Nvidia Chief Says AI Has Gone From Curiosity to Profitable Work
Nvidia's Jensen Huang argues profitable tokens justify more chips, even as chip stocks slump.
Bloomberg Daybreak: US Edition · Zelenskyy Defies Trump; Leakers 'Should Be Shot' Says PresidentIn Jensen Huang's telling, artificial intelligence has just become profitable work. Huang, the chief executive of Nvidia, whose chips power many AI systems, says AI spent its early life as something curious and full of potential but unprofitable and not very useful, and that in the last six months it has become productive work that generates profit.
His example is Anthropic, the AI company now preparing to start promoting its public offering in the week beginning November 9 at the earliest. Anthropic put its third-quarter revenue at $16.8 billion and said it was profitable on an adjusted basis. That revenue figure and adjusted-profit disclosure come from Anthropic's own account to prospective investors, not from a public earnings filing, and the company did not say what it excluded to reach profitability on an adjusted basis. Huang, speaking about Anthropic's plans, argued that its revenues are real and its profits from those revenues are real.
Huang's explanation turns on two plain words: tokens and compute. A token is a small chunk of text an AI model reads or writes, roughly a word or part of a word. Compute is the processing power needed to create those tokens, supplied by specialized chips packed into large data centers.
On his reasoning, the chain runs like this. Customer demand creates demand for tokens. Making tokens requires compute. Huang said "the tokens are now profitable." My plain-language reading of his word profitable is that each piece of text can be sold for more than the computing power costs to make it. If so, Huang argues, then buying more compute lets the company make still more tokens and more profit. He described that loop as a flywheel, like a heavy wheel that turns faster once it gets going.
That is his case for why profitable AI would sustain chip demand, not reduce it. That same week, investors were voting the other way. The semiconductor index known as the SOX was down 20 percent from where it stood in June, a drop widely described as a bear market. South Korean stocks closed almost 8 percent lower, with Samsung Electronics and SK Hynix each down more than 9 percent. The AI cloud company CoreWeave fell 17 percent as higher financing costs came into focus.
If Huang's reading of one company's numbers proves general, each profitable answer helps pay for the chips to supply the next one. If the market's worry about the cost of building all that computing power proves right, the flywheel could slow before it spins faster.
Questions people ask
What is a token in plain terms?
A token is a small piece of text, often a whole word or part of one, that an AI model processes. A sentence might contain a dozen tokens. Companies track tokens because they measure both how much work the system does and what customers can be charged for.
What does compute mean here?
Compute means processing power, mostly from specialized chips running in large data centers. Training a model takes a huge one-time burst of compute. Answering users takes steady compute every day, which is why rising use means rising chip demand.
Does profitability guarantee more chip buying?
That is Huang's argument, not a settled outcome. He says profitable tokens warrant buying more capacity. Stock investors, by contrast, have been selling chip and AI infrastructure shares on concern about the high borrowing and building costs needed to add that capacity.