Amodei: compute is Anthropic's deepest business dilemma

Days before Anthropic's IPO, its CEO said the hardest call is how much computing power to buy, and that it buys on the low side.

What happened

Days before its stock market listing, Anthropic's chief executive said his hardest problem is not building AI. It is buying the computers to run it.

On Friday, Dario Amodei sat down with JPMorgan Chase chief executive Jamie Dimon for a livestreamed conversation at JPMorgan's Manhattan headquarters. It was one of the last stops in Anthropic's push to build investor demand ahead of its IPO. The company expects to price the shares on Wednesday and start trading on Thursday, 19 November, raising up to $90.4 billion through the listing and a private share sale ($64.4 billion and $26 billion respectively, per Stocktwits). Its IPO website points to a valuation of about $1.4 trillion, below the $2 trillion some investors expected weeks earlier.

"Compute is the deepest business dilemma that the company faces," Amodei said. The problem, as he explained it, is timing. Data centre capacity has to be secured years in advance, but Anthropic's revenue reportedly grew roughly tenfold in each of 2024 and 2025, which makes demand very hard to forecast. Buy too little and growth stalls. Buy too much and the costs weigh on the company.

His answer so far is caution. Yahoo Finance reported him saying Anthropic tends to "buy on the low side", and Investing.com reported that the company tops up with extra capacity on the spot market when demand runs ahead. He said more cash on hand means more buffer, and Investing.com reported that strong gross margins and a large cash balance already provide some. He also said the company is not chasing the highest possible valuation and wants to promise less than it delivers. Anthropic's filings show commitments of nearly $901 billion, mostly for chips and computing infrastructure, of which about $71 billion is due in the year to 30 September 2027.

Amodei also spoke about customers. According to Investing.com, he said businesses in regulated industries need assurances that AI will not expose sensitive information or behave unpredictably, and he rejected the idea that falling model prices mean AI is becoming a commodity, arguing that better models open new markets.

Why it matters

When the head of one of the leading AI labs says capacity is his deepest dilemma, every business that relies on that lab should listen. A supplier that deliberately buys on the low side will sometimes run short. In practice that can show up as rate limits, slower responses or outages when demand spikes.

It also says something about pricing. AI is getting cheaper per task, but the companies selling it are making huge, long-term bets on hardware. Those costs have to be paid back somewhere.

What it means for your business

If you use Claude, or any single AI model, inside a workflow that touches customers or revenue, treat it like any other critical supplier. It can run short, slow down or go offline. That is not a reason to avoid it. It is a reason to have a fallback.

For most mid-sized firms this is a modest piece of work. Know your rate limits and usage tiers. Make sure the tools you build can switch to a second model if the first is unavailable. Keep a human path for the few processes that cannot stop.

What to watch: any changes to pricing or usage limits after the listing.

What to do this week

List the workflows where an AI model is in the critical path for customers or revenue. For each one, check whether it can switch to a second model, and how long that would take. Ask your AI vendor or partner what your current rate limits are and how they behave at peak times.