Last Wednesday, Alex Karp went on CNBC’s Squawk Box and said, on camera, that OpenAI and Anthropic are stripping enterprises of their competitive edge through token pricing. His exact words: “These people are stealing the weights and alpha of my business.” Strong stuff. But the part that should actually worry you didn’t happen on television. It happened that same week, with no cameras, when Marty Kausas, CEO of the startup Pylon, posted that his company’s Anthropic bill is about to jump from $400,000 to $1.4 million a year. Without using a single extra token. The bill didn’t go up because usage grew. It went up because his company crossed a seat count, and on the other side of that line sits exactly the token-billing business Karp attacked on national television.

Your AI bill doesn’t go up because you used more

Kausas said it plainly: usage didn’t explode, his team is just about to cross 150 seats. That number isn’t arbitrary. Claude’s Team plan runs from 5 to 150 seats with a hard ceiling. Hiring employee 151 isn’t a quota adjustment, it’s an eviction: the company gets forced onto Enterprise, where the seat fee stops bundling any usage and every token bills separately at standard API rates.

Pylon didn’t burn more tokens. It hired one more person, crossed a seat it wasn’t tracking, and the entire plan changed rules underneath it. That’s the pattern almost no company budgets for. The risk was never really how much AI you’ll use. It’s how many seats you’ll have the day your contract renews. Nobody puts that variable in a financial model. Nobody puts it in the proposal a CFO signs off on. Which is exactly why, when it hits, it hits as a surprise.

Karp said it on CNBC, and it wasn’t just an outburst

Karp never mentioned Pylon or the 150-seat ceiling. He aimed at something broader, with a question any CFO should be asking: “If it was so valuable, let’s say I can make you $1 billion tomorrow. Wouldn’t I say I’ll make you $1 billion and I want 30 percent? Why are they charging for tokens if it’s so valuable?” By his account, “every single enterprise in this country, these people are livid.” They’re paying for tokens that create no value, he argued, while handing over their data and operating patterns to the same labs that later sell that edge back as a product to their competitors.

It’s a loaded accusation coming from a CEO with skin in the game. Palantir competes directly with those same labs for enterprise AI budget. But the question stands on its own even with the bias attached: if the value is as large as the labs claim, why does the pricing model stay consumption-based instead of outcome-based? And the irony of the week is that two days before this interview aired, AWS announced a $1 billion investment in its own Forward Deployed Engineering unit, built on the exact model Karp’s own company pioneered over a decade ago. While Karp goes after token billing on camera, Amazon is betting a billion dollars that the future of enterprise AI gets sold as a service, not by the token.

How the 150-seat ceiling actually works

Worth understanding the full mechanism, because this isn’t a rumor circulating on social media. Anthropic confirms it in its own support documentation: on new Enterprise plans, the seat fee covers platform access, and usage bills separately at standard API rates, pooled across the whole organization. No seat carries an individual token allowance.

The Register reconstructed the timeline in April: Anthropic started renewing accounts onto the metered-usage model in November 2025, introduced the flat $20-per-seat monthly fee in February 2026, and by March 8 the legacy bundled-usage enterprise plans stopped being available to renewing accounts. If your company signed its Anthropic contract before that date, the next renewal can land under different terms than the ones you negotiated.

There’s also a number mix-up worth clearing up, because it trips people constantly: 50 seats is the minimum for sales-assisted Enterprise, not the Team ceiling. A company sitting at 80 or 140 seats can stay comfortably on Team. The only number that forces an automatic migration is 150. Crossing it isn’t a purchasing decision. It’s a side effect of hiring the wrong person at the wrong point in your billing cycle.

What we do about this at IQ Source

This connects directly to something I’ve already argued: list price per token lies, cost per job is what matters. That post was about measuring consumption wrong. This one is worse, because it isn’t consumption at all. It’s a seat threshold that almost no budget spreadsheet accounts for. Both problems land in the same place I described when I wrote about governance as the real moat: anyone can rent the model, but knowing exactly which contract terms you’re standing on, and what happens the day you cross a threshold, is on you.

When we run the discovery phase of AI Maestro, one of the first things we map is a client’s seat and consumption trajectory against the real thresholds of the vendor they’re evaluating or already under contract with. It’s not an abstract exercise. It’s asking: how many seats do you have today, how many will you have in twelve months based on your hiring plan, and what happens to your bill at each of those thresholds along the way. That projection happens before signing, not after the invoice lands. I made a related point in a spending cap that treats the symptom: a cap doesn’t fix the cause, it manages the symptom. The cause is not knowing when your contract changes its own rules.

Karp might be right, bias and all, that consumption-based pricing is a bad match for what AI is actually worth inside an enterprise. But as long as that model exists, the question in front of you isn’t philosophical. It’s operational: do you know which seat your contract sits on today, and what happens the moment you reach the next one?

Map your AI contract’s thresholds before you cross them