Google's chief executive described a trade-off that most AI customers never see. The cloud division could sell more computing power than it does, and it chooses not to, because the same machines are needed to build Gemini 4 Argon, the company's newest model. Argon went out in September, and Google is still short of the capacity to serve it.
The maker is first in line, then the signed contracts
Google's rule is that its AI lab, Google DeepMind, gets the computing power it needs first. Everything else is planned after that, and the cloud business is planned around contracts customers have already signed. The machines are built years ahead, so what Google has now follows decisions made in 2024. A customer who asks for capacity in 2027 is asking for something Google's chief executive called "incredibly difficult" to provide.
That makes three kinds of AI customer: the maker itself, customers with signed commitments, and everyone else. In a shortage the third group feels it first, as slower answers, usage limits and waiting lists. It is also the biggest group.
Car makers lost their place in the queue
In 2020, car makers cut their chip orders when showrooms emptied. Months later they wanted the chips back and found that the factories had moved their capacity to consumer electronics. A chip needed about 26 weeks to make, so cutting orders was easy and undoing it was not.
The shortage then forced a choice about what to build. Ford stopped three plants for a time so that the scarce chips would go into its more profitable pickups and large SUVs. Any business that runs short of a key input ends up ranking its own products.
Agents make the shortage last
Argon used about three times as many tokens in its first 72 hours as Gemini 3.0 did. The bigger reason the shortage will last is how the next products work. Google plans assistants that run around the clock in the cloud, with a small virtual computer started for each request. A growing share of what people ask Gemini is already a task rather than a chat.
A chat answer needs a few seconds of machine time. An assistant that watches your inbox and books your appointments holds on to computing power for hours. That is hundreds of times more per request, and it explains why Google's personal agent, Spark, is for paying subscribers in select countries for now. Free users will probably get the always-on version last.
What a careful business does with this
Take a hospital group's finance team that checks 20,000 supplier invoices overnight. At 3,000 tokens an invoice, that is 60 million tokens a month. At Argon's introductory price of $2 per million input tokens, reading them all costs $120, and $240 when the price doubles. Output tokens cost extra, and what tasks will cost is a separate question we covered earlier.
The money is small. The danger is the morning the queue is slow and the close is late. So a cheaper model, like the Flash models Google itself builds because its biggest models are slower and cost more, should read and sort every invoice. Argon would see only the 5% that do not match, which costs about $6 at the introductory price. And the team should write down in advance what happens when Argon is unavailable.
Ford had to decide which plants got the scarce chips, and it chose what earned the most. Teams that depend on AI will make the same choice, and it is better made before the shortage than during it: decide which work deserves the best model, and let the rest run on something cheaper.