# Revolut says self-hosted AI cut coding costs 90% _Revolut says 56 B200s and GLM 4.5 cut coding AI costs to 30 cents per million tokens for 2,000 engineers, but only after heavy tuning._ **Published:** 2026-10-29 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/revolut-says-self-hosted-ai-cut-coding-costs-90 --- [AI Engineer](https://www.youtube.com/watch?v=Ok6QdZHCcI8) host David Abitbol said [Revolut](/startups/revolut) cut AI coding costs 90% by moving 2,000 engineers off metered APIs onto self-hosted open models on 56 Nvidia B200 GPUs. The system was demoed as a production coding assistant for Revolut's engineers, not a lab test. Abitbol said the bank serves GLM 4.5 at about 30 cents per million tokens, or one tenth of its prior Anthropic and [OpenAI](https://www.startuphub.ai/startups/openai) invoice. He put the performance target at sub 3 seconds to first token and about 200 tokens per second for interactive use, with bursts to 800 tokens per second after 5 p.m. when load drops. The setup is OpenAI compatible and routed through tools like OpenCode and Codex. The punchline is the payoff. Abitbol said Gemini alone delivered roughly 10% faster lead time and 20% faster coding time in early 2025, while early Claude Code with strict budget caps showed little impact. After self-hosting GLM in April, Revolut saw about 60% improvement in coding time and 70% in lead time on a 2,000 engineer sample, with time to first token falling from over 10 seconds to 1.6 seconds even as requests per second rose from 3 to over 7. He attributed the larger lead time gain to tasks outside pure coding, like inbox triage and meetings, that engineers now automate when tokens are not rationed. The skeptic case is what it takes to get there. Abitbol said out of the box inference is terrible and the team improved performance 88% for 76% more demand without adding GPUs by tuning context limits, quantization and parallelism. He called long context a scam for this workload and capped it at 256k, said NVFP4 on Blackwell saves about 40% GPU footprint versus FP8 with no precision loss for this use, and noted GLM needs TP8 not TP4 to fit KV cache. The bank, which now serves more than 80 million customers and lets users hold up to 36 currencies and send to 160 plus countries per its site, as of July is Europe's largest private company, but the talk left cloud spend, staffing and observability costs outside the 30 cent token math. For security, Abitbol framed the threat as local not remote. The affected system is the coding agent harness used by employees, not customer banking cores. He said Revolut enforces a sandbox where every agent run is human in the loop, with allow lists like permitting git add but blocking broader destructive commands to satisfy EU regulated bank requirements and to limit rogue insider use. Prompt injection remains a local data access risk, he said, and the mitigation is jurisdictional: keep inference in countries covered by the data equivalency framework such as the US, EU, UK and Israel, rely on GPU providers with zero data retention by default, and instrument strong OpenTelemetry for incident forensics. He added that image tasks are routed to [Gemini](https://www.startuphub.ai/ai-news/ai-figures/2026/figure-sundar-pichai-google-io-2026-keynote-recap-2026-06-01) or similar models instead of forcing a vision capable open model, which keeps the self-hosted path text only. What remains unproven is portability. The numbers assume 56 B200s, steady tuning and a bank that can dictate sandbox rules and work hours to shape load. Smaller teams without that leverage will still face bill shock or slow first tokens, and regulators will still want proof that human in the loop is more than a config flag. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.