Revolut moves 2,000 engineers to self-hosted GLM, cuts coding AI cost ~90%
joecole · x · 2026-11-04
- Per David Abitbol, who leads AI ops at Revolut, 2,000 engineers were moved onto self-hosted GLM for coding, with Claude Code and Codex pushed to the exception list.
- Setup and cost: 56 Nvidia B200s, blended $0.30 per million tokens — roughly a tenth of what they paid Anthropic/OpenAI. Engineering lead time improved 70%.
- The gains showed up once tokens stopped being rationed: engineers stopped saving AI for hard tickets and pointed it at everything.
- Quoting karinanguyen's bigger thesis: every company should turn itself into a giant RL env, building self-improving AI that learns from every interaction; "frontier models become the exception, open weights the default" — a pricing problem for labs heading to IPO.
More from coding & agent
- First external PR to Void lets static pages skip the Worker, served free by Cloudflare — cnakazawa · 2026-11-04
- POA network hits scaling challenges as its AI agents debate and fix problems on their own — adamamcbride · 2026-11-04
- Freelancer Seeks a Production-Ready AI Agent Workflow for Client Websites — Downtown-Average-812 · 2026-11-04
- As new code cost nears zero, engineers stop collaborating and rebuild the same systems — alex_frantic · 2026-11-04
- Baseten's Philip Kiely: agents need gateways, sandboxes and post-training, not just inference — AccBalanced · 2026-11-04
- LlamaIndex launches OpenDocRouter, a unified API that parses PDFs into Markdown — llama_index · 2026-11-04