OpenAI’s $1.15 Trillion Compute Bill: Who Actually Holds the Bag

Published: · Updated:

OpenAI's cash burn for 2026 is already covered four times over by a single financing, so the unmarked risk sits with Oracle's bondholders and the private credit funds that paid for buildings the tenant is already outgrowing.

The cash burn is the funded part, and the exposure sits with the bondholders and private credit funds who paid for the buildings.

Disclosure: Impersonal research, not personal investment advice (see the disclaimer). Figures last refreshed 2026-09-06; sources linked inline. Evergreen piece, refreshed quarterly.

Oracle sold OpenAI the largest cloud contract anybody in the software business has ever signed, and then Oracle's own shareholders spent the year getting out. $ORCL trades at $158.78 this week, down about 33% over twelve months from a high of $344.21. And over that same stretch the company's contracted order book went from $137.8 billion to $638 billion.

OpenAI's burn is already paid for, because Sam Altman has raised more cash than his company plans to lose this decade, at a valuation nobody will mark down. The concrete is a different question, though. Every dollar of the roughly $1.15 trillion OpenAI has committed across seven vendors turns into a building that somebody finances years before OpenAI pays a cent of rent on it. That somebody is Oracle's bond investors, plus a short list of private credit funds who wrote long loans against halls for a tenant whose chips go stale in four years.

What OpenAI's Cash Burn Actually Buys

The burn gets quoted more than anything else in this story. OpenAI pencils roughly $27 billion of cash burn for 2026 and about $63 billion for 2027, and doesn't expect to turn cash-flow positive until 2030. But it closed a $122 billion round at an $852 billion valuation in March. This round of financing covers the 2026 hole four times over.

Annualized revenue is running near $40 billion, roughly double where it sat at the start of the year, but the price didn't move. OpenAI ran a $7 billion employee tender in August at the same $852 billion mark it set in March, so the private market watched revenue nearly double and declined to pay a dollar more for it. CFO Sarah Friar told staff at an August all-hands that the company goes public in 2027, sooner if growth holds, and she framed the IPO as a milestone rather than a finish line, just another fundraise after the $122 billion.

Then there's what OpenAI signed on the other side. The company has committed to about $1.15 trillion of vendor deals running to 2035, and the composition matters more than the headline. Roughly $1.05 trillion of it is compute and chips OpenAI has agreed to buy, while the other $100 billion runs the opposite direction. $NVDA agreed to invest up to that amount in OpenAI as part of a 10-gigawatt deployment deal, so a chip vendor is helping fund its own customer's purchase order, and that kind of arrangement earns a paragraph in every credit memo written about this sector.

The Map: Where the $1.15 Trillion Sits

Six vendors carry the purchase side, and you can see the split in the chart below. $AVGO has the largest at $350 billion for custom accelerators, while Oracle's is a $300 billion, five-year cloud contract that starts in 2027 and underwrites 4.5 gigawatts of new capacity. $MSFT holds $250 billion of Azure purchases from the October 2025 restructuring that ended its right of first refusal. $AMD, Amazon and $CRWV make up the rest.

Chart: OpenAI Compute and Chip Purchase Commitments ($B, 2025-2035)

Five of those six sell OpenAI something they were already building, since Microsoft has Azure, Amazon has AWS, Broadcom and AMD ship parts, and CoreWeave rents GPUs to whoever signs (we got into that after its second-quarter print).

Oracle had to become a different company to take the order. Capital spending hit $55.66 billion in fiscal 2026 against $67.36 billion of revenue, so 83 cents of every revenue dollar went into plant and free cash flow went negative. Total debt is now around $130 billion, with $248 billion of new lease commitments sitting alongside it. Oracle got there by selling $18 billion of bonds in a single day in September 2025 and then coming back for $25 billion more in an eight-part deal in February 2026.

That $638 billion order book is remaining performance obligations, which is contracted revenue Oracle hasn't delivered yet, so it's a promise to sell rather than cash in the door. (Our company page for $ORCL carries the quarterly series if you want to watch it convert.)

The Mechanism: Oracle Borrowed, OpenAI Didn't

You can see the mechanism when you look at Oracle's numbers for fiscal 2026. The order book went up more than four times while their debt went up by $37 billion to pay for the buildings. Capex more than doubled. But none of this comes out of OpenAI's pocket yet.

Chart: Oracle, twelve months on ($B)

In March, OpenAI decided against expanding its Stargate campus in Abilene with Oracle, because it wants clusters built around newer Nvidia generations than the Blackwell parts going into that hall. $ORCL fell 23% on the news, which makes sense when you remember that Abilene, in Texas, is the flagship site, the first Stargate build and the one every press release opened with.

Abilene is also financed, and not off Oracle's balance sheet. Blue Owl and JPMorgan hold roughly $13 billion in the vehicle that owns it, part of a $38 billion package covering two campuses in Texas and Wisconsin, plus an $18 billion loan against a New Mexico site. Those are special purpose vehicles, so the building sits in its own company with its own debt, off the parent's balance sheet and secured by the lease. That makes the lease the collateral, and the tenant just said out loud that the hall is a generation behind before its power is even energized.

None of this would matter if the buildings and the chips aged at the same speed, but they don't. Hyperscalers depreciate servers over five to six years, and Amazon cut a slice of its fleet back to five in February 2025 because of the pace of AI hardware. Nvidia ships a new architecture every 18 to 24 months, while the buildings get financed on much longer paper, with 15-year take-or-pay leases now standard on large hyperscale halls. So Oracle borrows against the long life of a shell while its tenant makes decisions on the short life of the silicon, and Abilene is what that looks like in practice.

So who ends up wearing it? Three groups have exposure to that $1.15 trillion, and only two of them have repriced so far.

Equity has processed this, since Oracle's shareholders already took a third of the value out. Credit has processed part of it, at least in the corners where prices are public. Oracle's five-year credit default swaps, the contracts investors buy to insure against a default, closed at their highest level since 2009 and now serve as the market's standing proxy for AI debt nerves, and Boaz Weinstein's Saba Capital has been selling that protection to the lenders who wanted it.

Nobody prices the third group daily, because that's where the money actually went. Total AI-related debt outstanding is tracking toward $570 billion in 2026, and the marginal dollar keeps moving out of public bonds into private credit and off-balance-sheet vehicles where there's no screen and no daily mark. Morgan Stanley puts private credit at roughly $800 billion of the $1.5 trillion data center financing gap through 2028, and Blackstone, Blue Owl, Apollo, Pimco and BlackRock originate most of it.

The Bank for International Settlements put it more plainly than the sell side has. Private credit lenders price AI infrastructure loans at essentially the same spreads as their non-AI loans, while AI equities trade at multiples that imply far higher expected returns, so one of those two markets has the risk wrong.

What to Watch

Oracle's next print comes first. Watch capital spending against operating cash flow, and whether management gives a delivery schedule for the $300 billion contract instead of restating the order book, because RPO is easy to grow when you sign one customer.

Then the mortgage bonds, which are the only public read on how the lenders feel. Two of the last three data center commercial mortgage deals had to widen pricing from initial talk to clear the market. Top-rated paper on single-borrower data center deals was quoted at 168 basis points over SOFR on July 28, against 153 a year earlier, so that's fifteen basis points of repricing on an asset class whose largest tenant just walked away from its flagship hall.

Then OpenAI's own listing. A 2027 IPO turns the largest private credit exposure in the market into a disclosure document, and every counterparty on that $1.15 trillion gets to read the same audited numbers at the same time. (Anthropic's version of this arrives first, which we covered in the rent bill piece.)

The View

OpenAI gets through this cycle, because it has the cash and the access to raise more, and its equity holders have already been told what they own at a flat $852 billion mark.

Oracle's shareholders were told too, at a 33% discount. But the lenders holding paper against Stargate halls haven't been told anything, because nobody has asked them to sell yet, and that's the last unmarked piece of the $1.15 trillion. Fifteen basis points is the whole adjustment so far.

What we're watching

Q1 FY2027
Whether Oracle gives a delivery schedule for the OpenAI contract instead of another RPO headline.
Q4 2026
Spreads on new single-borrower data center CMBS, last quoted at 168bp over SOFR for AAA paper.
2027
The OpenAI S-1, which turns private compute commitments into audited public disclosure.

Tags: pillar, openai, oracle, orcl, private-credit, compute-commitments, stargate