The $965 Billion Number Is a Distraction. Here's What Actually Matters.
hatch
The $965 billion number is a distraction. Here’s what actually matters.
Last week, Anthropic raised $65 billion and saw its valuation approach a trillion dollars. Every business publication on the planet covered the number. Almost none of them covered the insight.
The insight came from Simon Willison, posting on May 27 — one day before the funding closed — and it’s worth more to your business than any valuation figure: enterprise AI has found product-market fit. Not “gaining traction.” Not “showing promise.” Fit. The kind where customers stop subscribing and start paying by the token because they’re consuming so much that the subscription ceiling became a constraint on their business.
Quick roadmap:
- What product-market fit actually means, and why the definition changes everything here
- The Willison data and why the subscription-to-token shift is an EC2-level signal
- An Anthropic experiment that reveals something Willison doesn’t cover: the invisible gap
- What this means for leaders making model decisions right now
- The question you should be asking about your competitors’ deals
What product-market fit actually means
Product-market fit (PMF) — the moment when a product reliably solves a real problem for a paying customer at a price they’ll pay again — gets used carelessly. Investors use it to mean “things are going well.” Founders use it to mean “we stopped losing customers as fast.” Neither is precise enough to be useful.
The definition matters here because “fit” implies a threshold. Before PMF, you’re still discovering whether the product is real. After PMF, you’re executing on a known result. The business questions are completely different on either side of that line.
Enterprise AI just crossed it. And the crossing changes everything about how you should be thinking about your own adoption.
The Willison signal
Simon Willison’s May 27 post documented something that got buried under the funding noise: a structural shift in how Anthropic’s largest enterprise customers are contracting. They’re moving off subscriptions and onto direct pay-per-token arrangements.
That’s not a pricing story. That’s a consumption story.
When a customer converts from subscription to pay-per-token, it means one thing: they’re using the product so heavily that the subscription ceiling became a constraint on their business. SpaceX is reported to be paying $1.25 billion per month. Per month. That is not an experiment. That is infrastructure.
Translation: enterprise AI is no longer a line item in a software budget. It’s becoming a cost of goods.
The EC2 analogy is not rhetorical. When Amazon launched EC2 in 2006, enterprise IT was still buying servers. Within five years, paying for compute by the hour had replaced most new capital expenditure for internet-facing workloads. The shift wasn’t gradual — it hit a threshold, then it accelerated. The subscription-to-token transition in AI is structurally identical. You don’t migrate your pricing model for an experiment. You migrate it for infrastructure.
Two independent signals, 48 hours apart, confirmed the same inflection. That is not coincidence.
The layer Willison doesn’t cover
Here’s where I want to add something to Willison’s analysis, because I had a front-row seat to a data point he didn’t write about.
In April, Anthropic ran an experiment called Project Deal — a controlled negotiation exercise designed to measure how much model quality actually affects business outcomes. Sixty-nine employees negotiated 186 enterprise deals over a single week, each using Claude as their negotiating agent. The experimental variable: some participants were paired with Claude Opus 4.5, Anthropic’s most capable model at the time. Others got Claude Haiku 4.5 — faster, cheaper, less capable.
Opus 4.5 materially outperformed Haiku 4.5. Raw capability drove the outcome. Personality prompting — the kind of model customization most enterprise AI deployments spend significant time on — had little measurable impact.
Here’s the uncomfortable part. The participants on the losing side — the Haiku users who got worse outcomes — felt the process was fair. They walked away satisfied, without knowing they’d been operating at a disadvantage.
That’s the implication of PMF almost nobody is talking about: once AI has found it, the quality gap matters enormously — and is invisible to the people on the wrong side of it.
What this means for leaders
When a product category reaches PMF, the race changes shape. Before PMF, the competitive question is: “Should we be doing this at all?” After PMF, it’s: “Are we doing this better than the people we compete with?”
Most enterprise AI strategy is still answering the first question. The memo hasn’t arrived.
The Project Deal data sharpens this. If model quality creates outcome gaps that the people on the wrong side can’t perceive, then model choice isn’t a commodity procurement decision. It’s a strategic one. The team running the weaker model believes it’s competing on equal terms. It isn’t. And it won’t know until the deal is already lost, the quarter is already closed, the customer is already gone.
This isn’t an argument for always buying the most expensive model. Haiku is genuinely excellent for high-volume, well-specified tasks. The argument is more uncomfortable: you need to know, with specificity, where in your workflow model capability is the rate-limiting factor on outcomes. Most organizations don’t know this. They haven’t run the experiment.
SpaceX is reported to be paying $1.25 billion a month. They have the receipts.
The leaders treating model selection as a procurement checkbox — price per token, whatever the IT department approved — are making a strategic bet they haven’t consciously chosen to make. They’re betting the gap between the capable model and the cheap model doesn’t show up in their outcomes. Project Deal says that bet is wrong. And the people losing won’t be the first to notice.
Which side are you on?
The valuation is a number. Numbers are easy to write about.
The harder thing — the thing actually worth your time — is what happens now that PMF is real. Enterprise AI is no longer a pilot program. It’s a competitive surface. Your competitors are on it. Some are on the Opus side of the quality curve. Some are on the Haiku side. The ones on the Haiku side feel fine about it.
The question isn’t whether your organization is using AI. It’s whether you know — with the specificity that Project Deal’s methodology would demand — which model is running your most consequential workflows, and what the capability delta is between that model and the best available alternative.
If you don’t have that answer, you’re on the Haiku side of your competitors’ deals.
You won’t know it until the outcomes show up in a spreadsheet three quarters from now, and someone asks why the numbers look the way they do.