New Article: Wiring Capital to Compute
America possesses more capital than any other nation in history and the most compute demand in the world. It is still waiting for a market to wire them together.
This article by Daniel Painter was published at Palladium Magazine on July 20, 2026.
Over a thousand acres of land in Abilene, Texas, lie flattened with graded earth and poured concrete. The foundation pits are filled with conduit bundles and switchgear, and in some quadrants, fully operational substations have been erected. However, other quadrants sit empty, waiting on a schedule that keeps sliding to the right. Although the supercomputer intended for the site is humming with activity, what registers to a visitor is that the campus is only half-alive.
Abilene was the intended flagship of the Stargate consortium. Announced from the White House days after President Trump’s 2025 inauguration, SoftBank’s Masayoshi Son, Oracle’s Larry Ellison, and OpenAI’s Sam Altman stood beside the president as he declared the project’s ambition to pour $500 billion into a combined data center and supercomputer over a four-year horizon. While this signaled Washington’s blessing on a scale once reserved for projects like the Apollo program, those had the federal government as their underwriter and customer. Stargate would have neither. Private capital alone was set to accelerate machine intelligence through GPUs, land, and power, with the White House only lending its podium.
Today, the grounds of Stargate’s flagship site in Abilene suggest the fruits of this effort. Through the first half of 2026, stalled negotiations shrank the question from when the site would finish to how much power it would ever draw. To the initial 1.2 gigawatts planned—enough power to supply roughly a million homes—Abilene’s Stargate was slated to add nearly another gigawatt. Papers were drawn up and financing was secured, but the deal did not actualize. Originally projected to finish in March 2026, the site is now expected to be fully energized as late as mid next year, and operational and financial issues have canceled its ambitious expansion.
However, it is a mistake to interpret this by simply saying that “Stargate failed,” and the way in which that reading is wrong holds a valuable lesson for the future of the American AI buildout. The main campus still serves OpenAI compute. But the broader buildout that served to scale the multi-gigawatt arrangement between OpenAI, Oracle, and Crusoe collapsed. Abilene leaned heavily on two counterparties and their capacity to shoulder compute infrastructure risk. Stargate attempted to separate out risk according to whoever could best bear each specific burden, but when OpenAI’s internal cash-flow forecasts moved far enough, the deal’s structure could not absorb the swing. This led to OpenAI and Oracle abandoning nearly an extra gigawatt of capacity next to an existing supercomputer.
Is Stargate a failure on a national scale? Not yet. But this example is emblematic of a larger failure that looms on the horizon. The American AI buildout rests unsustainably on the same few corporate treasuries, and it is neither standardized nor repeatable in a manner that lets American capital markets fund frontier compute. This is unlike the way assets ranging from mortgages to traditional power plants are funded. Instead, urgency and a lack of an existing playbook mean capital comes in through side doors.
The question of interest is not whether America has money. By one measure, the United States holds roughly 40 percent of the world’s equity and a comparable share of fixed income. This is the largest concentration of hungry capital ever assembled. So what happens when the conversion machinery decays exactly as we face the buildout that will test it the hardest?
American Capital Can’t Reach Compute
America’s wager has never been laissez-faire in the strict sense. It is that capital markets, properly institutionalized, turn private capital into public goods. The lazy version of the story is that America built its great physical infrastructure by getting out of the way, but this is an incorrect reading.
Historically, America has developed the institutions that let private capital grasp and fund public infrastructure. In the nineteenth century, transcontinental railways were stood up and operated via land grants and public charters, but institutions and assets were developed in tandem. No deep market for industrial securities existed when the first rail promoters, long before Vanderbilt, started laying track. Rail bonds and shares were the instruments on which early American securities markets cut their teeth. Standardized paper, ratings, reporting, and, later, exchanges were developed while tracks were being laid. The funding apparatus that turned a distant, disparate, and initially underfunded project into something that counterparties could confidently fund at arm’s length was called into being by the infrastructure itself.
Thanks to the repetition of this dynamic, the United States possesses the deepest and most liquid capital markets on the Earth across pension funds, insurers, and private credit. The problem is that the instruments and standards that would make AI infrastructure projects bankable do not yet exist. A large portion of this liquid capital is hunting for long-duration and high-demand assets embodied in AI infrastructure. Capital is abundant, but bankable assets are not.
Corporate finance provides funds exclusively based on the balance sheet and expected financials of a company, providing a useful vehicle to raise debt for asset expansion. Project finance, by contrast, funds a predictable, contracted, and long-lived asset, especially one with high capital needs or a public-goods character, such as toll roads, pipelines, and traditional power plants. Project finance does this by walling the asset off into its own entity and lending against its cash flows through non-recourse or limited-recourse debt: because the lenders’ claim runs only to what the asset alone will generate, future revenues serve as both the basis for the loan and the lenders’ only real recourse if things go wrong.
The compute buildout is awkwardly positioned between corporate and project finance because it has the risk profile of the former and the scale of the latter, adding billions of dollars of debt to a company’s balance sheet. As such, there is currently no capital-conversion machinery—no templates—that fit this asset class cleanly. There are a few aspects of its asset dynamics that explain why it is so exotic compared to past infrastructure buildouts.
A compute infrastructure asset is actually several assets with differing lifespans stacked on top of each other. The shell, land, substation, and cooling plants can last decades, though even this “durable” layer is repriced by the chip cycle, as substation sizings and cooling standards designed around one hardware generation lose value against the next. The bulk of the volatile value lies in GPUs, which possess a short, uncertain economic life. GPUs and their associated parts, like memory or storage, are repriced not only by wear and conventional depreciation schedules, as ordinary server fleets are, but by product cycles so rapid that each chip generation reprices the last.
It is difficult to fund data centers as one long-dated asset whose most valuable layer obeys a much shorter, stochastic clock. It would be much more efficient to slice the supercomputer by layer, with long debt on the shell and more aggressive amortization against the silicon. This already exists in practice to an extent, but it is not standardized or predictable in the market as it stands. Additionally, the expansive needs and complexity of newer deployments often demand codevelopment—and thus parallel financing—of all layers.
For the buyer, mature infrastructure financing prices risk by distributing it over many uncorrelated customers, as this shields the project from a single offtaker tanking the asset’s viability. Today, frontier compute inverts this. Projects like Stargate are built for the use of a single, dominant offtaker, such as OpenAI for the Abilene expansion. In this case, the buyer was young, conventionally unprofitable, and prone to wildly revising its internal demand forecasts. When OpenAI’s forecast moved, no instrument stood ready to absorb this change. There was no termination pricing to compensate the asset for the demand revision, no resale market where a departing tenant’s capacity could clear, and no demand insurance to bridge the gap.
Standardized compute offtake contracts that tranche and distribute the various risks, robust rating methodologies purpose-built for the challenges of frontier compute, and securitization channels for assets under construction do not exist. Compute has fragments of similar instruments employed improvisationally, deal by deal. The closest existing template is the financing of a single power plant against a long-term, stable buyer like a municipality, but even that analogy is problematic. Compute is inherently more difficult than financing power infrastructure due to aggressive hardware obsolescence, buyer concentration, unstable demand shocks, and export controls. No infrastructure model has had to face this level of complexity to date, and no party is prepared to justify or explain the total cost recovery of AI buildouts as they readily can for a power plant.
According to Morgan Stanley, the global data center buildout should run around $2.9 trillion through 2028. Hyperscalers, for all their cash, can fund only half of this even deploying their entire balance sheets. Some $1.5 trillion must come, reliably and repeatedly, from external capital. Each deal is negotiated and priced from scratch. If U.S. markets are to fund AI infrastructure at a scale beyond any prior asset class, the process must first be standardized.
The Symptoms
A project meant to withstand massive operational and demand uncertainty must be bankable in the first place. A highly developed, mature market can price this uncertainty and push through a lot of complexity. Increasingly, compute capital has shifted away from straightforward spending through balance sheets and towards a gaggle of other experimental structures for compute—joint ventures, long-term leases, private credit, offtake mixes, securitized lease cash flows sold to institutional investors, and more. This is evidence that the market is hungry for ways to turn capital into resilient compute assets. But most of these structures only serve to tighten the link between hyperscalers, non-bank lenders, and insurers, rather than inject new blood.
Stargate’s Abilene deployment was structured as a joint venture. The site developer and operator, Crusoe, raised billions in construction debt—a $7.1 billion loan from a J.P. Morgan-led consortium for the campus’s second phase, atop an earlier $2.3 billion facility—secured against Oracle’s eventual ownership of the cluster, with OpenAI as the planned lessee in multibillion-dollar tranches. This was financeable almost solely on faith in OpenAI’s single tenancy, abstracted under several layers.
On paper, these capital structures appear sophisticated, but in practice they are rather fragile due to how they legally separate risks that remain economically correlated. A demand forecast revision from a tenant cuts straight through every layer of a joint venture. Companies like CoreWeave have emerged to build compute sans the trillion-dollar balance sheets of the hyperscalers and they have raised large sums of capital in forms far more preferable to the aforementioned structures, and against a broader base of offtakers.
But there is still something sour about the source of the money. The year before CoreWeave’s IPO, roughly 62 percent of its revenue came from Microsoft, while its second-largest customer accounted for another 15. Later deals, including many large OpenAI commitments, broadened their customer base but did not meaningfully dilute against this concentration. The development of American compute is effectively coordinated by a select group of hyperscaler treasuries and lab demand forecasts.
The underlying risk from this arrangement is never dispersed. For Morgan Stanley’s $2.9 trillion projected investment into AI infrastructure, roughly half must come from outside hyperscalers, yet nearly every external dollar today enters through a structure whose credit resolves back to this same handful of names.
When capital flows between a few giants who self-finance, some of this capital then spills over to fiscally thinner companies like neoclouds, who are incredibly dependent on those same giants. This is extremely unlike the capital-conversion machinery that set America apart in the past. One does not need to be a doomer to notice that it eerily resembles past moments in financial history, such as the 2008 financial crisis, when instruments appeared to diversify risk while actually concentrating it behind the scenes.
The malady of capital only reaching deployment through a handful of correlated balance sheets, and the largest technology companies comprising roughly a third of the S&P 500, means that pension funds and retail portfolios carry concentrated compute exposure whether they like it or not. But this deep pool of capital also has no proper channel to invest in the buildout. The hyperscalers can take buildout risk on the chin in a way a rules-bound pension fund cannot. There is no instrument that lets them underwrite a data center directly, price its own tranche of the risk, and be compensated for bearing it. But today, it holds the risk anyway.
China, the U.S.’s sole competitor in the AI race, is not waiting for the market to discover the true market price of compute and its returns. By coordinating its compute infrastructure finance via its own state apparatus, Beijing is converting national resources into compute at an accelerating pace.
China is Eating the Gap
As America’s compute market struggles to be born, China seeks to coordinate a centralized buildout. Beijing will spend roughly $295 billion over five years for a grid of nationalized AI data centers. More interestingly, China plans to produce 80 percent of its AI chips domestically by 2028.
This should be seen in light of the Eastern Data Western Computing (EDWC) initiative launched in 2022. Beijing has spent years developing and organizing the country’s compute resources around eight national hubs and ten data center clusters, intending to match eastern Chinese compute demand to the abundant power and land of western China, where electricity prices run 40 percent below those in the east.
China has not forgone the counterparts to America’s private institutions, but instead runs them in lockstep with centralized planning. State banks and state-owned enterprises work together with provincial governments to meet procurement mandates, mediating pricing and economics through compute vouchers and credits—in Zhengzhou, operators have gone as far as giving compute vouchers away to conjure demand for capacity built ahead of any customer.
The nail in the coffin for Abilene came when OpenAI, its internal forecasts revised, found that the expansion no longer fit its needs. The Chinese version of this process does not balance on such tenuous dealmaking because the parties are not independent. A provincial government allocates land according to plan while a state-owned or state-directed bank disburses the funds at CCP discretion. The project is underwritten against the state’s priorities and not against the cash flows of the corporate entities involved. The system is effective at getting infrastructure into the ground and powered on.
There are payoffs to this approach, but also costs. The high-speed rail buildout, which laid more track in a decade and a half than the rest of the world combined, was in many ways a dress rehearsal for the AI push. It included the same provincial land allocation, the same state-bank credit, the same underwriting against national priorities rather than project cash flows. However, this rehearsal also revealed a different set of failure modes. Many lines were built to political priorities as opposed to justifiable ridership, with debt buried across myriad local-government financing vehicles. Compute will prove a harder problem than rail: the technology obsolesces faster, and its demand curve is far less understood.
Under the EDWC, more than 500 data center projects were announced. Only 150 of these were completed by the end of 2024. Since then, local publications have reported that roughly 80 percent of the new capacity sits idle, and more than 100 projects have been scrapped outright. Provinces that were not designated as hubs built capacity anyway to chase political standing and subsidies. Additionally, EDWC’s poster child, Guizhou, remains near the bottom of the provincial GDP rankings. It is saddled with heavy provincial debts and is facing major corruption issues in the big-data industry. Beijing’s response to this has been to institute utilization minimums, bans on locally funded and small-scale facilities, and purchase mandates set before construction begins. The state is manually reconstructing the price signals and underwriting standards of a functioning market.
The most binding problem facing China’s buildout, however, is its goal of domestically producing 80 percent of its chips. Designing a high-performance chip and effectively producing it at scale are two different beasts. Should Huawei or the other domestic producers fail to meet this mandate, export controls on American chips will bite hard. Additionally, the question of chip efficiency compounds this issue. The EDWC hubs have electricity in abundance, but domestic accelerators deliver less performance per watt, so abundant electricity buys less intelligence per gigawatt.
China’s approach lends it a faster theoretical time-to-power for many facilities, but has exposed it to a completely different and potentially more severe set of failure modes. Central planning may yield better short-term progress, but the conversion machinery of a mature market would allow the United States to maintain its hardware and frontier-model edge. However, with every month that American markets fail to convert capital into buildout progress, the chorus calling for a nationalized buildout grows—even though America’s best strength has always been its savvy in using markets to address strategic concerns.
The Strategic Test
The solution is a bankable buildout, not a nationalized one. Pouring more capital onto the fire, as the U.S. has done thus far, will not suffice. Rather, the physical buildout needs to be made repeatable, reliable, and resilient.
Looking to the rail era once more, the Panic of 1873 began in railroad finance—Jay Cooke’s failure took the market down with it—but the crash did not end the buildout. The financial technologies that the boom had forged, such as standardized bonds, ratings, and disclosure, carried the next wave of construction on sounder terms, and became load-bearing financial mechanisms for later eras. A financial panic caused by compute would not mean the buildout was a mistake, the same way the dot-com bubble does not invalidate internet shopping or telecommunications.
That being said, the United States of 1873 was diversified and fast-growing enough to absorb this blow and continue forward. A compute panic today would not stay confined to a single site in Texas. At one third of the S&P 500, index portfolios and pension funds ride on hyperscalers by default, meaning a vast amount of the economy carries compute exposure that was never properly priced or underwritten. In the worst case, were the bubble to inflate further, a pop could even reach bondholders and index-trackers. This is all the more reason to develop the market machinery today so that proper pricing and tranching of risk can harden the broader economy against systemic events.
Every few generations, some cohort has stood present for what was, to that point, the most important buildout in history. However, the nature of AI as an asset stands apart from its predecessors. It has the capital intensity of the electrical grid, the depreciation clock of consumer electronics, and the strategic stakes of aerospace. Fiber-optic cable laid before the dot-com bust stayed useful for decades, but an AI infrastructure buildout done with the wrong equipment gets no such second life.
The output here is intelligence: and it is intelligence that feeds directly into military, industrial, and scientific advantage. This is why both Beijing and Washington treat it as a strategic priority, and why the problem is not just a single gigawatt in Texas. If the deepest capital markets ever assembled cannot fund this buildout, then we must ask whether the society that developed them is truly equipped to wield the future.
Daniel Painter is the COO of Squaretower, a compute economics firm based out of San Francisco. You can follow him at @NecessaryForm.