The Guarantee Gap

Methodology: AI-assisted evidence infrastructure · Human-directed thesis · Primary-source verified

In August, Nvidia agreed to guarantee up to $105 billion in lease and power payments for a single data-center campus in Ohio.¹

This isn't a $105 billion investment. It's a capped residual-value guarantee: if a defined trigger event occurs, Nvidia may have to cover a contractually specified shortfall, subject to mitigation and recovery provisions, up to the cumulative cap. Which is a fancy way of saying Nvidia is not contributing capital, only the promise to commit capital if the data center goes bust.

On the surface, this seems clear enough. Nvidia has disclosed that the guarantees had no significant reported fair value at the time of its filing,² but it has not disclosed the assumptions — default probability, recovery value, discount rate, the terms of the underlying lease — that produce that conclusion. This means investors are flying somewhat blind with no ability to reconstruct the estimate that turns a $105 billion contractual cap into a present-dollar risk number. In other words, we know the ceiling. We do not know the probability-weighted value of the liability beneath it.

As the AI buildout increasingly finances itself through guarantees, warrants, and conditional backstops, who is supposed to tell us what any of it really costs, and is anyone actually building the tool to do that?

When it comes to speed and scale, the AI race is unprecedented. When it comes to financing, the degree of risk is rapidly compounding, but is it actually following a playbook we've seen before? As the buildout's grid and compute layers become increasingly intertwined and influence the global economy, these questions become more critically load-bearing for the Grid-Silicon Order.

Nvidia's Ohio guarantee wasn't a one-off. Also within 2026, AMD handed OpenAI a warrant for up to 160 million AMD shares, worth roughly 10 percent of the company with vesting tied to both how many chips OpenAI actually deploys and how high AMD's own stock climbs.³ Anthropic signed a $45 billion, six-year compute deal with Nscale.⁴ Meta structured a $27 billion data center vehicle through Blue Owl and kept only a minority stake.⁵ Google backed a smaller AI power developer's lease obligations in exchange for its own warrants.

Each of these five deals landed as a headline number. Not one of them included sufficient public disclosure for outsiders to normalize economic exposure. For guarantees, that means estimating probability-weighted loss and recovery. For warrants, it means valuing contingent dilution and milestone-linked option value. For project vehicles and compute contracts, it means assessing long-dated lease, demand, refinancing, and residual-value risk. These are not the same instrument wearing different labels, which also means they should not be priced as if they were.

The absence of a common public framework is more glaring in light of recent regulatory developments. On August 26, the White House signed an executive order restricting foreign-made transformers and grid equipment from roughly two dozen countries, citing a national-security emergency in the power grid.⁶ The financing problem cannot be separated from the physical build problem: new restrictions on some foreign-made bulk-power equipment may increase sourcing complexity for transformers, inverters, batteries, and related systems. Where project economics depend on delivery dates, interconnection timing, and power availability, those constraints can widen the gap between a headline commitment and a credible estimate of downside exposure. More simply, lack of tools coincides with increasing physical infrastructure uncertainty.

Each deal below is disclosed by its ceiling. None publishes a probability-weighted expected-exposure figure.

Source: Nvidia/SB Energy guarantee and AMD/OpenAI warrant terms, SEC-filing-sourced coverage, Aug 2026; Anthropic/Nscale, Aug 2026; Meta/Blue Owl “Hyperion,” Reuters, Oct 2025, disclosed Q2 2026. Aroko synthesis (Sep 2026).

A conventional infrastructure loan usually has a stated principal balance, payment schedule, and contractual seniority. But the economics can still be complex. In project finance, repayment often depends on construction completion, contracted revenues, operating performance, and recovery values. A guarantee differs because it creates a contingent obligation for the guarantor rather than an upfront advance of cash. Again, Nvidia isn't handing over cash today. It's promising that if OpenAI's project can't cover certain lease and power payments, Nvidia will step in, up to a cap. If nothing goes wrong, Nvidia pays nothing.

A warrant works differently again: closer to a bonus tied to performance. AMD isn't selling OpenAI chips for cash alone; it's also handing over the right to buy AMD stock cheaply, in increasing amounts, the more compute OpenAI actually buys. The vendor's upside gets fused directly to the customer's spending. That is economically contingent, but it is not a credit guarantee — a warrant is best valued through option pricing, milestone-probability weighting, and dilution, not through a default-probability-times-exposure model.

Both instruments make the headline number a poor proxy for the economic cost. But they must be valued differently: a guarantee triggers when something goes south; a warrant triggers when things go well. The analytical machinery exists for both, though not in one universal template — credit models for guarantees, option models for warrants, and project-finance stress tests for leases and special-purpose vehicles. What's missing publicly isn't the method. It's the disclosed deal-level inputs needed to apply that method independently, and a common framework that would let an outsider compare the resulting risks across issuers.

KBRA, one of the rating agencies that tracks AI investment, estimated that gross disclosed contractual commitments and maximum contingent support across the seven biggest AI companies rose from roughly $575 billion at the end of 2024 to about $3.2 trillion by mid-2026.⁷ That's a real number, but it's a gross aggregation of commitments and maximum support, not a common risk measure or a portfolio expected-loss estimate. It tells you the size of the promises. It doesn't tell you what they'd actually cost if called.

Meanwhile, the tone from systemic-risk watchers has sharpened. The Bank for International Settlements flagged the scale of this buildout against prior technology bubbles.⁸ The Federal Reserve's spring stability report and a Chicago Fed note both raised bank exposure to generative-AI financing as a watch item.⁹ In January, a group of senators pushed the Financial Stability Oversight Council to formally study AI-debt risk. While this signals a concern in the aggregate sum, the underlying risk value remains opaque.

The structures across this deal set vary — guarantees, warrants, project joint ventures, compute offtakes — but they share a common feature: a large portion of the economic exposure is contingent, long-dated, and difficult for outside investors to normalize across issuers. Each deal increases the risk of financial contagion.

The closest precedent isn't the dot-com bust. It's Britain's Railway Mania.

At its 1847 peak, railway capital investment reached roughly 7 percent of British national income in a single year,¹⁰ an extraordinary concentration of capital into infrastructure that took years to pay off. The economist Carlota Perez has a name for the arc that pattern follows:¹¹ an installation phase, where capital floods a new technology faster than returns can be verified; a crash, when the mismatch between promised and actual returns finally gets priced; and a deployment phase, where the surviving infrastructure — now cheap, because the mania overbuilt it — quietly becomes the backbone of the next several decades of growth. Britain's railways followed exactly that arc.

The real parallel here isn't that Victorian railway shares were transparent and today's guarantees are intrinsically unpriceable. Publicly traded shares could be priced, and today's contingent instruments can, in theory, be valued too. The parallel is capital formation under radical uncertainty: transformational infrastructure creating genuine long-run value while financing and speculation run ahead of proven near-term cash flows, with the winners unevenly distributed.

What differs today is that modern structures can transfer economic exposure away from straightforward debt and equity since the relevant contract inputs are often private. The same AI counterparties frequently occupy multiple roles in the capital stack, and the exposure may be correlated through the common variable of AI-demand realization. As is already well established in the AI buildout, the guarantor, the equipment seller, and the party who benefits if the stock goes up are frequently the same handful of companies, cross-collateralized against each other. That concentration is actually the more notable difference from the railway era.

The railway mania's risk hid itself without anyone having to invent a new financial instrument. While the financing strategies used aren't novel per se, the scale of investment is. Still, estimates put current AI infrastructure capital spending somewhere between roughly 2 and 5 percent of U.S. GDP depending on what you count as "AI-related." In other words, while the potential risk is expanding, there is still very likely room for the AI buildout to run if history teaches us anything. And most likely, we will see an acceleration of financing deals like Nvidia's for the foreseeable future.

I expect the guarantee gap to widen, not narrow, through the end of 2027. Here, the gap is not a claim that no pricing tool exists — it is the distance between gross disclosed maximum contingent exposure and what a diligent outsider can actually estimate as probability-weighted exposure given what's currently public.

Two things have to remain true for the gap to keep widening through 2027. First, that the volume of capped, conditional guarantee- and warrant-structured AI infrastructure deals keeps growing quarter over quarter. That is a high likelihood given we are still short infrastructure and long frontier model capabilities. I count nine independent, primary-sourced deals inside a single recent quarter, plus a demand base (Nvidia's own data-center revenue, up over 100 percent year over year) that gives no sign of slowing. Second, that no rating agency or systemic-risk regulator publishes a generalizable, cross-issuer framework for translating these instruments' ceilings into comparable expected-exposure figures, backed by enough disclosed deal-level information for outsiders to apply it independently. I put that also at high likelihood because both legs of this trend reinforce each other: the faster deal volume grows, the less incentive any guarantor has to be the first to publish the assumptions that would let an outsider price their own exposure.

What is difficult to reconcile is Nvidia's going from a $30 billion direct investment into OpenAI its own CEO called possibly his last to a $105 billion Ohio guarantee months later. This can be read two ways:

Both are plausible and both play out in the very deal that opens this letter. There is no published number that would let anyone adjudicate which reading is right.

Every infrastructure gold rush finances itself with whatever instrument the moment allows, and the bill always arrives somewhere other than where the promise was made.

Railway shareholders found that out when branch lines built on optimistic traffic projections went bankrupt in the 1840s crash, dragging down banks that had never laid a single rail but had lent against railway paper as collateral. Telecom bondholders found it out in 2001, when "dark fiber" that had been capitalized as an asset turned out to be a stranded one.

The pattern isn't that these booms were fraudulent. It's that the instrument financing the boom is always a step or two removed from the people who eventually absorb the loss, and the market doesn't find out who that is until something breaks.

The AI buildout is running the same play with a new instrument. Guarantees and warrants let Nvidia, AMD, and their peers finance a trillion-dollar infrastructure build without it showing up as debt or capex on their own books today. That's a real innovation in how you fund a buildout this size this fast, and it may well be a smarter structure than the plain equity and lending that financed the last two infrastructure booms.

But it also means the risk is being distributed through channels that are new enough that the public infrastructure to track where it ends up has not yet caught up. History's version of that gap was geographic: a shareholder in Edinburgh had no way to know if the branch line they bought into would ever carry a paying passenger. Today's version is structural: a pension fund holding AI-adjacent credit three funds removed has no way to know if the guarantee underneath it will ever get called, because the entity that wrote it hasn't said what would trigger it.

That's the "so what" for anyone whose capital touches this buildout, directly or more tangentially: the risk is real, it is probably manageable, and it is currently invisible to almost everyone except the small number of companies writing these instruments to each other. It becomes visible at the worst possible moment, such as when a guarantee is actually called, or when audit and accounting rules force it onto a balance sheet. Contrast this with being priced in gradually, the way a bond spread or a public loss estimate would allow.

None of this argues that the AI buildout is a bubble or that these instruments are unsound. Railways and fiber both were built, and perhaps more relevant, both became indispensable infrastructure once the excess was burned off. The lesson from those cycles isn't "don't finance the boom this way", it's that the entities who came out ahead were the ones who understood which piece of the risk they were actually holding before the market forced them to find out.

This forecast breaks on either of two conditions. Either one is enough to make the legibility claim wrong, though the strongest version of that claim requires both.

It breaks if a named rating agency or systemic-risk regulator publishes a generalizable expected-loss or probability-weighted-exposure methodology for capped AI-infrastructure vendor guarantees, applied to at least two distinct issuers, before the end of 2027. A single disclosure by one guarantor about its own deal doesn't count: that would be self-interested and narrow, not the generalizable tool the gap is actually missing.

It also breaks if a majority of the next several major tracked deals shift toward transparently priced structures: publicly disclosed default-probability pricing, insurance-style premium disclosure, securitization with published loss tranches, rather than capped, conditional, undisclosed-pricing guarantees and warrants.

I expect deal volume to keep outrunning methodology, not the other way around, because the incentive runs the wrong direction for voluntary disclosure and the technical work of standardizing a new instrument class typically lags the instrument itself by years, not quarters. At this stage in the AI buildout, the opportunity far outweighs the risk, and capital will keep chasing the prize: none of the hyperscalers writing these guarantees believe they can afford to be the lab that finishes second in frontier capability, and that competitive pressure is a stronger force right now than the incentive to slow down and disclose.

About Aroko: Aroko provides strategic advisory and capital allocation intelligence at the intersection of energy transition, technology infrastructure, and geopolitical risk. Our analytical process combines proprietary evidence infrastructure with human-directed thesis formation. Every keystone claim is verified against primary sources, and all editorial judgment and capital allocation framing is conducted by Aroko’s team. The Letter is published biweekly for institutional allocators.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, legal, or tax advice. The opinions expressed regarding macro trends and infrastructure investments are solely those of the authors. Past performance does not guarantee future results. Readers should consult with a qualified financial professional before making any investment decisions.