Revolut has achieved a 90% reduction in AI costs by shifting from commercial OpenAI and Anthropic services to Chinese GLM models running on dedicated infrastructure, according to David Abitbol, its head of AI operations. The setup serves about 2,000 engineers—and the reported savings came alongside faster responses and shorter development times.
Key takeaways
Revolut runs GLM 5.3 using 56 NVIDIA B200 GPUs.
Average processing costs stand at $0.30 per million tokens.
Initial response latency fell 88% without adding GPUs.
Internally managed models operate with access controls and activity tracking.
How Revolut achieved its AI cost reduction
Revolut’s configuration centers on GLM 5.3 , developed by the Chinese team Z.ai, formerly Zhipu AI. Rather than relying on commercial model services as its default choice, Revolut runs the model on dedicated infrastructure under its control.
Abitbol said 56 NVIDIA B200 GPUs , based on the Blackwell architecture, handle workloads for about 2,000 engineers. The infrastructure is tuned for quick responses and multiple simultaneous requests.
The shift does not mean every external AI tool has disappeared. Claude Code, Codex and Gemini remain available, primarily for tasks where they offer clear benefits over internally operated models.
Faster responses from the same hardware
Between April and August 2026, Revolut reduced its time to first token, or TTFT, from 14.26 seconds to 1.68 seconds while keeping its GPU count unchanged. That was an 88% improvement in initial latency, even as requests per second increased by 76%.
TTFT measures the wait before a model starts responding. Abitbol considers it the most important performance measure, with a target below three seconds; beyond five seconds, responses begin to feel slow. He said generation speeds of around 200 tokens per second are sufficient for normal engineering work, which increasingly involves AI agents whose results developers review later.
One major optimization is FP4 quantization , which stores values using four bits rather than the eight used by FP8, reducing memory requirements. According to Abitbol, moving to FP8 would require about 40% more GPU resources without meaningful accuracy gains for Revolut’s developer workloads.
Revolut also targets a KV cache hit rate above 90% for programming tasks. The Key-Value cache retains intermediate results so the system can reuse previously processed information, cutting latency and GPU load when developers work with recurring code patterns.
Cache capacity, context length and concurrent requests must be balanced. Abitbol noted that some models support up to a million tokens of context, but longer contexts consume more memory and can restrict simultaneous requests. Revolut limits the information processed when extra context is unnecessary. Other techniques considered include speculative decoding and optimization of request-processing phases.
Lower token costs and shorter development times
According to Revolut, the company’s per-token processing expense averages $0.30 per million tokens , while coding speed has risen 60% and development lead times have shortened by 70%. Compared with what it formerly paid for OpenAI and Anthropic services, its total AI expenditure now sits at roughly one-tenth of that amount.
These gains build on a prior deployment of Gemini in January 2025, when early internal data showed coding time dropping by 20% and lead time falling by 10%.
The economics of Claude Code were less compelling at first, as heavy usage led the company to impose spending limits. Revolut managed to cut AI costs further by shifting to internally managed GLM models, applying techniques such as quantization, memory management and inference optimization to squeeze more performance from its existing accelerators before investing in additional hardware.
Internal execution keeps control over the data path
Using a Chinese-developed model does not necessarily mean Revolut sends requests to its producer. Model weights are downloaded and run on selected infrastructure, allowing Revolut to control the data path used for inference.
Abitbol described an architecture built around sandboxing, authorization rules and human oversight . Policies determine which agent operations are permitted or blocked, while telemetry records activity to help reconstruct what the system did and identify anomalies.
Revolut combines internally executed models with controlled access and traceable operations. Its programming agents operate under authorization policies, with telemetry supporting oversight of their actions.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
