Two warnings landed this week that cut through the AI hype. Benchmark’s Bill Gurley says the industry will “trip and run out of money.” Norway’s $2.1 trillion sovereign wealth fund is stress-testing scenarios where AI stocks crash 53%.
What ties them together: the growing realization that over $800 billion in AI financing is circular—chip makers investing in AI companies that buy their chips—and nobody knows what happens when the music stops.
The Circular Financing Problem
Here’s how it works. Nvidia pledged up to $100 billion to invest in OpenAI last September. OpenAI committed to deploy at least 10 gigawatts of data center capacity using Nvidia systems. Nvidia’s money flows to OpenAI, which flows back to Nvidia as chip purchases.
Amazon put $50 billion into OpenAI’s February funding round. SoftBank added another $30 billion. OpenAI has signed infrastructure deals with Oracle ($300 billion), AMD ($90 billion), and AWS ($38 billion).
The deals create financial statements that look impressive on paper. But strip out the circular arrangements and the picture changes.
According to Bain & Company analysis cited by Calcalist, AI companies need $2 trillion in annual revenue by 2030 to justify current infrastructure spending. Projected revenue: roughly $800 billion short.
OpenAI’s actual revenue—$13 billion projected for 2025—doesn’t match its $500 billion valuation or the $1.15 trillion in infrastructure commitments it’s signed. The gap gets filled by investors who are also customers who are also suppliers.
What Gurley Sees Coming
“One day, I just think we trip and run out of money on those things,” Gurley told Fortune. “I do think that moment stands in front of us.”
Gurley points to capital expenditure ratios that now exceed dot-com era levels. Capex-to-sales ratios hit 34% this year and are projected at 37% by 2028. Five major hyperscalers have accumulated nearly $1 trillion in undisclosed future lease commitments for data centers.
The burn rates are staggering. OpenAI needs an estimated $207 billion in additional funding through 2030, with total projected cash burn of $280 billion. Anthropic has spent over $10 billion training models that generated only half that in cumulative revenue.
Gurley compares this to Uber’s $2 billion annual burn rate, which gave him “high anxiety” as a board member. “The thing I’m learning is the larger a company you can scale to, the more money you can afford to burn getting there.”
His advice to investors: wait for the correction, then buy discounted SaaS stocks. Salesforce and ServiceNow have already lost over 20% since early 2026 as the market prices in AI disruption that may not materialize.
Norway’s Stress Tests
Norway’s $2.1 trillion sovereign wealth fund, the world’s largest, ran the numbers on what an AI correction would look like.
The results were brutal. In the fund’s “AI correction” scenario—where massive capex fails to produce real productivity gains—equities plummet 53%. The fund’s total value drops 35%. Fixed income rises 10% as investors flee to safety.
“Market concentration has increased, and AI-related capital expenditure has grown large and concentrated,” fund CEO Nicolai Tangen said. “If AI capex fails to deliver productivity gains, growth expectations could revert sharply.”
The fund identified an AI bubble as one of its top risk scenarios alongside geopolitical tensions. A worst-case geopolitical scenario—investment restrictions and severe tariffs—could wipe out 37% of value.
The Enterprise Reality Gap
Underlying all this is a fundamental question: is AI actually delivering returns for the companies deploying it?
MIT research found that 95% of AI agents developed within companies don’t work, representing roughly $40 billion in failed deployments with no measurable ROI.
A National Bureau of Economic Research study from February 2026 found that despite 90% of firms reporting no impact from AI on workplace productivity, executives still projected AI would increase productivity by 1.4%. Hope, not evidence.
Sam Altman himself acknowledged the contradiction last year: “Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes.” He still took the money.
What Makes This Different
The standard response to bubble talk is to invoke the dot-com era: some companies failed, but the technology was real and survivors thrived. True as far as it goes.
But AI infrastructure is harder to unwind than software. Dot-com excess produced abandoned websites and worthless stock. AI excess is producing physical data centers that consume 14% of U.S. electricity by 2040, according to projections. When Nvidia-powered servers become stranded assets, you can’t just delete them.
The concentration is also more extreme. Five companies—Nvidia, Google, Microsoft, Apple, Amazon—now account for 30% of S&P 500 value. Eighty percent of 2025 market gains came from AI-related stocks. When bubbles pop, they tend to take related sectors with them.
For now, the music plays on. The question isn’t whether AI is valuable—it clearly is in specific applications. The question is whether current valuations bear any relationship to near-term cash flows, or whether we’re watching the most sophisticated vendor financing scheme in history.
Bill Gurley thinks the reset is coming. Norway’s fund is modeling the crash. The circular deals keep spinning, but eventually someone has to generate actual revenue.