Broadcom has spent the past weeks seeking more than $50 billion in financing for the chips it is building with OpenAI under an internal project called “Project Nexus”, while Apollo and Blackstone held early talks about joining the package. The transaction, reported by The Wall Street Journal, could close as soon as the end of the year and would cover several gigawatts of compute. This is not a one-off deal on the margins of the industry. It is the third gigawatt-scale debt package this month that is turning AI from cash-financed construction into debt-financed construction.
Broadcom builds and deploys systems under multi-year contracts, which means Broadcom, not OpenAI, fronts most of the capital and recovers it through contracted deliveries. What Broadcom holds instead is the vitality of the guarantee: if OpenAI stops buying compute for any reason, the debt sits on Broadcom.
Three Deals, One Logic
The same pattern repeats every month, more times each time. Broadcom is seeking more than $50 billion for chips for OpenAI. Oracle is in talks with Apollo and Goldman Sachs for its own chip purchase, targeting a single gigawatt-scale data center. SpaceX, until recently a company with almost no influence on the cloud market, is raising $40 billion to buy processors from Nvidia, including roughly $10 billion in bank financing and $30 billion in corporate bonds, with Apollo leading the financing and Pimco as a potential lender. SpaceX financing may not close until 2027, but it has already reshaped the market.
All three deals share one foundation. Instead of buying hardware with cash on the balance sheet, each of these companies is looking for a way to have someone else finance the asset today and repay it from future revenue streams. When Nvidia wants to sell more GPUs and the buyer lacks cash, the only realistic path is debt.
| Company | Financing size | Who is leading | What is being bought | Status |
|---|---|---|---|---|
| Broadcom for OpenAI | more than $50B | Apollo, Blackstone in talks | custom ASICs with OpenAI, several GW | early talks, close possible by end of 2026 |
| Oracle | talks on a gigawatt data center | Apollo, Goldman Sachs | Nvidia chips and infrastructure | negotiations gaining pace |
| SpaceX | $40B (10B banks + 30B bonds) | Apollo, Pimco in play | Nvidia GPUs, mainly Vera Rubin | targeted close 2027 |
| Broadcom/Anthropic (earlier) | $60B package in distribution | BofA, Citi, Morgan Stanley | Google/Broadcom TPU chips for Anthropic | senior secured $42B + $18B subordinated |
Credit Prices Are Already Reacting
The credit market sees these deals before equities do. Broadcom’s five-year CDS widened by 3 basis points to a record 136 bps, and the pressure continues: the market prices the probability of an Oracle default above 20 percent over five years, SpaceX around 16 percent, and even Nvidia, the world’s most valuable company, trades above 7 percent on a five-year horizon.
“If we get a few more days like this one, we could see real panic in AI-related stocks,” said Kristina Hooper, chief market strategist at Man Group, commenting on the selloff after OpenAI’s revenue revelation.
The CDS widening is not yet a faint. The credit market reacts to volume: Broadcom alone may need to raise roughly $600 billion in coming years, according to Bloomberg calculations, to fund the entire compute program. That is a number the size of the entire stressed-credit market. As volume grows, protection gets more expensive even when fundamentals are stable, because the supply of risk grows faster than the supply of capital.
The Accounting Needs Explaining
The FT reported on Thursday that OpenAI told investors its annualized revenue run-rate reached roughly $50 billion by the end of September, about $20 billion less than the figure circulating a month earlier. The difference, as CNBC explained, stemmed from how cloud partners counted revenue. In practice, OpenAI may have been reporting $70 billion in aggregated revenue from contracts with partners such as Nvidia, and now reports $50 billion in cleaner, contracted revenue excluding infrastructure.
This distinction is both accounting and economic. OpenAI’s annualized revenue is not the same as realized annual revenue, and in practice the two kinds of income come from different periods and different sources. When Broadcom, Oracle and Nvidia sign multi-year contracts with OpenAI for compute deliveries, their underwriting relies on future revenue streams, not on what OpenAI shows today as recognized revenue.
A Risk Nobody Has Priced Yet
These deals share one requirement and one assumption on which all the rest rests. The requirement: a relatively stable AI revenue stream must be sufficient to service the debt on the acquired assets. The assumption: chip and compute prices will not fall sharply over the next three to five years, as the financing assumes.
The FT revenue revision indicates that at least one of these assumptions is less certain than it seemed a week ago. If OpenAI builds 8 gigawatts of data centers in Ohio with SB Energy, and Broadcom builds several gigawatts of chips under Project Nexus, both sides currently assume OpenAI will buy the entire compute for ten or twenty years. If it stops, the first to feel it will not be OpenAI. It will be Broadcom, with $50 billion of debt secured by custom chips nobody else can use.
Here lies the supplier-financing asymmetry. Nvidia can sell GPUs to many customers. Broadcom can sell custom chips to several clients at once, but Project Nexus is an exclusive arrangement: chips designed for OpenAI have no realistic other buyers. This works as long as OpenAI keeps growing. It becomes a credit catastrophe if it stops.
There is an engineering angle too: Jules Urbach, chief executive of Render Network, has reported that GPUs responsible for much of the chip cost have been moved onto a ship serving as a floating data center, which will run for three years. This is not only a metaphor: the ASICs from Broadcom for OpenAI sit in data centers for the life of the OpenAI contract, as long as those chips make sense.
A Funding Cycle, Not a Business Cycle
The comparison with the last great infrastructure funding wave, telecoms and fiber from 1995 to 2001, is obvious and simplified, but instructive. Back then banks provided tens of billions of dollars to build optical networks, at least half the capacity, before demand appeared. When demand failed to keep up, the debt was left with the banks and the fiber, not the suppliers. Chipmakers of that cycle had less protection than today: their chips were one-off series, mostly in single variants, not designed to order.
Today, an analogy of one year building, ten years of payback would have to assume OpenAI and Anthropic grow revenues by tens of trillions of dollars more per year than today. If OpenAI realistically reaches the revenue implied by building 8 GW of data centers in Ohio, and Anthropic its own 3.5 GW of Google TPU chips, the financing becomes self-sustaining. If not, the whole structure falls on lenders.
Until the market sees the first consolidated financials from these packages, the first closed deals and the first real debt-service impact on Broadcom and Oracle balance sheets, the market will speculate. Speculation alone does not create a crash, but it changes expectations for the future. And in the AI infrastructure funding cycle, that is already enough for CDS spreads to widen.