How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models - David Abitbol, Revolut

Summary

David Abitbol, who leads AI operations at Revolut, describes how the company expanded AI coding tools to roughly 2,000 engineers and reduced costs by self-hosting open models.

Productivity gains and lead time

Revolut’s internal tracking compares coding time with total feature lead time—from when an engineer starts work until a feature is merged and ready to deploy.

• Gemini initially improved coding time by about 20% and execution time by 10%.
• Later, self-hosted GLM models were associated with improvements of roughly 60% in coding time and 70% in execution time.

The larger improvement in end-to-end execution suggests coding is not the only bottleneck. AI can also help with surrounding work, such as searching documentation and handling information in tools like Jira and Confluence. Abundant, lower-cost tokens may also make engineers less inclined to ration AI use.

Cost, capacity, and performance

Revolut’s setup uses about 56 NVIDIA B200 GPUs to serve a large GLM model, described as having around 800 billion parameters. Abitbol says self-hosting costs about $0.30 per million tokens—roughly one-tenth of the company’s previous spending on Anthropic and OpenAI. Revolut still uses frontier models selectively.

The team targets a first-token response time under three seconds; Abitbol says latency above five seconds frustrates engineers. Around 200 tokens per second is adequate for interactive work, while off-peak throughput can be much higher.

Loading model weights onto GPUs without tuning produced poor performance. Using the same 56 GPUs, Revolut increased throughput from roughly three to more than seven requests per second, while reducing time to first token from over ten seconds to about 1.6 seconds. Abitbol characterizes the improvement as 88% while demand rose 76%.

Inference optimization

Abitbol recommends not assuming that a model’s maximum context window is useful for everyday workloads; his team often needs far less than a million tokens. For Blackwell-generation GPUs, he favors FP4 quantization, which he says saves about 40% of GPU resources compared with FP8 without sacrificing accuracy.

He also discusses tensor parallelism, including TP8 when KV-cache capacity is a constraint, cache-hit rates above 90%, and techniques such as speculative decoding and PagedAttention.

Integration, compliance, and safety

The self-hosted coding model is text-only, so Revolut routes image-recognition tasks to other models when needed. Its serving endpoints are OpenAI API-compatible and can be used with coding applications that support that interface.

For a regulated bank, Abitbol emphasizes human involvement, sandbox rules that restrict prohibited operations, protections against misuse and prompt injection, appropriate data-location controls, and detailed telemetry for incident analysis. He says providers used in the setup do not store data by default. Hosting model weights on a US-based service also avoids sending prompts to a dedicated endpoint in China.

Talk overview

This is a technical implementation and operations talk, not a product review. It offers practical guidance on GPU sizing, latency targets, inference optimization, integration, and compliance.

Main speaker and source: David Abitbol, AI Operations at Revolut. The productivity and deployment figures are presented as Revolut’s internal experience.