Who's Actually Paying the AI Bill? The Token Economy Got Real This Week

Last week I was telling a friend that the enterprise AI story has quietly stopped being about which model is smarter. This week kind of proved me wrong and right at the same time — the money data is now public, and the numbers are weird.

The money story nobody saw coming

Anthropic has apparently captured over 60% of business AI API spending. OpenAI is sitting around 35%. I had to read that twice, because the conventional wisdom for the last couple of years has been "OpenAI is the default." The nuance is that Anthropic is charging premium prices on a much smaller slice of token volume — it's not that everyone switched, it's that the customers who pay the most are buying from them. Enterprise buyers who run heavy agent workloads apparently decided the higher per-token cost buys them fewer bad outputs, fewer retries, less babysitting. To be fair, that's a defensible trade — my own API bills have taught me that a cheap model that needs three calls to get the answer right is the expensive model.

OpenAI isn't dead, obviously. 35% of a massive market is still a massive market. But the narrative flip is real: "default" no longer means "what enterprises pay for."

The tokenmaxxing hangover hit the balance sheet

Speaking of bills — the other big thread this week is the cost of the token consumption race. Remember when companies were bragging about how many tokens their engineers were burning through? Jensen Huang literally said a $500k engineer who doesn't consume at least $250k worth of tokens would alarm him. That quote aged like milk.

Uber's CTO admitted they blew through their entire annual AI budget in four months. Four. Months. Meta is reportedly planning per-employee token limits. Microsoft cancelled Claude Code licenses and consolidated everyone under Copilot. There were literal leaderboards tracking token spend per employee — one startup's single engineer racked up over $7k in a month — and those leaderboards are being torn down. Harvey's monthly token spend hit 12 trillion tokens a month, up 12x.

Here's my take: tokenmaxxing was always a cargo-cult metric. More tokens ≠ more productivity, and the companies that treated tokens like they were free are now discovering the meter was running the whole time. The ones that will survive this are the ones tracking outcomes, not token counts. Measuring how many tokens your team burns is like measuring success by how much fuel your car uses — you can do it, it's just not the number that matters.

Sovereignty is now a product pitch

Over in India, Gnani launched something called Artha — a "sovereign AI stack" with an open-weight foundational LLM trained from scratch, plus an agentic platform (Evon v3.3 and Plexus) on top. This is aimed squarely at Indian enterprises and public institutions that don't want their data flowing through foreign clouds. I find this genuinely interesting because "sovereign AI" went from a buzzword to an actual product category over the last year, and now countries with the talent pool are shipping their own stacks. Whether the model quality holds up is another question — building from scratch is a flex, but it's also a bet on compute and data you can actually control.

Debian said no to the AI ban

Smaller but worth a mention: Debian developers voted down proposals to ban or strongly discourage AI use in the project, and instead chose a "Responsible Use of Generative AI" position. That's a nuanced outcome from a community that could have gone full reactionary. It's not an endorsement — it's a "let's not pretend this doesn't exist, let's set ground rules" move. From my perspective that's the most realistic position a community can take in 2026.

A local-LLM setup that actually works

And on the personal-experiment front, someone on XDA hooked two mini PCs (DGX Sparks, 128GB unified memory each, linked over ConnectX-7) to a local LLM — DeepSeek V4 Flash, 284B params — that reads their entire codebase and opens GitHub issues for code review automatically. I love this because it's the anti-tokenmaxxing story: no API meter, no leaderboard, just hardware you own doing the work locally. The catch, as always, is the setup cost and the fact that running 284B params at home isn't exactly a $10/month hobby. But the pattern — owning your own inference for your own code — is going to keep eating at the "everything through the cloud" model, at least for developers who can afford the iron.

The practical bit

If you're a dev or a founder reading this, the takeaway isn't "use Anthropic" or "go local." It's: get a number for what your AI actually costs per outcome, not per token. That's the single most useful habit I've picked up this year. The companies that know their cost-per-shipped-feature are going to make very different decisions than the ones still staring at token dashboards.

There's a tool I keep coming back to when I need to sanity-check these numbers — PayCalc — it's been handy for quick cost math while I've been tracking my own API spend. Anyway, that's the week. The token economy stopped being theoretical, and the bills are here. Curious to see who blinks first.