Revolut's self-hosted GLM costs 30 cents per million tokens: the math

Revolut now runs most of its AI coding on an open-weights model it hosts itself, and its bill is a tenth of what it was. David Abitbol, who leads AI operations at the bank, told the AI Engineer NYC conference that Revolut started self-hosting Z.ai's GLM 5.1 around April and now serves GLM 5.3 to about 2,000 engineers on roughly 56 Nvidia B200 GPUs. He said the invoice is now "one-tenth" of what Revolut had been paying Anthropic and OpenAI. The blended cost is 30 cents per million tokens. The talk went up on October 29 and The Register reported it on November 3.

The reason for the switch will be familiar to anyone who has rolled out Claude Code or Codex. Abitbol said engineers were spending "a thousand a month, 10,000 a month," which pushed Revolut into strict budget limits. Self-hosting, he said, means "the billing is flat." Codex and Claude Code are still used, but as exceptions. We found no Revolut statement or blog post; the figures here are from the talk.

What the alternatives cost at list price

Seat fees are the floor, before any usage. Claude's Team plan stops at 150 seats, so a 200-developer company is on Enterprise.

The usage line is where Revolut's money went. Claude Enterprise bills usage at API rates on top of the seat, and GitHub warns that developers running agentic workflows will probably use more than the included credits. Z.ai's public pages we checked do not list team seat prices, so the GLM row uses its individual plan ($56 a month if billed yearly).

Per token, Revolut's 30 cents compares with these API list prices per million tokens:

At Opus 5.5's output price, one engineer's $1,000 a month buys 50 million output tokens. At Revolut's blended rate it buys about 3.3 billion. The comparison flatters self-hosting: StartupHub.ai notes that the 30 cents leaves out cloud spend, staffing and observability, and the conference site says it is "not a portable price."

What self-hosting takes

GLM-5.3 is a 753-billion-parameter model, according to its Hugging Face model card. Its license lets companies run, deploy, fine-tune or modify it. Companies that run a model-as-a-service business and have more than $10 billion in revenue over 12 months must pass a Z.ai security review before any commercial use. The Register puts the earlier GLM 5.1 at roughly 380 GB of GPU memory for the weights alone, plus 6 to 26 GB per user for context.

Revolut's ratio works out to about 35.7 engineers per B200. Getting there took tuning. Abitbol advised capping context at about 256K tokens and using FP4 for a "40% saving in GPU footprint," and noted that although GLM fits on four GPUs, eight may work better because it leaves more room for cache. On the same 56 GPUs, he reported performance up 88% while demand rose 76%.

Abitbol said Revolut's coding tasks always run with a human in the loop, enforced by sandbox approval rules. Addressing what he called the stigma around Chinese open models, he said companies are only taking the weights and running them on a US-based service.

Related: Goldman limits Claude Code and Copilot to engineers and GitHub Enterprise now includes Copilot.