OpenAI's $18B Chip Deal Can't Find the Money to Pay For Itself

OpenAI's $18 billion chip deal is quietly falling apart.

The financing arrangement that was supposed to fund a fleet of custom artificial intelligence accelerators — OpenAI's answer to reducing its dependence on Nvidia — has hit a snag. The Information reported Thursday that Broadcom, the chip designer managing the program, will only finance the first phase if Microsoft buys roughly 40 percent of the processors, or if OpenAI finds other major customers willing to commit. Broadcom shares fell about 4 percent on the news.

The deal, announced by OpenAI and Broadcom in October 2025, was framed as OpenAI's entry into the custom silicon club — a group that already includes Google, Amazon, and Microsoft, all of which built their own chips to control costs and reduce reliance on a single supplier. The ambition was real: enough computing power to run the largest AI training workloads the world had seen. The problem is that ambition and balance sheet are not the same thing.

"You can design the chip, but you still have to pay for the factory that builds it, the power plant that runs it, and the customer that makes the economics work," said one semiconductor analyst briefed on the deal, speaking without authorization to discuss it publicly.

The numbers behind the deal are large even by AI standards. The accelerators in development would consume 10 gigawatts of power at full load — roughly the output of five Hoover Dams, according to The Information's reporting. That is not a chip problem. That is an energy problem. The semiconductor is the easy part.

The structural mismatch has been visible for months to anyone reading the fine print. OpenAI announced the Broadcom partnership in October 2025, one week after confirming it had hired Intel's former chief technology officer, Sachin Katti, to run its infrastructure organization. The message was that OpenAI was taking direct control of its hardware stack — designing the silicon, owning the supply chain, reducing what it pays Nvidia and AMD at market rates.

The message and the reality have diverged. Custom silicon programs are capital-intensive on a scale that requires either a massive balance sheet or committed customers willing to prepay. Google began developing its Tensor Processing Units in 2013. Amazon launched its Inferentia training chip in 2018. Microsoft launched its Maia AI chip in 2023, according to Toms Hardware. All three had years of Azure revenue, customer commitments, and existing infrastructure to absorb the cost. OpenAI is arriving at the same destination on a different road — one that runs through someone else's financing.

Broadcom had originally required OpenAI to match every dollar Broadcom provided in financing — a standard arrangement in semiconductor deals to limit the chip vendor's exposure. Broadcom recently relaxed that demand and agreed to invest more capital up front than originally required, Sherwood News reported, an apparent concession to make the numbers work. The concession was not enough to close the gap.

Microsoft, meanwhile, has been quietly reducing its financial entanglement with OpenAI. Microsoft recently revised its agreement with OpenAI to end revenue-sharing payments from Microsoft to OpenAI, stripping out a cushion that had been a reliable source of cash for the AI company. The simultaneous end of that payment stream and the Broadcom financing pressure are not unrelated.

There is no shortage of entities that would consider filling the gap. GPU-equity swaps, chip-backed infrastructure funds, and compute-as-a-service prepayment models have all emerged as mechanisms for spreading the capital cost of AI hardware beyond a single company's balance sheet. Several infrastructure funds and sovereign wealth vehicles have shown interest in data-center debt as an asset class, according to people familiar with the market. Whether any would commit to a specific OpenAI program at the scale required — tens of billions of dollars, multiple years of construction risk — is an open question. The infrastructure fund that would write that check has not yet stepped forward.

The deal has not collapsed. Broadcom and Microsoft have not commented on the specifics of the negotiating positions. The financing pressure may be a tactic: Broadcom has incentive to appear accommodating while holding the line, and OpenAI has incentive to signal urgency to attract alternative buyers. What the reporting makes clear is that the gap between OpenAI's stated hardware ambitions and its current financing structure is real — and it is not going to close itself.

The outcome matters beyond OpenAI. If the financing structure cannot close, the fallback is buying more chips from Nvidia and AMD at whatever the market charges. That is not a crisis. It is the status quo. But it means the vertical integration story that made the custom chip program strategically important is not yet real — and the window to make it real is narrowing.

The race to own the silicon underneath AI is real. The question is who can afford to run it.