DSP families, silicon-photonics light engines, and linear-drive optics chipsets, positioning the company across every link of the connectivity stack from inside the package to between the cities.[21][22]
Ayar Labs represents the second front: optical I/O brought as close as physically possible to the computing package itself. Its $500 million Series E in March 2026, led by Neuberger Berman with participation from ARK Invest, Insight Partners, Qatar Investment Authority, Sequoia Global Equities, and strategic investors including Nvidia, AMD, MediaTek, and Alchip, raised total funding to $870 million at a $3.75 billion valuation—the largest financing round any photonics company has ever completed, and explicitly a production round rather than a research round, funding high-volume manufacturing and test capacity for the company’s TeraPHY optical engines.[26][28] Chief executive and co-founder Mark Wade stated the thesis of this paper more compactly than this paper can.[26]
“As bandwidth demands explode, copper becomes the bottleneck.”
— Mark Wade, CEO and Co-Founder, Ayar Labs [26]
And the capital allocators writing the checks have adopted the same language. The managing director at Neuberger Berman who led the round described the AI buildout as one of the largest capital-deployment opportunities of the generation, and located the constraint exactly where this paper locates it.[27]
“Data center interconnect has quickly emerged as the most critical bottleneck.”
— Gabe Cahill, Managing Director, Neuberger Berman [27]
The emerging competition will therefore involve both giant incumbent semiconductor firms and specialized photonic companies, with the open standards bodies—UALink for memory-semantic scale-up, the Ultra Ethernet ecosystem for scale-out—functioning as the neutral ground on which merchant silicon challenges vertically integrated platforms. The UALink Consortium’s board alone now includes Alibaba, AMD, Apple, Astera Labs, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft, and Synopsys, which is to say nearly every major buyer and builder of AI infrastructure that is not Nvidia.[12][13]
“UALink is the only memory semantic solution for scale-up AI.”
— Kurtis Bowman, Board Chair, UALink Consortium [12]
2.5 The Market Evidence: Optics Revenue Is Compounding Like Compute Revenue
The demand curves confirm what the corporate announcements imply. LightCounting, the optical industry’s most closely followed research house, estimates that sales of Ethernet optical transceivers and CPO for AI clusters reached $16.5 billion in 2025 and will reach approximately $26 billion in 2026—a 60 percent growth rate sustained for two consecutive years—with the firm openly debating whether the market reaches $100 billion by 2030.[33] Its broader Ethernet transceiver tracking shows the market up 93 percent in 2024 and roughly 82 percent in 2025, with 65 percent growth forecast for 2026, constrained not by demand but by the production capacity of indium-phosphide EML and laser chips, where demand has been exceeding supply by roughly 30 percent.[34] Dell’Oro Group, meanwhile, raised its 2026 optical transport forecast from 10 percent growth to 16 percent after first-quarter results, projecting the market past $18 billion for the first time since 2000, with datacenter-interconnect purchases up an estimated 40 percent year-over-year and key optical suppliers reporting growing backlogs as lead times stretch.[32]
“We raised our full year 2026 forecast from 10 percent to 16 percent.”
— Jimmy Yu, Vice President, Dell’Oro Group [32]
A market does not sustain 60 percent annual growth against 30 percent supply shortfalls because it is a peripheral accessory to the real story. It does so because it has become part of the real story—and because, as the next subsection argues, the bill of materials of intelligence itself has changed.
2.6 The New AI Bill of Materials
The traditional popular description of an AI server focuses on its accelerators, and the traditional financial model of AI infrastructure allocates most of its attention to accelerator unit counts and average selling prices. The Photon Constraint framework requires a broader accounting, because the machine that produces intelligence now depends upon accelerators; high-bandwidth memory; advanced 2.5D and 3D packaging; switch ASICs; network interface cards and DPUs; SerDes; optical DSPs; continuous-wave, EML, and VCSEL lasers; silicon-photonic engines and optical I/O chiplets; fiber, connectors, and cable-management systems engineered to survive a decade of thermal cycling; liquid cooling; power conversion advancing toward 800-volt architectures; and orchestration software capable of scheduling work across the entire fabric while routing around the component failures that statistics guarantee at million-part scale. Remove any single category from that list and the theoretical FLOPs of the remaining categories cannot be fully converted into revenue. The strategic lesson is that the AI bill of materials is becoming a systems bill of materials, and this is precisely why apparently obscure semiconductor components—a laser driver, a transimpedance amplifier, a photonic engine the size of a fingernail—can suddenly become multibillion-dollar AI businesses attracting the strategic capital of the largest company on Earth. The value has not left the GPU; the value has spread outward along every path a bit must travel to reach one.
Section 3: The Economics of Moving Intelligence
Economic frameworks lag industrial reality, and the frameworks currently used to value AI infrastructure were built for the scarcity that is ending rather than the scarcity that is beginning. This section constructs the economics of the Photon Constraint from first principles: what the correct unit of analysis becomes when communication binds, why small utilization losses at cluster scale represent stranded capital measured in billions of dollars, how the power budget converts networking efficiency directly into compute capacity, why a distinct photonic capital-expenditure cycle is now forming, and what new metrics investors and operators will need in order to see any of this clearly. The consistent theme is that bandwidth is ceasing to be a supporting cost line and becoming a factor of production—an input that, like electricity or labor, directly determines output.
3.1 From Cost per FLOP to Cost per Useful Token
The first generation of AI infrastructure economics emphasized computational efficiency through three deceptively clean questions: how many FLOPs can a processor perform, how much does each accelerator cost, and what is the energy consumed per operation? These questions were adequate when clusters were small enough that communication overhead was a rounding error, in the way that a village economy can ignore logistics. Photon Constraint requires another dimension: how much communication infrastructure must be purchased, powered, and operated to turn those processors into productive systems? The relevant unit of economic analysis therefore migrates from cost per FLOP to total system cost divided by useful model output—cost per useful token generated, cost per training run completed, cost per agentic workflow executed end to end. Under this accounting, a less expensive processor attached to an inefficient communication fabric may ultimately produce more expensive intelligence than a costlier processor embedded inside a highly optimized system, because the denominator of the calculation—useful output—is set by the fabric, not the chip. This inversion explains procurement behavior that otherwise looks irrational: why hyperscalers pay premium prices for integrated rack-scale systems, why Amazon negotiates optics and inference silicon in a single agreement with Qualcomm, and why Nvidia can describe its platform advantage in terms of agent throughput at the system level rather than benchmark performance at the chip level.[5][8]
3.2 The Stranded GPU Problem and the Photon Utilization Gap
Imagine a cluster containing 100,000 advanced accelerators, representing on the order of $4 to $6 billion of installed silicon before a single building, transformer, or chiller is counted. On paper, its theoretical computational capacity is enormous. But suppose network contention, communication overhead, synchronization delays, stragglers, or component failures reduce effective utilization by only five percentage points below what a better fabric would achieve. Five percent of a five-billion-dollar asset is $250 million of capital that produces nothing—silently, continuously, invoice after invoice—and unlike a factory that can warehouse unsold inventory, a GPU-hour that passes idle is destroyed forever. At the million-GPU scale that Nvidia, Broadcom, and their customers now describe as the design point, the same five-percentage-point gap strands billions of dollars of capability per cluster per year.[8][18] Failure statistics compound the arithmetic: a network built from hundreds of thousands of discrete pluggable transceivers presents hundreds of thousands of individually replaceable failure points, and every link flap can stall a synchronous training step across the entire machine, which is why Nvidia markets its co-packaged optics as delivering five times longer sustained AI uptime and ten times higher network resiliency, and why hyperscalers treat those reliability claims as economics rather than marketing.[8][10]
This paper therefore proposes a named concept for the quantity at stake. The Photon Utilization Gap is the difference between theoretical accelerator output and realized accelerator output attributable to communication limitations—bandwidth ceilings, latency, congestion, synchronization overhead, and interconnect failures. As clusters become larger, closing that gap becomes economically equivalent to purchasing thousands of additional GPUs, with one decisive difference: the additional GPUs would require additional megawatts, additional floor space, and additional cooling, while closing the gap requires none of them. This changes procurement priorities in a way that boards and finance committees have only begun to internalize. Sometimes the cheapest way to obtain more AI computation will no longer be to buy more processors; it will be to improve the network connecting the processors already owned.
3.3 Power Becomes the Bridge Between Energy and Photonics
Photon Constraint integrates directly with Layer 1 of the Five-Layer AI Economy, because every optical engineering decision is simultaneously an energy allocation decision. Hyperscalers are already struggling to secure gigawatts of electricity against interconnection queues measured in years, inside a global trajectory in which the IEA projects datacenter consumption rising from roughly 415 TWh in 2024 to about 945 TWh in 2030 and toward 1,200 TWh by 2035, with accelerated-server electricity growing 30 percent annually in the base case.[35] Within any single facility, every subsystem competes for a finite power envelope: GPUs, HBM, CPUs, fans and liquid-cooling loops, network switches, transceivers, storage, and power-conversion equipment all draw from the same feeder. Networking can no longer be treated as electrically free. In conventional architectures, the optical transceivers and the retimer silicon supporting them can consume a meaningful share of total cluster power—power that produces no tokens—and that share rises with every speed generation as electrical signal conditioning grows more desperate.
This is why the co-packaged-optics claims of 2026 matter beyond engineering pride. Nvidia’s integrated photonics uses four times fewer lasers to deliver 3.5 times more power efficiency at the network level, with the Vera Rubin generation claiming five times better power efficiency than transceiver-based fabrics; Broadcom’s TH6-Davisson reduces optical interconnect power by roughly 70 percent against traditional pluggables.[8][9][14] Whether every vendor achieves identical figures is less important than the structural direction: network efficiency is becoming part of compute efficiency, because every watt liberated from moving bits is a watt available for generating them. In a power-capped facility—and by 2027 essentially every frontier facility will be power-capped—the photonic upgrade and the GPU purchase are competing uses of the same megawatt, and the photonic upgrade increasingly wins.
3.4 The Photonic Capex Cycle
The AI infrastructure boom has proceeded through successive, overlapping capital cycles, each triggered when the previous cycle’s success exposed a new binding constraint. First came accelerators. Then high-bandwidth memory. Then advanced packaging. Then datacenter shells and the land beneath them. Then power generation, transmission, and grid interconnection. Photon Constraint identifies the next cycle, and the evidence assembled in Section 2 shows it is already funding itself: $4 billion of Nvidia strategic capital into Coherent and Lumentum with purchase commitments and a new fab behind it; $3.25-to-$5.5 billion of Marvell acquisition capital into Celestial AI; $500 million of growth equity into Ayar Labs’ production ramp; STMicroelectronics customer agreements carrying capacity commitments and cash advances; Broadcom expanding indium-phosphide capacity out of its own record cash flows; and LightCounting documenting a component market growing 60 percent annually against a 30 percent supply shortfall in laser chips.[1][17][19][23][26][34] Capital in the coming cycle flows toward silicon-photonics fabrication, optical-engine assembly and test, laser diode manufacturing, fiber and connector production, CPO packaging, optical DSP development, and the metrology and manufacturing-test infrastructure without which none of it ships at hyperscale quality levels. Investors, hyperscalers, and governments that continue defining AI infrastructure only as “GPU capacity” will misidentify where incremental economic rents emerge during 2027–2030, exactly as investors who defined the internet buildout only as “routers” missed the fiber, the towers, and the datacenters that captured the durable returns.
| Capital Cycle | Approximate Onset | Binding Constraint Addressed | Representative Beneficiaries |
| Accelerators | 2022–2023 | Raw training and inference compute | Nvidia, AMD, custom XPU programs |
| High-bandwidth memory | 2023–2024 | Feeding data to the accelerator die | SK Hynix, Micron, Samsung |
| Advanced packaging | 2023–2025 | Integrating logic and memory (CoWoS, SoIC) | TSMC, OSAT ecosystem |
| Datacenters and power | 2024–2026 | Shells, land, gigawatts, grid interconnection | Utilities, developers, turbine and transformer makers |
| Photonics (the Photon Constraint) | 2025–2027 | Moving intelligence between accelerators, racks, and campuses | Broadcom, Marvell, STMicro, Coherent, Lumentum, Corning, Ayar Labs, TSMC (COUPE) |
Table 3. Successive capital cycles of the AI buildout. Each cycle began when the prior cycle’s success exposed the next binding constraint.
3.5 A New Metric: Photon Productivity
Finally, this paper introduces a second analytical term to complement the Photon Utilization Gap. Photon Productivity measures how effectively an AI infrastructure system converts network bandwidth, optical power, and communication capacity into useful computational output. Conceptually, Photon Productivity equals useful AI work divided by communication cost, where communication cost incorporates bandwidth provisioned, energy consumed by the fabric, latency and its scheduling consequences, interconnect equipment cost, failure rates, and the utilization losses those failures impose. The definition does not need to harden into a single standardized engineering formula immediately, any more than “total factor productivity” needed a single formula to reorganize how economists thought about growth. Its value is conceptual: it forces investors, operators, and policymakers to treat information movement as an input into intelligence production rather than as overhead, and it gives them a language for comparing architectures—pluggable versus co-packaged, copper pod versus optical pod, single-site versus distributed campus—on the dimension that will actually differentiate them. A cluster with modest FLOPs and high Photon Productivity can outperform, in tokens delivered per dollar and per megawatt, a nominally larger cluster that strands its capability behind an inadequate fabric. Between 2027 and 2030, this paper expects Photon Productivity, under whatever name the industry ultimately settles on, to migrate from research papers into earnings calls, because it measures the thing the money is actually buying.
Section 4: Photon Geopolitics and the New Semiconductor Chokepoints
Technology policy is always fighting the last shortage. The export-control architecture constructed between 2022 and 2026 was built around the scarcity of that era—advanced accelerators and the lithography equipment that manufactures them—and it achieved genuine leverage precisely because those items were identifiable, countable, and produced at a small number of controllable points. The Photon Constraint complicates that architecture, because it relocates a meaningful share of frontier-AI capability into a component ecosystem that is more distributed, more internationally entangled, and more deeply fused with civilian telecommunications than the accelerator supply chain ever was. This section examines what happens to semiconductor statecraft when the strategic question shifts from “who can obtain the chips?” to “who can make the chips operate as one machine?”
4.1 Export Controls May Eventually Move Beyond GPUs
U.S.–China semiconductor competition has largely centered on advanced accelerators and semiconductor-manufacturing equipment, on the theory that controlling the computational engine controls the capability built upon it. Photon Constraint raises a broader question: if optical systems become indispensable to constructing enormous frontier-AI clusters—if the difference between a hundred thousand accelerators and one coherent hundred-thousand-accelerator machine is the photonic fabric between them—can advanced photonic components eventually acquire strategic importance comparable to accelerator chips themselves? A country might possess significant quantities of compute silicon, whether legally acquired, domestically produced, or accumulated through gray markets, and yet struggle to combine those processors efficiently at very large scale without leading-edge optical DSPs, co-packaged optical engines, high-power laser arrays, and the packaging processes that integrate photonics with logic. Effective AI capacity therefore depends upon an ecosystem rather than any single controlled component, which means that both the leverage and the blind spots of export policy are migrating outward from the processor into the fabric. Policy institutions that model national AI capability as an accelerator count are measuring installed compute; the Photon Constraint insists they begin measuring usable synchronized compute, which is a different and harder number.
4.2 The Optical Supply Chain Is International
The photonics ecosystem crosses national boundaries even more thoroughly than the logic-semiconductor ecosystem it now joins. American companies contribute semiconductor design, optical DSPs, networking silicon, laser manufacturing, and the hyperscale demand that anchors the entire market: Broadcom, Marvell, Nvidia, Coherent, Lumentum, Corning, and Ayar Labs are all headquartered in the United States, and Nvidia’s 2026 investments explicitly fund the expansion of U.S.-based laser fabrication, including Lumentum’s new facility.[23][24] Taiwan remains critical not only for advanced logic manufacturing but increasingly for photonic integration itself, since TSMC’s COUPE compact photonic engine process underlies Broadcom’s Davisson optical engines and Nvidia’s co-packaged switches, while Alchip and MediaTek have taken strategic positions in Ayar Labs and Ayar has opened operations in Hsinchu.[15][26][28] European companies such as STMicroelectronics manufacture the silicon-photonics and BiCMOS content inside the transceivers themselves, under long-term agreements with American hyperscalers.[1][36] Japanese firms supply indispensable optical materials, precision components, and connector technology—SENKO appears by name among Nvidia’s photonics collaborators—and Corning’s fiber and Foxconn’s assembly capacity round out a chain that no single jurisdiction contains.[9] China, meanwhile, remains deeply embedded in optical-module assembly, fiber production, and the broader electronics supply chain, with Chinese transceiver manufacturers historically among the world’s largest by volume. An AI-security strategy focused solely on GPUs risks ignoring this broader industrial architecture entirely—and a trade conflict that disrupted it carelessly would damage allied AI capacity as surely as adversary capacity.
4.3 The Next Export-Evasion Question
Export-control policy also becomes mechanically harder as the strategic frontier moves into optics. Accelerators are relatively identifiable, high-value, low-volume devices whose movements can be tracked, licensed, and audited; a single controlled GPU carries a five-figure price and a traceable serial number. A giant optical system, by contrast, is assembled from numerous components—lasers, modulators, DSPs, fibers, connectors—that move through global telecommunications and semiconductor supply chains in enormous volumes, most of them serving entirely civilian purposes in metro networks, broadband systems, and conventional cloud datacenters. Policymakers will eventually face genuinely difficult questions regarding which technologies constitute frontier-AI-enabling infrastructure and which constitute the ordinary plumbing of the internet. Controls drawn too narrowly may leave alternative bottlenecks untouched, allowing a determined actor to assemble frontier-class fabrics from uncontrolled parts. Controls drawn too broadly could disrupt ordinary telecommunications, cloud computing, and global internet infrastructure, imposing costs on the world economy far out of proportion to any security benefit, and accelerating the very indigenization efforts they aim to prevent. Photon Constraint therefore introduces a new layer to technological statecraft: the governance of dual-use light.
4.4 Optical Sovereignty
Governments already speak the language of semiconductor sovereignty and compute sovereignty; national AI strategies from Washington to Brussels to Tokyo to Riyadh are organized around securing domestic or allied access to accelerators and the fabs that make them. The next concept in that sequence may be optical sovereignty: the ability to manufacture, obtain, and operate the communication technologies necessary to connect nationally significant AI systems at frontier scale. For the United States, an optical-sovereignty assessment would inventory domestic or allied access to silicon-photonics fabrication; laser sources across CW, EML, and VCSEL families, together with the indium-phosphide capacity that LightCounting already identifies as the binding shortage of 2026; optical DSPs; the advanced packaging that co-integrates photonics with logic; fiber and connector manufacturing at hyperscale volumes; switching silicon; and the test-and-measurement base underneath all of it.[34] Nvidia’s $4 billion into Coherent and Lumentum, with its explicit emphasis on U.S.-based manufacturing buildout, can be read as the private sector conducting exactly this assessment ahead of the public sector and acting on the result with its own balance sheet.[23][24] The same analytical framework applies symmetrically to China, Europe, Japan, South Korea, Taiwan, and the Gulf states now purchasing their way into national AI capacity: each will discover that sovereign compute without sovereign connectivity is an archipelago of processors, not a machine.
4.5 From Datacenters to Intelligence Corridors
Optics also changes geography, and geography is where infrastructure meets statecraft. If future AI architectures increasingly use high-capacity optical links to connect buildings and campuses—the scale-across domain whose commercial vanguard is the 1.6T coherent DSP and ZR/ZR+ pluggable market that Marvell leads and that Dell’Oro measures growing 40 percent annually—then the economically meaningful AI unit may become larger than a single datacenter.[21][32] Individual facilities can evolve into AI campuses; AI campuses can evolve into regional compute zones; and regional compute zones, stitched together by dedicated fiber routes engineered for AI traffic, can eventually become connected intelligence corridors whose siting logic is set jointly by power availability, fiber geography, land, water, and political stability. This links Photon Constraint to the physical geography of the Five-Layer AI Economy without duplicating the argument of Synthetic Geography: the earlier framework asked where the AI economy would locate; this one asks how far it can stretch. The central open question becomes: how far can computation be physically distributed before latency, synchronization overhead, and communications economics overwhelm the benefits of geographic expansion—before the corridor stops behaving like one machine and degrades into several? The answer is not fixed; it moves outward with every improvement in optical bandwidth, every reduction in per-bit energy, and every advance in training algorithms tolerant of distance. Between 2027 and 2030, that moving answer may influence where enormous AI campuses are built, which regions capture them, and which nations discover that their most strategic asset is neither their chips nor their power plants, but their dark fiber.
Section 5: Photon Constraint in the 2027–2030 AI Economy
Forecasting infrastructure is easier than forecasting models, because infrastructure obeys lead times: the fabs being funded, the standards being ratified, and the fiber being trenched in 2026 largely determine what can physically exist in 2029. This section projects the Photon Constraint forward through the end of the decade along five threads—the communication intensity of future model architectures, the goodput economics of agentic workloads, the arithmetic of million-GPU factories, the reconceptualization of the campus as a computer, and the vertical integration of the constraint through all five layers of the AI economy. The consistent conclusion is that every visible trend in model design and deployment increases the ratio of communication to computation, which means the constraint tightens on its own even if compute growth merely continues on schedule.
5.1 Models Will Generate More Communication, Not Less
Future AI models are unlikely simply to resemble today’s models with larger parameter counts, and every direction in which they are actually evolving is a direction that intensifies data movement. Mixture-of-experts systems activate different portions of a model for different tokens, converting what was once local computation into continuous expert-routing traffic across the interconnect, with all-to-all exchanges whose latency sensitivity makes them the canonical stress test of any fabric. Multimodal models simultaneously process text, images, audio, and video, multiplying the volume of activations that must move between stages of heterogeneous pipelines. Reasoning models perform longer internal computational sequences per query, holding and shuttling large key-value caches across devices for the duration of extended inference. World models ingest enormous streams of environmental and simulation data whose bandwidth requirements resemble scientific computing more than chat. And agentic systems continuously interact with external tools, retrieval systems, and other models, converting a single user intention into cascades of network round trips. Each development increases the amount, the complexity, or the timing sensitivity of data movement—usually all three at once. The future AI workload is therefore not simply compute-intensive; it is structurally, increasingly, and irreversibly communication-intensive, and hardware roadmaps that project only FLOPs growth are projecting the wrong axis.
5.2 Agentic AI Makes Goodput More Valuable
Agentic workloads change the economics of the datacenter in a way that deserves careful statement, because it is the demand-side mirror of everything argued in Section 3. A chatbot answers one user request with one inference pass, and its economics are dominated by the cost of that pass. An autonomous business agent may perform hundreds of computational steps to complete a single task: planning, searching, coding, validating, consulting other agents, invoking external tools, checking results, and revising—each step an inference call, each call a traversal of the fabric, and the total task latency the sum of every traversal. The economic output of the datacenter therefore increasingly depends upon how many complete agentic workflows it can execute per hour rather than how many isolated operations its chips could theoretically perform, and any per-step latency or reliability penalty compounds geometrically across the workflow. This is why Nvidia emphasizes system-level agent throughput for the Rubin generation rather than isolated chip benchmarks, reporting ten times the agent throughput at scale of the prior Grace Blackwell platform, and why its uptime and resiliency claims for co-packaged optics are framed as application-runtime multipliers rather than networking statistics.[8][10] The AI factory becomes a production system in the literal industrial sense, and the Photon Constraint determines how smoothly the production line runs: a factory whose conveyor stalls between stations does not miss its targets by the duration of the stalls—it misses them by everything downstream that never happened.
5.3 Million-GPU AI Factories
The phrase “million-GPU AI factory” once sounded like conference hyperbole; it now appears verbatim in vendor roadmaps, product documentation, and shipping-product specifications. Nvidia describes Spectrum-X Ethernet Photonics as the foundational fabric for million-GPU AI factories, with named early adopters; Broadcom specifies Tomahawk 6 for clusters scaling to more than one million XPUs, and its two-tier flat topology for 128,000-accelerator networks exists precisely to remove the switching layers whose count, power, and failure surface explode at that scale.[8][18] At a million components of any kind, tiny inefficiencies become enormous, and scaling becomes multiplicative rather than additive: more GPUs require more networking; more networking consumes more electricity; more electricity produces more heat; more heat demands more cooling; more components raise the probability that something, somewhere, is always failing; and more physical distance stretches every synchronization guarantee the software stack depends upon. A network architecture that is perfectly acceptable at ten thousand accelerators may be economically or physically untenable at one million, which is why the vendors building for that scale converged, independently and near-simultaneously, on the same answer: collapse the electrical distance to millimeters, convert to light as early as possible, remove the discrete transceiver as a unit of failure, and spend the reclaimed watts on computation. Photonic architectures matter at million-GPU scale because they attack several terms of the multiplication at once—power, reach, reliability, and density—rather than trading one against another.
5.4 AI Campuses Become Distributed Computers
Between 2027 and 2030, hyperscalers may increasingly design datacenter campuses from the beginning as single gigantic distributed computing machines rather than as collections of buildings that happen to share a substation. Under that design philosophy, the physical building becomes secondary to the computational fabric: individual racks become components; buildings become modules; the fiber plant becomes the nervous system carrying the machine’s internal state; the power infrastructure becomes the metabolic system setting its sustainable intensity; and the AI models become the intelligence operating across the whole structure. The interpretation changes how planners should think about datacenter design at every stage. Site selection begins with fiber geography and latency budgets, not only with land and megawatts. Building placement within a campus becomes a network-topology decision, because hundreds of meters of extra path length are a latency tax paid on every synchronization for the life of the facility. Commissioning ceases to be a real-estate milestone and becomes the boot sequence of a machine. A future AI campus is not simply real estate containing computers; the campus itself is the computer, and the market evidence of Section 2—coherent DSPs, datacenter interconnect growing 40 percent annually, an optical transport industry restored to revenue levels last seen at the millennium—is the sound of that computer’s backplane being manufactured.[32]
5.5 Extending the Five-Layer AI Economy
Photon Constraint does not require creating a sixth layer of the Five-Layer AI Economy, and resisting that temptation is analytically important, because the constraint’s defining property is that it acts vertically across the existing layers rather than horizontally beside them. In Layer 1, Energy, optical efficiency determines how much scarce electricity remains available for computation, converting photonic engineering directly into effective generation capacity inside every power-capped facility.[35] In Layer 2, Chips, accelerators, HBM, switch ASICs, SerDes, optical DSPs, and silicon photonics are fusing into integrated computational systems whose competitive unit is the platform, not the die—the transformation that Broadcom’s networking-grows-as-fast-as-XPUs guidance and Marvell’s acquisition program both express in the language of money.[17][19] In Layer 3, Datacenters, fiber topology, rack layout, cooling, and physical distance become architectural decisions that directly set computational output, which is the campus-as-computer argument of the previous subsection. In Layer 4, Models, architectural choices—expert counts, context lengths, modality mixes, reasoning depth—determine communication patterns and bandwidth requirements, meaning model designers are now, whether they acknowledge it or not, network engineers. And in Layer 5, Applications and Agents, the ultimate economic value of the entire stack depends upon how efficiently the infrastructure converts capital and energy into useful inference, reasoning, and autonomous work—the goodput that agentic economics multiplies in importance. Photon Constraint is therefore not another layer; it is a constraint that passes through all five, and its severity at any moment is set by the layer adapting most slowly. That is precisely what makes it dangerous to ignore and valuable to solve: bottlenecks that live inside one layer get solved by one industry, while constraints that thread through every layer reprice all of them.
Section 6: What Have We Learned? Seven Pillars
A framework earns its keep by compressing evidence into principles that survive after the quarterly numbers fade. The preceding five sections assembled the disclosures, the physics, the economics, the geopolitics, and the forward trajectory of the Photon Constraint; this section distills them into seven pillars—five carried forward from the original architecture of this paper and two added because the evidence of 2026 demanded them. Together they constitute the operating logic of AI infrastructure for the remainder of the decade.
Pillar 1 — AI Compute Is Becoming a System, Not a Chip
The first lesson is that the semiconductor industry’s unit of competition is expanding. The GPU remains extraordinarily important—nothing in this paper diminishes it—but the performance of an advanced accelerator increasingly depends upon the memory feeding it, the packaging integrating it, the switching connecting it, the software orchestrating it, and the optical infrastructure surrounding it. Nvidia’s Vera Rubin platform is marketed as five purpose-built racks operating as one supercomputer; Amazon buys inference silicon and 1.6T optics in a single negotiation; Broadcom sells the accelerator, the switch, the DSP, and the laser from one portfolio.[5][8][16] The winning AI architecture of the late 2020s may therefore not necessarily contain the individually fastest processor. It may be the architecture capable of combining the largest quantity of processors into the most efficient single system. This is the central intellectual move behind Photon Constraint: stop measuring the processor alone and begin measuring the machine created by the network.
Pillar 2 — Bandwidth Is Becoming Productive Capacity
The second lesson is that bandwidth should be treated as an economic production input rather than as overhead. Electricity powers computation; semiconductors perform computation; but bandwidth is what makes distributed computation usable, and if information cannot reach an accelerator fast enough, a portion of its installed processing capacity becomes economically stranded—the Photon Utilization Gap of Section 3. Bandwidth therefore begins functioning like factory logistics: a manufacturing plant cannot maximize output simply by purchasing more machines if raw materials cannot reach those machines, and an AI factory cannot maximize intelligence output simply by purchasing more GPUs if data cannot reach those GPUs. The market has already delivered its verdict on this proposition in prices and growth rates—a $16.5 billion AI-optics component market in 2025 growing 60 percent annually, an optical transport industry restored past $18 billion, four-fifths of a major chipmaker’s AI revenue arriving through the fiber.[1][32][33] Bandwidth is no longer supporting infrastructure. Bandwidth is productive capacity, and it should be planned, financed, and valued as such.
Pillar 3 — Every Photon Saved Can Become a Compute Watt
The third lesson connects photonics directly to the energy crisis surrounding AI infrastructure. Datacenters have finite power envelopes set by interconnection agreements that take years to expand, inside a global trajectory heading from 415 TWh toward 945 TWh by 2030; within those envelopes, networking, memory, cooling, and compute all compete for the same electrons.[35] More efficient optical architectures therefore generate value twice: they improve communication performance, and they release electricity that can be reassigned to productive computation—which is exactly how the vendors now sell them, with Nvidia describing co-packaged optics as “freeing more power for compute” and Broadcom quantifying a roughly 70 percent reduction in optical interconnect power.[8][14] The race for more intelligence is simultaneously becoming a race for better allocation of every watt inside the AI factory. The Photon Constraint and the Power Constraint are converging into a single optimization problem, and the institutions that solve them jointly will outbuild the institutions that solve them separately.
Pillar 4 — Photonics Becomes Strategic Infrastructure
The fourth lesson is geopolitical. As optics becomes indispensable to frontier AI systems, governments will need to understand optical supply chains with the same seriousness they now apply to GPUs, HBM, advanced lithography, and semiconductor fabrication—mapping laser capacity, photonic-integration processes, optical DSP design, fiber and connector manufacturing, and the packaging that binds photonics to logic, across a supply chain that runs through the United States, Taiwan, Europe, Japan, South Korea, and China simultaneously. The relevant sovereignty question is no longer only “can a country acquire advanced GPUs?” It becomes “can a country assemble all of the technologies required to make those GPUs operate efficiently at frontier scale?” Nvidia’s $4 billion into Coherent and Lumentum, with its explicit U.S.-manufacturing emphasis, demonstrates that the most sophisticated private actor in the industry already answers procurement questions in sovereignty terms; public policy will follow, and the concept of optical sovereignty introduced in Section 4 will shape AI industrial strategy through 2030.[23][24]
Pillar 5 — Reliability Is the Hidden Currency of Scale
The fifth lesson—added to the original architecture of this paper because the 2026 disclosures made it unavoidable—is that at million-component scale, reliability becomes an economic variable of the same magnitude as bandwidth and power. A synchronous training run across a hundred thousand accelerators moves at the speed of its most fragile link, and a fabric assembled from hundreds of thousands of discrete pluggable transceivers presents hundreds of thousands of individually replaceable failure points, each capable of stalling the entire machine. This is why Nvidia quantifies its photonics advantage not only in watts but in time—five times longer sustained AI uptime, ten times network resiliency, translated directly into sustained application runtime—and why Broadcom’s hyperscale customers cite millions of accumulated CPO field-test hours as the decisive evidence for adoption.[8][10][14] Co-packaging optics does not merely save energy; it removes an entire category of failure from the machine. In the goodput economics of Section 3, mean time between network interruptions is capital efficiency wearing an engineering disguise, and the vendors that can prove reliability at scale will command premiums that spreadsheet comparisons of bandwidth per dollar will never explain.
Pillar 6 — The Academy Predicted This, and Its Roadmap Still Points Forward
The sixth lesson is epistemic: the Photon Constraint was foreseen, in remarkable detail, by an academic community that spent two decades being politely ignored by mainstream infrastructure economics. Stanford’s David A. B. Miller argued in the mid-2010s that interconnect energy, not logic energy, would become the binding limit on information processing, and that optics might be the only scalable escape.[31] Columbia’s Keren Bergman identified data movement as the singular roadblock to system scalability for AI workloads years before the first hundred-thousand-GPU cluster existed, and her laboratory’s roadmap—petabit-per-second chip-escape bandwidths at sub-picojoule-per-bit energies, with photonics packaged directly against compute and memory—describes precisely the generation of technology now moving through Ayar Labs, Celestial AI, and TSMC’s COUPE toward production.[26][29] She has called co-packaged optics the field’s long-sought “holy grail” of energy efficiency, and the industry has now spent tens of billions of dollars agreeing with her.[30] The practical implication for investors and policymakers is forward-looking rather than merely historical: the same laboratories are already publishing the next decade of the roadmap—deeper integration, optical memory fabrics, reconfigurable photonic topologies—and this time the market has every reason to read the papers early.
Pillar 7 — The Next AI Winner May Be the Company That Moves Intelligence Best
The final lesson is the most important. The first phase of generative AI rewarded companies capable of producing extraordinary computational processors; the next phase may disproportionately reward companies capable of moving information among those processors. That does not mean Nvidia becomes less important—Nvidia’s aggressive investment in photonic networking, from Spectrum-X Ethernet Photonics to its Coherent and Lumentum commitments, demonstrates precisely the opposite: the company itself recognizes that maintaining leadership requires controlling more of the complete AI system.[8][23] But it expands the competitive field. STMicroelectronics, Qualcomm, Broadcom, Marvell, Ayar Labs, Coherent, Corning, Lumentum, TSMC, and numerous less visible suppliers of lasers, connectors, and test equipment become participants in the economics of frontier intelligence, each capturing rents at a different point along the path a bit travels from one accelerator to another. The next trillion dollars of AI infrastructure therefore cannot be understood simply by asking who makes the GPUs. We must ask: who makes the GPUs communicate?
Conclusion: Intelligence Has to Travel
The artificial-intelligence boom trained investors, governments, and the public to think about computational scarcity through a remarkably simple visual symbol: the GPU. More GPUs meant more AI. A larger cluster meant a more powerful model. A larger semiconductor order meant greater strategic capacity. That simplification was genuinely useful during the first stage of the generative-AI infrastructure buildout, when the accelerator was the binding constraint and everything else could be procured from catalog. It will become progressively less useful as artificial-intelligence systems expand toward hundreds of thousands and potentially millions of tightly coordinated accelerators, because at that scale, adding another processor does not automatically add an equivalent amount of intelligence production. The processor must receive data. It must obtain model parameters. It must exchange activations. It must communicate with memory near and far. It must coordinate with other accelerators across a rack, a hall, and eventually a campus. It must participate in collective computations whose completion is gated by the slowest of its hundred thousand peers. And it must perform all of those functions quickly enough, and reliably enough, that billions of dollars of installed silicon do not spend economically meaningful periods waiting.
This is why STMicroelectronics’ September 9 disclosure matters far beyond one company’s revenue forecast. When approximately 80 percent of a major semiconductor company’s expected 2027 AI-datacenter revenue is projected to come from fiber-optic data-link chips—under long-term agreements carrying capacity commitments and cash advances from the largest cloud buyers on Earth—the market is providing an early, quantified indication that value creation is migrating outward from processors into the infrastructure connecting them.[1][3] Amazon and Qualcomm provide a second signal: their multigenerational collaboration joins custom AI silicon with advanced optical connectivity reaching 1.6 terabits per second inside a single ten-year commercial framework, implicitly recognizing that future accelerators and future networks must be designed together because they will succeed or fail together.[5][6] Nvidia provides perhaps the strongest signal of all. The company that became synonymous with GPU scarcity is itself building co-packaged photonics into the Rubin-era architecture intended to scale AI factories toward million-GPU dimensions, and it has committed $4 billion of strategic capital to the laser manufacturers beneath that architecture, because it concluded that its own future could be constrained by someone else’s optics capacity.[8][23] And the measurement layer of the industry confirms the pattern from the outside: AI-cluster optics revenue compounding at 60 percent annually against laser-chip shortages, optical transport restored to revenue levels last seen at the turn of the millennium, and the research community that predicted all of it now watching its twenty-year roadmap ship.[29][32][33][34]
The historical progression is therefore becoming visible in a form clean enough to state as sequence. The first constraint was access to advanced processors. Then came high-bandwidth memory. Then advanced packaging. Then power generation, transmission, and datacenter construction. Now another boundary is appearing between the processors themselves. That boundary is communication—and as electrical transmission encounters compounding challenges of reach, heat, power, and signal integrity at every new speed grade, communication increasingly becomes a problem solved with light. Each prior constraint followed the same arc: dismissed as a component detail, discovered as a shortage, repriced as strategic infrastructure, and finally institutionalized as a permanent planning variable. Photonics entered the third stage of that arc during the ten months chronicled in this paper; it will spend 2027 through 2030 completing the fourth.
This is why Photon Constraint is the appropriate title for this paper. Photon identifies the physical carrier becoming indispensable to the next generation of AI infrastructure—the particle that does not heat the board it crosses, does not care whether its journey is three millimeters or three kilometers, and does not degrade the signal it carries in the ways that electrons crossing copper unavoidably do. Constraint identifies the economic consequence that emerges when communication capacity fails to grow as rapidly as computational capacity: stranded FLOPs, widening utilization gaps, and intelligence that exists on invoices but not in output. Together, the two words describe a structural transformation in the Five-Layer AI Economy rather than a networking trend within one of its layers.
The term does not suggest that GPUs cease to matter. Quite the opposite. GPUs become so numerous, so powerful, and so valuable that connecting them efficiently becomes one of the principal conditions for realizing their value, and the irony of the next stage of artificial intelligence may therefore be that success in manufacturing ever more powerful silicon makes silicon itself less useful as the sole measurement of AI capability. A country can own processors. A hyperscaler can build datacenters. A frontier laboratory can train increasingly sophisticated models. But at sufficient scale, intelligence must continuously travel among all of those computational resources, and the speed, cost, distance, reliability, and energy efficiency of that movement increasingly determine how much theoretical computation becomes usable intelligence. Within the Five-Layer AI Economy, this makes the Photon Constraint unusually consequential precisely because it connects every layer at once: electricity provides the energy; semiconductors create computational capability; datacenters provide physical scale; models generate communication patterns; and applications and agents consume the resulting intelligence. The network is what allows those five layers to operate as one industrial system rather than five adjacent industries.
And this is ultimately the reason the title fits the paper so well. Artificial-intelligence scaling is approaching the point where owning more GPUs is no longer sufficient; the next frontier is making those GPUs communicate fast enough, far enough, and efficiently enough to behave as one machine. When that happens—and the evidence assembled here suggests it is happening now, on earnings calls and in fab groundbreakings rather than in laboratory previews—the decisive scarcity of the AI economy is no longer merely how many computations silicon can perform. It is how much intelligence light can carry. That is the Photon Constraint.
Footnotes / Endnotes:
[1] Reuters (via Global Banking & Finance Review). “STMicro’s 2027 AI revenue to be four-fifths optics, CFO says (remarks of Lorenzo Grandi, Citi Global TMT Conference, September 9, 2026).” Reuters, September 9, 2026. https://www.globalbankingandfinance.com/stmicros-2027-ai-revenue-four-fifths-optics-cfo/
[2] STMicroelectronics. “STMicroelectronics raises its revenue ambition for Data Centers amidst continued strong demand for AI infrastructure.” STMicroelectronics Newsroom, June 2, 2026. https://newsroom.st.com/media-center/press-item.html/c3396.html
[3] Investing.com. “STMicroelectronics at Citi’s 2026 Global TMT Conference: margin lift ahead (transcript).” Investing.com, September 9, 2026. https://www.investing.com/news/transcripts/stmicroelectronics-at-citis-2026-global-tmt-conference-margin-lift-ahead-93CH-4894576
[4] Morningstar Equity Research. “STMicro: Company Lifts Revenue Forecasts Amid AI Boom.” Morningstar, June 2026. https://www.morningstar.com/stocks/stmicro-company-lifts-revenue-forecasts-amid-ai-boom
[5] Qualcomm Technologies, Inc.. “Qualcomm and AWS Collaborate on AI Inference Chips and 1.6T Connectivity.” HPCwire, September 8, 2026. https://www.hpcwire.com/off-the-wire/qualcomm-and-aws-collaborate-on-ai-inference-chips-and-1-6t-connectivity/
[6] MLQ News. “Qualcomm, Amazon Announce Multi-Generation AWS Inference Silicon Deal (SEC 8-K warrant details).” MLQ.ai, September 9, 2026. https://mlq.ai/news/qualcomm-amazon-announce-multi-generation-aws-inference-silicon-deal/
[7] Futurum Group. “Qualcomm Lands $60 Billion Custom Silicon and 1.6T Optics Deal with AWS, Backed by a $4 Billion Warrant.” Futurum, September 2026. https://futurumgroup.com/insights/qualcomm-lands-60-billion-custom-silicon-and-16t-optics-deal-with-aws-backed-by-a-4-billion-warrant/
[8] NVIDIA. “NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide.” NVIDIA Newsroom, May 31, 2026. https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory
[9] NVIDIA (statements of Jensen Huang and C. C. Wei). “NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs.” NVIDIA Newsroom, March 18, 2025. https://nvidianews.nvidia.com/news/nvidia-spectrum-x-co-packaged-optics-networking-switches-ai-factories
[10] NVIDIA. “Silicon Photonics Networking for Agentic AI (co-packaged optics product overview).” NVIDIA.com, 2026. https://www.nvidia.com/en-us/networking/products/silicon-photonics/
[11] UALink Consortium. “UALink 200G 1.0 Specification (and subsequent 2026 specifications: In-Network Compute, Chiplet Definition, Manageability).” UALink Consortium. https://ualinkconsortium.org/specification/
[12] UALink Consortium (statements of Kurtis Bowman and Sameh Boujelbene). “UALink Consortium Releases the Ultra Accelerator Link 200G 1.0 Specification.” HPCwire, April 8, 2025. https://www.hpcwire.com/off-the-wire/ualink-consortium-releases-the-ultra-accelerator-link-200g-1-0-specification/
[13] SDxCentral. “UALink Consortium releases 200G 1.0 specification for AI accelerator interconnects.” SDxCentral, April 2025. https://www.sdxcentral.com/news/ualink-consortium-releases-200g-10-specification-for-ai-accelerator-interconnects/
[14] Broadcom Inc.. “Broadcom Announces Tomahawk 6 – Davisson, the Industry’s First 102.4-Tbps Ethernet Switch with Co-Packaged Optics.” Broadcom Investor Relations, October 8, 2025. https://investors.broadcom.com/news-releases/news-release-details/broadcom-announces-tomahawkr-6-davisson-industrys-first-1024
[15] ServeTheHome. “Broadcom Tomahawk 6 – Davisson 102.4T Switch with Co-Packaged Optics Shipping (TSMC COUPE photonic engine details).” ServeTheHome, October 2025. https://www.servethehome.com/broadcom-tomahawk-6-davisson-102-4t-switch-with-co-packaged-optics-shipping/
[16] The Motley Fool. “Broadcom (AVGO) Q2 2026 Earnings Call Transcript (remarks of Hock E. Tan).” Fool.com, June 3, 2026. https://www.fool.com/earnings/call-transcripts/2026/06/03/broadcom-avgo-q2-2026-earnings-transcript/
[17] Investing.com. “Earnings call transcript: Broadcom tops Q3 2026 estimates as AI sales surge (remarks of Hock E. Tan).” Investing.com, September 2026. https://www.investing.com/news/transcripts/earnings-call-transcript-broadcom-tops-q3-2026-estimates-as-ai-sales-surge-93CH-4886849
[18] Tech Times. “Broadcom Custom AI Chip Revenue Surges 221% to $16.7B, Q4 Guidance Disappoints.” TechTimes, September 3, 2026. https://www.techtimes.com/articles/326405/20260903/broadcom-custom-ai-chip-revenue-surges-221-167b-q4-guidance-disappoints.htm
[19] Marvell Technology, Inc. (statement of Matt Murphy). “Marvell to Acquire Celestial AI, Accelerating Scale-up Connectivity for Next-Generation Data Centers.” Marvell Investor Relations, December 2, 2025. https://investor.marvell.com/news-events/press-releases/detail/1000/marvell-to-acquire-celestial-ai-accelerating-scale-up-connectivity-for-next-generation-data-centers
[20] Marvell Technology, Inc.. “Marvell Completes Acquisition of Celestial AI.” Marvell Newsroom, February 2, 2026. https://www.marvell.com/company/newsroom/marvell-completes-acquisition-of-celestial-ai.html
[21] Marvell Technology, Inc.. “Marvell Ushers in the 1.6T Era with Expanded Optical DSP Platform Portfolio, Redefining AI Data Center End-to-End Connectivity (including XConn acquisition, January 9, 2026).” StorageNewsletter, April 3, 2026. https://www.storagenewsletter.com/2026/04/03/marvell-ushers-in-the-1-6t-era-with-expanded-optical-dsp-platform-portfolio-redefining-ai-data-center-end-to-end-connectivity/
[22] Counterpoint Research. “Celestial AI Acquisition Perfectly Positions Marvell For Upcoming Multi-Rack Scale-Up Boom.” Counterpoint Research, December 2025. https://counterpointresearch.com/en/insights/Celestial-AI-Acquisition-Perfectly-Positions-Marvell-For-Upcoming-Multi-Rack-Scale-Up-Boom
[23] CNBC. “Nvidia to invest $4 billion into photonics companies Coherent and Lumentum.” CNBC, March 2, 2026. https://www.cnbc.com/2026/03/02/nvidia-investment-coherent-lumentum.html
[24] Compound Semiconductor (statements of Jensen Huang and Michael Hurlston). “NVIDIA to invest $4b in Lumentum and Coherent.” CompoundSemiconductor.net, March 2026. https://compoundsemiconductor.net/article/123669/NVIDIA_to_invest_4b_in_Lumentum_and_Coherent
[25] Optics.org (statement of Jim Anderson). “Nvidia backs Lumentum and Coherent with $4BN cash investment.” optics.org, March 2026. https://optics.org/news/nvidia-backs-lumentum-and-coherent-with-4bn-cash-investment
[26] Ayar Labs (statement of Mark Wade). “Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics.” Ayar Labs Press Release, March 3, 2026. https://ayarlabs.com/news/ayar-labs-closes-500m-series-e-accelerates-volume-production-of-co-packaged-optics/
[27] StorageNewsletter (statement of Gabe Cahill, Neuberger Berman). “Ayar Labs Closes $500M Series E, Accelerates Volume Production of Co-Packaged Optics.” StorageNewsletter, March 13, 2026. https://www.storagenewsletter.com/2026/03/13/ayar-labs-closes-500m-series-e-accelerates-volume-production-of-co-packaged-optics/
[28] Berkeley Wireless Research Center. “Ayar Labs Closes $500M Series E (investor and valuation details).” BWRC, University of California, Berkeley, March 2026. https://bwrc.berkeley.edu/news/ayar-labs-closes-500m-series-e-accelerates-volume-production-co%E2%80%91packaged-optics
[29] Keren Bergman (Columbia University), quoted in Primeur Magazine. “Keren Bergman wins $4.8 million DARPA grant (on data movement as the roadblock to system scalability).” Primeur Magazine, 2019. http://primeurmagazine.com/weekly/AE-PR-10-19-135.html
[30] Photonics Online (citing Prof. Keren Bergman via Laser Focus World). “Silicon Photonics Tackles AI’s Data Bottleneck.” PhotonicsOnline. https://www.photonicsonline.com/doc/silicon-photonics-tackles-ai-s-data-bottleneck-0001
[31] David A. B. Miller (Stanford University). “Attojoule Optoelectronics for Low-Energy Information Processing and Communications: a Tutorial Review.” Journal of Lightwave Technology 35(3), 2017; arXiv:1609.05510. https://arxiv.org/abs/1609.05510
[32] Dell’Oro Group (statement of Jimmy Yu). “Optical Transport Equipment Market Forecast to Grow 16 Percent in 2026.” Dell’Oro Group, May 19, 2026. https://www.delloro.com/news/optical-transport-equipment-market-forecast-to-grow-16-percent-in-2026/
[33] LightCounting. “AI creates a new wave in demand for optical transceivers and accelerates CPO adoption (Optics for AI report).” LightCounting, January 2026. https://www.lightcounting.com/newsletter/en/january-2026-optics-for-ai-clusters-366
[34] LightCounting. “Demand for optical connectivity continues to surprise (April 2026 Market Forecast).” LightCounting, April 2026. https://www.lightcounting.com/newsletter/en/april-2026-market-forecast-379
[35] International Energy Agency (IEA). “Energy and AI — Executive Summary (global data centre electricity to ~945 TWh by 2030).” IEA, Paris. https://www.iea.org/reports/energy-and-ai/executive-summary
[36] Investing.com. “STMicroelectronics Breaks Out as AI Data-Center Revenue Nears $1 Billion (multi-year AWS commercial agreement of February 9, 2026).” Investing.com, June 2026. https://www.investing.com/analysis/stmicroelectronics-breaks-out-as-ai-datacenter-revenue-nears-1-billion-200681347
