AI compute demand outruns supply: price by the outcome

If you've been waiting for AI to get cheap because the market "corrects," stop waiting. On July 12, CNBC reported that the executives selling compute are seeing demand they can't fill, and the enterprises buying it are quietly changing how they spend. The takeaway for anyone running a small business on AI: the price of raw compute isn't falling from oversupply, and the days of "use as much AI as possible" are over. What replaces it is boring and correct — measure the return, or stop paying.

What actually happened

CNBC talked to the people on both sides of the AI compute trade. Marc Boroditsky, chief revenue officer at Nebius — which builds data centers on Nvidia GPUs — said there's "much more demand than we're able to fulfil," and that it's been that way for a while. Cerebras CEO Andrew Feldman echoed it: demand for compute far outstrips capacity, and there's a shortage of data centers to run it in.

Meanwhile the stocks got jumpy. Samsung, one of the largest memory-chip makers, forecast a big jump in profit — and its shares still fell, after a 360%+ run over the prior year left the market wondering how much higher it could go. None of that volatility dented actual demand for compute.

The most useful detail for operators is the spending shift. CNBC describes a fading era of "tokenmaxxing" — companies telling employees to use as much AI as they could — giving way to a hard focus on return on investment, precisely because frontier models stay expensive relative to open-source alternatives. Nebius's own pitch to enterprises now is "value-maxxing," not token-maxxing.

Why it matters for your business

Read the two facts together: compute is scarce, and buyers are turning ruthless on ROI. That means your AI bill has a floor set by physics and demand, not a ceiling that vendor competition will kindly lower for you. Waiting for prices to drop is not a strategy.

The lever you actually control is architecture. The savings come from routing each task to the cheapest model that clears the quality bar, keeping an open-weights option in the loop so you're not captive to frontier pricing, and — most of all — measuring cost per completed outcome, not per token. A refund resolved, an invoice reconciled, a lead qualified: that's the unit that pays for itself. Tokens burned is the unit that quietly bankrupts a pilot. If your AI stack can't tell you the cost of a finished job, you're token-maxxing by accident.

Sources: CNBC — AI demand 'almost unlimited' amid stock volatility, CNBC — Nebius CRO on 'value-maxxing' in AI.