Revolut Cuts AI Coding Costs 90% With Self-Hosted GLM-5.3

Revolut has cut its AI coding inference costs by approximately 90% in less than a year by replacing commercial AI coding services with a self-hosted deployment of Z.ai's GLM-5.3 model. The London-based fintech now serves roughly 2,000 engineers through 56 Nvidia B200 GPUs, delivering inference at $0.30 per million tokens compared to monthly commercial seats that previously cost $1,000 to $10,000 per engineer.

David Abitbol, Revolut's operations and infrastructure lead, told The Register the economics are stark. Engineers who had been spending thousands monthly on Claude Code and OpenAI's Codex now access a 754-billion-parameter mixture-of-experts model with 40 billion parameters activated per token, achieving sub-three-second time to first token and throughput near 200 tokens per second during business hours, rising to 800 tokens per second on evening shifts.

Release Date and Deployment Scale

Officially launched on November 3, 2026, the deployment spans seven eight-GPU DGX systems running Nvidia's B200 architecture. The model requires approximately 380 GB of GPU memory for weights, with an additional 6 to 26 GB per user allocated for context windows. Revolut caps recommended context length at 256,000 tokens, a deliberate constraint Abitbol defends bluntly: "Long context is a scam. You really need to be mindful of that."

Performance Optimization Strategy

The infrastructure stack achieves its performance through several compression techniques. NVFP4 quantization reduces memory overhead while maintaining output quality, and Revolut employs disaggregated prefill and decode operations to improve resource utilization. Hyperparameter tuning and strict context-length limits keep latency predictable. Abitbol maintains a hard service-level objective: "As soon as you exceed 5 seconds, my Slack inbox starts to blow up." The team treats sub-three-second TTFT as a non-negotiable standard.

Business Impact and ROI

Revolut credits the GLM-5.3 rollout with a 60% improvement in coding time and a 70% improvement in lead time across its engineering organization. The company eliminated the unpredictable monthly bills that characterized its commercial AI tooling contracts, replacing variable costs with flat infrastructure expenses. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI," Abitbol said. The self-hosted approach delivers what Revolut describes as "favorable economics" with zero bill shock at month-end.

Open-Source Infrastructure Priority

Abitbol positions the deployment as part of Revolut's broader shift toward open-source AI infrastructure. "Open source is now a priority at Revolut, and our economics are favorable," he told The Register. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception." The statement reflects a pattern visible across enterprise AI deployments in 2026, where large organizations increasingly run inference workloads in-house rather than relying on third-party API providers.

What This Means

Revolut's deployment offers a concrete data point in the ongoing debate over commercial versus self-hosted AI infrastructure. For organizations with sufficient engineering scale and GPU access, the cost differential between API-based services and in-house inference can exceed an order of magnitude. The company's willingness to accept a 256K context ceiling and invest in operational discipline around latency targets suggests that performance optimization at scale remains as much a systems engineering challenge as a model selection problem. As enterprises evaluate 2026 AI budgets, Revolut's 90% cost reduction provides a benchmark for what infrastructure investment can deliver when inference volume justifies the upfront GPU capital expenditure.