OpenAI: Reportedly $0 debt, $665B in commitments*

OpenAI reportedly had no debt as of March 31. Its long-term contracts tell a more complicated story.

Deep Dive: OpenAI's $665 Billion in Commitments

Photo by NASA on Unsplash

Anthropic signed a reported $10 billion deal to buy computing power from Volta, an Nvidia-backed cloud company. Blackstone has reportedly pitched investors on a separate loan package tied to Anthropic's purchases of Google chips. Nvidia is also reportedly negotiating a guarantee for companies set up to finance a 10-gigawatt campus in Pike County, Ohio, that SoftBank's energy arm would build and OpenAI would lease. The reported amount is as much as $250 billion and would cover construction and the lease rather than the chips. Nvidia declined to comment on that proposed arrangement.

Only the Volta agreement has been announced. The Blackstone package was reportedly pitched, and the Nvidia support remains under negotiation.

OpenAI shows why those labels matter. The company reportedly had no debt, less than $750 million in obligations under leases, and $46 million in spending on buildings and equipment for the first quarter of 2026. Those figures come from The Information's review of company financial statements, as reported by TNW, rather than from a public filing. The same reporting put OpenAI's purchase commitments at roughly $665 billion, covering long-term access to chips, power, and data-center capacity from Microsoft, Oracle, Amazon, and projects including Stargate.

The asterisk in today's subject line matters. The figures are as of March 31, 2026, are unaudited as far as the cited reporting indicates, and have not been independently verified by Augment.

Purchase commitments are long-term contracts, not borrowed money. That is a real distinction, and there is no suggestion that OpenAI accounted for them improperly. (Augment has not reviewed OpenAI’s underlying financial statements or accounting policies.) But the debt line does not show how much the company has agreed to spend on chips, power, and data centers.

Reported private-company financials and secondary-market indications may be unaudited, incomplete, non-standard, or based on limited transaction activity. They should not be relied upon as fair value, executable pricing, or a basis for any investment decision.

How Meta paid for Hyperion

Meta's Hyperion campus in Richland Parish, Louisiana provides a public example of how companies are paying for these projects. Meta and Blue Owl Capital announced a jointly owned company in October 2025. Blue Owl-managed funds own 80 percent; Meta owns 20 percent. Blue Owl contributed roughly $7 billion, while Meta contributed land and assets under construction. The venture reportedly raised approximately $27 billion through investment-grade bonds in a financing involving PIMCO

Meta operates the campus and leases it from the venture under an initial four-year term. It also guaranteed the campus's value for the first 16 years of operations. The venture borrowed the money. Meta still owes rent and has made a long-term guarantee.

Other companies have used similar, though not identical, arrangements to pay for AI infrastructure. A separate company created for xAI's Colossus 2 was reportedly funded with roughly $12.5 billion in borrowed money and $7.5 billion from its owners, including as much as $2 billion from Nvidia. It owns chips that it leases to xAI. If the loans are not repaid, lenders can reportedly claim the hardware. CoreWeave has reportedly borrowed $18.8 billion through several separate companies, with GPUs backing the loans.

The details differ. One company may guarantee the debt while another does not. One separate company may appear on the operator's financial statements while another may not. The common feature is simple: some of the borrowing belongs to the company that owns the data center or chips, not the company that uses them.

For more company information, recent coverage, and available market data, visit OpenAI on Augment.

What private buyers may not see

Public-company investors can usually examine filings, footnotes, lease schedules, and credit reports. That record can still be difficult to interpret, but it exists.

Private companies share financial details with investors on their own terms. Some buyers may receive detailed information under confidentiality agreements; others may see far less. A quoted price may not reveal whether a company owns its computing equipment, leases it from a separate entity, has guaranteed part of the financing, or has signed long-term purchase contracts.

Two companies with similar reported prices or valuations may have very different obligations. One may own its equipment. Another may rent it. A third may have promised to cover losses if the equipment is worth less than expected. The price alone does not tell you which is which.

Paul Kedrosky, a partner at SK Ventures, has argued that putting debt in separate entities can limit what rating agencies see when they examine the operating company. More broadly, the company using the hardware may not be the company that borrowed the money. Understanding the arrangement means knowing who owes what, what property backs the loan, and who covers any shortfall.

Reported private-company financials and secondary-market indications may be unaudited, incomplete, non-standard, or based on limited transaction activity. They should not be relied upon as fair value, executable pricing, or a basis for any investment decision.

There isn't much history for used GPUs

Companies have separated ownership from use for decades. Airlines lease aircraft. Telecom companies have sold towers. Hotel brands often manage properties owned by others. Investors willing to wait longer for repayment can be a good match for assets built to last decades.

The harder question is what the equipment will be worth later. Aircraft lenders have established ways to estimate value, along with maintenance records, registries, and decades of resale and loss data. The market for high-end GPUs is much younger. The companies have reportedly used different accounting useful lives for certain computing equipment. Accounting depreciation periods do not necessarily represent expected resale values or physical useful lives

Reported rental prices have also moved sharply. H100 rental rates reportedly fell from nearly $8 an hour in early 2024 to about $1.70 by late 2025, before rising to roughly $2.35 by March 2026. Those observations come from a limited set of third-party reports and may not represent all contracts, configurations, service levels, locations, or transaction volumes.

Chip suppliers are participating in this financing market. Nvidia has reportedly invested $30 billion in OpenAI and has separately been reported to be discussing financing for as much as $350 billion of chip sales to the company. Nvidia has also reportedly supported $860 million of lease obligations for a partner's data center and entered a $1.5 billion arrangement to lease its GPUs from Lambda.

In July, Nvidia announced a program under which it may rent unused GPUs from participating cloud providers at a fixed rate in exchange for a share of cloud revenue. Nvidia described the arrangement as a "revenue-sharing and credit-support model." Google has separately guaranteed lease payments for Fluidstack, according to DataCenterDynamics.

Nvidia knows a great deal about how the hardware performs and what customers will pay to use it. In some of these arrangements, it also has money at risk if the hardware goes unused or loses value faster than expected. That is not necessarily a problem. But the terms should show who estimated the equipment's future value, who takes the loss if that estimate is wrong, and whether a guarantee sends some of the risk back to Nvidia or the AI company.

The lenders will decide whether the interest and other protections are enough for those risks. A buyer of private shares faces a more basic problem: the company's reported debt may not include all the money it has promised to spend on infrastructure.

📊 Data point of the day

$46 million

That was OpenAI's reported spending on buildings and equipment for the first quarter of 2026, based on a third-party review of company financial statements. Salesforce reported spending more, but it is a rough comparison: the two companies have different businesses, are different sizes, and pay for infrastructure in different ways. OpenAI's reported purchase commitments totaled roughly $665 billion.

Reported private-company financials and secondary-market indications may be unaudited, incomplete, non-standard, or based on limited transaction activity. They should not be relied upon as fair value, executable pricing, or a basis for any investment decision.

🎓 Manual

Structural subordination

Say a company puts a data center in a separate subsidiary, and that subsidiary borrows the money to build it. The subsidiary's lenders generally get paid from its assets before value can flow back to the parent company. Guarantees and other contracts can change who ultimately bears the loss.

‍