Revolut moved 2,000 engineers off Claude Code and Codex — and cut AI coding inference costs by 90%. The numbers come from a GTC Berlin presentation on October 22, 2026 by David Abitbol, the company’s infrastructure lead. The model is GLM-5.3 from Chinese firm Z.ai. The cost: $0.30 per million tokens, against roughly $3/M for the commercial tools they replaced. That’s not a pilot. That’s a production system serving a fintech with $6 billion in annual revenue.
The Numbers That Matter
Before the switch, Revolut engineers were spending anywhere from $1,000 to $10,000 per person per month on AI coding tools. The bills were unpredictable and the ROI was murky. Self-hosting flipped the model: instead of variable SaaS charges, Revolut now pays flat infrastructure costs and runs the inference on its own hardware.
The result is $0.30 per million input/output tokens on the self-hosted stack, compared to approximately $3/M on commercial alternatives. At 2,000 engineers with heavy daily usage, the difference compounds fast. Abitbol’s framing is worth noting: “When tokens are cheap, you stop worrying.” Cheap tokens didn’t just save money — they changed how engineers use the tool. Debugging, ticket work, documentation: tasks that felt too expensive to AI-assist became routine.
The Model: GLM-5.3
GLM-5.3 is a mixture-of-experts model from Z.ai (formerly Zhipu AI) with 753 billion total parameters and roughly 40 billion activated per token. Its architecture runs 78 decoder layers — three dense MLP layers plus 75 MoE layers — with 256 routed experts per layer, top-8 routing, and DeepSeek-style sparse attention. The maximum context window is 1 million tokens, though Revolut deliberately caps it at 256K in production.
The practical appeal for self-hosters: GLM-5.3 exposes an OpenAI-compatible API via vLLM or SGLang, meaning it’s a drop-in replacement for any codebase using the OpenAI SDK — just change the base URL. Getting it running on a vLLM cluster:
python -m vllm.entrypoints.openai.api_server --model ./glm5-1-fp8 --served-model-name glm5-1-fp8 --tensor-parallel-size 8 --quantization fp8 --enable-expert-parallel --max-model-len 32768 --port 8000Note the --enable-expert-parallel flag. Without it, tensor parallelism copies every expert’s weights to every GPU — a memory waste that kills batch capacity on MoE models.
The Infrastructure: 56 B200 GPUs
Revolut deployed 56 Nvidia B200 GPUs (192GB HBM3e each), disaggregated prefill and decode operations, and used NVFP4 quantization to fit the 753B model into available memory. An 8-GPU B200 server runs approximately $390,000 — Revolut’s total hardware investment is estimated around $2.7 million before power, networking, and engineering time.
The latency target was strict: responses exceeding 5 seconds triggered an escalation. After deployment, time-to-first-token on coding requests dropped from over 10 seconds to 1.6 seconds. Throughput reached 200 tokens per second during business hours and 800 tokens per second in the evenings. Requests per second went from 3 to over 7.
The Model Router: Hybrid by Design
Revolut didn’t try to replace everything. The self-hosted GLM stack handles text and code generation — the high-volume workload. Vision and image tasks stay routed to external providers (Gemini, GPT Luna). This is a deliberate hybrid architecture: use self-hosted inference where the token volume justifies it, and commercial APIs where the capability gap is real and the volume is manageable.
It’s also the more honest framing for most organizations: self-hosting is not an all-or-nothing decision.
The Context Length Trap
Abitbol’s most quotable line from GTC Berlin: “Long context is a scam. You really need to be mindful of that.”
GLM-5.3 supports 1 million tokens. Revolut uses 256K. The reason is operational: longer context means slower time-to-first-token and more GPU memory consumed per concurrent user. At 2,000 engineers sending requests simultaneously, optimizing for throughput and latency across the fleet matters more than maximizing individual context windows. Theoretical capability and production performance are different things.
The Productivity Gains
Across the 2,000-engineer cohort, Revolut reported a 60% improvement in coding time and a 70% improvement in lead time. These are significant numbers, though they come from the company itself — treat them as directional rather than independently verified benchmarks.
What’s not in dispute: the cost reduction is real, the latency improved, and the system is running in production.
Who Should Care
The break-even point for self-hosted LLM inference sits between 10 and 30 million tokens per day, depending on model size and infrastructure. Revolut, with 2,000 engineers using AI coding tools heavily, is well past that threshold. For teams under 100 engineers, the operational overhead likely outweighs the savings — for now.
But the trajectory is clear. Open-weight coding models are closing the gap with frontier commercial models. Infrastructure tooling (vLLM, SGLang) has matured to the point where production deployment is achievable without a dedicated ML platform team. And as more companies publish their real numbers — not just benchmark comparisons but actual cost and productivity data — the case for self-hosting will only get stronger at lower scales.
Revolut’s GTC Berlin presentation is one of the first detailed public accounts of what this looks like in production at scale. It won’t be the last. For a deeper look at how AI coding tool costs compare across the market, the numbers make the case on their own.
