How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models - David Abitbol
B · Useful, limited First-hand talk by Revolut's head of AI operations with concrete serving and cost figures, but the productivity gains are the company's own metrics with no control group and no quality comparison.
Summary based on the full transcript·Original: about 12 minutes to watch
In short
A 12-minute AI Engineer talk (uploaded 29 October 2026) by David Abitbol, who leads AI operations at Revolut. The bank capped Claude Code spending in early 2026 and saw little effect. From April it self-hosted open GLM models on dedicated GPUs and gave engineers unrationed tokens. It reports large gains in its own delivery metrics and a token bill about a tenth of the earlier frontier invoice. Frontier tools remain in use "as the exception".
Key numbers
- 56 B200GPUs serving GLM-5.3 to about 2,000 engineers
- 30 cents per millionblended cost per million tokens; invoice about 1/10 of previous Anthropic and OpenAI spend
- 60% and 70%reported improvement in coding time and lead time after self-hosting from April 2026
- over 10 s to around 1.6 stime to first token, April to August, on the same 56 GPUs while requests per second rose from about 3 to over 7
As reported by the source.
English·Video·opens youtube.com
What it reports
- Gemini (January 2025) brought 10% better lead time and 20% better coding time. Claude Code at scale under strict budget limits showed little measurable impact; some engineers had been spending a thousand to ten thousand dollars a month.
- Self-hosted GLM-5.1 from April 2026: 60% better coding time, 70% better lead time. The speaker reads the larger lead-time gain as time saved on inbox, Jira and Confluence work, not coding.
- Sizing: 56 B200 GPUs for about 2,000 engineers on GLM-5.3. Targets: first token under 3 s (complaints start above 5 s), 200 tokens a second; about 800 tokens a second after 5 pm.
- Blended cost 30 cents per million tokens; invoice 1/10 of previous Anthropic and OpenAI spend.
- Tuning on fixed hardware, April to August: "88% improvement in performance for 76% increased demand".
- Parameters: 256K context is enough despite 1M support; NVFP4 on Blackwell; TP8 over TP4 for KV-cache headroom; cache hit above 90% for coding.
- Images routed to other models since GLM is text-only; endpoints are OpenAI-compatible and used in OpenCode and Codex.
- Compliance: weights run on non-Chinese infrastructure; a sandbox with per-operation allow and deny rules keeps a human in the loop for the regulator.
How they got it
Internal engineering metrics for a sample of 2,000 engineers over a year (lead time from pick-up to merge, coding time), plus serving telemetry. Presented as charts in a conference talk; transcript from auto-captions with mis-hearings fixed.
Limits
- Productivity numbers are self-reported before-and-after figures with no control group; other changes over the same period are not separated out.
- No quality comparison between GLM and frontier models on Revolut's own work.
- The cost figure excludes staff, observability and whether GPUs are owned or rented.
- The speaker's 40% GPU-footprint saving from FP8 to FP4 "for zero loss" is not supported with data.
Why it matters for companies
Per-seat or per-token rationing can suppress the use that produces value; a flat-cost internal service changed behaviour more than access to the best model did. At this scale, open-weight models on dedicated GPUs can cut token cost by an order of magnitude, but the saving depends on serving expertise most firms do not have. Regulated firms can borrow the compliance framing: a sandbox that enforces human approval makes coding agents a supervised tool rather than an autonomous system.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com