The AI infrastructure boom has plenty of capital. What it increasingly lacks is time-to-power.
Executives talk their book and politicians say whatever the room wants to hear, but credit markets have to put actual money against a physical outcome. When bankers start demanding a premium for something, they have usually noticed a constraint before the rest of us finished arguing about it.
That is what makes the summer of 2026 worth a closer look. The story getting attention is the sheer scale of AI infrastructure borrowing. The story underneath it is more interesting: credit markets are beginning to draw a sharper distinction between debt backed by operating infrastructure and debt funding the promise of infrastructure still to be built. That distinction is essentially a price on physics.
The Spread Tells You Where the Risk Actually Lives
The financing requirements behind the AI buildout have become enormous. Hyperscalers are accumulating hundreds of billions of dollars of capital commitments, leases, and purchase obligations as they race to secure computing capacity. But not all of that infrastructure risk looks the same.
A securitized data center deal can be backed by a building that already exists. The concrete is poured, the interconnection agreement is signed, the transformers showed up, the switchgear is energized, and a creditworthy tenant signed a long lease. Most of the hard steps have cleared. An unsecured corporate bond ultimately relies on the company’s ability to execute an enormous future capital program, building facilities somewhere, eventually, using equipment that may not yet have been ordered and electricity that may not yet be available.
Anyone who has lived through a capital project knows these are fundamentally different risks. It is the difference between lending against a refinery running at rates and lending against a permit application for one. Same nameplate on the drawing, wildly different odds of producing cash flow next year. Credit markets are beginning to recognize the difference.
The Queue Is Not the Grid
The physical evidence backs them up. Berkeley Lab’s latest interconnection study found roughly 8,200 projects waiting to connect to the U.S. grid at the end of 2025, representing about 1,312 gigawatts of generation plus another 749 gigawatts of storage. The generation queue alone is roughly comparable to the generating capacity of the entire existing U.S. power system — before you add another 749 gigawatts of proposed storage.
That is not a pipeline. It is a wish list.
Filing an interconnection request is relatively cheap. Energizing a facility is not. Between those two events sit transmission studies, permits, transformers, switchgear, generation equipment, construction crews, financing, local approvals and, increasingly, questions about who pays for all of it. That gap is where a lot of AI capex assumptions quietly go to die.
Texas makes the point with more drama. ERCOT is sitting on roughly 474 gigawatts of large-load interconnection requests, about ninety percent of it associated with data centers, against an all-time system peak of roughly 91 gigawatts. The queue is more than five times everything Texas has ever consumed at once. Nobody seriously analyzing the grid believes all of it gets built. The hard question is which tenth does, on what schedule, and who eats the cost of transmission built for the nine tenths that evaporate.
Credit markets are putting a price on physics. Texas just showed how quickly politics can put a price on it too.
Texas Just Proved the Point in Eight Days
On August 3, Governor Abbott directed the Public Utility Commission and ERCOT to pause new data center grid connections pending a comprehensive audit of their power and water consumption. The trigger was almost comically mundane: state law already required these facilities to report usage, and fewer than one in ten had bothered to comply.
Then, on August 11, the EIA published its Short-Term Energy Outlook and cut its forecast for Texas electricity load growth in 2027 from fourteen percent to six percent, citing the pause directly. One administrative action, eight days, and a federal statistical agency revised a major state’s expected electricity demand growth by more than half.
It would be easy to read that as a story about one governor, but it is not. Political consent is becoming a first-order variable in the AI buildout, sitting right alongside chips, capital, land and electricity. Communities are starting to ask what they get in return for the water, transmission infrastructure and megawatts consumed by enormous data centers, and they are discovering that “jobs” can be a thin answer for facilities that may employ relatively few people once construction is finished. That conversation gets louder from here, county by county and state by state.
The Other Queue Nobody Talks About
Even political approval does not repeal physics. Once you have the permit and the interconnection, somebody still has to manufacture the machine.
Gas turbine order books are the tightest I can remember. GE Vernova entered the summer with roughly 116 gigawatts of gas equipment backlog and slot reservations and expects to have at least 125 gigawatts under contract by year-end. The company is expanding manufacturing capacity aggressively, with plans to move annual turbine output toward 30 gigawatts by 2030, but manufacturing capacity takes years to build too. Siemens Energy is facing a similar surge in demand, with its gas turbine backlog approaching 70 gigawatts. For practical purposes, access to large-frame gas turbines through the end of the decade has become an increasingly scarce commodity.
That has a consequence worth flagging for anyone who cares about ratepayers rather than just shareholders. Hyperscalers and data center developers are reserving turbine slots and competing with regulated utilities for the same equipment. A technology company pursuing very high returns on incremental AI capacity can justify paying considerably more to secure a turbine than a regulated utility earning a single-digit allowed return. In that competition, the hyperscaler has an obvious advantage.
The utility that loses the slot, however, still has an obligation to serve its customers. It will meet that obligation somehow , potentially later and more expensively , and ultimately some portion of that cost can land on residential and commercial ratepayers who never participated in the auction. The AI infrastructure race is therefore beginning to affect not just the companies building data centers, but everyone competing with them for the physical equipment required to produce electricity.
The Winning Technology Is the One That Shows Up
The generation mix tells the same story from another direction. Natural gas still supplies about forty percent of U.S. electricity, and EIA expects it to remain near that level through 2027. Meanwhile, solar generation grew roughly twenty-one percent in the first half of 2026 while gas generation grew about two percent. That is not a contradiction. It is what happens when a power system suddenly places an enormous premium on time-to-power.
Solar and storage can often arrive fastest. Gas provides dispatchable power but increasingly has to wait for equipment. Nuclear may eventually play a larger role, particularly as hyperscalers sign agreements around existing plants, restarts and next-generation reactors, but new nuclear development cycles are still measured in years rather than quarters. AI does not particularly care which technology wins the ideological argument. It needs electrons, and it will consume whatever reliable electrons developers can actually get to the meter.
That may be the most important shift of all. For years, the electricity debate was organized largely around the cost of energy. AI is forcing the market to think about something different: the cost of waiting for energy.
What I Would Actually Watch
If you want to know whether the AI infrastructure buildout is real, stop counting announcements and start watching what has to happen before the electrons actually flow. Watch the credit spreads and financing structures, because the growing distinction between operating assets and promises of future capacity is the market’s live estimate of execution risk. Watch the interconnection agreements, because a gigawatt in a queue is not a gigawatt on the grid.
Watch how the Texas audit resolves. If Texas starts imposing tougher standards around proof of power, water availability, financial commitments or cost allocation, other states will be watching closely. And watch the turbine and transformer order books, because you cannot finance your way past a manufacturing lead time.
Money is not the scarce resource. Time-to-power is.
Interconnections take time. Transformers take time. Turbines take time. Transmission takes time. Water infrastructure takes time. Political consent takes time. Wall Street is beginning to price that difference because lenders eventually have to distinguish between a data center someone announced and a data center capable of consuming an electron.
You can finance an announcement. You cannot finance your way around physics.
Further Reading
- EIA Short-Term Energy Outlook, August 2026 — Texas load revision and U.S. generation mix data
- Berkeley Lab, “Queued Up: 2026 Edition” — U.S. interconnection queue through year-end 2025
- Utility Dive: Facing an estimated 474 GW of interconnection requests, Texas hits pause on data centers
- Governor Abbott’s directive ordering a comprehensive data center audit
- Power Engineering: Data centers drive record surge in GE Vernova orders as turbine slots tighten through 2030
- Utility Dive: Siemens Energy’s gas turbine backlog nears 70 GW
- Structured Finance Association: How Data Center ABS and CMBS Fit in a Broader Financing Ecosystem