The Plane No One Knows How to Land
Bill Gurley has spent twenty-five years at Benchmark identifying the exact moment when enthusiasm stops being an asset and becomes a liability. He did this with Uber in 2017, when he pushed to remove its founder before a toxic culture and unsustainable numbers dragged everything down. Now, in an interview with CNBC on March 16, 2026, he repeats the diagnosis on a scale ten times larger: the artificial intelligence sector is burning capital at a rate that no profitability horizon can justify.
His words are precise and unembellished: "One day we simply trip and run out of money." He is not talking about a technological collapse. The model works. Systems learn. Products exist. He is talking about something much more mundane and much more lethal: between thirty and forty AI startups are losing billions annually, and the volume of capital available to sustain those losses has a mathematical limit that the sector’s optimism has chosen to ignore.
The imagery Gurley uses is revealing: "It’s harder to land the plane" when so much burnt fuel has not reached cruising altitude. Companies like OpenAI and Anthropic have raised tens of billions in funding. But raising capital is not the same as building a business model. It is, at best, buying time.
Why Zero Marginal Cost Doesn't Save Those Who Can't Afford It
Here is the structural paradox that no sector analysis is naming clearly enough: artificial intelligence operates on a logic of decreasing marginal cost. Once the model is trained, the cost of generating the tenth million response is marginally lower than that of the first one. The technology, in theory, tends toward a state where producing more costs almost nothing. That is its deepest economic promise.
But the market’s current diagnostic error is to confuse the marginal cost of inference with the total cost of building the system. Training a frontier model costs hundreds of millions of dollars. Keeping it updated costs just as much. Building the data infrastructure for it to function at scale costs billions. Gurley points out that the seven major tech giants—Apple, Amazon, Alphabet, Meta, Microsoft, Nvidia, and Tesla—are investing hundreds of billions in data centers to sustain that infrastructure. That expenditure does not diminish marginally. It is fixed, massive, and accumulative.
The result is an inverted cost structure compared to what the market is discounting: revenues scale slowly because enterprise adoption is lagging behind media hype, while infrastructure costs are growing almost exponentially to sustain the arms race of parameters and computational capacity. The logic of zero marginal cost applies to the future of the sector, not its financial present.
This has direct consequences for business models. A startup charging for access to its language model faces brutal price pressure because its competitors, also funded by venture capital, are willing to sell below cost to capture market share. No one is building margins. They are buying users with losses subsidized by investors betting to be the last ones standing when capital becomes scarce.
The Bubble Bursts Not Due to Technology, but to Capital’s Patience
Gurley draws a parallel with the dot-com era that deserves more dispassionate analysis than is usually given. The easy comparison is that back then there were also companies without revenues with stratospheric valuations. But the mechanism of collapse is different and more instructive.
In 2000, capital dried up when public markets closed the IPO window and retail investors lost their appetite. Today, private capital has much deeper reserves, artificially extending the runway. But this also means that when the adjustment comes, it will arrive suddenly, not gradually. Michael Burry has warned about dangerously high levels of over-investment. Jeremy Grantham of GMO has systematically documented how tech bubbles burst precisely when the underlying technology begins to demonstrate utility, not before.
The pattern Gurley identifies as "interlopers"—actors entering attracted by momentum, without a rigorous investment thesis—is the most reliable indicator that a bubble has surpassed its formation phase and is in its terminal phase. When every venture capital firm declares it is only looking at AI opportunities, and when founders of fitness and language-learning apps reframe their pitches as AI companies, capital stops being allocated where it generates the most value and starts being allocated where there is the most narrative.
What Gurley recommends to investors in that context is operationally simple: identify software companies with proven subscription models, wait for the valuation correction the reset will produce, and buy with discipline. Do not bet on private AI startups that are, in his words, "enormously risky." The information asymmetry in those investments is too high and the path to profitability too uncertain.
Block, the parent company of Square led by Jack Dorsey, laid off nearly half of its workforce in a deliberate move to adopt AI. That is not marginal optimization. It’s a signal that even companies with established business models are rewriting their operational architecture on the premise that human capital can be partially replaced. If profitable companies are doing this, imagining that those not yet generating revenue will escape that pressure is an exercise in denial.
The Reset Reorders the Hierarchy, Not Eliminates the Technology
The misinterpretation of Gurley’s warning would be to conclude that artificial intelligence is an illusion. It is not. The tools have been demonstrating measurable utility in personalization, information synthesis, and automating repetitive tasks for the past two to three years. The problem is not the technology. The problem is the gap between the value that technology generates today and the value that the capital markets are discounting for tomorrow.
When the reset occurs—and financial logic indicates it will occur, not as a possibility but as an arithmetical consequence of the accumulated burn rate—artificial intelligence will not disappear. Dozens of companies that failed to build sustainable unit economics will vanish, those that prioritized user growth over margins, and those that assumed capital would continue to be available indefinitely because the narrative was large enough.
What will survive are models where the customer acquisition cost bears a reasonable relation to the value that customer generates over time, where the infrastructure does not require permanent subsidies to function, and where differentiation does not depend exclusively on having the biggest model, but on solving a specific problem better than any alternative.
Leaders who understand that value in this sector is built on unit economics, not on round funding valuations, are the ones who will be positioned to acquire assets and talent at rational prices when capital becomes scarce. The reset is not the end of the artificial intelligence cycle. It is the moment when technology stops belonging to the storytellers and starts belonging to the builders.
