The Albatross’s Bill: Data Center Debt
“And every tongue, through utter drought,
Was withered at the root;
We could not speak, no more than if
We had been choked with soot.
Ah! well a-day! What evil looks
Had I from old and young!
Instead of the cross, the Albatross
About my neck was hung.”
— Samuel Taylor Coleridge, The Rime of the Ancient Mariner
When the Bird Lands
The collapse will not arrive as a mushroom cloud. It will arrive as a missed rent check on a building that is still a hole in the ground.
A lab that has never posted a profit misses a lease payment on a campus that doesn’t exist yet. The landlord is a new public company whose stock gets its value from that lease. The landlord calls the chipmaker, who wrote a residual-value guarantee to cover for the missing rent. The chipmaker reminds the landlord that the guarantee was voided when the public company earned a credit rating upgrade.
Lawyers discover what the S-1 already admitted. The board that must enforce the lease sits next to the tenant that also controls the landlord. Bondholders learn that “guarantee” was a bit of an overstatement.
An albatross is a very, very big bird. Its weight drags down the body and the spirit.
Somewhere, a pension fund for midwestern teachers begins to sink.
Not fraud in the Enron sense. A stack of future promises that behave like bad debt the moment the tenant cannot pay - and where a breakdown in the underlying business of next-gen data processing would mean every “guarantor” going down together.
The Claim
In August 2026, the Wall Street Journal weighed the bird: about $3 trillion.
Future obligations, loaded onto the largest technology firms as the cost of securing future land, power, concrete, and chips. Only about $600 billion of that load has reached GAAP balance sheets. The rest is multi-year leases and purchase agreements - not on the balance sheet, but right there in the footnotes of documents like SB Energy’s new S-1.
The size isn’t the issue, exactly. Revenue is always a good answer to debt. The problem is duration, circularity, and businesses that have the same names as former cash engines - but have shifted their business model to setting dollars on fire.
Queries cost money. Chips age, and quickly. Tenants concentrate. Credit has already begun to flinch.
If those obligations convert on schedule and the models do not throw off cash, the firms that have carried U.S. equity returns become black holes. The panic would reach far beyond “AI,” beyond even tech. Because AI and tech are the entire market now.
Executive Summary
Public hyperscalers and private labs have pulled forward a buildout that turns high-margin software businesses into capital-intensive utilities. But the revenue to pay off that commitment is still speculative.
Reported free cash flow for Google, Microsoft, Amazon, Oracle, and Meta is on a path from an aggregate $200 billion profit in 2024 toward a projected $125 billion loss in 2027. That reversal is already visible. Off-book leases and special-purpose vehicles make the true claim on future cash larger still.
Even in an optimistic scenario, it seems unlikely that all of this debt will be productive. There will be winners and losers. But the belief that LLM technology will become even more ubiquitous and useful than it is now creates a prisoner’s dilemma for the most critical industry in the U.S. economy. Software firms must lock compute capacity now or wind up completely irrelevant. Competitive pressure means they’re each incentivized to get as much compute as possible - to build as if they will be the only survivor.
The financial model behind all of this is hilariously circular. Nvidia helps customers buy Nvidia. OpenAI’s forward leases and purchase agreements have been tallied near $665 billion while the company projects tens of billions in losses through 2028. SB Energy is being prepared for public markets on the back of a $430 billion OpenAI lease and a $105 billion Nvidia residual guarantee that OpenAI in turn indemnifies. Meta’s Hyperion project in Louisiana is funded through Beignet Investor LLC, a partnership with Blue Owl, with lease and value guarantees that analysts treat as near-debt.
Accounting watchdog Francine McKenna is right that this is not Enron. Future obligations sit off the balance sheet - that’s fine. However, the cash still has to show up at some point. If demand, duration, or counterparties fail, the obligations hang. Because a handful of these firms have been the ballast of U.S. indices, the equity-market question is not only credit. It is what happens to returns when the cash engines stop compounding.
Don’t think this matters? Go look at US equity returns with Google, Microsoft, Meta, Amazon, and Nvidia removed from the index over the past 10 years.
Two Faiths
There are two ways to understand why the largest technology firms (excluding Apple) took a model that had catapulted their stocks to the moon and lit it on fire. Both are operating at once. Pretending only one is in force is how analysts stay invited to mixers.
The first framing is the one that can be said out loud on a credit call. Artificial intelligence is treated as a winner-take-all market. Compute, power, and land are scarce. The firm that locks them early compounds an advantage that later money cannot buy. In that game, a massive forward commitment is individually rational even if the industry as a whole cannot earn its cost of capital.
The second framing is the sincere religion of a large share of lab staff and leadership. The spend does not need to clear a hurdle rate because they are building God. The Jubilee writes the debt to zero, ends the existing order, and renders the coupon a costume. What is a few trillion against the end of scarcity?
Markets were not really meant to price rapture. The problem is that the same capex schedule is being justified by both stories at the same time, and that the second one rests upon beliefs that are not testable, verifiable, or commercial, but rather religious in nature.
These ideas have been simmering for decades, going back to Ray Kurzweil’s 2005 coinage of “The Singularity.” The elaboration on those ideas, and the fervency with which they’re discussed by developers, have reached truly Talmudic levels. (You can read about some of the Singularity’s most extreme zealots, such as the Manson-like Zizians, in David’s book.)
The operational logic is narrower and more testable. First, that this generation of large language models creates durable demand for research help, images, and above all code. Usefulness is no longer the interesting argument - LLMs may still be philosophically probabilistic, but they’re good enough for government work.
Another assumption is that demand continues to outrun installed compute. That seems like a safe bet in the medium term - Anthropic and others are already throttling at peak hours. But much of that demand is still effectively subsidized - costs do seem to be coming down, but Anthropic’s brief experiment with per-token billing was catastrophic.
That means pricing changes could whipsaw demand, shrinking or stalling capacity demand. The compounding problem is that data centers are depreciation machines: every second a paid query isn’t going through those racks of chips, money is melting.
And data center builds take many years: if the market suddenly discovers there’s enough compute, there will still be plenty of debt-burdened concrete in the pipeline. Under a bear scenario, they would be completed, then sit underutilized.
The final assumption of this spending spree - and perhaps the most problematic - is that capacity will lead to contracts and customers, and that the most capacity will net the most business. That might be true for enterprises and governments with compliance mandates, but it’s already clear that there are few “moats” in the LLM business - Chinese open models, some of them frontier-grade, can be run locally. There’s actually a case that’s even more secure than a U.S.-based cloud. Power users, individual or corporate, may not join a members-only club that charges them extra for brand comfort.
These tenets produced a scramble for dirt, turbines, construction slots, and chips.
The third article of faith should stop a serious allocator. Many of the people signing the leases talk as if statistical models of human text are about to become conscious minds many times smarter than their makers.
That claim has been non-falsifiable since ELIZA in the 1960s. It still moves billions, even though you can technically build the technology in Age of Empires 2 using goats. If we think LLMs are conscious, then so is a video game.
You Keep Using that Word - “Utility.”
The public companies’ reported spending is already dramatic. Meta has to monetize a product that costs money every time a user asks a question. That is not the zero-marginal-cost feed that built the last PE multiple. Open-weight models such as Kimi and DeepSeek make the American high-cost stack harder to defend by the day, nipping at the heels of the builders even as they struggle to race ahead.
Then there is the part that does show up on the 10-K. Special-purpose vehicles take the construction loan. Joint ventures take the land. Lease commitments stretch years, in at least one Alphabet case toward 2056. Off-balance-sheet exposures also include promises to buy another firm’s stock later, or to backstop a lease for a tenant that is not you.
McKenna’s objection to the louder coverage is precise. These items sit off the balance sheet because they are future obligations, not hidden current debt. “This is not Enron. The numbers look really big, but these are the rules.”
The rules have never protected people from risk, however. They just tell you where to put the numbers. The risk is that the obligations are nearly as hard as debt once the data center is built and the chips are plugged in. Investors who treat “off-balance-sheet” as “not real” are reading the footnote as a lullaby. In reality, they should hear a ticking sound.
The Journal’s $2.42 trillion of off-balance-sheet commitments have mixed duration. Some die in a few years. Some crawl into mid-century. Add the on-book debt of hyperscalers and operators and the working total lands in the neighborhood of $4 trillion.
The Nvidia-OpenAI Circle
Circularity is a defining feature of the data center funding trade. Manufacturers funding their own customers. Labs paying leases to entities they own 20% of. Depreciating chips collateralized to fund infrastructure for more chips.
Another word for circular financing is “terrifying one-way risk.”
Nvidia has worked with large banks on a lending facility on the order of $500 billion so that customers can buy Nvidia chips. The vendor finances the demand that becomes its revenue. The customer’s “commitment” becomes the vendor’s growth story. Neither income statement has to show the full loop - just promise that it will be closed.
OpenAI has been the most aggressive of the pre-IPO startup labs. One tally of its forward lease and purchase agreements runs to $665 billion. At the same time, internal figures put 2025 losses near $38 billion, with the company’s own projections extending losses toward $78 billion through 2028. If the math does not change, and if public markets do not absorb an offering large enough to refill the tank, those leases break. The entities holding the paper take the hit. In more than one structure, that includes Nvidia.
Credit has started to say no. SoftBank’s attempt to pledge OpenAI equity as collateral was rejected. That is not a theory about 2030. That is a present-tense price on opacity.
And OpenAI is not unique here. The revenue must flow – all the time, everywhere – or it’s not going to catch up to the obligations.
The End of Tech Equity?
U.S. equity returns in this cycle were not a broad harvest. A thin group of mega-cap technology names produced a disproportionate share of S&P 500 and Nasdaq gains. Those firms funded buybacks and supported multiples with free cash flow from the model that made Silicon Valley Great: software margins, light physical plant, money that arrived before the next foundation had to be poured.
A swing from roughly $200 billion of combined free cash flow profit to a projected $125 billion combined loss is not a rounding error in that machine. The cash isn’t going to buybacks anymore: it’s going to substations and GPU racks. Dividends don’t seem likely to replace it. Valuation math that treated these balance sheets as perpetual cash compounders has to reprice duration, capex, and residual value.
There are few historical precedents for old, established entities like Google and Facebook upending their strategy so fundamentally, on the basis of a technology that hasn’t been proven yet. One notable exception was Facebook’s pivot to “the Metaverse.” We should check in on how that’s going.
Index investors do not hold “the market.” They hold a concentrated bet that these names remain cash-flow positive at scale. If the five hyperscalers and the chip vendor that feeds them become sinks rather than sources, passive flows keep buying the same names as their cash yield dies.
The drawdown does not require a default. Multiple compression on the ballast is enough to take the index with it. That is the hidden transmission into household wealth. Pensions and target-date funds are long this concentration whether they know the lease schedule or not.
A credit event at a neocloud would be loud. A silent five-year collapse in mega-cap free cash flow would be larger.
SB Energy: Big Contracts. No Buildings.
SB Energy is the latest emblem of this dynamic. SoftBank, OpenAI, and Nvidia sit behind a data-center and power developer that announced plans to list. Bank talk put the valuation near $50 billion. The company has not completed or operated a data center. 2026 power revenue exists. The equity story is future lease revenue from OpenAI.
The contracted figure attached to that tenant is about $430 billion. OpenAI does not have the cash as of today. SB Energy investors are offered a $105 billion residual-value guarantee from Nvidia on the Pike County, Ohio campus. OpenAI indemnifies Nvidia against paying that guarantee - and it’s void as soon as OpenAI gets a good credit rating. Risk is being remixed, transported, and shared - but it can’t be gotten rid of.
The S-1 language is not coy. The board’s will to enforce a lease against a tenant that is also a controlling shareholder may be “constrained in ways that would not apply with respect to an unaffiliated tenant.” That sentence is the collapse scene written in advance. The risk is spelled out: do you expect them to blow themselves up?
The Tenant Stack and the Residual Years
CoreWeave is the public version of the same shape. Expected revenue in the billions depends on a handful of long-term contracts with OpenAI, Meta, and Microsoft, each with pre-set performance conditions. The operator also carries its own debt. If a tenant can’t find enough customers, the revenue that was booked as good now becomes an open question.
Meta’s Hyperion project in Louisiana sits in Beignet Investor LLC. The vehicle was set up to issue on the order of $27 billion of bonds to build a data center that Meta plans and intends to use. But those bonds do not sit on Meta’s balance sheet.
What Meta gives the holders are lease and value guarantees. Analysts who have read the paper say those promises are nearly as tight as debt. Four years of direct obligation can be followed by sixteen years of residual-value obligation. The albatross may not be on the balance sheet, but it is there no less.
This is the pattern. A thin set of tenants. A vehicle that looks independent. A guarantee that’s comforting until you realize everyone in the deal shares an office.
The GPU Peg
Under the leases sits a physical question that finance keeps trying to treat as settled. Can you collateralize a depreciating asset when the issuer is motivated to make last year’s card obsolete? That’s Nvidia’s business. Lenders are being asked to treat the current chips as durable security even as the company creating them has every incentive to do the exact opposite.
There is a counter-argument. Some operators say processor-heavy data centers last longer than the old storage buildouts, and that infrastructure life is being understated. Amazon, in 2025, shortened the financial life it assigns to GPUs.
Those two facts can both be true. Hardware can run and still be worthless as collateral if the only buyer wants a newer card. It can be equally true that hardware is not excellent at all tasks, and so the relative shifts in demand can create or destroy fortunes held in silicon.
Francine McKenna notes that the better contracts contain adjustment schedules. Lenders can revalue collateral against new estimates. That is competent lawyering. It is not a source of new cash. A mark-to-model clause does not print watts or tokens.
The Ground Game
Even a clean model of demand still has to touch dirt and a grid. And at the moment, there are increasing obstacles to getting shovels in.
New York’s queue has been described as holding about 12 gigawatts of data-center energy demand, on the order of Portugal’s entire output - and Governor Kathy Hochul just implemented a one-year pause on new center approvals. Coverage in 2026 had fourteen states besides New York considering bans. Blackstone’s QTS killed a large project in Prince William County, Virginia. Virginia, which already hosts the densest cluster on earth, added a tax on the facilities’ energy use.
This is not a morality play - whether these activists are “right” is close to irrelevant. It is duration risk meeting misinformation meeting pitchforks meeting the American unwillingness to build, well, anything. A twenty-year lease that cannot get interconnect is a twenty-year claim on a building that never draws power. The ideology of the race assumes the concrete pours. Politics can stop the pour without reading the indenture, or, to be blunt, even being able to read.
There’s also the AI firms’ own seeming allergy to acting normal. Years’ worth of threats to take away everyone’s job, and maybe blow up the Earth in the process, have turned out to be kind of bad for public opinion of AI! Oh yeah, and as we write this, there’s the fresh allegation that OpenAI has been using Codex queries to steal research from scientists.
None of that seems exactly great for achieving those fat revenue goals.
Why This Can Blow
Assume, generously, that the industry as a whole becomes profitable. That still doesn’t mean the profits cover everyone’s car note.
The obvious place to look for comparable revenue is enterprise software. The guys Citrini thinks will be dead by the side of the road sooner rather than later, displaced by AI that can do everything. The five largest software firms – Microsoft, Oracle, SAP, Salesforce, Adobe – do on the order of $600 billion in annual revenue combined, with double-digit net margins that software earned because the next copy was nearly free.
If the entire AI stack merely matched that revenue, it would take years of full capture to cover commitments already signed. The comparison is worse than the arithmetic. Each query has a real marginal cost in power and silicon. Traditional software did not. Meanwhile, open-weight models push price toward the cost of running the weights, not toward the multiple the labs need.
The blowup does not require every name to fail. It requires one large tenant to miss, one residual guarantee to be tested, and one public vehicle whose only asset is that tenant to reprice. From there the chain breaks. Nvidia’s growth story is other people’s leverage. Hyperscaler cash flow is already turning. Indices are long both.
Would there be a bailout? Lab executives have learned the music of national security. That is a political option, not a covenant. Larry Fink’s 2026 line was that the gains accrue to holders of capital.
But the counterfactual is much more shadowy: If the buildout goes bad, we can’t be sure who takes the fall.
The Bird
Coleridge’s mariner killed a bird that had been guiding the ship. The crew hung it on him so he would not forget. The data-center complex did not shoot a guide bird: it enshrined the omen, special-purpose vehicle by special-purpose vehicle, and called the prediction a sure thing.
If data centers earn their keep, this essay is a period piece about a scare that passed. If they do not, the survivors will spend a decade climbing out of the wreckage of a silicon Tower of Babel. China, the boogeyman that partly motivated the frenzy, will have models 97% as good for 4% of the investment.
The albatross is not the $3 trillion.
The albatross is a promise that cannot be discounted when the tenant is also the landlord, the guarantor, and the story that held up the index. Put simply: the stakes are the entire market.