Revolut says self-hosted AI coding models cut its inference costs by about 90 percent

Revolut, speaking at the AI Engineer conference, says it moved its roughly 2,000 engineers' coding workload from commercial services such as Claude◆ Code and Codex to self-hosted GLM-5.1 and later GLM-5.3◆, served on 56 Nvidia◼ B200 GPUs, cutting inference costs by about 90 percent. Optimisations included NVFP4 quantisation, disaggregated prefill and decode, hyperparameter tuning and a roughly 256k-token working context, with sub-3-second time to first token. The firm notes that the privacy risk lies mainly in calling Chinese-hosted model APIs from outside China, and that on-prem or regional hosting mitigates it. No CVEs are mentioned.

For security analysts · Written by AI from 1 source · Wrong?