Revolut has moved roughly 2,000 engineers off metered AI APIs and onto self-hosted open-weight models, and says the switch cut its AI coding costs by 90%. David Abitbol, who leads AI operations at Revolut, said in a talk published by AI Engineer on October 29 that running Z.ai's GLM models in-house brought the company's blended cost to about 30 cents per million tokens, one-tenth of what it had been paying Anthropic and OpenAI.
The setup runs on about 56 Nvidia B200 GPUs and serves the full engineering group as a production coding assistant. Self-hosted GLM is now Revolut's default for coding work, with frontier tools such as Claude Code and Codex treated as the exception. Revolut serves about 80 million customers, and a secondary share sale that closed in July 2026 valued the company at $115 billion.
Revolut got there in stages. Gemini, adopted in early 2025, delivered roughly a 10% gain in lead time and a 20% gain in coding time across the engineer sample. Claude Code was trialed around February 2025 and launched at scale in January 2026, but showed little measurable impact at first, while engineers were spending $1,000 to $10,000 a month each on frontier APIs.
Key facts
- 01Revolut says self-hosted GLM models cut its AI coding costs 90%, to a blended 30 cents per million tokens.
- 02About 56 Nvidia B200 GPUs serve roughly 2,000 Revolut engineers as a production coding assistant.
- 03After self-hosting began in April 2026, coding time improved about 60% and lead time about 70%.
- 04Between April and August, time to first token fell from over 10 seconds to about 1.6 seconds with no added GPUs.
- 05The 30-cent figure excludes cloud spend, staffing and observability costs.
After Revolut began self-hosting GLM in April 2026, coding time improved about 60% and lead time about 70% on the same 2,000-engineer sample. Abitbol attributed the larger lead-time gain to work outside of writing code, such as inbox triage and meetings, that engineers now hand off to agents.
“What we find is the bottleneck for engineering is not necessarily coding”— David Abitbol, Leads AI operations at Revolut
Out-of-the-box inference performed poorly, according to the talk. Between April and August, Revolut held its GPU count at 56 while cutting time to first token from over 10 seconds to about 1.6 seconds and raising throughput from about 3 requests per second to more than 7. The gains came from tuning context limits, quantization and parallelism rather than buying more hardware.
“We've had 88% improvement in performance for 76% increased demand”— David Abitbol, Leads AI operations at Revolut
The specifics are aimed at coding workloads. Revolut capped context at 256,000 tokens, and Abitbol said coding work needs a cache hit rate above 90%. Running NVFP4 on Blackwell saves about 40% of GPU footprint compared with FP8 without a loss in precision, he said. The latency targets are under 3 seconds to first token and about 200 tokens per second for interactive use, with bursts to about 800 tokens per second after 5 p.m.
Because Revolut is a regulated bank, the deployment sits behind several controls. Inference runs only on GPUs located in the US, EU, UK or Israel, through providers that retain no data by default, and the team uses OpenTelemetry for forensics. Every agent run is human-in-the-loop inside a sandbox with allow lists, and the system covers the employee coding-agent harness, not customer banking systems. Image tasks still route to Gemini or similar models, keeping the self-hosted path text-only.
The 30-cent figure excludes cloud spend, staffing and observability costs, and Revolut has not disclosed whether it owns or rents the cluster. One estimate puts 56 B200 GPUs at roughly $2.7 million in hardware before power and networking, or $245,000 to $327,000 a month to rent at $6 to $8 per GPU-hour, about $120 to $165 per engineer.
Revolut is not the only company leaning on open models. Decagon co-founder Jesse Zhang wrote on X in August that the company's workflows are roughly split 90% open source and 10% closed source, while adding that it still needs a minimum amount of frontier tokens for new or novel projects.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
