The Power Behind the Power: Why the Real AI Bottleneck Isn’t What You Think
A new category is forming in the gap between hyperscaler self-build and the utility queue. Almost no one is watching it.
Quick Take:
- The two most visible solutions to the U.S. power crisis — Kevin O’Leary’s 40,000-acre off-grid Stratos project in Utah and Meta’s $27B+ Hyperion deal turning Entergy Louisiana into a captive contractor are structural mirror images — and neither is available to the rest of the market
- A new commercial category has quietly formed in the gap between hyperscaler self-build and the broken utility queue; what customers are actually buying is time-to-power, at premiums of 30–50%
- The U.S. competitive field consists of just two pure-play operators — both Houston-based, both racing upmarket — leaving an $80–120 billion mid-market opportunity substantially unattended
- FERC’s pending June 2026 large-load rule, VoltaGrid’s recent billion-dollar raise, and accelerating hyperscaler self-build will likely consolidate the category by 2028 — making the next 18 months the defining commercial moment
In early May 2026, the Box Elder County Commission in northern Utah voted unanimously to advance a 40,000-acre data center development called the Stratos Project. The project, backed by Shark Tank investor Kevin O’Leary and a Utah real estate venture called West GenCo, will eventually require nine gigawatts of electricity — more than double what the entire state of Utah currently consumes in a year. None of that electricity will come from the local grid. Every watt will be generated on site, drawn from a natural gas pipeline running through the property, in facilities built and owned by the developers themselves.
Six hundred miles east in rural Louisiana, Meta is taking the opposite approach to the same problem. Its Hyperion data center campus, sprawling across nearly four thousand acres in Richland Parish, will eventually require more than seven gigawatts of power — roughly half of all electricity Entergy Louisiana currently produces. Meta is not building its own power plants. It has instead written checks to Entergy to construct ten new gas-fired power plants on its behalf, with the utility owning and operating them under a twenty-year service agreement. The total project cost has been variously estimated at $27 billion to over $200 billion, depending on which figures one accepts and how the off-balance-sheet financing is counted.
These two projects represent the two most visible solutions to the most pressing infrastructure problem in the American economy. They are also, structurally, mirror images of each other. O’Leary’s project bypasses the utility entirely; Meta’s makes the utility into a contractor. Both are responses to the same underlying constraint: the U.S. electric grid cannot deliver new connections fast enough to support the build-out that AI and advanced manufacturing now demand.
What gets less attention is what happens to everyone who is not a hyperscaler, and is not a billionaire with the patience for a forty-thousand-acre permitting fight. The middle of the market — facilities needing one to fifty megawatts of new power, with capital budgets in the tens or hundreds of millions rather than the tens of billions — faces the same five-year wait, the same broken interconnection queue, and the same crushing cost of delay. They cannot self-build a private power plant the size of a small county. They cannot finance ten utility-owned gas plants. And there is, at the moment, almost no one selling them a packaged solution to the problem.
A new commercial category is forming in that gap. It does not yet have a settled name. The companies operating in it sometimes call it Microgrid-as-a-Service, sometimes Resiliency-as-a-Service, sometimes Behind-the-Meter Power Delivery. The phrase that best captures what is actually being sold, and what customers are actually buying, is time-to-power. The category is real. The competitive field is thin. And the window during which the category remains uncontested is probably narrower than three years.
The Bottleneck
The American grid is in trouble. As of the end of 2025, more than 2,060 gigawatts of new generation and storage capacity were waiting in interconnection queues to connect to the grid — roughly twice the total existing generating capacity of the United States. The median wait time for a project to reach commercial operation now approaches five years, and the withdrawal rate for projects entering the queue in the current environment is approaching eighty percent. The grid has effectively stopped accepting new customers at the speed required by the economy trying to connect to it.
This bottleneck has multiple causes, all of them mutually reinforcing. The most discussed is artificial intelligence: training and inference workloads have driven a roughly tenfold increase in rack densities at large data centers in three years, and U.S. data center power demand is now projected to double or triple by 2030. Less discussed but equally important is the parallel build-out of advanced manufacturing under the CHIPS Act and the Inflation Reduction Act — semiconductor fabs, electric vehicle plants, battery manufacturing, and the reshoring of industrial capacity that had quietly drifted offshore over thirty years. None of this was on utility planning sheets in 2020.
On the supply side, the picture is worse. Lead times for the large transformers required to expand the grid run eighteen to twenty-four months, and the global manufacturing capacity is concentrated in a handful of suppliers, several of them outside the United States. Permitting timelines for new transmission lines have grown longer, not shorter, as siting fights become more contested. Skilled labor is in chronic short supply. Equipment that was uneconomic when interest rates were near zero has become more expensive in real terms even as demand has surged.
The Federal Energy Regulatory Commission has spent much of the past three years working on reforms to the interconnection queue process. Those reforms, broadly, address how new generation connects. They do not address how new large loads — a fifty-megawatt semiconductor fab, a twenty-megawatt AI inference cluster — connect on the other side. That gap is now the subject of an active and unresolved federal rulemaking. Docket No. RM26–4–000, FERC’s large-load interconnection proceeding, was originally targeted for a final rule by Energy Secretary Wright’s April 30, 2026 deadline. The Commission missed that deadline. On April 16, 2026, it instead issued an Order Regarding Intent to Act, formally extending the timeline and committing to action by the end of June. Chairman Laura Swett has publicly attributed the delay to the volume of stakeholder feedback — over 3,500 pages of public comments — and to a deliberate prioritization of “legally durable” rules capable of surviving the federal-versus-state-authority court challenges that are widely expected.
In the absence of a national rule, FERC has acted in piecemeal fashion. In December 2025 and April 2026, it directed PJM, the nation’s largest grid operator, to implement transparent rules for co-located loads — the category that includes data centers sited at or adjacent to existing power plants. In January 2026 it approved the Southwest Power Pool’s High Impact Large Load initiative, a fast-track mechanism for connecting massive new power users in that region. These actions are likely to form the spine of the eventual national rule, but the larger architectural questions — who has jurisdiction over what, on what timeline, with what cost allocation — remain open. The regulatory environment during this transition is unusually permissive at the project level, precisely because no settled federal framework yet exists to constrain it.
Against this backdrop, two responses have become visible at the upper end of the market. The first is hyperscaler self-build: Amazon, Microsoft, Google, Meta, and a small number of similarly capitalized players are increasingly committing to fund their own generation, either directly or through arrangements like Meta’s with Entergy. The second is the O’Leary-style off-grid private development, where a wealthy backer or consortium assembles land, gas pipeline access, and political support to build something that resembles a small utility serving a single customer — or a captive customer base.
Both of these are reasonable responses to an unreasonable situation. Neither is available to the rest of the market.
The Field as It Actually Exists
The companies addressing the gap below hyperscaler scale, in the United States, can be counted on the fingers of one hand. Two of them dominate the conversation. Both are based in Houston.
Enchanted Rock is the older of the two, founded in 2006. It invented the phrase Resiliency-as-a-Service, and has spent two decades quietly building a portfolio that now totals around 300 microgrids and roughly 600 megawatts of installed capacity, with another 400 megawatts under construction. Its customer base is heavily weighted toward mid-market commercial and industrial buyers: Walmart, the H-E-B grocery chain, municipal water utilities, regional hospital systems, logistics operations. The financial structure is the interesting part. Enchanted Rock finances, owns, and operates the microgrid assets itself; customers pay only for the service of guaranteed power. Public disclosures suggest the company has raised approximately $36.5 million through October 2025, a strikingly lean capital footprint for a firm controlling close to a gigawatt of distributed generation. The implication is that the project-level economics are strong enough to support self-financing, with equity used primarily for the corporate platform.
VoltaGrid, founded only in 2020, has taken the opposite trajectory. In May 2026 it announced a $1 billion strategic equity investment from Blackstone Tactical Opportunities and Halliburton, alongside a 2.3-gigawatt agreement to power Oracle’s AI data centers — the largest single deal in the company’s history. VoltaGrid describes its target market as data centers, microgrids, and industrial applications, but the gravitational pull of its current capital structure and contract pipeline points clearly upmarket. The company is moving toward hyperscaler-scale work, not away from it.
Beyond these two pure-plays, the field consists mostly of large industrial equipment companies — Bloom Energy, Schneider Electric, Siemens, Caterpillar, Aggreko, ABB, Honeywell — offering Microgrid-as-a-Service as a product line rather than as a core identity. Their commercial motion is fundamentally different from Enchanted Rock’s. They sell equipment and adjacent services; they do not, generally, take principal risk on the asset or on the customer’s time-to-power outcome.
That is the universe. Two pure-plays focused on the U.S. market, both based in the same city, plus a handful of conglomerates treating the category as a side business. For a market characterized by five-year wait times and customer budgets measured in billions per facility, the competitive field is conspicuously thin.
The global Microgrid-as-a-Service market, as currently measured, reached approximately $3.5 billion in 2025 and is projected by industry analysts to grow to roughly $9.25 billion by 2033, a 12.9 percent compound annual growth rate. These figures, however, capture only existing service-model revenue. They do not capture the total addressable spend that could shift to the model if customers stopped trying to wait for the utility. By that broader measure — taking the 35 to 40 percent of new data center capacity not absorbed by hyperscalers, plus the industrial reshoring not yet served by self-build solutions — the addressable opportunity over the 2026 to 2030 window is more credibly estimated at $80 to $120 billion in cumulative infrastructure spending, with recurring service revenue layered on top.
The gap between the $9 billion projection and the $80 billion latent opportunity is essentially the gap between what the existing players have been able to capture and what the underlying demand would justify if a category-defining service operator emerged at scale. That operator does not yet exist.
What Is Actually Being Sold
It is worth being precise about what these companies do, because the terminology obscures the business.
A behind-the-meter service operator does not sell electricity in the way a utility does. It sells a packaged contract that bundles four things: on-site generation (typically natural gas turbines, increasingly augmented with batteries and solar), a smaller grid interconnection for backup and grid services, software-driven dispatch that optimizes when to generate, store, draw, or sell, and full project management — permits, vendor relationships, financing, regulatory work — handled on the customer’s behalf. The customer signs a fifteen- to twenty-year service agreement, pays a fixed rate per kilowatt-hour plus capacity and standby fees, and never deals with the underlying complexity.
What the customer is actually buying, in commercial terms, is time. A 100-megawatt AI inference facility that comes online in eighteen months instead of seven years generates roughly five and a half years of additional revenue at a cost of perhaps thirty to fifty percent above standard grid electricity rates. The math is almost embarrassingly favorable for the customer in any scenario where the underlying business is viable, which is why these customers are willing to pay the premium.
The revenue model on the operator side is layered. The largest stream is the power spread — the margin between what the customer is charged per kilowatt-hour and what the asset actually costs to run. Beyond that, the operator earns standby and capacity fees for having power available regardless of usage; project development fees for designing and shepherding the installation through permitting; demand response revenue from grid operators like PJM, ERCOT, and CAISO who pay large customers to reduce load during stress events; capacity market payments in regions where these exist; and tax credits for solar, batteries, and clean equipment, monetizable through transfer to companies with large tax appetites.
These streams compound. A well-structured project produces project-level returns in a band that has historically been attractive to infrastructure investors, with revenue characteristics that are highly recurring and contractually locked. The economics are not the constraint. The constraint is the operational complexity of actually executing one of these projects, on schedule, with the right equipment, in the right regulatory environment.
This is not, importantly, a software business. It is a project finance, regulatory coordination, vendor management, and operational discipline business. The moat is execution, not technology. Anyone with capital can buy turbines; very few people can install them on time, secure interconnection rights, navigate the relevant state PUC, and operate the resulting facility reliably for two decades. The category that is forming will be defined by who can do that at scale.
Where the O’Leary and Meta Stories Belong in the Picture
Return briefly to the two anchor projects.
O’Leary’s Stratos / Wonder Valley project in Utah is, in effect, an extreme version of the behind-the-meter strategy. The developer is sourcing land cheap enough and political backing strong enough to build something that resembles a utility, off-grid, to serve what the project’s promoters describe as “national defense” and contracted AI workloads. The scale is preposterous — nine gigawatts is more than twice the average state’s total power consumption — and the execution risk is correspondingly enormous. The project’s Alberta predecessor, also called Wonder Valley, was announced with a 2027 commissioning target; as of mid-2026, construction is not expected to begin until 2028, and the original site reportedly remains undeveloped. The Utah version is at an earlier stage and has yet to clear the litigation likely to follow the Box Elder County Commission’s controversial approval.
What the project demonstrates, regardless of whether it is ultimately built, is that the cost of delay in the AI economy is now high enough that serious capital is willing to attempt a forty-thousand-acre private off-grid solution rather than wait for the grid. That is the most expensive possible response to the bottleneck. It is available only to a vanishingly small number of buyers.
Meta’s Hyperion in Richland Parish, Louisiana, is the inverse strategy at similar scale. Meta has effectively turned Entergy Louisiana into a captive contractor, paying for the construction of ten gas-fired power plants, 240 miles of 500 kilovolt transmission lines, and battery storage across multiple parishes, all under a structure in which Entergy owns and operates the assets while Meta funds the cost. The Louisiana Public Service Commission has fast-tracked the approval through a “Lightning Initiative” specifically designed for projects of this scale. The structure relies on cooperation from a friendly state regulator, a willing utility partner, and political backing at every level. The final commission vote is scheduled for December 2026, and the first phase of the data center itself is expected to be operational by 2030.
What Meta’s project demonstrates is that even hyperscalers cannot solve their power problem alone. They require utility cooperation, state-level legislative carve-outs, and bespoke regulatory pathways. The fact that Mike Johnson, Speaker of the House, represents the congressional district in which the project sits is not incidental to its progress.
Both projects illustrate, by their extremity, what the rest of the market cannot do. A semiconductor fab needing forty megawatts cannot get a “Lightning Initiative” carve-out from a state PSC. A regional health system needing twelve megawatts of resilience cannot lobby the Speaker of the House. A Tier 2 colocation operator with a 75-megawatt expansion plan in Phoenix cannot privately fund $27 billion of utility infrastructure or assemble forty thousand acres of off-grid land. These customers exist in significant numbers; they have committed capital; they cannot wait. They constitute the addressable market for the category that is now forming.
Who Buys the Middle of the Market
A useful exercise is to identify, concretely, who occupies the gap between hyperscaler self-build and irrelevance. Five customer categories are visible at the moment, with quite different commercial dynamics.
The first and cleanest is AI inference colocation and the neocloud operators. Companies like CoreWeave, Lambda, RunPod, Crusoe, Nscale, Genesis Cloud, Together AI, and Cerebras Inference are building distributed inference capacity that needs to live near end users — near population centers, not near cheap stranded power in West Texas. Typical facility sizes run 500 kilowatts to 20 megawatts. Rack densities have climbed from 25 kilowatts to over 130 kilowatts in two years, with industry forecasts pointing toward 600 kilowatts per rack by 2027. These operators are growing fast, are not balance-sheet equipped to self-build power infrastructure, and have shown clear willingness to sign long service contracts in exchange for time-to-energization.
The second is secondary and Tier 2 colocation operators: NTT Global Data Centers, Iron Mountain, DataBank, Flexential, Aligned, QTS, and similar firms operating mid-size facilities (5 to 50 megawatts) in markets where hyperscaler dominance is incomplete. These companies have customers waiting for capacity but lack the means to bring new sites online without power. Their balance sheets are not designed to absorb $50 million-plus of generation capex per site. The service model fits their commercial structure naturally.
The third — and the most underdiscussed in the current conversation — is reshoring industrial manufacturing. Semiconductor fabs below the TSMC / Intel / Samsung scale, electric vehicle and battery plants below mega-scale, CHIPS Act and IRA-funded advanced manufacturing facilities, biopharmaceutical manufacturing, food processing at scale. Typical power needs run 10 to 100 megawatts. These customers face the same interconnection delays as data centers but receive less attention from the press and less attention from the existing behind-the-meter operators. Almost no one is specifically targeting them today.
The fourth is critical infrastructure motivated by resilience rather than capacity: hospitals, water and wastewater utilities, regional airports, port and terminal operations, county and municipal emergency operations centers. These customers have historically relied on diesel backup; converting them to dual-purpose natural gas microgrids that earn revenue during normal operations and provide resilience during outages is the original Enchanted Rock model and remains the most validated commercial case in the entire category.
The fifth, less attractive but real, is cryptocurrency mining and behind-the-meter compute arbitrage. Crypto miners have historically been willing to site directly at stranded gas or stranded power sources, and several behind-the-meter operators have used crypto loads as bridge revenue during construction of facilities later anchored by more reputationally durable customers. The category is volatile but it should not be excluded from a complete picture of the market.
These five segments do not require the same product. An inference colocation operator wants speed and density. A reshoring manufacturer wants reliability and a long predictable cost curve. A hospital wants resilience and a clean regulatory story. The category that captures this market will need to develop differentiated commercial templates for each, but the underlying engineering and financing infrastructure can be shared across all of them.
The Geography Is Less American Than It Looks
The U.S. story is well covered in the trade press; the international story is less so. It is worth a brief survey, because the structural conditions that created the American market exist almost everywhere, and in several places they are worse.
Europe is the most acute case. Within the European Union, wait times for new grid connections range from two to ten years, and in the primary data center hubs — Frankfurt, London, Amsterdam, Paris, Dublin — queues now average seven to ten years. German data centers face connection dates as far off as 2031. The United Kingdom has reported queues as long as thirteen years. The Netherlands has imposed strict new rules in Amsterdam, and Dublin has paused new facility construction until 2028. Pressure is now spreading to secondary markets including Madrid, Milan, and Warsaw, with over two hundred gigawatts of projects awaiting grid connections in Italy and Spain alone.
Behind-the-meter adoption in Europe was running at roughly five to ten percent of data centers eighteen months ago and has since risen to approximately twenty percent. The first true microgrid-powered data center in Europe — Pure Data Centre Group’s facility outside Dublin — only came online in early 2026, partnered with the Danish power-solutions firm AVK. The competitive field in Europe consists of equipment vendors with strong positions — Aggreko in temporary power, INNIO in gas engines, Wärtsilä in industrial generation, Siemens Energy, and the major utilities themselves — but no integrated multi-customer service operator at scale. The U.S. service model has not been replicated in Europe. This is a structural gap.
Southeast Asia and India represent the largest raw growth opportunity. More than fifty billion dollars is being invested in the Southeast Asian regional data center market, adding 8.5 gigawatts of capacity. Malaysia’s data center electricity use is projected to rise from nine terawatt-hours in 2024 to sixty-eight terawatt-hours in 2030 — roughly thirty percent of national power consumption. The grid problem in Asia differs from the U.S. and Europe: it is not slow interconnection queues but inadequate base capacity and unreliable supply. India in particular combines massive load growth, fragmented grid reliability, abundant natural gas, and aggressive industrial reshoring under the Make in India and Production Linked Incentive schemes. The customer base extends well beyond data centers, into semiconductor fabs, electric vehicle plants, and large commercial complexes. The competitive field includes local colocation operators self-developing power infrastructure but no pure-play behind-the-meter service operator.
The Middle East is smaller in absolute terms but better-funded per capita than anywhere else, with approximately $15.5 billion of new data center investment expected by 2027 and Saudi Arabia accounting for nearly half of upcoming capacity. The region has abundant cheap natural gas, abundant land, and strategic backing for digital sovereignty. The competitive field includes G42, Aramco Digital, and ACWA Power, but again no integrated service operator. VoltaGrid’s recent investment from Halliburton was partly motivated by the eastern hemisphere opportunity; the two firms have together secured manufacturing capacity for four hundred megawatts of modular natural gas systems for data centers outside the United States. The window for a new entrant in this region is narrowing faster than in Europe.
Industrial customers beyond data centers are the largest and most globally dispersed segment. Mining operations across Australia, Chile, Peru, South Africa, Indonesia, and the Democratic Republic of Congo run heavily on diesel that costs three to five times the price of natural gas or renewable alternatives. Stranded gas locations across the Permian, Marcellus, Vaca Muerta in Argentina, and offshore platforms represent potential behind-the-meter generation in the multi-gigawatt range. Reshoring manufacturing in Mexico — where nearshoring from China has driven enormous foreign direct investment into Monterrey, Saltillo, and Querétaro — faces a CFE grid notoriously unable to deliver new industrial connections at speed. Cement, steel, chemicals, and desalination plants in regions with strong carbon credit or decarbonization-grant regimes represent additional long-duration customer pools.
The honest assessment is that the U.S. is the most validated market, Europe is the most attractive structural gap, Asia is the largest growth opportunity but requires local execution capability, the Middle East has the capital but is consolidating quickly, and industrial verticals are underexploited everywhere.
What Closes the Window
A new commercial category, when it forms, tends to consolidate faster than its early observers expect. The behind-the-meter power category is unusually exposed to consolidation pressure for three reasons.
First, the two existing U.S. pure-plays are not standing still. VoltaGrid’s billion-dollar capital raise positions it to compete for hyperscaler-scale contracts, which will eventually push down-market pricing pressure into the mid-market. Enchanted Rock, with its long operating history and modest capital footprint, is an obvious acquisition target for a private equity infrastructure fund or a strategic acquirer wishing to enter the category quickly. Either outcome reduces the available competitive whitespace.
Second, hyperscaler self-build is expanding. Bloom Energy’s recent forecast projects that thirty-eight percent of data center facilities will use onsite generation for primary power by 2030, up from thirteen percent in 2025. The hyperscalers absorbing the top of the market reduces the relative attractiveness of serving them and concentrates the addressable opportunity into the mid-market that is currently underserved. That is an opportunity, but it is also a moving target.
Third, the regulatory environment is in flux at both the state and federal levels. State public utility commissions are increasingly aware of behind-the-meter generation and are beginning to write rules about standby charges, interconnection requirements, and the relationship between behind-the-meter assets and grid services. At the federal level, the FERC large-load rulemaking discussed earlier will, when it lands, change the baseline conditions under which all of this operates. The current period — 2026, possibly into 2027 — is unusually permissive across most U.S. jurisdictions because regulators are reacting to a power-supply crisis rather than to a behind-the-meter category crisis. That permissive window will narrow as the category becomes more visible and as the federal framework consolidates.
The most defensible read on the situation is that the category will exist in a recognizable form by 2028, with three or four scaled operators in the U.S., one or two in Europe, and the international markets either consolidating around U.S. expansion or developing local champions. The opening for a category-defining new entrant is now, and probably not for much longer.
What This Adds Up To
The headlines tell one story. Kevin O’Leary fights with Utah environmentalists about a forty-thousand-acre data center. Meta and Entergy negotiate ten gas-fired power plants in rural Louisiana under a “Lightning Initiative” fast-track. The Trump administration debates the future of clean energy tax credits. FERC wrestles with how to regulate the connection of large new loads to a grid that cannot accommodate them on the timeline the economy now demands.
These are the visible surface of a much larger underlying structural shift. The U.S. economy, and to varying degrees the global economy, has discovered that electricity delivery is the binding constraint on the industries it most wants to grow. Chips can be designed; data centers can be permitted; manufacturing capacity can be financed. None of it operates until the electrons arrive. The grid as currently configured cannot deliver them on the schedule the rest of the economy now requires.
The hyperscalers have responded by building their way around the constraint, either through utility partnerships or, in the most extreme cases, through private off-grid development. The rest of the market — the inference clusters, the Tier 2 colocation operators, the semiconductor fabs below the TSMC scale, the reshored manufacturing, the resilience-driven critical infrastructure — needs the same solution at a smaller and more standardized scale. A commercial category has formed to serve that need. It is currently under-capitalized, under-covered, and structurally underserved relative to the demand.
The trade press has spent the past eighteen months focused on data centers, training compute, and the chip supply chain. These are important stories. The story of how all of it actually gets powered, customer by customer, in the gap between Meta’s private utility contract and the public grid that cannot keep up, is the less glamorous version of the same story. It is also where, by the end of the decade, much of the actual money will be made.
The power behind the power is the part of the AI infrastructure build-out that nobody is watching. That probably will not remain true for long.
