Revolut cuts AI coding costs 10x by self-hosting GLM

November 3, 2026. The Register reports that Revolut, the British fintech, cut its AI coding inference costs tenfold in under a year by making self-hosted models the default instead of leaning on commercial services like Claude. October 2026. David Abitbol, its operations and infrastructure lead, lays out the numbers at the AI Engineer conference in New York. 2025–2026. The company had first tested Gemini in early 2025, then Claude Code in early 2026. Why it matters: this is the first hard-number, at-scale proof that self-hosting an open-weight model for coding becomes economically defensible — but only under specific conditions.

The switch: the default changed, not the menu

What Revolut did was not add a cheaper model alongside the others. It flipped the default. Self-hosted models now handle AI coding work first, and commercial services — Anthropic Claude, OpenAI Codex — sit behind them. This is not a pilot or an experiment: a default decides where most of the token volume goes.

The deployment serves roughly 2,000 engineers on 56 NVIDIA B200 GPUs, equivalent to seven eight-GPU DGX B200 systems. The model is GLM-5.1 from Z.ai, a China-developed open-weight family that Revolut later moved to GLM-5.3. Treating the stack as something to upgrade on your own schedule, without waiting for a vendor roadmap, is part of the payoff.

The point of comparison matters. Before the switch, some engineers were burning $1,000 to $10,000 a month on commercial APIs “with not so much ROI,” in Abitbol’s words. The initial rollout therefore ran with strict budget limits — and therefore token quotas. At $0.30 per million tokens self-hosted, the cost drops to about one-tenth of OpenAI Codex or Anthropic Claude Code.

The behavioral economics of cheap tokens

The least expected result is not the price but what the price does to usage. “When you have unlimited tokens, you’re free to do that,” Abitbol says: engineers started handing the models debugging, Jira tickets, and Confluence tickets — the tasks nobody delegated when you had to save your tokens for the end of the quarter or the most difficult task.

That is where self-hosting changes character. A metered service pushes you to ration; a fixed capacity, paid up front, pushes you to use. The productivity gains reported — 60 % less coding time, 70 % less lead time across a sample of about 2,000 engineers — should be read with caution: the baseline is not explained, and we do not know whether the same engineers would have matched those gains on Claude Code or Codex. But they suggest a cheaper model did not hurt throughput, which is already a signal.

The mechanics of “ten times cheaper”

The model is not small. GLM 5.1 weighs roughly 754 billion total parameters (40 billion active per generated token), which demands ~380 GB of GPU memory for the weights alone, plus 6–26 GB per user for context. Serving that to 2,000 engineers is partly hardware, but mostly engineering.

Abitbol is explicit: the biggest gains did not come from adding GPUs, but from tuning hyperparameters, switching to NVFP4 quantization, implementing disaggregated prefill and decode, and tuning context lengths. His verdict on giant contexts is blunt: “long context is a scam.” Even though you can serve one million tokens, “256 [thousand] is really what you need.”

The result: sub-3-second time to first token on contexts of about 256,000 tokens, and generation around 200 tokens per second — the “sweet spot” for his engineers. The target is explicit: “what you want to aim for is a sub-3-second time to first token. As soon as you exceed 5 seconds, my Slack inbox starts to blow up.” Engineers who sign in after 5 pm get closer to 800 tokens per second: the cluster, paid up front, never idles.

The accepted limit: GLM does not do everything

GLM has not replaced GPT or Claude everywhere. The model is text-only, so it does not handle vision or UI work. Revolut therefore put in place a model router that sends vision queries to Gemini or GPT Luna while keeping code generation on GLM. “Open source is now a priority at Revolut, our economics are favorable, the billing is flat, there is no bill shock at the end of the month, and the frontier is the exception,” Abitbol says.

On the sensitive question of Chinese open-weight models, he draws the line: the real risk is accessing these models through APIs hosted out of China. Run on-premises or on regional cloud resources, “there’s really nothing to worry about.” The corollary follows: self-hosting also keeps your prompts and your code inside your own network.

Verdict

Revolut’s result is a data point, not a universal law. It proves that at the scale of 2,000 engineers, with an infrastructure team able to keep 56 B200 GPUs busy, self-hosting an open-weight model like GLM can cut the bill tenfold — without breaking throughput. But “ten times cheaper” is a token-price comparison, not a total cost of ownership: cloud, personnel, hardware, and operations costs may be excluded.

If you exceed a thousand heavy users and have the team to keep a GPU cluster loaded, the self-hosted path holds up: pick an open-weight model, enable NVFP4 quantization, set up disaggregated prefill and decode, cap context at ~256,000 tokens, and keep a model router for vision. If you are a 50-person team, the math does not work: a GPU cluster that sleeps at night costs the same as one running flat out, and the cost advantage dissolves. In between, the right posture is hybrid — self-host what is textual and stable, keep a commercial model as fallback. The underlying rule: self-hosting inference is a bet on utilization, not on GPU price.

References

• The Register — British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted (November 3, 2026)
• TechRadar — Revolut cut Claude for self-hosted Z.ai models and slashed inference costs by 90 %
• Cybernews — Revolut Chinese AI model cuts coding costs by 90 %
• selfh.st — Self-Host Weekly (November 6, 2026)