Most AI systems aren't ready. Check yours in 15 min →
AI

Anthropic’s IPO: $17B Revenue, 11,900 Big Spenders, Compute Risks

Published on:
Published in:AI

This is either a triumph of modern business… or a flashing warning sign that we’re getting a little too comfortable with make-believe numbers.

Anthropic going from zero revenue to $17 billion in three years is the kind of story people repeat like gospel. And the “11,900 customers spending at least $100,000 a year” line is meant to land like a mic drop. Serious customers. Serious budgets. Serious demand.

But the part that sticks in my throat is that the same public filing also admits what anyone paying attention already suspects: the computing bill is not some small line item. It’s the whole game. This isn’t a normal software company where you build once and sell forever. This is closer to a factory that has to keep buying electricity just to keep the doors open. The product doesn’t exist without the burn.

That changes how I read those big numbers. $17 billion can mean “we found product-market fit at historic speed.” It can also mean “we turned on the money hose and now we can’t turn it off without customers noticing.”

And let’s talk about those 11,900 big spenders. If true, that’s impressive. It suggests AI isn’t just demos and experiments anymore. A company doesn’t casually spend $100,000 a year unless someone inside is betting their job on it. This is procurement, security reviews, internal champions, actual workflows.

But it also raises an uncomfortable question: how many of those customers are “real” in the sense we care about—using it because it’s essential—versus using it because the fear of missing out is essential?

Imagine you’re a VP at a big company. You don’t fully understand this tech, but you know your CEO keeps hearing about it. You can either be the person who pushed the company to adopt it early, or the person who waited and got blamed later. Spending $100,000 suddenly looks like cheap insurance. Not against a business problem, but against looking behind.

Now add the incentives on the other side. Anthropic has every reason to grow fast ahead of public market scrutiny. Once you’re public, “trust me” becomes quarterly math. Growth becomes expectation. And computing commitments—especially the “enormous” kind hinted at in the filing—don’t care about market moods. If you’ve locked yourself into massive capacity, you’d better keep feeding it with demand.

That’s where I get skeptical. Not because I think the product is fake. I think the product is real. The risk is that the business model gets forced into a treadmill: lower prices to win customers, bigger commitments to deliver performance, more usage to justify the spend, and less room to slow down when something breaks.

And something will break. Maybe not in a dramatic “the model goes down for a week” way. More likely in slow, annoying ways customers hate. Rate limits. Unexpected cost spikes. “We changed the model and now our support bot sounds different.” Or the scarier one: “It made a confident mistake and we didn’t catch it.” If your customers are spending six figures, they are putting this into places where mistakes have consequences—contracts, customer support, internal analysis, code.

Here’s the other tension: safety and risk management aren’t free. If Anthropic is serious about controlling harmful outputs and reducing misuse, that takes people, time, and sometimes turning down revenue. Public markets generally don’t love “we chose to grow slower on purpose.” They say they do. Then they punish you.

So you end up with a cultural fight inside the company. The people who want to ship faster versus the people who want to be careful. The people who want to sign every customer versus the people who want to say no to certain uses. It’s easy to talk about balance in an IPO document. It’s harder when the balance is between the stock price and doing the right thing.

A fair counterpoint is that big compute commitments can be smart, not reckless. If demand is real and keeps rising, locking in supply could protect margins later. If you can secure reliable capacity while everyone else fights over it, you may end up the stable provider enterprises trust. And enterprises love stability more than magic.

But that “if” is doing a lot of work.

Picture two futures. In one, Anthropic’s customers keep renewing, usage keeps growing, and the computing costs become manageable as models and hardware get more efficient. In the other, budgets tighten, AI projects get cut, and those $100,000 contracts turn out to be the easiest line item to “pause for now.” In that second world, huge commitments become an anchor. You can’t gently scale down a factory you already paid for.

I’m not rooting against them. I’m rooting against the story we tell ourselves: that revenue growth alone means we’ve solved the hard parts. In AI, the hard parts are often the costs you don’t see, the risks you don’t want to price in, and the human behavior that turns “useful tool” into “mandatory spend” before anyone can prove the return.

If this IPO is a test, it’s not just a test of Anthropic. It’s a test of whether public markets can value an AI company honestly—accounting for the reality that intelligence is expensive to produce, and even more expensive to produce safely.

So here’s the question I actually care about: when growth and computing costs pull in opposite directions, will Anthropic choose discipline even if it hurts the stock?

Frequently asked questions

What is AI agent governance?

AI agent governance is the set of policies, controls, and monitoring systems that ensure autonomous AI agents behave safely, comply with regulations, and remain auditable. It covers decision logging, policy enforcement, access controls, and incident response for AI systems that act on behalf of a business.

Does the EU AI Act apply to my company?

The EU AI Act applies to any organisation that develops, deploys, or uses AI systems in the EU, regardless of where the company is headquartered. High-risk AI systems face strict obligations starting 2 August 2026, including risk management, data governance, transparency, human oversight, and conformity assessments.

How do I test an AI agent for security vulnerabilities?

AI agent security testing evaluates agents for prompt injection, data exfiltration, policy bypass, jailbreaks, and compliance violations. Talan.tech's Talantir platform runs 500+ automated test scenarios across 11 categories and produces a certified security score with remediation guidance.

Where should I start with AI governance?

Start with a free AI Readiness Assessment to benchmark your current maturity across 10 dimensions (strategy, data, security, compliance, operations, and more). The assessment takes about 15 minutes and produces a prioritised roadmap you can act on immediately.

Ready to secure and govern your AI agents?

Start with a free AI Readiness Assessment to benchmark your maturity across 10 dimensions, or dive into the product that solves your specific problem.