David Abitbol, who leads AI operations at Revolut, describes its self-hosted coding-model deployment in this 12-minute AI Engineer talk.

Spend limits changed how engineers used AI

  • His internal series covers 2,000 engineers and distinguishes coding hours from lead time, measured from starting a feature to merge readiness.
  • Expensive frontier-model usage led Revolut to impose strict budgets before introducing self-hosted GLM models.
  • He associates the later rollout with roughly 60% better coding time and 70% better lead time.
  • Lower token pressure let engineers use agents for inboxes, Jira, Confluence, and other work around coding.

The reported economics require scale

  • Abitbol reports roughly 56 B200 GPUs serving the engineering group, with blended token cost around $0.30 per million.
  • He says the invoice became about one-tenth of previous Anthropic and OpenAI spending, the basis of the source title’s 90% claim.
  • Frontier tools remain available as exceptions. The deployment changes the default rather than eliminating those services.
  • Target time to first token is under three seconds; delays above five seconds prompt engineer complaints.

Serving configuration matters

  • On unchanged GPU count, Abitbol reports substantially lower first-token delay while request demand increased.
  • Tune context length, quantization, parallelism, and prompt-cache reuse against the actual workload.
  • Quantization stores model numbers at lower precision; tensor parallelism spreads a model across GPUs.
  • Fitting weights into memory leaves another question: enough room for the KV cache, the model’s stored prompt state, under concurrent load.
  • Route image work to a suitable model while retaining the chosen coding model for text-heavy tasks.

Put controls around the model

  • He recommends compatible APIs so existing coding tools can use the self-hosted endpoint.
  • His governance approach uses sandbox restrictions, human approvals, and detailed telemetry for incident investigation.
  • Hosting downloaded weights locally changes the data path, but retention and compliance still require checking the actual service and agreement.

“There’s no secret recipe.” — David Abitbol, 7:08

Caveat: these are Revolut’s reported internal results, without a controlled comparison or full operating-cost breakdown. They do not establish a universal 90% saving or regulatory approval.