ShanghAI - Revolut made the switch: Our analysis of the shock
The model was ten times cheaper. It’s not the most important.
Hey Lobsters, welcome to this new edition of ShanghAI.
You have seen the story everywhere: 2,000 engineers, 56 Nvidia GPUs, a Chinese model running on Revolut's own machines, and a 90% cut to the AI bill. Read the replies and you would think the lesson was "Chinese models beat Claude."
That is not what happened.
Before the switch, Revolut ran Claude Code with a cap of $300 per engineer per month. On stage at GTC Berlin, the man who runs AI operations there said the quiet part out loud: the cap "was not enough to carry an engineer through the whole month."
Engineers had been spending $1,000 to $10,000 a month before anyone capped them.
So they rationed. They kept the hard refactor and skipped the boring 60% - the Confluence search, the Jira ticket, the spec buried in Slack. Not because they were lazy. Because the meter made those tasks feel expensive.
Then Revolut removed the cap, and moved the default onto a model it runs itself. Coding time improved 60%. Lead time improved 70%.
Now the part you are right to bring up: the price. Serving that model in-house took the cost of a million tokens from roughly $3 on a frontier API to $0.30 - a tenfold cut, and it is the number everyone quotes. It is real. You do not get to a 90% smaller invoice without it.
But a cheaper token does not explain the productivity chart. A cheaper token makes the same decisions cheaper. It does not make an engineer decide that the Jira ticket is finally worth delegating. That is what the cap was doing.
Read the two numbers in order: coding time +60%, lead time +70%. Delivery got faster than code. A better model does not do that, and a cheaper one certainly does not. Only a budget without a meter makes the work around the code faster.
That is this edition: a budget cap does not reduce spend. It reduces the number of things your team thinks are worth spending on. And that number never appears on an invoice.
OUR PARTNER
CHATFEATURED’ AUTONOMOUS AEO/GEO AGENT
The AEO work, without the AEO hire.
Most AI visibility tools show you where you rank, then leave the work to you.
ChatFeatured is an AI search agent that does it: finds the questions competitors win, drafts the content in your voice, and tracks the lift.
Your team decides what ships. Month to month.
HOW IT WORKS
THE THREE THINGS A CAP ACTUALLY DOES
Revolut's talk is fifteen minutes of serving engineering. The part that transfers to any company is what a meter does to people.
It hands the decision to the most tired person in the room.
Give tokens a price and every engineer becomes a procurement officer. At 4pm on a Friday, a mid-level developer decides whether a task is "worth it".
That is your AI strategy, being executed by whoever is least paid to think about it, at the hour they have the least patience. Revolut's own numbers show what that produces: a $300 cap that saved the invoice and quietly cost the quarter.
It saves the number you can see and spends the one you cannot.
The invoice is a line item. The nine Jira tickets nobody delegated are not. That is why the lead-time number matters more than the coding number.
When the work around the code speeds up more than the code itself, you did not buy a better model. You bought back the tasks people had been skipping. Finance counted the saving. Nobody counted the queue.
It teaches a habit that outlives the cap.
Six months of rationing builds a reflex: ask the model only when stuck. Remove the cap and the reflex stays. This is the failure mode of most "we gave everyone Copilot" rollouts.
The tool arrived with a meter attached - a seat price, a message limit, and a confirm dialog before the good model. You shipped a discount and called it a capability.
WHAT TO USE
TOP 3 WAYS TO STOP METERING YOUR ENGINEERS - NOVEMBER 2026
The price spread is why this story keeps circulating. Per million tokens: top frontier is $5 in, $25 out. Hosted GLM-5.3 is $1.40 in, $4.40 out. The cheap tier starts at 10 cents in. Top to bottom, that is a 30x difference for a lot of routine coding work.
BEST FOR TEAMS UNDER 50: a flat seat, not a meter.
One fixed-price coding plan per engineer - Claude Code Max, Codex, or a GLM coding plan - plus an explicit instruction: stop asking permission. The point was never the model. The point is the absence of a counter next to the button.
BEST FOR 50-300 ENGINEERS: a menu, not a meter.
Route by task. Cheap hosted tiers do the mechanical 80%: GLM-5.3-Flash or GPT-6 Luna, both around 10 to 15 cents per million input tokens. Blended, that lands near 6 to 8 cents (estimate) - a third of Revolut's $0.30, with no GPUs to buy and no on-call rota. Keep a frontier model for the 20% that needs judgement, behind a written rule rather than a manager.
ShanghAI CHOICE: the two-envelope budget.
One flat interactive seat per engineer, no per-person quota. One shared budget for agents, tests and background work. Frontier access gated by a rule - "security review, production incident, architecture" - not by a spend limit. Flat on the humans, metered on the machines.
APPLY TO YOUR BUSINESS
THE 4-STEP REVOLUT PLAYBOOK
You do not need 56 GPUs to run this experiment. You need one week and a spreadsheet.
Step 1: Find every meter.
Seat caps, workspace spend limits, approval queues, "advanced model" toggles, the colleague who owns the API key. List them. Most companies have four to six and only know about one.
Step 2: Price the rationing.
For one week, ask five engineers what they did not ask the model to do. Multiply the hours by loaded cost. That is your shadow bill, and it is usually bigger than the invoice. Revolut's cap was $7.2M a year on paper (estimate: 2,000 x $300 x 12). Removing it paid because of everything the cap was suppressing.
Step 3: Replace per-person meters with one team envelope and one written rule.
Flat, visible, no per-task judgement. Write the escalation rule down so nobody has to guess whether today is a "good model" day.
Step 4: Measure delivery, not tokens.
Track lead time and cycle time, not spend per seat. Borrow Revolut's service target: under 3 seconds before the answer starts typing on screen. Their ops lead's own threshold: at 5 seconds, "my Slack inbox explodes." That is the only SLA in this story that survives contact with a 20-person team.
If your leadership team wants to audit your enterprise AI privacy, escape vendor lock-in, or build a cost-effective European sovereign AI stack:
Until next week,
The ShanghAI Guy
