Introduction: The Megawatts Between Now and 2030
On July 8, 2026, Meta broke ground on something considerably larger than another data center. In Sturgeon County, in Alberta’s Industrial Heartland north-east of Edmonton, the company began building a one-gigawatt, AI-optimized campus that it describes as its first data center in Canada and the thirty-third in its global fleet, representing an investment of more than C$13 billion.[1] Meta expects roughly 3,000 construction workers on site at the peak of development, more than 300 permanent jobs once the facility is operational, and approximately C$60 million of local road and water improvements, while the campus itself is designed around a closed-loop, liquid-cooled, dry-cooling system that the company says requires no operational water for cooling.[2] Reuters, reporting on the announcement in Calgary alongside Premier Danielle Smith, estimated that the finished facility will consume about as much electricity as 800,000 homes.[3] Alberta’s appeal is easy to understand: abundant land, a mature power industry, a cold climate that flatters cooling economics, discounted natural gas, established transmission corridors, and an increasingly explicit provincial ambition to become a North American center for artificial-intelligence infrastructure.
Yet the most revealing feature of the Sturgeon County project is not the servers, the buildings, the construction budget, or even the eventual gigawatt-scale demand. It is the gap between the moment the data center can begin consuming electricity and the moment the permanent energy infrastructure intended to support it becomes available. On July 2, 2026, Pembina Pipeline, Morgan Stanley Infrastructure Partners and Kineticor announced a positive final investment decision on the Greenlight Electricity Centre, a 932-megawatt gas-fired combined-cycle plant in Sturgeon County that the partners describe as the customer’s “long-term, behind-the-meter power provider.”[4] Trade reporting placed the plant’s cost at roughly C$4.6 billion and its expected commercial operation in the second half of 2030.[5] Six days later, Capital Power announced a long-term energy supply agreement of more than ten years for 250 megawatts of capacity and energy, with the load “anticipated to be in service in the back half of 2028,” backed by its existing Alberta generation fleet.[6] Data Center Frontier described the logic plainly: the 250 megawatts begin flowing from an existing fleet in the second half of 2028, before Greenlight enters service, while Meta works with Capital Power, the transmission owner AltaLink, and the Alberta Electric System Operator on the campus’s “earlier and grid-connected energy requirements.”[7] Invest Alberta, the province’s investment-attraction agency, adds the final piece: under Alberta’s bring-your-own-power framework the project will be completed in two phases with a combined total of roughly 970 megawatts of grid-connected power, with future on-site gas generation developed alongside Kineticor and Pembina.[8]
The sequence is therefore not a single power plant feeding a single building. It is a staircase. The server halls rise first. A slice of load is energised through the interconnected Alberta system. A contracted block of 250 megawatts from an existing independent power producer’s fleet carries the campus through the middle years. A dedicated combined-cycle plant, sanctioned in mid-2026 and not expected until late 2030, eventually assumes the role of primary supplier, with an expansion path that could take the site toward 1.8 gigawatts.[9] Between the first rack and the first turbine lies an interval of roughly two years, and that interval deserves its own vocabulary.
Capital Power’s chief executive, Avik Dey, framed the bargain in terms that make the underlying economics unusually explicit, and the Premier of Alberta framed it in terms that make the underlying politics equally explicit:
“This agreement is exactly the kind of opportunity we have been preparing for – AI infrastructure will be built where power is available, reliable and scalable, and with the support of Capital Power’s fleet, Alberta meets the mark.”
— Avik Dey, President and CEO, Capital Power [10]
“This agreement is another example of Alberta’s plan delivering results. Investments like this create thousands of jobs, generate new revenue that helps pay for the services Albertans rely on and help reduce transmission costs for consumers.”
— Danielle Smith, Premier of Alberta [10]
The artificial-intelligence infrastructure boom is increasingly governed by two clocks that no longer keep the same time. The compute clock is accelerating: hyperscalers can acquire land, pour foundations, drop in prefabricated electrical skids, deploy direct-to-chip liquid cooling, populate halls with accelerator racks, and energise portions of a campus in successive stages. The power clock moves differently and more slowly, because new transmission lines require routing studies, landowner negotiations and regulatory approvals; substations require large power transformers and high-voltage switchgear whose manufacturing queues stretch across years; gas turbines are now sold against production slots at the end of the decade; pipelines require engineering and rights-of-way; nuclear plants require licensing processes measured in years rather than quarters; and large renewable portfolios require both interconnection and the transmission to reach load. The result is a structural inversion in which entire generation facilities can take longer to complete than the buildings they are intended to power.
That mismatch creates a new category of electricity demand, one that is neither ordinary permanent utility load nor conventional emergency backup. These megawatts exist because the compute infrastructure is ready, or nearly ready, while its ultimate energy system is not. This paper calls them Bridge Megawatts, and defines them as the temporary or transitional megawatts supplied through a shared electrical system, utility arrangement, bilateral power contract, or interim generation configuration that allow a hyperscale AI facility to begin or expand operations before its intended long-term generation and energy architecture is complete.
The word “public” in the working title refers to the shared, interconnected grid rather than to public ownership of generation. Alberta runs an energy-only competitive market with privately owned generation; Texas runs a largely islanded competitive market; Virginia is served by a vertically integrated investor-owned utility participating in PJM; Georgia is served by a vertically integrated utility outside any organised market. What these very different systems increasingly share is the question of whether, and on what terms, the existing interconnected system should serve as an energy bridge between the completion of an AI factory and the completion of the infrastructure eventually intended to power it. The scale of that question is already visible in the numbers regulators themselves publish. In June 2025, Alberta’s system operator reported 29 proposed data-center projects seeking more than 16 gigawatts of grid connections against a city of Edmonton whose entire load is roughly 1,400 megawatts.[11] In August 2026, the Governor of Texas reported that ERCOT was considering approximately 474 gigawatts of connection requests, some 90 percent from data centers, in a system whose all-time peak had just reached 91.1 gigawatts on July 22, 2026.[12]
Figure 1. The Scale Mismatch Behind Bridge Megawatts
Sources: Office of the Texas Governor and Morrison Foerster for ERCOT request volume and the July 22, 2026 record peak;[12] AESO for Alberta requests, the Edmonton comparison, and the 1,200-MW interim limit.[11]
Meta’s Alberta development makes the mechanism unusually visible, but it is unlikely to remain unusual. Between 2027 and 2030, hyperscale campuses may routinely be designed around staged power architectures: first a hundred megawatts drawn from the grid, then several hundred through contracted generation, then dedicated turbines, batteries, fuel cells, nuclear offtake, renewable portfolios or microgrids, eventually producing a campus that remains connected to the wider system but depends far less on it for ordinary operation. If that model spreads, the organising question of AI-era electricity policy changes. It is no longer simply “where will the permanent power come from?” It becomes “who supplies the electricity between the moment the GPUs arrive and the moment their permanent power plant arrives, who pays for the infrastructure that makes that possible, and what happens to that infrastructure when the bridge is no longer needed?” That interval, this paper argues, could become one of the defining infrastructure problems of the Five-Layer AI Economy, the stack that runs from energy through chips and networks, through data centers, to models and applications.
Why the Title “Bridge Megawatts”
I chose the phrase Bridge Megawatts because it captures a largely overlooked temporal problem inside the AI infrastructure boom. A hyperscaler may know precisely where its future electricity will come from, may have signed the tolling agreement, sanctioned the plant, and reserved the turbines, and still face several years during which its data center becomes operational before that generation system is complete. The megawatts required during that interval form a bridge between two infrastructure states: dependence on the shared grid today, and increasingly dedicated, contracted or self-supplied energy tomorrow. The metaphor also carries an important caution, because bridges are not free, are not self-dismantling, and are usually engineered to outlast the traffic that justified them.
The term deliberately shifts the analysis away from the familiar and somewhat exhausted debate over whether AI consumes “too much” electricity and toward a more precise set of economic questions: how temporary hyperscale demand should be supplied, financed, contracted, guaranteed and eventually unwound. A bridge may last two years, five years, or considerably longer if the destination plant slips. At gigawatt scale, however, even a temporary bridge can require permanent transmission assets, new substations, reserve margins, generation commitments and multi-decade financial obligations. That contradiction, temporary load supported by long-lived infrastructure, is the analytical heart of this paper, and it is why a more generic title about data-center electricity demand would miss the point.
It is worth noting that the power industry already uses the phrase “bridge power,” but typically in the opposite direction. Equipment vendors and developers speak of bridge power as on-site gas engines, turbines, fuel cells or batteries that carry a campus until its utility interconnection is ready; the Texas developer Prometheus Hyperscale, for example, contracted Conduit Power for a five-year initial term of gas generators and batteries while its campuses await grid service, with the engines potentially remaining as backup or moving to other sites afterwards.[13] This paper argues that the AI economy now contains bridges running in both directions: generation-first bridges, where private machines carry load until the grid arrives, and grid-first bridges, where the shared system carries load until private machines arrive. Alberta is the clearest example of the second type, and it is the second type that raises the sharper questions of public cost, reliability obligation and intertemporal fairness.
Section 1. Two Clocks: When Data Centers Arrive Before Their Power Plants
Every infrastructure transition is ultimately a problem of synchronisation, and the AI build-out is no exception. The railways of the nineteenth century needed rolling stock, rails, stations and coal to arrive in roughly the same decade; the electrification of the twentieth century needed generators, wires and appliances to mature together. The AI economy of the 2020s differs in one crucial respect: the two most important inputs, compute and electricity, are produced by industries that operate on radically different cycles of capital formation, permitting and manufacturing. Compute is purchased in quarters and depreciated over roughly five to six years; the electrical assets that feed it are planned over five to ten years and depreciated over thirty to fifty. This section sets out the two clocks, explains why their divergence has become sharper in 2025 and 2026 rather than easing, and argues that the divergence converts time itself into an energy resource.
1.1 The Compute Clock
The first clock belongs to artificial intelligence, and it has never run faster. When Alphabet, Microsoft, Meta and Amazon reported second-quarter 2026 results between July 22 and July 30, their combined 2026 capital-expenditure guidance summed to roughly US$730 billion at the midpoints, nearly double the approximately US$410 billion they spent in 2025.[14] Amazon raised its 2026 plan to roughly US$220 billion, Alphabet to a range of roughly US$195–205 billion, and Meta lifted the floor of its range, while Microsoft was tracking toward approximately US$190 billion on a calendar-year basis.[15][16] The financial strain of that acceleration is visible: Alphabet posted its first negative free-cash-flow quarter since its 2004 initial public offering, and its shares fell sharply after it raised guidance, pulling the other hyperscalers down with it before they reported.[14][17] Yet none of the four reduced spending, and Amazon’s chief executive told investors that capacity constraints were likely to persist through 2027, while Meta’s chief financial officer described capital planning as “highly dynamic.”[18]
Figure 2. The Compute Clock: Big-Four Hyperscaler Capital Expenditure
Source: Futurex Capital AI Lab synthesis of Q2 2026 earnings for Alphabet (July 22), Microsoft and Meta (July 29), and Amazon (July 30).[14] Guidance bases differ across companies; figures are approximate.
The most analytically important disclosure of the Q2 2026 season was not a number but a method. According to one widely read synthesis of the calls, all four companies described the same de-risking playbook: commit early to the long-lived assets, meaning land, data-center shells and power, and decide on the short-lived assets, meaning the accelerators that account for most of the cost, only a few months before they are needed, once demand is visible.[15] That playbook is the compute clock’s answer to the power clock. It concedes that electricity and buildings are the slow, scarce, long-duration inputs, and that chips are the fast, abundant, short-duration input. It also means that hyperscalers now secure power options years before they know exactly how many racks they will install, which is precisely the condition under which bridge arrangements become attractive.
AI infrastructure is also increasingly modular. Buildings are constructed in phases; electrical rooms arrive as prefabricated skids; cooling loops are expanded hall by hall; and accelerator clusters are commissioned incrementally rather than waiting for an entire campus to be finished. A planned one-gigawatt campus therefore does not need a gigawatt on its first operational morning. Its load is more likely to rise through a sequence such as fifty, then one hundred and fifty, then two hundred and fifty, then five hundred, then seven hundred and fifty megawatts, before reaching its full design capacity. That ramp is economically decisive because every energised rack produces useful compute long before the campus is complete, and for a company spending tens of billions of dollars a quarter on accelerators, leaving installed hardware dark while a distant power plant is finished represents an enormous and compounding opportunity cost. The incentive is therefore overwhelming: energise whatever can be energised, as soon as it can be energised, from whatever source is available.
No executive has described that incentive more candidly than Microsoft’s chief executive, whose remarks on the BG2 podcast in late 2025 have become the canonical statement of the two-clock problem:
“The biggest issue we are now having is not a compute glut, but it’s power – it’s sort of the ability to get the builds done fast enough close to power. So, if you can’t do that, you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today. It’s not a supply issue of chips; it’s actually the fact that I don’t have warm shells to plug into.”
— Satya Nadella, Chairman and CEO, Microsoft [19]
“We are, and have been, short now for many quarters. I thought we were going to catch up. We are not. Demand is increasing.”
— Amy Hood, Executive Vice President and CFO, Microsoft [19]
Nat Bullard, chief strategy officer of the research group Halcyon, translated the same point into the language of cloud growth in mid-2026, noting that Amazon Web Services grew 28 percent, Microsoft Azure 40 percent and Google Cloud 63 percent between the first quarter of 2025 and the first quarter of 2026, growth rates that double revenue in two years or less in already mature businesses.[20]
“That revenue can only be serviced with compute, and that compute can only serve when energized.”
— Nat Bullard, Chief Strategy Officer, Halcyon [20]
1.2 The Power Clock
The second clock belongs to the electricity system, and in 2026 it is not merely slow but in several respects slowing. Power infrastructure is physically larger than compute infrastructure, geographically dispersed, regulated by different institutions at different levels of government, and dependent on longer and more concentrated industrial supply chains. The most telling indicator is the gas turbine, the workhorse of dispatchable new capacity. GE Vernova ended the second quarter of 2026 with a combined gas-turbine backlog and slot-reservation position of 116 gigawatts, up from 100 gigawatts a quarter earlier and 83 gigawatts at the end of 2025, and management expects at least 125 gigawatts under contract by the end of the year.[21][22] One industry analysis of the July 22 earnings call summarised the practical implication bluntly: a heavy-duty turbine ordered today will not arrive until 2031.[22] Earlier in the year, the company had explained why customers were skipping 2029 altogether:
“We sold a lot of 2030 slots because the reality is we had a lot of customers looking at planning with EPC schedules and other dynamics needed the ’30 slot more than ’29.”
— Scott Strazik, CEO, GE Vernova [23]
Pricing tells the same story as timing. Power Engineering reported that GE Vernova’s new gas-turbine orders in the first half of 2026 were pricing ten to twenty points higher per kilowatt than orders booked only a quarter earlier, and cited Wood Mackenzie’s projection that gas-turbine prices could reach roughly US$600 per kilowatt by the end of 2027, nearly triple 2019 levels.[23] The same company’s electrification segment, which manufactures the substations, switchgear and transformers without which no generator or grid connection can serve a data center, booked more data-center equipment orders in the first quarter of 2026 than in all of 2025.[23] In other words, the bottleneck is not a single component but the entire chain from turbine to transformer to transmission line.
At the system level, the most important institutional acknowledgement of the power clock came from PJM, the largest wholesale market in North America. Its July 2026 capacity auction for the 2028/2029 delivery year cleared at the administratively capped price of US$325 per megawatt-day and procured 6,831 megawatts less than its reliability requirement, the second consecutive auction in which the entire regional transmission organisation fell short, prompting PJM to seek approval for a special backstop procurement in September 2026.[24] PJM’s own investigation of the previous shortfall used language that could serve as the epigraph of this paper:
“Given the demand forecasts, a reliability deficit was practically unavoidable, as the required volume of new generation could not be built fast enough to match the surge in consumption.”
— PJM Interconnection, 2027/2028 BRA Reserve Target Shortfall Report [25]
PJM called that condition a “transition gap,” and its executives have been equally direct about its cause. After the December 2025 auction, PJM’s executive vice president for market services observed that data centers’ demand for electricity continued to far outstrip new supply, and after the July 2026 auction its chief executive framed the same imbalance in general terms.[26]
“These auction results show that demand for electricity continues to grow faster than electricity supply.”
— David Mills, President and CEO, PJM Interconnection [24]
Rob Gramlich of Grid Strategies captured the problem in a phrase that has become common currency in 2026 policy debate, describing a “mismatch in timing” between data-center demand and the grid’s ability to serve it, while the North American Electric Reliability Corporation’s 2025 long-term assessment projected more than 224 gigawatts of summer-peak growth over the coming decade, 69 percent more than it had forecast only a year earlier.[27] The academic literature arrives at the same conclusion from a different direction. Stanford’s Sally Benson situates data centers within a broader, multi-source return of load growth after two decades of stagnation, and draws the operational lesson:
“This wasn’t much of a surprise because it was really a national priority to bring manufacturing on shore. We were beginning to electrify the vehicle fleet. But at the same time, something else emerged: demand for electricity to power data centers. … Adding huge loads very quickly is always challenging.”
— Sally Benson, Precourt Family Professor of Energy Science and Engineering, Stanford University [28]
Table 1. The Two Clocks Compared
| Dimension | Compute clock | Power clock |
| Core assets | Accelerators, servers, networking, liquid-cooling loops, prefabricated electrical rooms | Transmission lines, substations, large power transformers, turbines, pipelines, reactors, storage |
| Typical decision horizon | Months; chips increasingly ordered only a few months before deployment[15] | Years; heavy-duty turbine orders now land in 2030–2031 slots[22] |
| Asset life | Roughly five to six years for servers | Thirty to fifty years or more for wires and plants |
| Governing institutions | Corporate boards, capital markets, chip suppliers | System operators, utility commissions, FERC, provincial and state governments, municipalities |
| Scaling mode | Modular, hall by hall, rack by rack | Lumpy; a substation or a combined-cycle unit arrives whole or not at all |
| 2026 signal | ~US$730B big-four capex guidance[14] | PJM 6.8-GW shortfall; 116-GW turbine backlog[24][21] |
Author’s synthesis of the sources cited in each cell.
The consequence is a reversal of traditional industrial development. Historically, energy-intensive factories followed established power availability: aluminium smelters went to hydroelectric valleys, and chemical plants to gas fields and industrial grids. Increasingly, AI developers choose attractive sites first, for reasons of land, fibre, climate, talent or political welcome, and then assemble the ultimate energy system around those sites afterwards. That reversal is what makes a bridge necessary.
1.3 Alberta as the Prototype
Meta’s Sturgeon County project illustrates the mismatch with unusual clarity because every stage of the staircase has a named counterparty and a public date. The table below reconstructs the sequence from company and regulatory disclosures.
Table 2. The Sturgeon County Bridge: A Timeline of Infrastructure States
| Date | Milestone | Bridge significance |
| June 2025 | AESO allocates a one-time 1,200-MW interim large-load connection limit for projects seeking service in 2027–2028[11] | Creates the scarce near-term grid capacity from which early bridge megawatts are drawn |
| July 2, 2026 | Positive FID on 932-MW Greenlight Electricity Centre (Pembina 47.5%, MSIP 47.5%, Kineticor 5%)[4] | Destination plant is sanctioned but four years from service |
| July 8, 2026 | Meta breaks ground on 1-GW, >C$13B campus[1] | Compute clock starts |
| July 2026 | Capital Power ESA: 250 MW, >10 years, from existing Alberta fleet[6] | Contracted bridge delivered over the shared system |
| H2 2028 | Anticipated in-service date for the 250-MW load[6] | Bridge interval begins in earnest |
| H2 / late 2030 | Greenlight expected in service; Meta holds long-term tolling agreement[5][3] | Dedicated behind-the-meter supply becomes primary |
| Beyond 2030 | Site capable of scaling toward 1.8 GW; Greenlight expandable beyond 1.8 GW[7][9] | Bridge may re-open with each expansion phase |
Dates and quantities as disclosed by the cited parties; the exact physical allocation of grid, contracted and dedicated supply at any given moment is not public.
The important analytical point is not whether any single contract can be classified neatly as temporary or permanent. Capital Power describes its agreement as long-term and designed to deliver stable contracted cash flows, and it may well outlast the arrival of Greenlight as a firming or backup product.[6] The point is the sequence itself: data-center construction, then initial grid service, then contracted supply delivered over the shared system, then dedicated generation, then a mature hybrid energy system. That sequence is likely to become common across AI infrastructure, and it is worth noticing who celebrated it. Pembina’s chief executive described dedicated gas-to-power infrastructure as a new growth platform for a midstream company, and Morgan Stanley’s infrastructure fund framed dispatchable power as the foundation of the AI economy.
“Dedicated, contracted gas-to-power infrastructure represents a promising new growth platform, through which we are also helping to catalyze new natural gas demand that will provide additional benefits throughout our business.”
— Scott Burrows, President and CEO, Pembina Pipeline [4]
“Reliable, dispatchable power is the foundation of the AI and cloud economy and Greenlight will deliver it at scale to one of Canada’s most important new data centre developments.”
— Chris Ortega, Head of the Americas, Morgan Stanley Infrastructure Partners [4]
1.4 The Hidden Economic Value of Time
Suppose a hyperscaler completes enough server halls to host 250 megawatts of compute two years before its dedicated generation becomes available. Those 250 megawatts are not merely an entry in a utility load forecast; they represent accelerated revenue from cloud customers, earlier training runs for frontier models, earlier inference capacity for products already in the market, and potentially an earlier competitive response to a rival’s model release. Electricity available today therefore carries a strategic value that an otherwise identical megawatt available in 2030 does not. Bridge power, in short, contains a time premium.
Economists would recognise this as an option value problem layered on top of a conventional capacity problem. The hyperscaler is not simply buying energy; it is buying the right to convert sunk capital, in the form of installed or soon-to-be-installed accelerators, into revenue earlier than it otherwise could. The IMF’s 2025 working paper on AI and energy documented that electricity costs for vertically integrated AI companies nearly doubled between 2019 and 2023, which suggests that electricity is becoming a material line item for these firms, yet the time premium implies that the price of a bridge megawatt may matter less to the buyer than its availability.[29] That asymmetry has consequences for regulators. When a customer values timing far more than price, the conventional regulatory tool of setting a just and reasonable rate may not ration scarce near-term capacity efficiently, and the terms of access, including collateral, curtailment obligations and commitments to exit, become the real price.
The Five-Layer AI Economy consequently acquires another scarce commodity. It is not simply the megawatt but the megawatt available at the correct moment, at the correct location, with the correct reliability characteristics. Recognising that commodity changes how one reads corporate energy announcements: a sanctioned plant arriving in 2030 and a contracted 250-megawatt block arriving in 2028 are not substitutes, because they serve different moments in the life of the same campus.
1.5 From Power Availability to Power Sequencing
Data-center site selection has historically asked whether a location has sufficient power. The two-clock problem suggests that the more useful question is whether a location has sufficient power sequencing. A region may credibly advertise two gigawatts of future generation, but that headline says little about whether one hundred megawatts will be available in 2027, another three hundred in 2028, another five hundred in 2029, and the remainder in 2030. The competitive geography of AI may therefore be determined increasingly by the shape of a region’s power ramp rather than by the ultimate size of its power supply.
Figure 3. Illustrative Bridge-Megawatt Profile for a 1-GW AI Campus
Author’s stylised model, not any company’s actual schedule. The shape is informed by the Sturgeon County sequence of initial grid service, a 250-MW contracted block from H2 2028, and dedicated combined-cycle supply from H2 2030.[6][5] The shaded band marks the bridge interval in which the campus draws on the shared system before its destination plant arrives.
Alberta’s regulatory design already reflects this shift. AESO’s interim limit was not an abstract statement of how much load the province could eventually serve; it was a calculation of the additional large load that could be connected reliably by 2028 using a conservative scenario for available dispatchable generation, applied only to projects of at least 75 megawatts that did not require new transmission reinforcements.[11][30] All 1,200 megawatts were allocated, and all remaining requests, including additional megawatts for projects that had received a share, were deferred to the second phase of the program.[31] The interim limit is, in effect, a published bridge-capacity number, and the fact that it was fully subscribed while more than sixteen gigawatts waited behind it is the clearest possible evidence that sequencing, not aggregate supply, is the binding constraint.
Section 2. The Shared Grid as Launch Platform for Private AI Infrastructure
If the first section established why bridges are needed, this section examines what the shared grid actually does when it becomes a bridge, and why that role is both valuable and hazardous. The central claim is that hyperscale private generation and utility service, long debated as alternatives, are becoming sequential complements. A campus may begin as an ordinary grid customer, progressively add contracted and on-site generation, and end as a largely self-supplied facility that retains its grid connection for balancing and resilience. Understanding that trajectory requires distinguishing the two directions in which bridges can run, examining the institutional machinery that Alberta and other jurisdictions have built to govern them, and confronting the reliability obligations that a physical grid cannot wish away merely because a contract labels a load as temporary.
2.1 Two Directions of the Bridge
The AI economy now contains two mirror-image bridge architectures. In a generation-first bridge, private on-site machines carry the campus until the utility interconnection is ready, after which the machines become backup or move elsewhere; the Prometheus Hyperscale arrangement in Texas, with up to 300 megawatts of gas generators and batteries per site on a five-year initial term, is a representative example, and its developer noted that hyperscalers tend to prefer hybrid or co-located solutions that pair on-site resources with a transmission connection.[13] In a grid-first bridge, the shared system carries the campus until dedicated generation is ready, after which the grid becomes a partner, backstop and market interface; Sturgeon County is the archetype. Bloom Energy’s 2026 industry survey suggests how rapidly the first model has spread: roughly one in three hyperscalers and colocation providers now expect to operate entire campuses on on-site power, and 73 percent report actively evaluating or selecting on-site power providers.[32]
Table 3. Generation-First and Grid-First Bridges
| Feature | Generation-first bridge | Grid-first bridge |
| What carries early load | On-site engines, turbines, fuel cells, batteries | Shared system, often via a contracted block from an existing fleet |
| Destination state | Utility interconnection, with on-site assets as backup | Dedicated or behind-the-meter generation, with grid as backstop |
| Who bears near-term risk | Mainly the developer (fuel, permits, emissions, equipment) | Shared system and its other customers, unless contracts reallocate it |
| Principal public concern | Local air quality, noise, permitting | Reliability, reserve margins, cost shifting, stranded wires |
| Representative case | Prometheus Hyperscale / Conduit Power, Texas[13] | Meta / Capital Power / Greenlight, Alberta[6][4] |
The distinction matters because the policy questions differ. Generation-first bridges raise questions of local environmental permitting and the efficiency of small, fast-deployed machines, while grid-first bridges raise questions of who bears the reliability and financial risk of serving a customer that intends to leave. The remainder of this paper concentrates on the grid-first variety because it is the one in which the public, through the shared grid and its ratepayers, becomes a counterparty.
2.2 Grid First, Private Power Later
Bridge Megawatts invert the long-standing assumption that hyperscale private generation and utility power are competitors. A project can begin fully grid-connected and gradually add contracted generation, on-site turbines, battery storage, fuel cells, renewable power-purchase agreements, nuclear offtake or behind-the-meter supply. The final architecture may differ substantially from the one used during commissioning, and the transition between them may take several years. Alberta has formalised the destination explicitly: the Sturgeon County project is described as proceeding under the province’s bring-your-own-power framework, and AESO’s second-phase working group is designing a “bring-your-own-generation” process that includes generator qualification requirements and “tethering agreements” linking loads to specific generation.[8][33]
That design is important because it treats the bridge not as an exception but as a phase within a governed process. The first phase allocated scarce near-term grid capacity pro rata among qualified projects that could post financial security, estimated at roughly C$14 million per 100 megawatts through two months of demand transmission service and a payment in lieu of notice, and that could demonstrate municipal support and completed power-flow studies.[30] The second phase, according to AESO, is exploring a long-term framework spanning connections, system planning, operations, markets, tariffs and reliability.[33] Bridge Megawatts touch all six simultaneously, which is why large-load policy can no longer be separated into isolated questions of generation, transmission, tariffs and interconnection. The system operator’s chief executive explained the underlying constraint when the interim approach was announced:
“Alberta has never seen this level and volume of load connection requests. As the system operator, we are responsible for ensuring that new project connections do not compromise grid reliability.”
— Aaron Engen, CEO, Alberta Electric System Operator [11]
2.3 The Grid as an Infrastructure Incubator
Electricity grids are usually understood as permanent infrastructure supplying customers continuously for decades. The AI boom adds a new function: the grid as an incubator for new industrial loads, enabling a campus to cross from construction into operation before its mature energy architecture exists. In that role, shared-grid capacity resembles the temporary infrastructure that accompanies any large construction project, such as a site generator or a temporary water main, but the analogy breaks down at scale. A 250-megawatt electrical bridge is not a construction convenience; it is a load comparable to a mid-sized city, and at hyperscale, “temporary” can resemble the permanent electricity demand of a metropolitan region.
The incubator function also changes the political economy of the utility relationship. A utility that serves a customer for thirty years can amortise connection costs across a long revenue stream; a utility that incubates a customer for four years before that customer shifts most of its consumption behind the fence faces a fundamentally different risk profile. This is one reason why every serious large-load tariff adopted in North America since 2024, examined in Section 3, contains some combination of minimum bills, long contract terms and exit fees. Those provisions are, in effect, the fee the incubator charges for accepting a tenant who intends to move out.
2.4 The Reliability Problem
The physical grid cannot distinguish between a bridge megawatt and a permanent one. If a data center requires 250 megawatts at four o’clock on a stressed summer afternoon, the system operator must accommodate that demand unless contractual flexibility exists; generation reserves must be available, transmission must remain secure, voltage must remain stable, and contingencies must be planned. Temporary AI demand can therefore create permanent reliability responsibilities for as long as it exists. Large computational loads also introduce reliability risks that ordinary industrial customers do not. Harvard’s Belfer Center recounts that in July 2024 a voltage fluctuation in northern Virginia caused sixty data centers to disconnect simultaneously, suddenly leaving roughly 1,500 megawatts of surplus generation on the system and forcing emergency adjustments to prevent cascading outages.[34] In Texas, ERCOT’s large-load working group reported in May 2026 that it had identified four groups of large loads whose simultaneous tripping could exceed 3,200 megawatts during severe disturbances, even as roughly 9,062 megawatts of large load had received approval to energise over the previous twelve months.[35]
The key regulatory question therefore becomes less philosophical and more contractual: what obligations should accompany access to scarce bridge capacity? The menu that has emerged across jurisdictions includes interruptible or non-firm service, curtailment provisions, minimum payments, collateral, distinctions between firm and non-firm service, demand-response obligations, customer-funded substations, transmission contributions, on-site standby generation, voltage ride-through requirements and staged increases in contracted load. Alberta’s interim program already signalled that its tariff redesign would consider new interruptible rate classes and terms for load shedding, demand response and backup generation.[30] The appropriate combination will vary by market, but in every market the obligations are the price of the bridge.
2.5 The Value of Flexible AI Load
Artificial-intelligence campuses possess a characteristic that most traditional industrial customers do not: parts of their workload can be delayed, slowed, relocated or shifted onto on-site storage without destroying economic value. Training jobs tolerate interruption differently from latency-sensitive inference; batch workloads differ from interactive ones; and operators with multiple campuses can in principle move work across regions. This raises a consequential possibility, namely that Bridge Megawatts could evolve from an electricity burden into a dispatchable relationship between compute and the grid, in which the question is not only whether the grid can serve AI but whether AI workloads can help the grid manage scarcity.
The most influential quantitative case for that possibility came from Duke University’s Nicholas Institute. Tyler Norris, Tim Profeta, Dalia Patiño-Echeverri and Adam Cowie-Haskell analysed twenty-two balancing authorities serving about 95 percent of U.S. peak load and found that roughly 76 gigawatts of new load, about ten percent of national peak demand, could be integrated with an average annual curtailment rate of only 0.25 percent, rising to 98 gigawatts at 0.5 percent and 126 gigawatts at one percent, with average curtailment events lasting roughly 1.7 to 2.5 hours.[36][37]
“Our study demonstrates that existing U.S. power system capacity—intentionally designed to handle extreme peak demand swings—could accommodate significant load additions with modest flexibility measures.”
— Tyler Norris, Nicholas School of the Environment, Duke University [36]
“There’s gold in the hills, and we should be hunting for it.”
— Tim Profeta, Senior Fellow, Nicholas Institute, and Associate Professor of the Practice, Duke University [36]
The Duke findings are directly relevant to the bridge because they imply that much of the “headroom” a hyperscaler needs during its transitional years already exists in the system, provided the load does not insist on firm service during the relatively few hours of peak stress. Norris emphasised that curtailment need not mean switching a campus off, and that in most curtailment hours the majority of the load would continue to run.[38]
“You might only need to curtail 10% or 20% or 50% [of a data center] because that’s all the flexibility you need right to stay below the existing system peak.”
— Tyler Norris, Duke University [38]
Subsequent work has extended the case. Emerald AI’s chief executive has cited a March 2026 paper led by Boston University’s Ayse Coskun that found an 18 to 55 percent power-flexibility opportunity across representative AI workloads spanning training, fine-tuning and inference, and reported five field demonstrations with partners including NVIDIA and EPRI’s DCFlex initiative.[39] An April 2026 preprint on AI data-center flexibility and interconnection formalised the trade-off between deferring and shifting load as a route to faster connection.[40] An August 2026 simulation reported that pre-identifying grid-compatible sites and requiring flexible use during grid stress increased the number of viable locations for one-to-two-gigawatt facilities by nine to twenty-one percent in a Texas-scale system without materially raising average power prices.[41] MIT researchers are pursuing the same agenda at the level of the individual facility.
“By looking at the system as a whole, our hope is to minimize energy use as well as dependence on fossil fuels, while still maintaining reliability standards for AI companies and users.”
— Deepjyoti Deka, Research Scientist, MIT Energy Initiative [42]
Policy has begun to follow the evidence. In June 2026, FERC directed every regional grid operator to justify or reform its rules for, among other things, transmission service for flexible large loads, co-located loads and loads with behind-the-meter generation.[43] Utility Dive reported that a 2026 Duke study found that a reduction of just one to two percent in data-center peak demand could lower electricity rates by 0.5 to 2.8 percent.[20] The Congressional Research Service noted in September 2026 that NERC had lowered its summer-2026 peak forecast for the ERCOT area by 3.7 gigawatts because more data centers could be curtailed by grid operators when needed to prevent emergencies.[44] Flexibility, in other words, is ceasing to be a research proposition and becoming a tradable attribute of large-load service, and it is the attribute most naturally exchanged for access to the bridge.
Section 3. Who Pays for the Bridge?
Every bridge eventually raises the question of the toll. In electricity, that question is especially acute because the infrastructure that makes a bridge possible is long-lived, lumpy and, once built, largely immovable, while the customer it serves may intend to reduce its reliance on the system within a few years. This section examines that central economic paradox, surveys how six North American jurisdictions and two federal institutions have begun to answer it between 2025 and September 2026, and identifies the case that none of them has yet fully addressed: the bridge that ends early because the customer’s private energy system succeeds.
3.1 Temporary Electricity, Permanent Infrastructure
A hyperscaler may rely on grid power for only a few years, yet serving it can require infrastructure that lasts for decades. A high-voltage transmission line may remain in service for half a century; a substation may operate for generations; transformers and generation investments may persist long after the original customer’s energy architecture has changed. If a 500-megawatt campus leaves most of its reliance on the shared grid after five years because its dedicated plant has become operational, someone must carry the remaining cost of assets constructed specifically to serve it. At that moment Bridge Megawatts stop being an engineering concept and become a problem of intertemporal cost allocation.
Legal scholarship published in 2025 sharpened the stakes. In their paper for the Harvard Electricity Law Initiative, Eliza Martin and Ari Peskoe documented how utilities competing for data-center load can offer discounted special contracts and lopsided tariffs while socialising the associated costs through rates charged to the general public, and they identified co-location arrangements between data centers and existing power plants as a third channel through which wholesale and delivery prices may be distorted.[45][46]
“Without systematic changes to prevailing utility ratemaking practices, the public faces significant risks that utilities will take advantage of opportunities to profit from new data centers by making major investments and then shifting costs to their captive ratepayers.”
— Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative [46]
“We’re all paying for the energy costs of the world’s wealthiest corporations.”
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [47]
The macroeconomic literature points in the same direction. The International Monetary Fund’s 2025 working paper on AI and energy demand concluded that the AI boom would produce manageable but uneven increases in energy prices and emissions, contingent on policy and infrastructure constraints, and estimated that under scenarios of constrained renewable growth and limited transmission expansion U.S. electricity prices could rise by 8.6 percent.[29] The bridge interval is precisely when infrastructure constraints bind most tightly, which is why the allocation of its costs has become a first-order political issue rather than a technical footnote.
3.2 Alberta: “Bring Your Own Power” Meets the Grid
Alberta’s policy rhetoric emphasises that large data centers must bring their own power and pay for their own infrastructure, and Meta explicitly states that it is fully funding the new generation and grid infrastructure required for the Sturgeon County site, which it argues will improve reliability for the entire provincial grid.[1][3] At the same time, the campus’s earlier requirements are grid-connected, supplied through a contracted block from an existing fleet, and served through AltaLink’s transmission network under AESO’s oversight.[7] That combination makes Alberta a particularly useful laboratory, because it contains both ends of the transition, a shared system and a dedicated system, within a single project.
The unresolved question is how the transition itself should be priced. Capital Power’s release notes that the agreement preserves the company’s ability to pursue further commercial optimisation of its Genesee site, a reminder that the 250 megawatts are drawn from an existing fleet whose output would otherwise serve the competitive market.[6] In an energy-only market, a large, inflexible block of new demand drawing on existing supply during the bridge years can tighten reserve margins and raise wholesale prices for everyone until the dedicated plant arrives. The Premier’s claim that the investment helps reduce transmission costs for consumers may well prove correct over the life of the project, but the bridge years are where that claim will be tested.[10] It is also notable that Capital Power’s chief executive argued the project differs from those that have met local opposition because of how the proponent has accepted responsibility as a partner.
“I do think this one’s different. … It comes down to how a project proponent is taking on the responsibility as a partner and as a player in a local community.”
— Avik Dey, President and CEO, Capital Power [9]
3.3 Virginia: Long-Term Commitments for Large Loads
Virginia, home to the densest data-center cluster in the world, approached the problem through rate design. In its November 25, 2025 final order in Dominion Energy Virginia’s biennial review, the State Corporation Commission created a new GS-5 rate class for customers of 25 megawatts or more, effective January 1, 2027, and required certain large-scale customers to pay a minimum of 85 percent of contracted distribution and transmission demand and 60 percent of generation demand in order to help insulate other ratepayers from the costs of rapid infrastructure build-out.[48] The order imposes fourteen-year contracts and collateral requirements, and reporting on the case noted that the typical Dominion data-center customer had grown from an average of about 13 megawatts in 2013 to around 300 megawatts, with some requests reaching several gigawatts.[49][50] The Commission’s deputy director described the minimum as a floor that applies regardless of consumption:
“[Customers must] pay at least 85% of the transmission and distribution cost incurred to serve them each month regardless of their actual electricity consumption.”
— Brian Pratt, Deputy Director, Virginia State Corporation Commission [51]
Two further provisions make GS-5 directly relevant to Bridge Megawatts. First, the order allows a customer to include within the fourteen-year term “a load ramp period not to exceed four years,” an explicit regulatory acknowledgement that large loads arrive in stages. Second, a customer that ceases operations or defaults during the term must pay an exit fee covering outstanding minimum charges over the remaining duration.[52][50] Those rules address the heart of the bridge problem: a customer cannot consume enormous quantities of utility infrastructure for several years and then vanish from the system without consequence, because the length of the physical investment and the length of the customer’s commitment must be reconciled. What GS-5 does not yet address is the customer that remains in operation but migrates most of its consumption to private generation, a case examined in Section 3.8.
3.4 Ohio: Contracting the Load Ramp
Ohio produced the first widely copied template. On July 9, 2025, the Public Utilities Commission of Ohio approved a negotiated data-center tariff for AEP Ohio requiring new data centers above 25 megawatts to pay for at least 85 percent of their subscribed capacity even if they use less, for contract terms of up to twelve years including a four-year ramp-up period, with proof of financial viability and exit fees for cancelled or under-performing projects.[53] Reporting on the tariff described the ramp as a gradually increasing percentage of contracted load followed by an eight-year minimum commitment, and later coverage noted that the exit fee equals three years of minimum charges.[54][55] AEP Ohio’s president framed the settlement as alignment between demand and infrastructure cost:
“We are glad the PUCO agrees that it is critical to align data centers’ demand for energy with the infrastructure costs needed to support their growth in Ohio.”
— Marc Reitter, President and COO, AEP Ohio [53]
The Ohio template is closely related to the bridge thesis because the utility is not merely pricing electricity; it is pricing the transition toward full load. In an economy where both data-center capacity and generation capacity arrive incrementally, ramp schedules may become as economically important as the headline rate. The tariff has also drawn a substantive economic critique that deserves attention. The Buckeye Institute argued in March 2026 that an 85 percent take-or-pay obligation lasting twelve years charges the most price-elastic customers the highest mark-ups, in effect inverting the Ramsey pricing rule, and warned that if large, flexible loads go elsewhere, residential customers may ultimately shoulder a larger share of grid expansion costs.[56] That critique matters for bridge design because it implies that protective terms which are too rigid can drive away the very customers whose contributions would otherwise help pay for shared infrastructure.
3.5 Texas: Separating Real Demand From Speculative Demand
Texas demonstrates a different facet of the problem: distinguishing credible future load from projects that may never materialise, and doing so under intense political pressure. On June 18, 2026, the Public Utility Commission of Texas approved ERCOT’s Batch Zero process, a one-time, system-wide study of new large loads and expansions of 75 megawatts or more, replacing a utility-by-utility approach that ERCOT itself said had produced repeated restudies and growing backlogs. ERCOT reported more than 438,000 megawatts of large-load requests, nearly 89 percent from data centers.[57] Roughly 35 gigawatts were expected to qualify as committed base load and another 65 gigawatts to be evaluated within the batch, a combined volume exceeding ERCOT’s historical peak.[58] The financial screen was deliberately steep: applicants faced financial security of US$100,000 per megawatt at the intermediate-agreement stage, and projects pairing new generation with load could elect a treatment in which the generation carries part of the load and ERCOT allocates an annual megawatt amount for the incremental portion served from the grid, which is, in substance, a formally rationed bridge.[59]
“Our existing process was really not designed for the volume of large load interconnection requests that we have been experiencing.”
— Jeff Billo, Vice President of Interconnection and Grid Analysis, ERCOT [60]
“We could potentially be solving a national issue on how to do this… reliably, stably.”
— Pablo Vegas, President and CEO, ERCOT [60]
“The load interconnection changes coming with ERCOT’s implementation of Batch Zero represent a fundamental shift from prior processes.”
— Beth Garza, Senior Fellow, R Street Institute, and former Director, ERCOT Independent Market Monitor [58]
Then politics overtook process. On August 3, 2026, Governor Greg Abbott directed the PUC and ERCOT to conduct a comprehensive verification and audit of every data center advancing through interconnection before any further project could proceed, citing approximately 474 gigawatts of requests and the failure of some developers to comply with a state survey of water and power use.[61][12] ERCOT suspended the Batch Zero classification notices it had planned to deliver on August 7, and legal advisers noted that Senate Bill 6, signed in June 2025, had already established disclosure and curtailment obligations for large loads of 75 megawatts or more.[62] On September 21, 2026, the Governor broadened the pause by directing the state environmental regulator to halt data-center permits until the audit is complete.[63]
“Our top priority is to protect Texans’ safety and quality of life. Any project that fails to comply with the requirements set forth by the PUCT and ERCOT, and by state law, must be denied connection to the Texas grid. Simply put, Texans must come first.”
— Greg Abbott, Governor of Texas [61]
“Because the information sought by the PUC, ERCOT, and TWDB is necessary to make informed decisions by each of those agencies, no other state agency shall move forward with regulatory approvals related to data centers until this information is acquired.”
— Greg Abbott, Governor of Texas, letter to TCEQ [63]
The industry’s response was notably conciliatory, framing the audit as a tool for separating credible developers from speculative ones. For Bridge Megawatts, the Texas lesson is straightforward: before a system commits scarce transitional electricity to a project, it must know whether that project is financially committed enough to deserve it. Deposits, collateral, site control, credible load ramps and long-term commitments therefore function not merely as financial protections but as mechanisms for allocating scarce near-term electricity, and in 2026 disclosure and community legitimacy were added to that list.
“The Data Center Coalition and its members strive to be good actors where we operate, which is why we are hopeful this directive from the Governor will help separate those who are responsible water and energy stewards from those who are not.”
— Dan Diorio, Executive Vice President, Data Center Coalition [64]
3.6 Georgia and Pennsylvania: Contracts, Collateral and the “But-For” Standard
Georgia adopted a contract-centred model. Since February 1, 2025, any new Georgia Power customer expecting 100 megawatts or more must take service under a customised contract with financial guarantees, terms of up to fifteen years rather than five, minimum monthly payments and Commission review at least thirty days before signing, and early termination can require the customer to repay distribution, transmission and generation investments already made on its behalf.[65] In December 2025 the Commission approved nearly 10,000 megawatts of new generation, about 80 percent expected to serve data centers, and in April 2026 it approved a Customer Identified Resource program allowing large customers to propose and fund their own clean-energy projects for bill credits, up to three gigawatts through 2035.[65] That program is significant for the bridge because it creates a sanctioned path by which a large customer can move from utility-supplied to customer-identified resources within a regulated framework.
“The amount of energy these new industries consume is staggering.”
— Jason Shaw, Chairman, Georgia Public Service Commission [66]
Pennsylvania chose a model-tariff approach. The Public Utility Commission’s final order of May 12, 2026 established guidance for customers above 50 megawatts individually or 100 megawatts in aggregate, covering interconnection studies, infrastructure-cost responsibility, collateral, minimum contract terms, load-ramp schedules, minimum billing demand, early-termination obligations, public queue transparency and customer self-construction of certain facilities.[67][68] Its most consequential principle is a “but-for” standard under which large-load customers pay for upgrades that would not have been needed but for their interconnection, irrespective of whether other customers might also benefit.[69] The order also requires emergency-response protocols consistent with U.S. Department of Energy Section 202(c) orders authorising PJM to direct the deployment of data centers’ backup generation during a system emergency, an early example of private bridge assets being enlisted as public reliability resources.[68]
“This is one of the most important infrastructure and consumer protection issues facing utility regulators across the country. Pennsylvania is confronting a level of electric load growth that has not been seen in generations, driven largely by data centers and advanced manufacturing. Rather than waiting for these challenges to overwhelm the system, this Commission chose to lead.”
— Steve DeFrank, Chairman, Pennsylvania Public Utility Commission [70]
3.7 The Federal Layer: FERC and the Ratepayer Protection Pledge
Above the states, two federal developments reshaped the bridge between December 2025 and July 2026. On December 18, 2025, FERC unanimously found PJM’s tariff unjust and unreasonable because it lacked clear rates, terms and conditions for generators serving co-located load and did not offer transmission services for customers willing and able to limit their use of the system, and directed PJM to create options for such customers.[71] Commissioner David Rosner, noting that PJM’s auction had just fallen short of its reliability requirement, wrote that the Commission needed to “skate to where the puck is going.”[72] On June 18, 2026, FERC extended that logic nationally through show-cause orders to all six regional transmission organisations, directing them to justify or reform their rules on large-load and co-located-load interconnection, transmission service for flexible loads and behind-the-meter generation, and to report by July 20, 2026 on how they would ensure adequate generation for existing and new large loads.[43][73]
“Clarifying new rules will help release the bottleneck of large load investments across the PJM footprint.”
— Laura Swett, Chairman, Federal Energy Regulatory Commission [71]
The executive branch acted through persuasion rather than regulation. On March 4, 2026, Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed the White House Ratepayer Protection Pledge, committing to build, bring or buy new generation, to cover the cost of all power-delivery infrastructure upgrades required for their data centers, and to negotiate separate rate structures with utilities and states that they would pay whether or not they use the power.[74] By late July 2026 the White House said the pledge had been extended to utilities, developers and governors and, by its own account, covered 80 percent of all power delivered to U.S. homes and businesses.[75] Energy-industry observers noted from the outset that the pledge contained no enforcement mechanism of its own, leaving implementation to the state proceedings described above.[76]
“If a hyperscaler wants the power, they pay for the power — the generation, the delivery, and the grid upgrades that come with it.”
— The White House, Ratepayer Protection Pledge [77]
Read through the lens of this paper, the pledge is significant less for its enforceability than for the vocabulary it normalises. “Build, bring, or buy” is a description of the destination state of a grid-first bridge, and “pay whether used” is a description of the minimum-bill mechanism that Virginia, Ohio and Georgia had already adopted. What the pledge does not specify is how the interval between “buy from the grid now” and “build our own later” should be priced, which is precisely the gap a Bridge-Power Tariff would fill.
Table 4. How Jurisdictions Are Pricing the Transition (as of September 2026)
| Jurisdiction | Threshold | Term and minimum | Ramp / exit provisions | Bridge relevance |
| Alberta (AESO) | ≥75 MW (interim phase) | Financial security ≈C$14M per 100 MW; 1,200-MW cap to 2028[30] | Phase 2 BYOG and tethering agreements under design[33] | Explicit published bridge capacity |
| Virginia (GS-5) | ≥25 MW, ≥75% load factor | 14 years; 85% T&D and 60% generation minimum[48] | Ramp ≤4 years; exit fee for remaining minimums[52] | Reconciles asset life and commitment |
| Ohio (AEP) | >25 MW | Up to 12 years; 85% minimum[53] | 4-year ramp; exit fee of three years’ minimums[55] | Prices the ramp explicitly |
| Texas (ERCOT) | ≥75 MW | US$100k/MW security; SB 6 curtailment duties[59][62] | Batch Zero; annual grid allocation for generation-paired loads; audit pause Aug–Sep 2026[61] | Rations the bridge and screens speculation |
| Georgia | ≥100 MW | Up to 15 years; minimum bills; guarantees[65] | Termination repays investments; CIR program[65] | Sanctioned path to customer resources |
| Pennsylvania | >50 MW or ≥100 MW aggregate | Model tariff with minimum terms and billing[67] | Load-ramp schedules; early termination; “but-for” upgrades[69] | Enlists backup generation in emergencies |
| Federal (FERC / White House) | Varies | Pay-whether-used pledge[74] | Flexible-load and co-location transmission services[43] | Normalises the destination state |
Author’s compilation from the regulatory and legal sources cited in each cell. Terms summarised; consult the underlying orders for precise conditions.
3.8 The Stranded-Bridge Problem
The most difficult case arises when permanent private energy infrastructure arrives earlier than expected, or when it simply works as designed. Imagine that the shared system has financed transmission and substation capacity on the assumption of eight years of data-center demand, and that after four years the hyperscaler shifts most of its campus onto an on-site generation complex. The data-center project has succeeded; the private energy system has succeeded; yet the utility may now own under-utilised assets built for a customer that no longer needs them at the expected level. Minimum-bill and exit-fee provisions protect against the customer that leaves or fails; they are much less clear about the customer that stays connected, pays a modest standby charge, and buys very little energy.
Critics of the Ratepayer Protection Pledge have already made a version of this argument, contending that large-scale behind-the-meter self-generation by hyperscalers could raise residential rates by shrinking the base over which fixed network costs are recovered.[78] Whether or not that critique proves correct, it identifies the logical gap in current frameworks. Bridge-power contracts will eventually require an explicit answer to a question that existing tariffs rarely ask: who pays when the bridge ends early? A defensible answer would tie the customer’s financial obligation to the depreciation schedule of the assets built for it, not to its volumetric consumption, and would allow that obligation to be reduced only when those assets are demonstrably re-used by other load. Pennsylvania’s “but-for” standard and Virginia’s exit fees point in that direction, but neither yet treats voluntary migration to private generation as a distinct contingency with its own price.
Section 4. From Grid Connection to Energy Island
If the bridge is the interval, the island is the destination, or at least the destination that many hyperscalers now describe. This section argues that the binary vocabulary of “grid-connected” versus “off-grid” obscures more than it reveals, because the architectures now being built occupy a continuum in which the campus progressively assumes responsibility for more of its own supply while retaining a connection whose economic function changes over time. Understanding that continuum is essential to pricing the bridge, because what the grid is asked to provide at the end of the transition determines what it should be paid during the transition.
4.1 The Energy Architecture Is a Continuum
It is misleading to divide future AI campuses into two categories. The architecture is better described as a sequence of states, each with a different balance of shared and private supply and a different set of obligations running in each direction. Different workloads, and different phases in the life of the same campus, may occupy different points along that sequence.
Table 5. The Continuum from Grid Customer to Island-Capable Campus
| Architecture state | Role of the shared grid | Principal contractual question |
| Fully grid-supplied | Sole supplier of energy, capacity and reliability | Standard large-load tariff; minimum bills and term |
| Grid plus contracted generation | Delivery network for a bilateral block (e.g., 250 MW from an existing fleet) | Who bears reserve and congestion costs during the bridge |
| Grid plus on-site generation | Supplementary supplier and backstop | Standby charges; netting rules for behind-the-meter output |
| Grid plus microgrid | Balancing partner; occasional importer or exporter | Two-way service terms; export caps; ride-through |
| Predominantly islanded campus | Insurance and emergency supply | Pricing optionality rather than kilowatt-hours |
| Island-capable campus with grid exchange | Market interface; campus may sell capacity or services back | Reciprocal obligations; emergency dispatch of private assets |
Author’s typology.
4.2 Why Hyperscalers Move Toward Partial Islanding
The incentive to move along the continuum goes well beyond the price of electricity. Large AI operators increasingly care about certainty of capacity, speed to power, exposure to regional outages, power quality, the right to expand generation as the next accelerator generation arrives, the carbon characteristics of supply, and insulation from regional grid politics of the kind that halted Texas interconnections in August 2026. Owning or controlling more of the energy stack converts electricity from a purchased commodity into a component of data-center architecture. That mirrors a broader pattern throughout the Five-Layer AI Economy, in which hyperscalers have sought progressively greater control over chips, networking, data centers, models and applications; energy is becoming another vertically integrated layer.
The International Energy Agency’s 2026 update documents the scale of the movement. It reports that, constrained by slow grid connections, U.S. data-center developers are pushing forward projects with on-site gas-based generation, and that satellite tracking indicates roughly one-fifth of such projects have already begun land clearing.[79] The same update projects data-center electricity use roughly doubling from about 485 terawatt-hours in 2025 to about 950 terawatt-hours in 2030, broadly consistent with its 2025 central case, while noting that bottlenecks across the value chain are making the most aggressive near-term scenarios less likely.[79] When the IEA published its original special report, its executive director framed the scale in terms any policymaker could grasp:
“Global electricity demand from data centres is set to more than double over the next five years, consuming as much electricity by 2030 as the whole of Japan does today.”
— Fatih Birol, Executive Director, International Energy Agency [80]
U.S. estimates have risen with each revision. Lawrence Berkeley National Laboratory’s June 2026 assessment concluded that data centers could account for 9.5 to 15.3 percent of U.S. electricity consumption in 2030, compared with the 6.7 to 12 percent it had projected for 2028 in its 2024 report.[81][34] The IEA also estimated that roughly 20 percent of planned data-center projects globally could face delays unless grid risks are addressed, which is another way of stating the demand for bridges.[82] MIT researchers have warned about the carbon implications of meeting that demand at speed:
“The demand for new data centers cannot be met in a sustainable way. The pace at which companies are building new data centers means the bulk of the electricity to power them must come from fossil fuel-based power plants.”
— Noman Bashir, Computing and Climate Impact Fellow, MIT Climate and Sustainability Consortium [83]
4.3 Behind-the-Fence Power and Its Hidden Overbuild
A behind-the-fence system changes the relationship between hyperscaler and utility. Electricity is generated adjacent to the load without every megawatt traversing the wider transmission network, and such systems may combine natural-gas generation, batteries, solar, wind, fuel cells, nuclear supply, demand management and a retained grid interconnection. The campus remains connected while reducing its dependence on external electricity. This is not energy independence; it is selective interdependence.
The economics of that interdependence are subtler than they first appear. The IEA’s 2026 analysis found that providing reliable on-site gas-fired electricity to meet critical and variable data-center load requires overbuilding on-site generation by 30 to 70 percent relative to demand, because AI training and inference induce large and rapid power swings and because on-site systems must cover their own outages.[79] The same analysis projected that 20 to 25 gigawatts of battery storage could be installed in data centers globally by 2030, potentially making them grid assets if incentives are right.[79] Those findings reveal why most hyperscalers do not in fact want to leave the grid entirely: the shared system is the cheapest source of the reserve margin that an isolated campus would otherwise have to build for itself. Greenlight is itself described as behind-the-meter, yet the campus it serves will remain connected to Alberta’s system and contracted to Capital Power’s fleet, which is exactly the hybrid the continuum predicts.[4][6]
There is also a carbon-accounting distinction that the bridge makes visible. Meta states that the Sturgeon County campus’s electricity use will be matched with 100 percent clean and renewable energy, while Data Center Frontier observed that the physical electricity serving much of the campus will be generated from natural gas and that renewable contracts or certificates may offset consumption on an annual accounting basis without making a co-located gas plant carbon-free.[1][7] During the bridge interval, the physical source of supply is the marginal generator of the shared system, which in many markets is also gas-fired. Distinguishing physical supply from contractual matching is therefore essential to any honest account of transitional electricity.
4.4 The Grid Becomes Insurance
At the far end of the continuum, the grid’s economic role changes. Today the grid is the primary electricity supplier; tomorrow, for a subset of campuses, private generation may be the primary supplier while the grid serves as balancing mechanism, backup, market interface and reliability insurance. That creates a difficult tariff problem. A campus might purchase relatively little energy each year yet still expect the grid to stand ready to supply hundreds of megawatts if its private generation fails, and traditional volumetric pricing does not capture the value of that standby capability. The grid is no longer selling only kilowatt-hours; it is selling optionality.
FERC’s December 2025 order anticipated this shift when it required PJM to create transmission service options for customers that are able to limit their withdrawals under certain conditions, in effect recognising that a co-located load buying firm service for its full demand and one buying non-firm or conditional service are purchasing different products.[71] The pledge’s commitment that hyperscalers will, where possible, make backup resources available to the grid during scarcity points to the reciprocal obligation, and Pennsylvania’s alignment with federal emergency orders allowing PJM to direct data-center backup generation shows how that obligation can be operationalised.[74][68] In a mature bridge relationship, the insurance runs in both directions.
4.5 AI Campuses as Energy Systems
By 2030, some large AI developments may resemble industrial energy systems more than conventional commercial buildings, containing dedicated substations, multiple transmission feeds, gas generation, batteries, fuel infrastructure, microgrid controls, flexible compute scheduling and the ability to operate independently for extended periods. The campus will still consume electricity, but it will also become an active participant in producing, storing, shifting and potentially returning electricity to the system. MIT’s Energy Initiative, which launched a dedicated Data Center Power Forum in 2025, has made that integration its explicit research agenda.
“One of the biggest challenges today is to find the energy to power all the new data centers that are being constructed in the United States and around the world. At MIT we’re developing a clear understanding of the problem and finding the solutions that will work for everyone.”
— William H. Green, Director, MIT Energy Initiative, and Hoyt C. Hottel Professor of Chemical Engineering [84]
Within the framework of the Five-Layer AI Economy, this development has a structural meaning. The data-center layer is moving downward into the energy layer, and the boundary between the two is becoming porous. The AI factory is becoming partially responsible for operating its own energy system, which is why bridge arrangements are not a temporary curiosity but an early expression of a new industrial form.
Section 5. The Emerging Geography of Bridge Megawatts
Bridges are built where rivers are wide and crossings are scarce, and Bridge Megawatts follow a similar geography. The jurisdictions examined in Section 3 were described in terms of how they allocate costs; this section examines them as distinct starting points on the continuum described in Section 4, and asks what each reveals about how the transition from shared to private supply may unfold. It then assembles those lessons into a forward-looking proposal: a Bridge-Power Tariff designed around time as well as capacity.
5.1 Alberta: The Explicit Bridge
Alberta is the clearest conceptual example because the transition is visible and sequenced. The campus is planned at one gigawatt with room to grow; its initial requirements are grid-connected under a province-wide interim limit; a contracted 250-megawatt block from an existing fleet begins in the second half of 2028; and a dedicated combined-cycle plant follows toward the end of 2030.[1][6][5] The province therefore illustrates the full sequence from grid to bridge to dedicated generation to hybrid campus. If the model succeeds, it may attract additional hyperscalers seeking similarly staged paths to power, and the long-run test will be whether Alberta’s second-phase framework can turn a one-time interim allocation into a repeatable, priced product. Canada-wide data offer a note of caution: The Globe and Mail, citing the tracking firm Aterio, reported that of roughly seventy data-center proposals announced in Canada since 2024, only a handful had begun construction.[9]
5.2 Texas: Bridge Megawatts at Gigawatt Scale
Texas poses the same question under far larger prospective load and far greater political volatility. It possesses natural attributes favourable to hybrid models, including abundant gas infrastructure, large renewable fleets, industrial-scale land, a competitive market and long experience with on-site industrial generation. Batch Zero’s option allowing generation-paired loads to receive an annual allocation of grid megawatts for their incremental needs is, conceptually, the first formally rationed bridge product in a major U.S. market.[59] Yet the August and September 2026 pauses show that the bridge can be closed by executive action even for projects that have posted security and progressed through study.[61][63] The central research question for Texas is therefore not simply whether it has enough power, but how much transitional grid capacity it can safely and legitimately allocate while hundreds of gigawatts of proposed demand wait behind a political as well as a physical queue.
5.3 Virginia and PJM: The Mature Grid-Dependent Model
Virginia represents the opposite starting point. Its data-center ecosystem grew primarily around established utility and transmission infrastructure, and customers there have grown from tens of megawatts to hundreds.[49] The question Virginia poses is whether a mature, grid-dependent market can gradually evolve toward greater dedicated generation, and what happens to infrastructure constructed on the expectation of continued grid demand if it does. The wider PJM market in which Virginia sits is already operating with slimmer reserves: PJM reported a 14.7 percent reserve margin for the 2028/2029 delivery year, below its target, and is pursuing a special backstop procurement.[24] In that environment, co-location and behind-the-meter supply are not merely private preferences but potential relief valves for the system, which is why FERC’s co-location order and GS-5’s long-duration commitments must eventually be read together.
5.4 Ohio, Pennsylvania and Georgia: Different Answers to the Same Problem
Ohio has written the load ramp into contract; Pennsylvania has built a model tariff around cost causation, collateral, self-construction and emergency coordination; and Georgia has combined long contracts and minimum bills with a sanctioned path for customers to fund their own resources.[53][67][65] These approaches differ in instrument and emphasis, but all respond to the same structural fact: AI load arrives faster than the electricity system historically planned for industrial demand. They also share a limitation. Each was designed primarily to protect the system against a customer that arrives and then fails to materialise or leaves; none was designed explicitly for a customer that arrives, succeeds, and then migrates most of its consumption to private supply while retaining the right to call on the grid.
5.5 The 2030 Possibility: A Bridge-Power Tariff
This leads to the paper’s most forward-looking proposition. By 2030, regulators and utilities may conclude that hyperscalers moving from shared-grid electricity toward dedicated power require a distinct service category, which this paper calls a Bridge-Power Tariff. It is not a standardised tariff in any jurisdiction today; it is a conceptual framework for what a transitional large-load contract might contain, assembled from components that already exist in scattered form across the orders surveyed above. Rather than treating a hyperscaler as an ordinary permanent customer, a Bridge-Power Tariff would recognise explicitly that the customer expects its electricity architecture to change, and would price that expectation.
Table 6. Components of a Bridge-Power Tariff
| Component | Design question | Existing precedent |
| 1. Bridge capacity | Maximum temporary megawatts reserved from the shared system, and on what firmness | AESO 1,200-MW interim limit[11]; ERCOT annual grid allocation for generation-paired loads[59] |
| 2. Bridge duration | Expected period before private or dedicated generation assumes a larger share | Implicit in Capital Power’s >10-year ESA and Greenlight’s 2030 date[6][5] |
| 3. Ramp schedule | How quickly consumption may rise, and what is owed during the ramp | Ohio 4-year ramp[53]; Virginia ramp ≤4 years[52]; Pennsylvania ramp schedules[67] |
| 4. Exit schedule | How quickly grid demand may fall once permanent generation arrives | No explicit precedent; the key missing element |
| 5. Infrastructure contribution | How transmission, substations and dedicated facilities are funded and depreciated | Pennsylvania “but-for” standard[69]; Georgia upstream cost recovery[65] |
| 6. Reliability charge | Compensation for the grid’s obligation to stand behind the facility | FERC-directed service options for loads that limit withdrawals[71] |
| 7. Curtailment rights | When transitional loads reduce consumption in emergencies, and how often | Texas SB 6[62]; Duke curtailment-enabled headroom[36] |
| 8. Collateral | Protection against cancellation, delay or early migration | ERCOT US$100k/MW[59]; Virginia and Georgia collateral[48][65] |
| 9. Private-generation milestones | Disclosure of FID, turbine slots and commercial-operation dates for dedicated assets | Texas audit disclosure demands[61] |
| 10. Residual grid rights | Capacity the campus retains after private generation is operational, and its price | Emerging via FERC June 2026 show-cause orders[43] |
Author’s proposal. Precedents are partial analogues rather than complete implementations of each component.
The conceptual change would be significant. Today’s tariff generally asks how much electricity a customer will consume. A Bridge-Power Tariff would additionally ask how long the customer will need the shared system, how quickly that need will grow, how quickly it will shrink, and what happens to the infrastructure when the customer no longer needs the same amount. Two components deserve particular emphasis. The exit schedule is the missing piece in every framework reviewed here, and its absence is what creates the stranded-bridge risk described in Section 3.8. The private-generation milestone, meanwhile, would convert the destination plant from a press-release promise into a contractual obligation, so that a customer whose dedicated generation slips from 2030 to 2032 would automatically extend its bridge commitments rather than leaving the shared system to absorb an open-ended transitional load.
Such a tariff would also make the time premium described in Section 1.4 visible and tradable. A customer seeking firm bridge capacity at short notice would pay more, post more collateral and accept fewer curtailment exemptions; a customer offering meaningful flexibility, credible private-generation milestones and a short, contracted exit schedule would pay less and connect sooner. That is an electricity contract designed around time as well as capacity, and it would align the private incentive to move quickly with the public interest in not being left with the bill.
Section 6. What Have We Learned? Nine Pillars of Transitional Electricity
The preceding sections moved from a single campus in Alberta to the regulatory architecture of an entire continent. This section distils those findings into nine pillars. The first five restate and deepen the propositions with which the paper began; the final four are new, and emerged from the evidence of 2026 rather than from the original outline. Together they describe what might be called the transitional electricity economy: the set of contracts, institutions and physical assets that govern the years in which compute has arrived and its permanent power has not.
Pillar 1 — Time Is Becoming an Energy Resource
The AI infrastructure race is not determined solely by the number of megawatts available; it is determined by when those megawatts become available. A 500-megawatt plant arriving in 2030 cannot energise accelerators installed in 2028, and a turbine ordered in 2026 will, on current manufacturing schedules, arrive around 2031.[22] That timing mismatch creates economic value for transitional electricity and explains why hyperscalers have adopted a strategy of committing early to power and shells while deferring chip purchases until demand is visible.[15] The next electricity scarcity is therefore simultaneously a scarcity of megawatts, of location, of reliability and of timing, and any analysis that collapses those four dimensions into one will misprice the bridge.
Pillar 2 — Temporary Demand Can Create Permanent Obligations
Calling electricity demand temporary does not make its infrastructure temporary. A hyperscaler that relies on grid electricity for several years can still require permanent substations, transformers, transmission expansion, generation reserves, grid-control equipment and system-reliability investments. The fundamental economic challenge is to ensure that the lifespan of the financial obligation reflects the lifespan of the infrastructure, even when the customer’s dependence on the grid is shorter. Virginia’s fourteen-year GS-5 contracts with exit fees, Ohio’s twelve-year tariff with a four-year ramp, and Georgia’s fifteen-year contracts with termination repayment are early and partial solutions to this problem, each attempting to stretch the customer’s commitment toward the life of the assets it causes to be built.[52][53][65]
Pillar 3 — The Grid Is Becoming the Launchpad Rather Than the Destination
The historical assumption was that connecting a data center to the grid completed its energy-development process. Bridge Megawatts suggest the opposite: for future hyperscale projects, grid connection may increasingly be the beginning, with a campus moving from shared grid to contracted power, to dedicated generation, to microgrid, to hybrid energy system. The interconnected grid remains crucial throughout, but its role changes from supplier to partner, backup, balancing mechanism and insurer. Alberta’s bring-your-own-generation design and FERC’s insistence on transmission services for loads that can limit their withdrawals are both institutional recognitions of that changing role.[33][71]
Pillar 4 — The Future AI Data Center Is Also an Energy Campus
As compute reaches hundreds of megawatts and then gigawatts, energy infrastructure can no longer remain a peripheral utility service. Generation, storage, transmission, cooling and power management become integrated components of the AI factory, and the hyperscaler that once bought electricity at a meter increasingly finances or contracts generation, substations, transmission, batteries, microgrids, fuel infrastructure and grid services. The IEA’s finding that on-site gas systems must be overbuilt by 30 to 70 percent to serve variable AI loads reliably, and its projection of 20 to 25 gigawatts of data-center batteries by 2030, describe a campus that is becoming a small utility in its own right.[79] Layer 3 of the Five-Layer AI Economy, the data center, is moving downward into Layer 1, energy, and the boundary between them is becoming porous.
Pillar 5 — The Next Large-Load Tariff May Be Temporal
Existing tariff debates concern who pays. The next debate will add when they pay, for how long, and what happens when their energy architecture changes. Traditional tariffs concentrate on connecting customers; future hyperscale tariffs will also have to manage migration. The most important innovation may therefore not be a cheaper rate but a contractual architecture capable of recognising that a one-gigawatt AI campus can change from a grid customer into a partially self-supplied energy system within a few years, which is what the Bridge-Power Tariff in Section 5.5 is designed to do.
Pillar 6 — Flexibility Is the Currency of the Bridge
The evidence assembled in Section 2.5 points to a conclusion with direct commercial implications: the most valuable thing a hyperscaler can offer in exchange for early access to scarce bridge capacity is not money but flexibility. Duke’s estimate of 76 gigawatts of curtailment-enabled headroom at a 0.25 percent annual curtailment rate, Boston University’s finding of an 18 to 55 percent flexibility opportunity across AI workloads, NERC’s downward revision of ERCOT’s 2026 summer peak because data centers can be curtailed, and FERC’s June 2026 orders on flexible-load transmission service all point toward a market in which firmness is priced and flexibility is rewarded.[36][39][44][43] A campus willing to shed a portion of its load for a few dozen hours a year may be able to cross the bridge years earlier than one that insists on firm service, and at lower cost to everyone else. The Buckeye Institute’s critique of rigid take-or-pay tariffs reinforces the point from the opposite direction: rules that ignore the flexibility of large loads risk driving away the customers best able to help the system.[56]
Pillar 7 — Credibility Is the Price of Admission
Scarce near-term capacity must be rationed, and 2026 showed that the rationing device is increasingly credibility rather than price. ERCOT’s US$100,000-per-megawatt security requirement, AESO’s pro rata allocation to projects with financial security and municipal support, the collateral provisions in Virginia and Georgia, and the Texas audit’s demand for disclosure of power, water, ownership and community impacts all serve the same purpose: to separate committed projects from speculative queue positions before the system commits its bridge.[59][30][61] For hyperscalers, this means that a sanctioned destination plant, a named counterparty and a public commercial-operation date have become assets in their own right, because they make the bridge request believable. Meta’s Alberta package, which arrived with a final investment decision, a signed supply agreement and a tolling contract, is an example of credibility deployed as an interconnection strategy.[4][6]
Pillar 8 — Legitimacy Is Now as Scarce as Capacity
The events of August and September 2026 in Texas demonstrated that a bridge can be closed not only by physics but by politics. Governor Abbott’s audit halted interconnection approvals in the largest large-load queue in North America and was later extended to environmental permits; industry groups responded by welcoming the chance to separate responsible actors from others; and the White House pledge, the Pennsylvania model tariff and Georgia’s contract review all testify to a political environment in which the cost of AI electricity to ordinary households has become a first-order concern.[61][63][64] Stanford’s 2026 data-center symposium reflected the same tension, with PG&E estimating that every gigawatt of new data-center demand could lower all customer bills by one to two percent even as other analyses predicted rising bills in data-center-heavy states.[85] The international institutions have framed the stakes more broadly still:
“The rise of technologies such as artificial intelligence and cryptocurrency mining has significantly increased energy consumption.”
— Rebeca Grynspan, Secretary-General, UN Trade and Development (UNCTAD) [86]
The IMF’s conclusion that AI-driven price increases are manageable but depend on policy and infrastructure constraints is, read carefully, a statement about legitimacy as much as economics: the bridge is politically sustainable only if the public can see that its costs are being borne by those who benefit.[29] Social licence has become a capacity constraint, and bridge contracts that ignore it will be vulnerable to exactly the kind of executive intervention Texas has now demonstrated.
Pillar 9 — Physical Supply and Contractual Matching Must Be Distinguished
The final pillar concerns honesty in accounting. During the bridge interval, a campus draws physically on the marginal generators of the shared system, while its sustainability claims typically rest on annual contractual matching with renewable purchases elsewhere. After the bridge, its physical supply may be a dedicated gas plant, while its claims still rest on matching. Data Center Frontier’s observation that renewable certificates do not make a co-located gas plant carbon-free, and MIT’s warning that the pace of construction means the bulk of new supply must come from fossil sources, both point to the need for transitional electricity to be reported in physical as well as contractual terms.[7][83] A Bridge-Power Tariff that records the physical source of bridge megawatts would give regulators, investors and communities a far clearer picture of what the transition actually costs in emissions.
Table 7. Nine Pillars of Transitional Electricity: A Summary
| Pillar | Core proposition |
| 1. Time as an energy resource | Scarcity is measured in megawatts, location, reliability and timing simultaneously. |
| 2. Temporary demand, permanent obligations | Financial commitments must track asset lives, not customer intentions. |
| 3. Grid as launchpad | Interconnection is the start of a campus’s energy evolution, not its end. |
| 4. Data center as energy campus | The data-center layer is merging into the energy layer of the AI economy. |
| 5. Temporal tariffs | Future tariffs must price entry, ramp, duration and exit. |
| 6. Flexibility as currency | Early access to scarce capacity is best purchased with curtailable load. |
| 7. Credibility as admission price | Collateral, disclosure and sanctioned destination plants ration the bridge. |
| 8. Legitimacy as capacity | Public consent can close a bridge as decisively as a transformer shortage. |
| 9. Physical vs. contractual supply | Transitional electricity must be reported by its physical source, not only by matching. |
Conclusion: The Bridge Between the AI Factory and the Power Plant
The next phase of the artificial-intelligence infrastructure boom will not be decided simply by who acquires the most accelerators or announces the largest data center. It will be decided increasingly by whether electricity infrastructure can arrive at the speed of compute infrastructure, and, where it cannot, by how intelligently societies govern the interval in between. Meta’s Sturgeon County project offers an unusually clear glimpse of that future. A one-gigawatt AI campus is being built alongside an evolving energy architecture in which initial requirements rely on Alberta’s grid, a 250-megawatt contracted block from an existing fleet begins in the second half of 2028, and a dedicated 932-megawatt combined-cycle plant is expected toward the end of 2030.[1][6][5] What appears from a distance to be one data-center project is, on closer inspection, a sequence of infrastructure states unfolding on different timetables.
That sequence is why the phrase Bridge Megawatts fits this paper. It describes the electricity located between two moments: the moment the AI factory is ready, and the moment its mature energy system is ready. Those megawatts may come from an existing grid, a utility portfolio, an independent power producer, temporary on-site generation, contracted capacity or a combination of all of these. They may last for two years or considerably longer; they may gradually disappear as private generation enters service, or remain permanently as backup and balancing capacity. What they cannot be is economically insignificant merely because they are transitional. At gigawatt scale, even the bridge becomes infrastructure.
The years between 2027 and 2030 are likely to be when this matters most. AI campus construction schedules are compressing while generation and transmission schedules remain constrained by equipment, permitting, interconnection, financing and physical construction, and the result is a widening temporal gap between compute readiness and power readiness. PJM has already named that gap a transition gap and fallen short of its reliability requirement in consecutive auctions; turbine manufacturers are selling slots at the end of the decade; and the largest queue in North America has been paused pending a political audit.[25][24][21][61] Alberta is exposing the mechanism early; Texas is testing it at enormous scale and under intense political pressure; Virginia shows what happens when data-center load becomes deeply embedded in a utility system; Ohio has contracted the load ramp; Pennsylvania has built a model of cost causation and emergency coordination; Georgia has tied infrastructure obligations more directly to the customers that drive them; and FERC and the White House have begun to write the vocabulary of the destination state.
The longer-term possibility is that these fragmented approaches converge toward something resembling a Bridge-Power Tariff: a large-load electricity contract organised not only around how many megawatts a campus requires, but around when those megawatts arrive, how quickly demand ramps, how long dependence on the grid lasts, how infrastructure costs are secured, how much flexibility the customer offers in exchange for early access, and what happens when the customer migrates toward private generation. That would represent an important change in the political economy of the Five-Layer AI Economy. Energy planning has traditionally assumed that major industrial loads connect to the grid and remain customers. Artificial-intelligence infrastructure challenges that assumption. Tomorrow’s hyperscaler may arrive asking for 250 megawatts, grow to 500, build a dedicated gigawatt-scale energy complex, reduce its ordinary purchases, retain its transmission connection for resilience, deploy batteries for balancing, shift computational workloads during emergencies, and expand again when the next generation of accelerators arrives. The grid relationship becomes dynamic rather than permanent, and the megawatt becomes temporary, contractual and strategically timed.
Some scholars doubt that the underlying demand can be slowed at all, which only raises the importance of governing its transitional phase well. The University of Chicago’s Andrew Chien put the point with characteristic candour, and it is a fitting note on which to close:
“I think it’s probably forlorn hope that we could stop or slow the increase of power consumption from AI. At this point, I think that horse has left the barn.”
— Andrew Chien, Professor of Computer Science, University of Chicago [87]
If the horse has indeed left the barn, the task is not to close the door but to build the bridge well. The subject of this paper is therefore not simply power, but the power between power systems; not simply the grid, but the grid during transition; not simply whether the AI economy can obtain another gigawatt, but whether the Five-Layer AI Economy can synchronise the extraordinarily fast clock of artificial intelligence with the much slower clock of physical infrastructure, and do so in a way that the public will continue to accept. Between those two clocks lies an increasingly valuable, increasingly contested new commodity, and it deserves its own name.
Bridge Megawatts.
Footnotes and Endnotes:
[1] Meta Platforms, Inc.. “Breaking Ground on Meta’s First Data Center in Canada.” Meta Newsroom, July 9, 2026. https://about.fb.com/news/2026/07/breaking-ground-on-metas-first-data-center-in-canada/
[2] Meta Data Centers. “Hello, Sturgeon County!” datacenters.atmeta.com, July 8, 2026. https://datacenters.atmeta.com/2026/07/hello-sturgeon-county/
[3] Reuters (via EnergyNow). “Meta to Build C$13 Billion Alberta Data Center, its First in Canada.” EnergyNow.com, July 2026. https://energynow.com/2026/07/meta-to-build-c13-billion-alberta-data-center-its-first-in-canada/
[4] Pembina Pipeline Corporation, Morgan Stanley Infrastructure Partners & Kineticor. “Pembina Pipeline Announces Positive Final Investment Decision on the Greenlight Electricity Centre.” Pembina Media Centre, July 2, 2026. https://www.pembina.com/media-centre/news/details/3403fb77-2257-466b-a8e2-f043f4cd4650
[5] Power Technology. “Pembina moves forward with $3.2bn Greenlight Electricity Centre.” Power Technology, July 7, 2026. https://www.power-technology.com/news/pembina-fid-greenlight-electricity-centre/
[6] Capital Power Corporation. “Capital Power Enters Long-term Energy Supply Agreement with Meta in Alberta.” Capital Power Media Release, July 2026. https://www.capitalpower.com/media/media_releases/capital-power-enters-long-term-energy-supply-agreement-with-meta-in-alberta/
[7] Data Center Frontier. “Meta’s Canadian AI Data Center: A New Model for Infrastructure and Energy Integration.” Data Center Frontier, July 29, 2026. https://www.datacenterfrontier.com/hyperscale/article/55391399/metas-canadian-ai-data-center-a-new-model-for-infrastructure-and-energy-integration
[8] Invest Alberta. “Meta chooses Alberta for Data Centre Campus.” Invest Alberta, July 2026. https://investalberta.ca/meta-data-centre-alberta/
[9] The Globe and Mail. “Meta to spend $13-billion to build AI data centre in Alberta.” The Globe and Mail, July 10, 2026. https://www.theglobeandmail.com/business/article-meta-ai-data-centre-sturgeon-county-alberta/
[10] Power Technology. “Meta and Capital Power seal 250MW energy supply agreement.” Power Technology, July 9, 2026. https://www.power-technology.com/news/meta-capital-power-deal/
[11] Alberta Electric System Operator (AESO). “AESO Announces Interim Approach to Large Load Connections.” AESO Newsroom, June 4, 2025. https://www.aeso.ca/aeso/newsroom/aeso-announces-interim-approach-to-large-load-connections/
[12] Morrison Foerster. “Texas Governor Orders Data Center Moratorium – Disrupting Projects Under Construction and Parties’ Contractual Expectations.” Morrison Foerster Client Alert, August 11, 2026. https://www.mofo.com/resources/insights/260811-texas-governor-orders-data-center-moratorium-disrupting
[13] Utility Dive. “Batteries, gas generators to supply ‘bridge power’ for Texas data centers.” Utility Dive, September 10, 2025. https://www.utilitydive.com/news/conduit-power-bridge-power-texas-data-center-engie/759759/
[14] Futurex Capital AI Lab. “Q2 2026 Earnings: Big Tech’s AI Capex Race Hits $730 Billion (Microsoft, Meta, Amazon, Google).” Futurex Capital, August 2026. https://futurex.capital/en/ai-lab/reports/big-tech-ai-capex-2026q2
[15] UncoverAlpha. “Amazon, Google, Microsoft, Meta Q2 earnings: The AI CapEx ROIC is bad thesis is DEAD.” UncoverAlpha, August 3, 2026. https://www.uncoveralpha.com/p/amazon-google-microsoft-meta-q2-earnings
[16] ValueAdd VC. “$205B Google, $200B Amazon — AI Capex (2026).” ValueAdd VC, August 20, 2026. https://valueaddvc.com/blog/big-tech-ai-capex-in-2025-microsoft-google-meta-amazon-and-the-spending-race
[17] CNBC. “Amazon, Meta and Microsoft face skeptical investors this week after Google report sparked sell-off.” CNBC, July 28, 2026. https://www.cnbc.com/2026/07/28/hyperscalers-face-higher-capex-scrutiny-after-alphabet-report-panned.html
[18] TMT Finance. “2026 hyperscaler capex tops US$700bn – analysis.” TMT Finance, August 18, 2026. https://www.tmtfinance.com/intel/2026-hyperscaler-capex-tops-us700bn-analysis
[19] DatacenterDynamics (quoting Satya Nadella, BG2 Pod, and Amy Hood, Microsoft earnings call). “Microsoft has AI GPUs ‘sitting in inventory’ because it lacks the power necessary to install them.” DatacenterDynamics, 2025–2026. https://www.datacenterdynamics.com/en/news/microsoft-has-ai-gpus-sitting-in-inventory-because-it-lacks-the-power-necessary-to-install-them/
[20] Utility Dive (quoting Nat Bullard, Halcyon). “Data centers are ready to negotiate flexibility for speed.” Utility Dive, June 26, 2026. https://www.utilitydive.com/news/data-centers-flexibility-utilities-speed-to-power/822588/
[21] Utility Dive. “GE Vernova gas turbine backlog climbs to 116 GW.” Utility Dive, July 23, 2026. https://www.utilitydive.com/news/ge-vernova-gas-turbine-backlog-climbs-to-116-gw/826039/
[22] Energy News Beat. “GE Vernova’s Gas Turbine Backlog Hits 116 GW. What Does This Mean for the AI Market?” Energy News Beat, August 2026. https://energynewsbeat.co/ai/ge-vernovas-gas-turbine-backlog-hits-116-gw-what-does-this-mean-for-the-ai-market/
[23] Power Engineering (quoting Scott Strazik, CEO, GE Vernova). “Data centers drive record surge in GE Vernova power equipment orders as turbine slots tighten through 2030.” Power Engineering, April 23, 2026. https://www.power-eng.com/gas/turbines/data-centers-drive-record-surge-in-ge-vernova-power-equipment-orders-as-turbine-slots-tighten-through-2030/
[24] PJM Interconnection (quoting David Mills, President and CEO). “PJM Capacity Auction Procures 138,318 MW of Generation Resources as Work Continues To Address Growing Electricity Demand.” PJM Inside Lines, July 14, 2026. https://insidelines.pjm.com/pjm-capacity-auction-procures-138318-mw-of-generation-resources-as-work-continues-to-address-growing-electricity-demand/
[25] PJM Interconnection. “2027/2028 Base Residual Auction Reserve Target Shortfall Report.” PJM, February 2026. https://www.pjm.com/-/media/DotCom/markets-ops/rpm/rpm-auction-info/2027-2028/2027-2028-bra-reserve-target-shortfall-report.pdf
[26] Power Engineering (quoting Stu Bresler, PJM). “PJM capacity auction hits price cap again as region falls short of reliability target.” Power Engineering, December 18, 2025. https://www.power-eng.com/business/pjm-capacity-auction-hits-price-cap-again-as-region-falls-short-of-reliability-target/
[27] Fortune (quoting Rob Gramlich, Grid Strategies). “AI wants electricity now. The electric grid needs years to catch up.” Fortune, September 3, 2026. https://fortune.com/2026/09/03/ai-data-centers-demand-electric-grid/
[28] Stanford Report (quoting Prof. Sally Benson, Stanford Doerr School of Sustainability). “Stanford fuels solutions for a power-hungry world.” Stanford University, April 2026. https://news.stanford.edu/stories/2026/04/energy-innovation-clean-power-research
[29] International Monetary Fund. “Power Hungry: How AI Will Drive Energy Demand (IMF Working Paper No. 2025/081).” IMF, April 2025. https://www.imf.org/en/publications/wp/issues/2025/04/21/power-hungry-how-ai-will-drive-energy-demand-566304
[30] McCarthy Tétrault LLP. “Alberta Faces a Surge in AI Data Centre Power Demand: AESO Responds with Phased Connection Plan.” Canadian Energy Perspectives, 2025. https://www.mccarthy.ca/en/insights/blogs/canadian-energy-perspectives/alberta-faces-a-surge-in-ai-data-centre-power-demand-aeso-responds-with-phased-connection-plan
[31] Alberta Electric System Operator (AESO). “Large Load Projects.” AESO, 2026. https://www.aeso.ca/grid/connecting-to-the-grid/large-load-projects/
[32] Utility Dive (sponsored content, Bloom Energy, 2026 Data Center Power Report). “Redefining data center power strategies in the AI era.” Utility Dive, March 30, 2026. https://www.utilitydive.com/spons/redefining-data-center-power-strategies-in-the-ai-era/815634/
[33] AESO Engage. “Phase 2A: Large Load Integration | BYOG Process.” AESO Engage, 2026. https://aesoengage.aeso.ca/pre-engagement-phase-ii-large-load-integration
[34] Belfer Center for Science and International Affairs, Harvard Kennedy School (Mural et al.). “AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment.” Belfer Center Policy Brief, February 2026. https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
[35] Zero Emission Grid. “ERCOT LLWG 05/21 Meeting Summary: Batch Zero, Large Loads, LLIS.” Zero Emission Grid, May 26, 2026. https://www.zeroemissiongrid.com/iso-rto-meeting-summaries/ercot-llwg-05-21/
[36] Norris, T. H., Profeta, T., Patiño-Echeverri, D., & Cowie-Haskell, A. (Duke University), as reported by POWER Magazine. “Duke Researchers: Grid Flexibility Key to Accommodate Load Growth.” POWER Magazine, February 19, 2025. https://www.powermag.com/duke-researchers-grid-flexibility-key-to-accommodate-load-growth/
[37] Norris, T. H., Profeta, T., Patiño-Echeverri, D., & Cowie-Haskell, A.. “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems.” Nicholas Institute for Energy, Environment & Sustainability, Duke University, February 2025. https://www.ourenergypolicy.org/resources/rethinking-load-growth-assessing-the-potential-for-integration-of-large-flexible-loads-in-us-power-systems/
[38] Latitude Media (interview with Tyler Norris, Duke University). “The US grid may have over 100 GW of load to spare.” Latitude Media, February 11, 2025. https://www.latitudemedia.com/news/the-us-grid-may-have-over-100-gw-of-load-to-spare/
[39] Latitude Media (interview with Varun Sivaram, Emerald AI; citing Prof. Ayse Coskun, Boston University). “The rise of flexible data centers.” Latitude Media Catalyst, April 10, 2026. https://www.latitudemedia.com/news/catalyst-the-rise-of-flexible-data-centers/
[40] arXiv preprint 2604.05376. “To Defer or To Shift? The Role of AI Data Center Flexibility on Grid Interconnection.” arXiv, April 2026. https://arxiv.org/html/2604.05376v1
[41] TechXplore. “Making room for data centers without derailing the energy transition.” TechXplore, August 18, 2026. https://techxplore.com/news/2026-08-room-centers-derailing-energy-transition.html
[42] MIT News / MIT Sustainability (quoting Deepjyoti Deka, MIT Energy Initiative). “Responding to the climate impact of generative AI.” Massachusetts Institute of Technology, 2025. https://sustainability.mit.edu/article/responding-climate-impact-generative-ai
[43] McGuireWoods LLP. “FERC Issues Section 206 Show Cause Orders Directing All Six RTOs/ISOs to Justify or Reform Large Load Integration Rules.” McGuireWoods Client Alert, June 2026. https://www.mcguirewoods.com/client-resources/alerts/2026/6/ferc-issues-section-206-show-cause-orders-directing-all-six-rtos-isos-to-justify-or-reform-large-load-integration-rules/
[44] Congressional Research Service. “Data Centers and the Electricity Grid (R49326).” CRS Report, September 1, 2026. https://www.everycrsreport.com/files/2026-09-01_R49326_52044d1339d5734a5792df728c67e8a966d12438.html
[45] Martin, E., & Peskoe, A. (Harvard Law School Electricity Law Initiative). “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power.” Harvard Environmental and Energy Law Program, March 5, 2025. https://eelp.law.harvard.edu/extracting-profits-from-the-public-how-utility-ratepayers-are-paying-for-big-techs-power/
[46] Utility Dive (quoting Eliza Martin and Ari Peskoe, Harvard Law School). “Utilities may subsidize data center growth by shifting costs to other ratepayers: Harvard Law paper.” Utility Dive, March 10, 2025. https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif-harvard-peskoe/742001/
[47] Ohio Capital Journal (quoting Ari Peskoe, Harvard Law School). “Power for data centers could come at ‘staggering’ cost to consumers.” Ohio Capital Journal, March 19, 2025. https://ohiocapitaljournal.com/2025/03/19/power-for-data-centers-could-come-at-staggering-cost-to-consumers/
[48] Virginia State Corporation Commission. “SCC Issues Order on DEV Biennial Review 2025.” Virginia SCC News Release, November 25, 2025. https://www.scc.virginia.gov/about-the-scc/newsreleases/release/scc-issues-order-on-dev-biennial-review-2025/scc-rules-in-dev-biennial-review-case.html
[49] Inside Climate News. “Virginia Regulators Approve New Dominion Rates, Assign More Costs to Data Centers.” Inside Climate News, January 2026. https://insideclimatenews.org/news/07012026/virginia-regulators-approve-new-dominion-rates/
[50] DatacenterDynamics. “Virginia regulators approve new rate class for data centers and other large loads.” DatacenterDynamics, 2025–2026. https://www.datacenterdynamics.com/en/news/virginia-regulators-approve-new-rate-class-for-data-centers-and-other-large-loads/
[51] Citizen Portal (testimony of Brian Pratt, Deputy Director, Virginia SCC). “State Corporation Commission orders new GS-5 rate class for large data-center customers.” Citizen Portal, February 4, 2026. https://citizenportal.ai/articles/7371359/virginia/2026-legislature-va/state-corporation-commission-orders-new-gs5-rate-class-for-large-datacenter-customers-sets-85-minimum-charges-and-14year-contracts
[52] MTS AI Capex Wiki. “Virginia SCC GS-5 Rate Order (PUR-2025-00058), Annotated.” MTS, 2026. https://drops.mts.now/ai-capex/wiki/documents/dominion-gs5-order/
[53] AEP Ohio (quoting Marc Reitter, President and COO). “AEP Ohio Proposal on Data Centers to Protect Ohio Consumers Adopted by PUCO.” AEP Press Release, July 9, 2025. https://seekingalpha.com/pr/20161398-aep-ohio-proposal-on-data-centers-to-protect-ohio-consumers-adopted-by-puco
[54] Data Center Frontier. “Ohio Sets New Precedent: AEP’s Power Rules Shift Data Center Cost Burden.” Data Center Frontier, 2025. https://www.datacenterfrontier.com/energy/article/55304787/ohio-sets-new-precedent-aeps-power-rules-shift-data-center-cost-burden
[55] Ohio Capital Journal. “Ohio Manufacturers’ Association challenges new utility billing for data centers.” Ohio Capital Journal, November 13, 2025. https://ohiocapitaljournal.com/2025/11/13/ohio-manufacturers-association-challenges-new-utility-billing-for-data-centers/
[56] The Buckeye Institute. “Undermining Ohio’s Competitive Edge (Policy Brief).” The Buckeye Institute, March 16, 2026. https://www.buckeyeinstitute.org/library/docLib/2026-03-16-Undermining-Ohio-s-Competitive-Edge-policy-brief.pdf
[57] Electric Reliability Council of Texas (ERCOT). “PUCT Approves ERCOT’s Batch Zero Process for Connecting Large Electricity Users While Protecting System Reliability for Texans.” ERCOT News Release, June 18, 2026. https://www.ercot.com/news/release/06182026-puct-approves-ercots
[58] Data Center Knowledge (quoting Beth Garza, R Street Institute). “Texas Approves ‘Batch Zero’ Study as Data Center Demand Soars.” Data Center Knowledge, June 24, 2026. https://www.datacenterknowledge.com/build-design/texas-creates-a-credibility-test-for-gigawatt-scale-data-center-demand
[59] Keentel Engineering. “ERCOT Batch Zero Guide: Large Load & Ride-Through Rules.” Keentel Engineering, July 6, 2026. https://keentelengineering.com/ercot-batch-zero-guide-large-load-interconnection
[60] The Dallas Morning News (quoting Pablo Vegas and Jeff Billo, ERCOT). “ERCOT advances ‘batch zero’ process for data center, large loads.” The Dallas Morning News, June 2026. https://www.dallasnews.com/business/energy/article/batch-zero-ercot-data-center-22290855.php
[61] Office of the Texas Governor, Greg Abbott. “Governor Abbott Directs Comprehensive Data Center Audit.” Office of the Governor, August 3, 2026. https://gov.texas.gov/news/post/governor-abbott-directs-comprehensive-data-center-audit
[62] Troutman Pepper Locke. “Texas Hits Pause on Data Center Grid Connections Amid Growing Oversight Push.” Troutman Pepper Locke, August 4, 2026. https://www.troutman.com/insights/texas-hits-pause-on-data-center-grid-connections-amid-growing-oversight-push/
[63] The Texas Tribune. “Gov. Greg Abbott broadens moratorium on data center approvals to include environmental permits.” The Texas Tribune, September 21, 2026. https://www.texastribune.org/2026/09/21/texas-data-center-moratorium-water-energy/
[64] Houston Public Media (quoting Dan Diorio, Data Center Coalition). “Gov. Greg Abbott pauses new data centers until ERCOT, PUCT audit energy, water usage.” Houston Public Media, August 3, 2026. https://www.houstonpublicmedia.org/articles/news/energy-environment/2026/08/03/558529/gov-greg-abbott-pauses-new-data-centers-until-ercot-puct-audit-energy-water-usage/
[65] Compute Law Blog. “Georgia Power and Public Service Commission rules for AI data centers.” Compute Law Blog, May 23, 2026. https://computelaw.blog/power/georgia-power-psc-data-center-rules/
[66] DatacenterDynamics (quoting Jason Shaw, Chairman, Georgia PSC). “Georgia PSC approves new billing rules for data centers and large load customers.” DatacenterDynamics, 2025. https://www.datacenterdynamics.com/en/news/georgia-psc-approves-new-billing-rules-for-data-centers-and-large-load-customers/
[67] Pennsylvania Public Utility Commission. “PUC Releases Final Order Establishing First-of-Its-Kind Large Load Model Tariff Framework.” PA PUC Press Release, May 13, 2026. https://www.puc.pa.gov/press-release/2026/puc-releases-final-order-establishing-first-of-its-kind-large-load-model-tariff-framework-05132026
[68] K&L Gates LLP. “Pennsylvania Public Utility Commission Adopts Model Interconnection Tariff for Large Load Customers.” K&L Gates HUB, May 29, 2026. https://www.klgates.com/thought-leadership/Pennsylvania-Public-Utility-Commission-Adopts-Model-Interconnection-Tariff-for-Large-Load-Customers-5-29-2026
[69] Utility Dive. “Pennsylvania releases ‘first-of-its-kind’ large-load model tariff.” Utility Dive, May 18, 2026. https://www.utilitydive.com/news/pennsylvania-releases-first-of-its-kind-large-load-model-tariff/820456/
[70] Pennsylvania State Association of Township Supervisors (quoting Steve DeFrank, Chairman, PA PUC). “PUC Issues Final Order for New Tariff Framework to Protect Ratepayers, Guide Data Center Growth.” PSATS, May 15, 2026. https://www.psats.org/puc-issues-final-order-for-new-tariff-framework-to-protect-ratepayers-guide-data-center-growth/
[71] Federal Energy Regulatory Commission (quoting Chairman Laura Swett). “FERC Directs Nation’s Largest Grid Operator to Create New Rules to Embrace Innovation and Protect Consumers.” FERC News, December 18, 2025. https://www.ferc.gov/news-events/news/ferc-directs-nations-largest-grid-operator-create-new-rules-embrace-innovation-and
[72] Federal Energy Regulatory Commission, Commissioner David Rosner. “E-1: Commissioner Rosner’s Concurrence on PJM Co-Location.” FERC, December 18, 2025. https://www.ferc.gov/news-events/news/e-1-commissioner-rosners-concurrence-pjm-co-location
[73] RMI. “Understanding FERC’s Large Load Orders.” RMI, July 6, 2026. https://rmi.org/resources/understanding-fercs-large-load-orders/
[74] The White House. “Fact Sheet: President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge.” The White House, March 4, 2026. https://www.whitehouse.gov/fact-sheets/2026/03/fact-sheet-president-donald-j-trump-advances-energy-affordability-with-the-ratepayer-protection-pledge/
[75] The White House. “President Trump’s Ratepayer Protection Pledge Secures American AI Dominance, Protects Consumers.” The White House, July 28, 2026. https://www.whitehouse.gov/releases/2026/07/president-trumps-ratepayer-protection-pledge-secures-american-ai-dominance-protects-consumers/
[76] Latitude Media. “Energy industry greets White House data center pledge with a shrug.” Latitude Media, March 2026. https://www.latitudemedia.com/news/energy-industry-greets-white-house-data-center-deal-with-a-shrug/
[77] The White House. “Ratepayer Protection Pledge.” The White House, July 2026. https://www.whitehouse.gov/ratepayer-protection-pledge/
[78] Avanza Energy (Christopher Johnson). “Seven Tech Giants Signed a Pledge to Protect You From Higher Electric Bills. It Will Do the Opposite.” Avanza Energy Substack, April 1, 2026. https://avanzaenergy.substack.com/p/the-444-a-year-lie-how-the-white
[79] International Energy Agency (IEA). “Key Questions on Energy and AI: Executive Summary.” IEA, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
[80] International Energy Agency (quoting Fatih Birol, Executive Director). “AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works.” IEA News, April 10, 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[81] PolitiFact (citing Lawrence Berkeley National Laboratory, June 2026). “How big an energy drain will AI data centers be?” PolitiFact, July 14, 2026. https://politifact.com/factchecks/2026/jul/14/donald-trump/ai-data-centers-electricty-grid/
[82] International Energy Agency (IEA). “Energy and AI: Executive Summary.” IEA Special Report, April 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
[83] MIT News (quoting Noman Bashir, MIT Climate and Sustainability Consortium). “Explained: Generative AI’s environmental impact.” Massachusetts Institute of Technology, January 17, 2025. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117
[84] MIT Energy Initiative (quoting Prof. William H. Green, Director). “Data Center Power Demand.” MIT Energy Initiative, 2026. https://energy.mit.edu/current-initiatives/data-center-power-demand/
[85] Palo Alto Online (report on Stanford Precourt Institute Sustainable Data Centers Symposium). “‘We need more energy’: Stanford symposium explores data center growth.” Palo Alto Online, May 4, 2026. https://www.paloaltoonline.com/stanford-university/2026/05/04/we-need-more-energy-stanford-symposium-explores-data-center-growth/
[86] Grynspan, R., Secretary-General, UN Trade and Development (UNCTAD). “Launch of the Digital Economy Report 2024.” United Nations — UNCTAD, July 10, 2024. https://unctad.org/osgstatement/launch-digital-economy-report-2024
[87] University of Chicago News (interview with Prof. Andrew Chien). “Could data centers break our power grid? Big Brains podcast with Andrew Chien.” University of Chicago, May 2026. https://news.uchicago.edu/big-brains-podcast-could-data-centers-break-our-power-grid-andrew-chien
