Big Tech’s $3 Trillion AI Commitments Sit Off Balance Sheet

The reported capital-expenditure figures for AI infrastructure do not capture the full scale of the commitments that large technology companies have made. The discussion cites an estimate that nine top tech companies hold roughly $3 trillion in off-balance-sheet commitments, mostly related to AI—obligations growing faster than the approximately $600 billion in conventional capex they reported over the preceding year.

Those commitments are also described as roughly triple the companies’ outstanding lease liabilities and long-term borrowings. The implication is not simply that companies are spending heavily on data centers and chips; it is that a large portion of the economic exposure sits outside the conventional numbers investors tend to use to assess spending, leverage, and cash flow.

The on-screen chart covers a narrower group—Alphabet, Amazon, Meta, and Microsoft—and separates obligations already recognized on their balance sheets from commitments that have not yet appeared there. Its figures should not be treated as a breakdown of the nine-company $3 trillion estimate; they are a separate four-company view of the kinds of commitments involved.

Ranjan Roy explains the attraction plainly: these arrangements can let companies expand infrastructure without drawing down their own cash or holding the new asset directly on their balance sheet. Outside investors—private-equity funds, pension capital, or other institutional pools—can finance a project through a separate vehicle, while the technology company contributes some capital, agrees to lease capacity, or provides a guarantee. Roy says chips can also serve as collateral in some structures.

What makes the arrangements difficult to assess is that their elements appear in different categories. Current debt and recorded lease liabilities are recognized obligations. Uncommenced leases and purchase commitments bind future spending. A guarantee may never produce a cash payment, but can still underwrite the project’s financing by assuring creditors that a large technology company will support the asset if necessary.

Roy’s concern is not that outside capital physically came from the technology company. It did not. It is that ownership, construction financing, leases, customer demand, guarantees, and collateral can be dispersed among different entities and disclosures. The ordinary financial presentation may therefore offer an incomplete view of the company’s future exposure.

For the companies, the arrangement is locally rational. Roy cites Meta’s roughly $91 billion in cash and liquid investments and Alphabet’s roughly $187 billion, but argues that substantial cash does not eliminate the incentive to preserve it. If another investor will fund the asset, a company has a tactical reason to retain liquidity rather than finance every data center directly.

The larger concern is that the world’s biggest technology companies are pursuing variations of the same long-duration wager on AI demand, with substantial commitments distributed among outside capital providers.

Alex Kantrowitz focuses on Meta’s Hyperion data-center project in Louisiana as a concrete example of how one of these structures works. Hyperion covers the equivalent of about 1,700 football fields. Meta is building it, but neither the campus nor the $27 billion in debt financing its construction appears on Meta’s balance sheet.

Instead, funds managed by Blue Owl Capital hold the majority stake in a joint venture that owns the campus. A holding company for Blue Owl’s stake, Beignet Investor, raised construction financing through a bond sale. Meta is a minority partner and tenant; its lease payments are intended to provide cash flows that help service the bonds.

The arrangement illustrates Roy’s point that the technology company can be both the principal customer and the party whose commitment makes financing possible. Hyperion is being built so Meta can use it, Roy says, not as a general-purpose project undertaken for other customers. The external vehicle owns the asset and raises the debt, but Meta’s presence as tenant gives bondholders a reason to expect revenue.

Roy nevertheless resists describing the arrangement as simply Meta’s spending in another form. The initial funding comes from outside investors, not directly from Meta’s cash balance. The distinction is conditional: if AI-infrastructure demand is strong and the facility generates sufficient cash, Hyperion can meet its obligations without Meta having to provide additional support.

But Meta also agreed to lease Hyperion for an initial four-year term beginning in 2029, with options to renew for as long as 20 years. Kantrowitz says Meta guaranteed it would make bondholders whole if it did not stay for the full two decades. Because Meta does not view payments under that guarantee as probable, it has not recorded a related liability on its balance sheet.

The reported aggregate initial Hyperion lease commitment is about $12.3 billion. Meta had also disclosed $347 billion in total obligations for leases that had not yet commenced as of June. Kantrowitz compares that figure with Meta’s cited market capitalization of roughly $1.39 trillion and treats the magnitude as extraordinary.

The point is not that every dollar of those lease obligations is an immediate cash demand, or that every externally financed data center uses the same arrangement. It is that a company’s exposure can include future leases, guarantees, and commitments that are economically important without appearing as direct capex or current debt.

The central uncertainty is not whether the structures exist; it is what turns them from future commitments into losses for the companies, lenders, and institutional investors involved. The optimistic case is that AI demand rises enough for data centers to produce the cash flows needed to service their financing. In that scenario, Meta’s customer commitment supports the project without requiring the company to fund a rescue.

The anxious case is that demand, prices, utilization, or the economics of compute fall short. Then guarantees, long-term lease commitments, and financing structures that looked remote or contingent become economically material. The costs would not necessarily appear as a sudden $347 billion bill for Meta. Ranjan Roy emphasizes that these are obligations spread across years, contracts, lawyers, lenders, and separate pools of capital. A downturn could be slow and distributed: weaker returns for pension funds or other outside investors, stressed data-center projects, and eventually greater pressure on the companies that committed to use the capacity.

That makes the absence of a simple current-liability figure more consequential, not less. Morgan Stanley accounting analysts, quoted in the discussion, said the growing scale and complexity of these arrangements make it harder for investors to assess total potential leverage.

Kantrowitz argues that the structures also have an evident market benefit for the companies. If their full economic exposure appeared as direct spending in the most prominent financial figures, the investment profile could look riskier and potentially pressure their shares. Roy agrees that this is part of the tactical logic. He does not argue that the financing is unlawful; he assumes extensive legal work supports it. His concern is whether it is healthy for the broader economy when the largest companies are all making levered bets on the same demand story.

Whether Wall Street should already have priced that risk is a point of disagreement. Kantrowitz argues that if reporters can draw implications from footnotes in SEC filings, analysts managing large pools of capital should be able to do the same. Roy is less confident that markets reliably process such risks during a boom. Missing Nvidia or leading private AI companies has carried reputational costs, he says; an investor questioning data-center financing 12 to 18 months earlier could have looked foolish while AI-linked assets kept rising.

That is how a collective blind spot can persist even when details are technically disclosed. Roy credits Ed Zitron with raising concerns about data-center financing before the subject received broader attention, and notes that investment-bank research and Wall Street Journal reporting are now trying to total commitments that have been accumulating for at least 18 months.

The speakers agree that AI demand may grow substantially. Roy expects the economy to become increasingly “agentified,” while Kantrowitz calls present AI capabilities exceptional. But technical progress and investment returns are separate propositions. A useful AI system does not guarantee that every facility financed today will earn the returns assumed in its lease, debt, and utilization models.

Kantrowitz frames the spending as a call option on AGI: if systems achieve the transformative capabilities their advocates expect, the upside could be immense. Roy’s objection is that the data-center wager can fail even if AI eventually works. AI could become valuable on a longer timeline than current financing assumes. Efficiency gains could also reduce compute requirements enough to undermine the economics of some facilities while AI still delivers real scientific, commercial, or social value.

The speakers therefore reject the idea that a smooth outcome should be presumed. Their concern is not an immediate collapse in the cash-generating advertising or cloud businesses of Meta and Alphabet. It is the difficulty of modeling an unwind if expected IPOs disappoint, the AI narrative cools, and capital providers stop treating infrastructure demand as an almost unlimited option on future capabilities.

Kantrowitz initially proposes three broad outcomes for AI itself: it works and produces positive results; it works but enables destructive uses, such as hacking or biological threats; or it does not work. Roy argues that those outcomes do not resolve the financial question. They settle on another possibility: AI works, but too slowly to justify the capital structure built around it.

That distinction also shapes their disagreement over medical claims. Kantrowitz argues that AlphaFold and more personalized healthcare examples should not be dismissed, and that AI could have major value in managing or treating disease. Roy agrees that AI can drive meaningful progress in rare-disease research and related work. His objection is narrower: broad promises of medical breakthroughs do not establish why Anthropic or OpenAI specifically will capture that value, or why either company’s IPO should succeed.

The figures discussed make Anthropic’s reported growth striking ahead of a prospective public listing. Kantrowitz cites Bloomberg reporting that Anthropic had reached an annualized revenue run rate of more than $65 billion by the end of July, more than seven times its pace at the end of the prior year. The same report cites quarterly revenue of $11.5 billion, compared with $787 million in the corresponding quarter of 2025.

Kantrowitz treats the quarterly figure as more useful than a standalone run-rate number because it resembles a conventional financial-statement measure. Ranjan Roy remains wary of what the two numbers reveal together. A $65 billion run rate implies a pace beyond simply annualizing $11.5 billion in quarterly revenue, so readers need a clearer view of monthly revenue to understand the degree of acceleration embedded in the claim.

What Roy wants is straightforward: actual monthly or quarterly reporting, ideally showing the trajectory, rather than asking outsiders to reverse-engineer July and August from a combination of period revenue and an annualized run rate. He wonders whether an unusually hyped IPO could sustain vague run-rate disclosures even after going public, though he presents that as a concern rather than a prediction.

Both speakers treat the public appearance of private-company figures as potentially strategic, but neither establishes their origin. Roy draws on his own experience around potential IPO processes to say that bankers and companies generally protect material financial information aggressively. In his view, conveniently timed figures in the press are unlikely to be accidental. Kantrowitz speculates that Anthropic could have an interest in promoting its growth figures before an IPO, potentially to discourage OpenAI from listing first; that remains his interpretation, not a reported fact.

Anthropic and OpenAI had both reportedly filed confidential paperwork to go public, with Anthropic expected to make its Wall Street debut as soon as the fall. The discussion also references reporting that OpenAI chief financial officer Sarah Friar told employees the company would not go public until 2027.

OpenAI’s reported second-quarter figures make a less favorable contrast. Kantrowitz cites Wall Street Journal reporting that OpenAI told investors revenue grew 18% from the first quarter to the second while losses deepened, disappointing some shareholders who had hoped it would show more progress relative to Anthropic. He notes that $1 billion in quarterly revenue remains substantial. His assessment, however, is that slower reported growth versus Anthropic can produce organizational strain, particularly when code generation has become an important commercial battleground.

Roy sees the executive departures as especially consequential. Denise Dresser left the chief revenue officer role less than a year after joining, and the discussion also references departures including former chief technology officer Mira Murati. Dresser’s background as the former CEO of Slack made her a prominent choice to lead an enterprise push. Roy argues that moving a company of OpenAI’s scale toward enterprise revenue and a public-market valuation should not be a project that exhausts its leader in eight months.

Kantrowitz offers a structural explanation. Fast-growing startups often cycle through leadership teams as they move through stages: early builders, executives who impose growth processes, and more seasoned leaders who prepare a company for public markets. OpenAI is compressing those stages at unusual speed.

Roy accepts the description but argues that OpenAI’s valuation ambitions demand more operational maturity. A company seeking a $2 trillion market capitalization cannot, in his view, operate as though it is merely moving from a Series B to a Series C.

Kantrowitz replies that OpenAI and Anthropic remain startups in a crucial sense: neither has settled on a stable product. Their technology is changing quickly, and the question of what they will sell and to whom remains open. Even “enterprise” has not become a settled product category. Is the offering Codex for developers, a knowledge-work system, or something else? An executive who fits the scale of the revenue may not be suited to the continuing chaos of figuring out the offering itself.

Kantrowitz’s competitive interpretation is that Anthropic is currently putting distance between itself and OpenAI, particularly in code generation. He does not treat that lead as permanent. OpenAI could close or reverse it. But Anthropic’s reported revenue trajectory, set against OpenAI’s reported 18% quarterly growth, deeper losses, executive turnover, and later IPO expectation, makes the present contrast difficult to ignore.

Alex Kantrowitz argues that travel is an unusually useful real-world benchmark for AI because it tests context management, changing constraints, and execution under meaningful but recoverable stakes.

Travel planning involves where to stay, what to do, where to eat, what is open, how to get between places, and how to manage changes. The web is poorly designed for this task, Kantrowitz argues, because search results and travel content are distorted by affiliate incentives, fake review sites, and fragmented information. A generative AI system can potentially absorb relevant context and return a concise, tailored answer.

Travel also resembles knowledge work in structure. It is a cascading sequence of small and large tasks that change over time. That makes it a useful proxy for more economically valuable work while remaining a lower-cost environment in which users can test reliability. A mistaken hotel recommendation is frustrating and expensive, but usually recoverable; it does not destroy a company or compromise a critical project.

Kantrowitz calls this the “Goldilocks zone” of AI evaluation: important enough that accuracy matters, but not so high-stakes that experimentation is irresponsible. During a trip through several Indonesian islands and Dubai, he used an ongoing chat for planning and logistics, then asked it each morning to search the conversation, Gmail, and calendar and produce a briefing with the day’s plans, confirmation numbers, and relevant details. He says it worked well until he reverted to ChatGPT’s free version after changing credit cards, which he found markedly worse.

Roy, previously skeptical of travel as a headline AI use case, accepts the framing. He argues that there should be more “normie evals”—tests based on tasks ordinary people actually undertake—rather than relying solely on abstract benchmarks.

His trip to Spain supplied a practical case. He gave an AI system detailed requirements: a Barcelona stop, good Airbnb value within a specified price range, a pool, a sandy beach, shallow-water snorkeling appropriate for a child learning to swim, and a less tourist-heavy setting. The system recommended Denia, Spain, which Roy says met those constraints.

For Roy, the more significant effect is not merely convenience. Reliable assistance can increase people’s willingness to plan more ambitious logistics—more transitions, rentals, remote check-ins, and bookings—because the burden of keeping every detail organized no longer rests entirely on memory, folders, spreadsheets, or printed itineraries.

That does not establish that an AI system can reliably handle higher-stakes workplace tasks. It does make travel a useful test of a prerequisite: whether a system can retain context, integrate scattered information, follow changing constraints, and produce dependable help when mistakes still impose real costs. It is evidence of capability in a bounded setting, not evidence that the infrastructure commitments now being financed will earn adequate returns.