Over the past year, America's largest technology companies have provided up to $300 billion in guarantees for AI data centers and chips. According to reports from last week, these arrangements have underpinned the current construction boom, yet much of this exposure has not been immediately recorded as debt. This money deserves special attention because it does not appear in any major line item of a financial statement—it lives in the footnotes.

Where the debt actually sits

These guarantees carry the formal name "residual value guarantees" (RVGs). The mechanics are straightforward: a special purpose vehicle (SPV) borrows money to purchase computing equipment or build data centers, while the technology company backs a minimum future value. If these assets are ultimately sold or leased at a price below that floor, the tech company must cover the shortfall. The structure can be broken down into three layers: the SPV borrows to buy chips or server space, with the loan supported by cash flows from the AI company's usage contracts; if that AI company stops paying, assets are subleased or sold to service the debt; should a gap remain, the guarantor—typically a chipmaker or cloud provider—steps in to cover it. For investors, this structure allows tech companies to expand AI capacity while preserving credit ratings and conserving cash—at the cost of pushing potential liabilities off their balance sheets. If that floor gives way—demand disappoints, hardware depreciates faster than expected, or the industry builds out more computing power than customers need—those promises cease to be mere paper words.

Three case studies: Meta, Broadcom, and NVIDIA

Meta Platforms, Inc. pioneered this approach, first employing a residual value guarantee at its Hyperion campus in Richland Parish, Louisiana last year. Meta formed a joint venture with a fund under Blue Owl, with the partner holding 80% and Meta retaining 20%. Development costs totaled approximately $27 billion, with the SPV bond solely bookrun by Morgan Stanley and rated A+ by S&P—just one notch below Meta's own AA-. Meta contributed land and construction-in-progress in exchange for flexibility: upon completion, it would lease the campus in full for an initial four-year term with renewal options. In return, Meta provided approximately $28 billion in residual value guarantees covering the first 16 years of operations—should the lease not be renewed or be terminated early, Meta would pay cash to the joint venture. Per reports at the time, this arrangement allowed the project to raise $27 billion at costs only slightly above Meta's own debt, while keeping most obligations off Meta's balance sheet.

Broadcom has applied guarantees to selling chips: it provided approximately $29 billion in support to a special purpose vehicle that would purchase chips and then lease them to Anthropic. The vehicle belongs to the AI XPV platform Broadcom spearheaded in June, with an initial $35 billion private credit facility arranged by Apollo and Blackstone to purchase Google-designed TPUs for lease to Anthropic. In its latest quarterly report, Broadcom disclosed that an investor partner assumed rack purchase agreements containing its custom AI accelerators along with corresponding customer leases. Broadcom committed to covering payments on these leases for five years, with maximum exposure of $29 billion on the initial transaction. The company stated that given the expected profitability and equipment residual values of leading AI developers, the likelihood of losses is very low. Bank of America analysts estimate this financing vehicle could accumulate $370 billion in senior debt by mid-2029 at 20 gigawatts of capacity.

NVIDIA has gone even further. The company indicated it could support up to 25% of certain financing arrangements and provided $105 billion in guarantees to SB Energy for a data center campus in Ohio serving OpenAI. According to the 8-K filing from August 17, the guarantees cover leases for approximately 4.25 gigawatts of IT load at the campus, with the tenant being an OpenAI affiliate. NVIDIA's cumulative payment obligation is capped at $105 billion, contingent on the lessor meeting "ready for operation" conditions, with expected activation starting in 2028. The company also retains options to provide credit support for an additional approximately 3.8 gigawatts. The triggering conditions are quite narrow: only in the event of OpenAI bankruptcy causing lease default or failure to pay rent would NVIDIA be required to make payments covering the gap between the minimum lease guarantee value and proceeds from subleasing or sale. In return, the campus is expected to use NVIDIA hardware exclusively for 20 years, with NVIDIA additionally investing $1.5 billion in SB Energy. CEO Jensen Huang denies this constitutes "circular financing," stating OpenAI will pay its own rent.

Why giants prefer guarantees over issuing debt

The answer lies in cash flow. Apple, Alphabet, Meta Platforms, Inc., and Microsoft collectively project approximately $381 billion in 2025 capital expenditures, almost entirely funded from operating cash flow. For 2026, this figure is expected to grow over 60% to approach $700 billion, while 2025 AI revenue stands at only around $60 billion. The gap cannot be covered by cash flow, and issuing corporate debt of equivalent scale would crush even the most pristine investment-grade ratings—so guarantees replace bonds: tech companies pledge to cover losses if projects fail, giving private credit the credit quality it needs, while the companies never formally borrow. The costs are already appearing on balance sheets: Alphabet and Amazon both saw free cash flow turn negative in the second quarter, while Meta Platforms, Inc. managed only $784 million and required a $30 billion private credit lifeline. In June, Google completed its first-ever stock issuance since going public, raising approximately $85 billion net, alongside a $100 billion bond offering. Some institutions calculate that five cloud providers will generate approximately $3.5 trillion in external financing needs from 2026 to 2030.

Beyond $300 billion: the $3 trillion iceberg

Morgan Stanley estimated in an August 25 report that five hyperscale cloud providers plus NVIDIA and Broadcom have combined commitments exceeding $3.1 trillion through guarantees, leases, and other forms of financial support. Alphabet leads with the largest exposure at $890 billion, primarily from purchase commitments. These companies' 2027 cash capital expenditures are projected to exceed $1.2 trillion versus expected cash flows of approximately $1 trillion, with over 40% of revenue being reinvested into AI capital spending. Another analytical framework differs slightly: nine leading tech companies disclosed approximately $3 trillion in off-balance-sheet commitments in their latest filing footnotes, including roughly $1.2 trillion in "leases not yet commenced" and approximately $1.9 trillion in purchase commitments. This scale is about five times their combined capital expenditure of $600 billion over the past year, and grew about 50% in just two months from the previous approximately $1.8 trillion. Moody's calculations focus on physical assets: debt and lease commitments related to six leading cloud providers total approximately $1.3 trillion, with off-balance-sheet debt growing roughly eightfold in four years. Contingent liabilities carried through residual value guarantees currently stand at approximately $70 billion, not yet fully priced by markets. Other analyses suggest the hidden debt of five major US tech companies totals approximately $1.65 trillion, already exceeding their on-balance-sheet debt of approximately $1.35 trillion, with Meta Platforms, Inc. estimated highest at approximately $420 billion.

Why accounting rules allow it to stay hidden

The key lies in threshold settings. Under US GAAP, lease renewals must be recorded as liabilities only when they meet the "reasonably certain" standard—typically interpreted as greater than 70% probability. Residual value guarantee payouts triggered by non-renewal require accrual only when "probable"—meaning likelihood exceeds 50%. Between these sits a gray zone. Moody's analysts David Gonzales and Alastair Drake note that whether to renew depends partly on hyperscalers' willingness to invest further in hardware, and strict application of current guidance would render many renewals short of "reasonably certain." Moody's thus warns "disclosures may not show the full picture" and that reported debt figures "are unlikely to reflect certain plausible future scenarios." Meta Platforms, Inc.'s disclosures confirm this logic: citing low probability of payout, the company has recorded no liability to date for its approximately $28 billion guarantee.

Rating agencies and bond markets are already pricing it

Rating agencies have not turned a blind eye: they assess how much commitment values might exceed asset sale prices under stress scenarios and may incorporate the difference into leverage calculations. Moody's confirmed NVIDIA's Aa1 rating with a positive outlook in August, based on pro-forma figures showing approximately $106 billion in cash and marketable securities and adjusted total debt-to-EBITDA of just 0.2 times. However, the same report lists "concentration of multiple such transactions in a short period" as a key risk, stating that if Broadcom's contingent obligations increase substantially, even with current low debt leverage, its financial flexibility would be constrained. Moody's also expects NVIDIA's guarantee exposure to peak in 2031.

Bond investors use blunter language. CreditSights compares these residual value guarantees to "selling a put option": nearly costless in boom times, most damaging during severe recessions when customers default and hardware values fall—such pro-cyclicality amplifies both booms and busts. Maria Entina, portfolio manager at asset manager DoubleLine, says this resembles exploiting regulatory loopholes to seek special treatment from rating agencies: "When financial engineering prevails, it masks the true face of finance." Brian Gelfand of TCW notes this is not ordinary investment-grade credit underwriting; given the off-balance-sheet nature, tail risk is elevated. So far, however, analysts broadly conclude the collateral provided by these assets is sufficient to limit expected losses. Counter-evidence exists—CoreWeave claims it recently signed a contract at near-full price for NVIDIA's 2020-era A100 chips running through 2029, citing inference demand extending the useful life of older-generation GPUs.

What would expose this accounting

The risk pathways are straightforward: demand falling short, hardware depreciating faster than expected, or the industry building more computing capacity than customers need. Structural vulnerabilities include tenant concentration—OpenAI and Anthropic hold significant shares of hyperscaler order backlogs, with combined remaining performance obligations approaching $900 billion. More notable is the duration mismatch: guarantees are pro-cyclical while debt and leases are long-lived—one Bitcoin miner pivoting to compute hosting signed a 20-year lease with Anthropic, yet model generations turn over in months. Residual value assumptions rest on an asset class defined by rapid obsolescence.

This is precisely what investors should watch next: as AI hardware ages and financing commitments continue accumulating, whether these assets can continue providing sufficient collateral. Operationally, that means examining the guarantee schedule in the next quarterly report—is it shrinking or growing? Under the agreement structures, guarantees activate as leases commence (the Ohio project's tranches are expected to begin in fiscal 2029), and exposure should decline as tenants pay rent. Should a payout-triggering event occur, guarantee amounts rise rather than fall, or a platform quietly gets withdrawn, the true cost of this model will reveal itself.