Revolut moves coding work for about 2,000 engineers to self-hosted GLM, citing costs about one-tenth of before (community reports)

Verified (fact)

According to a GeekNews summary, UK fintech Revolut has switched the AI coding tool used by about 2,000 engineers to a self-hosted GLM model, relegating Claude Code and Codex to exceptional cases. David Abitbol, head of AI operations, said it runs on 56 NVIDIA B200s at a combined input-and-output cost of $0.30 per million tokens, reportedly about one-tenth of what it paid Anthropic and OpenAI. He also said engineering lead time fell 70%, without explaining how it was measured. He said the gains came after token usage limits were removed and engineers began using AI for every task.

Needs fact-check (claim)

Revolut claims costs fell to about one-tenth and lead time improved by 70%, without disclosing how these were measured.

Why it matters (interpretation)

This is a case of a major financial company sharply cutting costs by running an open-weight model itself instead of using frontier APIs. As coding agent usage grows, more companies are likely to consider self-hosting.

Sources

• Community Revolut, 자체 호스팅 GLM으로 전환 GeekNews

More about Revolut

• 2026-10-22 P2 Infrastructure Unconfirmed AI platforms & solutionsEfficiency & infraScience & bio NVIDIARevolut
  After GTC Berlin, NVIDIA shares European open-model adoption cases in finance, retail, public sector and weather, plus 'DSX' cases targeting the power bottleneck

  NVIDIA published five posts on European industry adoption cases to coincide with GTC Berlin. In finance, Revolut is rebuilding its...