Anthropic buys compute on the low side. The crunch is the plan

I build my website, my funnels and much of my content with Claude. So when the vendor's CEO says out loud that he buys compute on the tight side, I don't read it as an IPO story. I read it as a supplier notice.

That is exactly what Dario Amodei said on Friday. Six days before the first trading day.

What Amodei said

• Where: a livestreamed conversation with JPMorgan CEO Jamie Dimon at JPMorgan's Manhattan headquarters on Nov. 13.
• The line: According to Yahoo Finance and Bloomberg, Amodei called securing enough compute Anthropic's "deepest business dilemma." The company generally buys less capacity in advance than its growth projections would justify, then adds more when demand comes in higher. His words: "We actually tend to buy on the low side."
• The buffer: More cash on hand means more room between what you bought and what customers want. That is also part of the answer to why Anthropic is going public now.
• The IPO: pricing reportedly on Wednesday, trading from Thursday, Nov. 19. Between the offering and a concurrent private placement, Anthropic is seeking up to $90.4 billion at a valuation of about $1.4 trillion.

Scarcity is a decision, not an accident

Sounds like finance talk. It isn't. If you deliberately buy on the low side, you accept that peak demand will sometimes outrun supply. That's sound management. Idle chips burn billions.

For customers, though, it means crunches aren't incidents. They're priced in. And when supply is short, the vendor decides who gets served first. In my view, that makes priority the real product: big committed contracts up front, self-serve and small team plans behind them. No price list says so. It's just how every scarce good works.

The listing brings a second force: quarterly earnings. Every data-center dollar becomes visible to shareholders. If anything, I'd expect the incentive to buy lean to get stronger.

If AI is your operating system, the supply chain matters

I keep saying AI belongs in marketing as an operating system, not a nice assistant on the side. I still believe that. But an operating system has a supply chain, and right now it sits with a vendor that is publicly planning its capacity close to the edge.

Friday's Claude login outage showed what a hard failure feels like. The quiet version is worse: limits that kick in earlier, slower answers at peak hours, an agent that stalls at 5 p.m. before a campaign launch. Nobody calls a crisis hotline over that. The launch still slips.

Add this: reports say Google, Microsoft and Nvidia are also buying in alongside the IPO, companies that supply Anthropic with chips and cloud compute. Supplier, landlord and shareholder become the same names. That doesn't make the dependency chain any shorter.

What I'd do now

• Make a list: which team workflows run on Claude, and which of them have a deadline? Newsletter sends, campaign assets, client reports. Those are your peak-hour risks.
• Check your plan: what usage limits does your tier carry, and is there a tier with committed priority? If you have an account manager, ask exactly that.
• Test a second model on your single most critical workflow before you need it. Same brief, not improvised in an emergency.

Bottom lineWhen your vendor says it would rather buy too little than too much, plan as if you're not first in line.