Nvidia's $100 billion OpenAI deal stalls as CEO privately questions 'lack of discipline'
The gap between announcement and execution reveals deeper fragility in AI industry financing
In September, Jensen Huang called it 'the largest computing project in history'. Standing alongside Sam Altman at Nvidia's Santa Clara headquarters, he unveiled a memorandum of understanding worth up to $100 billion: ten gigawatts of data centre capacity, millions of processors, a partnership that would power the next generation of artificial intelligence. Nvidia's stock jumped four per cent on the news. Markets added $180 billion to the chipmaker's valuation before lunch.
Four months later, the deal has stalled. Huang has privately criticised OpenAI's 'lack of discipline' and expressed concern about competition from Google and Anthropic, according to people familiar with the matter. The negotiations Altman expected to finalise within weeks never progressed beyond early stages. The two companies are now rethinking their partnership entirely.
The arc from triumph to doubt in a single quarter tells a story larger than one corporate negotiation gone cold. It reveals a financing mechanism that has become central to the AI industry: announce massive non-binding commitments with fanfare, let stock prices rise, then negotiate actual terms far from public view. Or don't.
The anatomy of a non-binding commitment
The September deal carried no legal obligation. Nvidia would invest progressively as OpenAI deployed each gigawatt of capacity, with the first tranche expected in late 2026. OpenAI would lease the chips back, creating a loop where Nvidia's investment flowed directly back as revenue. The arrangement prompted immediate concern. 'The action will clearly fuel circular concerns,' wrote Stacy Rasgon at Bernstein Research. NewStreet Research estimated that for every ten billion dollars Nvidia invested, it would see thirty-five billion in purchases or lease payments.
This is not fraud. Vendor financing exists across industries. But the gap between announcement impact and execution uncertainty creates conditions where valuations inflate beyond fundamentals. A non-binding letter of intent moved Nvidia's market capitalisation by nearly two hundred billion dollars. The subsequent stalling barely registered.
Lucent's ghost
Those with long memories have seen this before. In the late nineties, Lucent Technologies became the darling of telecommunications, its stock rising tenfold as it supplied hardware to the new competitive carriers created by deregulation. To capture this growth, Lucent extended billions in financing to customers who used the funds to buy Lucent equipment, recording these transactions as revenue.
The parallels sting. Lucent committed eight billion dollars in customer financing; Nvidia's AI investments now exceed one hundred and ten billion. Lucent's customers were cash-burning startups building infrastructure ahead of demand; OpenAI burned two and a half billion dollars in the first half of 2025. Lucent's strategy seemed brilliant until the music stopped. When WinStar Communications collapsed, Lucent wrote off billions. Revenue crashed sixty-nine per cent. The company never recovered.
The telecom bubble taught a specific lesson: when a supplier lends money to customers who use it to buy the supplier's products, everyone wins until someone asks whether actual demand justifies the spending. Twenty-five years later, the question echoes.
The calculus changes
What makes the stalled deal remarkable is velocity. In September, ChatGPT dominated with 800 million weekly users. Altman's infrastructure ambitions seemed backed by an unassailable position.
By December, the ground had shifted beneath his feet. Google released Gemini 3, which topped benchmarks and drew rapturous reviews. Salesforce's Marc Benioff declared publicly he would not return to ChatGPT. Anthropic's Claude Code captured the developer market, generating over five hundred million dollars in revenue. Enterprise market share tells the story starkly: Anthropic rose from twelve to thirty-two per cent in a single year while OpenAI fell from fifty to twenty-five.
Altman responded with an internal 'code red' memo warning of 'rough vibes' and 'economic headwinds'. Advertising plans were shelved. Shopping agents delayed. Resources marshalled to fix ChatGPT's core experience. Prediction markets shifted Google to ninety per cent odds of having the leading model.
Nvidia watched from its privileged position as supplier to all parties. In November, the company committed ten billion dollars to Anthropic. Huang, according to people familiar with his thinking, still wanted to support OpenAI - a major customer whose failure would dent Nvidia's projections. But the hundred-billion-dollar framework no longer fit a world where OpenAI's dominance was no longer assured.
The circular problem
The ecosystem now exhibits structural characteristics that warrant scrutiny. OpenAI has committed to approximately $1.4 trillion in infrastructure over eight years: three hundred billion with Oracle, arrangements with Microsoft, Amazon, AMD, Broadcom, and the now-uncertain Nvidia partnership.
The arithmetic is merciless. Current annualised revenue: twenty billion dollars. Required growth to meet commitments from cash flow: roughly ninefold. HSBC estimates a two hundred and seven billion dollar funding gap by 2030, even assuming aggressive growth.
This creates a dependency loop with no obvious exit. OpenAI took on massive commitments to justify valuations that allow capital raising. But servicing those commitments requires continued raising, which requires maintaining the growth narrative that justifies confidence. The structure functions as long as new money flows. When it stops, the numbers become unforgiving.
Short sellers have noticed. Jim Chanos, who famously shorted Enron, argues that circular financing represents the 'real Achilles heel' of the AI market. Michael Burry wrote that 'true end demand is ridiculously small' and 'almost all customers are funded by their dealers'.
Bulls respond that today's giants differ from telecom-era disasters. The hyperscalers generated over four hundred and fifty billion in operating cash flow in 2024. They can fund AI from profitable businesses. Valuations remain more reasonable than dot-com multiples.
These defences may hold for profitable giants. They comfort less for a cash-burning startup whose survival depends on investor confidence during intensifying competition.
What the supplier sees
Perhaps the most telling detail is Huang's willingness to criticise at all. The Nvidia chief has built a reputation for careful diplomacy, praising partners and competitors alike while maintaining commercial relationships across the industry. Private criticism of a major customer's 'discipline' suggests he sees something concerning.
Suppliers often have the clearest view of customer health. They see order patterns, payment timelines, the gap between public statements and commercial reality. Lucent's executives knew before outside investors that their customers were struggling. Nvidia occupies a similar vantage today.
The conversation has shifted from a hundred-billion-dollar infrastructure partnership to a potential equity stake in OpenAI's current funding round - smaller, with different risk characteristics. Equity investment implies belief in upside. The original partnership implied certainty of demand. September's certainty has become January's doubt.
The fibre laid during the telecom bubble eventually proved useful. Dark fibre that seemed like waste in 2002 became essential by 2012. Perhaps today's data centres will find similar vindication.
But that is cold comfort for investors in companies whose valuations assume exponential growth, or employees whose jobs depend on the next funding round. The announcement-to-execution gap the Nvidia-OpenAI deal exemplifies is not a bug. It is a feature of an industry where ambition and arithmetic drift further apart with each quarterly call.
Jensen Huang, standing at the intersection of supplier and investor, can see further than most. What he sees gives him pause.