The Token Bill Comes Due

The Token Bill Comes Due

Status
Draft
Author
Created
Nov 8, 2026
Tags
Cost
Architecture
Agentic AI
Text
Apple, Uber and Revolut all touched the same dial this week. Your AI spend is now an architecture decision.
Somewhere in Cupertino, an engineer who used to have $300 a day of AI to spend now has $200. Per Bloomberg, it may drop to $100. Tesla reportedly caps staff at $200 a week. Uber limits each AI coding tool to $1,500 a month. Three of the most sophisticated engineering organizations on earth, in the same week, reached for the same lever.
None of them announced that AI coding doesn't work. They announced that it works so well people won't stop using it, and that the invoice has arrived.
I think this is the week the industry's mental model quietly changed. For two years we treated model spend as a rounding error against engineer salaries, something to be justified by vibes and a productivity slide. That framing is dead. Token spend is now a line item with its own owner, its own budget, and its own architecture review.

The same week, a very different answer

While Apple was tightening caps, Revolut went the other direction. According to this week's reporting, it moved roughly 2,000 engineers off paid coding tools onto GLM-5.3, an open model it hosts itself on 56 NVIDIA B200 GPUs. The claimed effective cost is about $0.30 per million tokens, roughly a tenth of what the paid tools cost, so around a 90% reduction.
Treat that number with the skepticism you'd apply to any vendor-adjacent figure. It's a reported estimate, and it almost certainly benefits from high utilization: 2,000 engineers hitting the same cluster smooths demand in a way a 20-person team never could. Idle GPUs are the silent killer of self-hosting economics. Revolut has the scale to keep those 56 cards busy.
But the direction matters more than the decimal. Two opposite reactions to the same pressure, rationing and self-hosting, tell you the market has no settled answer yet. Which means you get to make the decision deliberately instead of inheriting it.

The frontier is getting cheaper too

Meanwhile, the managed side moved. AWS's November 2 roundup announced OpenAI's GPT-6.1 Terra generally available on Bedrock, with a 1 million token context window, text and image input, and token prices AWS says are at least 50% lower than GPT-6.1 Sol and GPT-5.6 Terra. The performance claims come from OpenAI's own benchmark and the roundup notes they aren't independently verified, so run your own evals before you believe a word of it.
The pricing direction is still the point. Per-token cost is falling fast on the hosted side. If you locked in a self-hosting business case twelve months ago, it may already be stale. If you capped spend based on last year's price list, you may be rationing something that just got half as expensive.
So the honest answer to "should we self-host?" is: it depends on a number most teams don't actually have.

The number you need

Most teams can tell me their monthly AI bill. Almost none can tell me their cost per accepted unit of work: per merged pull request, per resolved ticket, per successful agent run. That's the denominator that turns a scary invoice into a decision.
A flat per-engineer cap, the Apple and Uber approach, is a blunt instrument precisely because it ignores this. It throttles the engineer doing high-value agentic refactors the same as the one burning tokens on retry loops. Budgets that don't distinguish productive spend from waste punish your best users first.
Here's what I'd instrument before touching any cap:
  • Tokens per successful outcome, not tokens per seat. Agent loops that retry, re-read the whole repo, or stuff 200K tokens of context to change three lines are where the money leaks.
  • Routing by task difficulty. A cheaper model like Terra for boilerplate, tests and summarization; the expensive tier only where it demonstrably changes outcomes. This is the highest-leverage change and requires no new hardware.
  • Caching and context discipline. Prompt caching and tighter retrieval often beat any model swap.
  • A utilization model for self-hosting. Before you buy GPUs, estimate sustained utilization honestly. Below some threshold, the cloud API wins even at list price. Revolut's economics depend on scale most companies don't have.
And there's a related infrastructure detail worth noting. AWS also shipped Step Functions workflows as agent tools through AgentCore Gateway. MoEngage reported turning about 400 internal API operations into agent tools, and found that listing every tool would have burned over 20,000 tokens per connection, so their agents search for tools by task instead. That's cost engineering in miniature: the cheapest token is the one you never send.

What this really signals

We spent 2025 and most of 2026 asking whether agents could do the work. The question for 2027 is whether they can do it at a unit cost the CFO will keep funding. Those are different skills, and the second one looks a lot more like classic FinOps and systems design than like prompt engineering.
The teams that will win aren't the ones with the biggest budgets or the loudest self-hosting stories. They're the ones who can say, to the dollar, what a merged pull request costs them, and who built the routing, caching and measurement to push that number down every quarter.
Caps are what you set when you can't see inside the bill. The goal is to stop needing them.