AI chip executives say demand is 'almost unlimited' even as companies get pickier about costs

Key points

• Pat Gelsinger, former Intel CEO, told CNBC this week he considers AI demand "almost unlimited", limited only by available energy.
• Nebius, a company building Nvidia-powered data centres, says it cannot fulfil current customer demand and has not been able to for some time.
• Cerebras Systems CEO Andrew Feldman called Meta and xAI selling spare computing capacity a "unique" case, not a sign of industry-wide oversupply.
• Lumentum, which makes optical networking components for data centres, says its products are sold out for the next five years.
• Enterprises are shifting from encouraging maximum AI use to asking what they actually get back for the money spent.

Chip stocks have swung sharply in recent weeks, and investors have been asking a straightforward question: is the world's hunger for AI actually slowing down?

Several company leaders say no. In interviews with CNBC Tech this week, they argued that demand for computing power is still running well ahead of what the industry can supply.

"What we're experiencing in terms of demand is extraordinary. There's much more demand than we're able to fulfil," said Marc Boroditsky, chief revenue officer at Nebius, a company building data centres packed with Nvidia GPUs (the specialised chips that do the heavy number-crunching AI needs).

Pat Gelsinger, the former CEO of Intel and now a venture capital partner at Playground Global, put it bluntly: "I somewhat think of AI demand as almost unlimited."

His reasoning is simple. More intelligence applied to any business problem tends to produce more economic value. Energy supply, he said, is "the only real limiter."

Is the industry building too much?

No, according to most of the executives who spoke up this week. The concern started when Meta announced it would sell spare AI computing capacity it was not using, and Elon Musk's xAI did the same. To some investors, that looked like a glut.

Cerebras Systems CEO Andrew Feldman pushed back. Meta and xAI, he said, are unusual cases. "For the industry as a whole, the demand for compute far outstrips available capacity."

Lumentum, which makes photonics and optical components (the hardware that moves data between servers at high speed inside data centres), is perhaps the sharpest illustration. Its CEO Michael Hurlston said the company's products are already spoken for through to 2030. Lumentum shares have risen roughly 600 percent in the past twelve months.

Now the tricky part: what are businesses actually getting for all this spending?

For a period, many companies told employees to use AI tools freely, measuring success by volume of use rather than results. Boroditsky at Nebius called this "tokenmaxxing", where a token is the basic unit of text that a large language model (the technology behind chatbots like ChatGPT) processes.

That phase is ending. Finance directors are now pressing for proof that AI spending produces real returns, a shift executives describe as moving towards "valuemaxxing."

Feldman offered a practical way to think about it. Not every task needs the most powerful AI model available. "You don't need a giant bus to go to the grocery store," he said. Simpler tasks will move to cheaper, smaller models; harder problems will stay with the expensive frontier systems.

For businesses using AI today, that means the sensible question is no longer "are we using enough AI?" but "is this particular tool worth what we are paying for it?"