Revolut moves to self-hosted GLM (twitter.com)

Revolut has transitioned approximately 2,000 of its engineers to a self-hosted General Language Model (GLM) for coding purposes, while relegating Claude Code and Codex to exceptions. This shift, led by AI operations head David Abitbol, utilizes 56 Nvidia A100 GPUs and achieves a cost-effective rate of 30 cents per million tokens—around one-tenth of the fees previously paid to Anthropic and OpenAI. The move has resulted in a remarkable 70% improvement in engineering lead time, particularly as engineers became less conservative with AI utilization when token limitations were lifted. This significant change underscores the potential for organizations to reduce costs and enhance efficiency by leveraging self-hosted models instead of relying on expensive frontier APIs. As the AI/ML community continues to explore cost-effective solutions, Revolut's successful implementation illustrates an emerging trend where companies might prioritize open weights over expensive proprietary models. This could challenge existing pricing structures in the AI industry, prompting companies to reconsider their strategies as they edge towards IPOs.