What Revolut Actually Did (and What It Didn’t)

David Abitbol, Revolut’s ops and infrastructure lead, presented the deployment at AI Engineer NYC on October 13, 2026 — the talk was titled “Shipping like a startup, regulated like a bank.” The key facts from that talk, as reported by The Register and subsequently by AI Weekly and CellCog:

• Revolut has roughly 2,000 engineers now defaulted to a self-hosted instance of Z.ai’s GLM-5.3 (the 754B parameter version — not the 320B Flash variant) as their primary AI coding assistant.
• The model runs on 56 Nvidia B200 GPUs — seven eight-GPU DGX systems — using NVFP4 quantisation, disaggregated prefill and decode, and strict context caps.
• Blended inference cost lands at roughly $0.30 per million tokens, Abitbol said. That is about one-tenth of their previous spend on Anthropic and OpenAI APIs.

One nuance the headlines flattened: Claude Code and Codex are not gone. Abitbol’s own framing was “the frontier is the exception.” High-stakes or genuinely hard tasks still route to frontier models. The self-hosted instance handles the routine coding volume that makes up the majority of requests — completions, refactors, inline questions — where token volume is high and latency matters more than raw capability.

The Self-Hosting Setup That Made It Work

Getting a 754B-parameter mixture-of-experts model to serve 2,000 engineers with acceptable latency is an infrastructure problem, not a model problem. Abitbol detailed four things that made it viable:

Quantisation. GLM-5.3 runs at NVFP4 precision on Blackwell. Abitbol said FP4 uses about 40% less GPU footprint than FP8, which is what made seven DGX systems viable instead of twelve.

Disaggregated prefill and decode. Splitting the prefill (reading the prompt) and decode (generating the response) phases across GPU pools improved throughput. Requests per second went from about 3 to over 7 on the same hardware.

Cache discipline. Cache hit rate for coding traffic runs above 90%. This is achievable because coding context — system prompts, file skeletons, project conventions — is largely repetitive. Abitbol’s team enforced a 256K context cap rather than allowing the model’s theoretical 1M-token maximum: “Long context is a scam. You really need to be mindful of that.”

Resisting demand growth. Demand grew 76% over the deployment period, but throughput scaled faster — about 88% improvement — without adding GPUs. Abitbol said the lesson was to tune before you scale hardware: “You really need to resist scaling.”

Time to first token tells the user experience story: it fell from over 10 seconds in April 2026 to about 1.6 seconds in August 2026, on the same GPU count. Abitbol’s operational rule: anything over 5 seconds triggers a Slack flood.

The Cost Math at Different Scales

Abitbol reported individual engineers previously spending $1,000 to $10,000 per month on frontier AI coding tools. The table below shows what the token-level economics look like, using Revolut’s operational figure alongside published API prices.

* Blended assumes 50/50 input/output split for cloud API prices. Revolut’s $0.30 is their stated operational figure inclusive of all costs (Abitbol, via The Register, November 2026). † A “1K-tok request” is 500 tokens in, 500 tokens out. Sources: Anthropic pricing page (Opus 5.5); CellCog blog citing OpenRouter (GLM-5.3 cloud); David Abitbol via The Register, cited by AI Weekly and CellCog (self-hosted figure).

The 40× gap between cloud-API Opus 5.5 and Revolut’s self-hosted cost is the number that caused engineers’ $1,000–$10,000 monthly bills to collapse. The intermediate option — routing to the cheapest GLM-5.3 cloud host — lands 16× cheaper than Opus 5.5, which many teams could do today without any infrastructure work.

What GLM-5.3 Actually Scores

GLM-5.3 is a 754B-parameter mixture-of-experts model from Z.ai, with roughly 40B parameters active per token. Z.ai released it on August 14, 2026. The benchmarks below are vendor-run by Z.ai unless noted otherwise.

On Z.ai’s own coding benchmarks, GLM-5.3 sits very close to Kimi K3 and GPT-5.6 Sol — within fractions on Terminal-Bench 2.1. The GDPval-AA v2 score of 1,769 Elo comes from Artificial Analysis, an independent evaluator, and is the most credible single data point. All other rows above are vendor-run and lack error bars; treat them as directional.

Throughput matters as much as accuracy for this use case. Artificial Analysis reports GLM-5.3 at roughly 115 tokens per second versus about 47 tokens/s for Qwen3.8-Max and 40 tokens/s for Kimi K3 — which partly explains why it was viable for a high-request production environment like Revolut’s.

For Swiss & EU teams

Revolut is London-based and serves its models on its own infrastructure in the US, EU, UK and Israel, with no data routed to Chinese endpoints. For regulated firms in Switzerland or the EU, this is the template: download the weights, deploy in a data centre covered by your data-processing agreements, and the FADP or GDPR question becomes a standard on-prem data-handling question rather than a cross-border transfer question.

The alternative — routing developer code context through the cheapest GLM-5.3 cloud host — raises questions Swiss and EU teams need to answer before deployment: where does that host process data, and does it have a data-processing agreement you can sign? Z.ai is a Chinese company; its cloud API endpoints sit outside EU/CH adequacy arrangements by default. Self-hosting, as Revolut did, sidesteps that entirely.

For teams exploring this path, see our coding agents pillar for a broader breakdown of EU availability and data-residency status across seven tools.

What to Do on Monday

1. Audit your AI coding spend. Pull the last 90 days of invoices from Anthropic, OpenAI, and GitHub Copilot. Find the per-engineer monthly average and the top-spending outliers. If the average is above $200/month per developer, the self-hosting economics start to make sense above a few hundred engineers.
2. Check what data flows through your coding AI. Does your IDE extension or CI pipeline send code that contains customer data, credentials, or regulated information to external APIs? If yes, you have a compliance risk today, regardless of your data-processing agreements — Revolut’s answer was to remove the question by keeping the data on their own GPUs.
3. Get a GPU quote. Seven 8-GPU DGX B200 systems (56 GPUs) is Revolut’s configuration. Cloud GPU providers list B200 configurations; your hardware vendor can quote on-prem CapEx. If the numbers suggest break-even under 18 months against your current API spend, escalate the conversation internally.
4. Pilot before you commit. Deploy GLM-5.3 or DeepSeek V4.1-Flash on whatever GPU capacity you have via vLLM or Ollama. Run one team for two weeks. Measure TTFT, cache hit rate, and whether code quality degrades on your actual tasks — not on benchmarks.
5. Cap context from the start. Abitbol capped at 256K tokens regardless of the model’s theoretical maximum. Set the same limit in your harness before launch, or you will spend CapEx on context you don’t need.

Further Reading

• Revolut’s David Abitbol at The Register — the primary source for Revolut’s cost and performance figures; worth reading for the operational nuance the aggregators compressed.
• AI Coding Agents in 2026: Benchmarks, Costs, and How to Choose — our pillar post comparing seven agents on Terminal-Bench 4.0, cost-per-task math, and EU data-residency status.
• GLM-5.3-Flash: The 320B Model That Hid in Plain Sight — background on the smaller GLM-5.3 Flash variant, useful context if you are evaluating the model family for a smaller GPU footprint.

Your turn: Has your team done the math on self-hosting a frontier-class open-weight model versus paying API rates? What stopped you — CapEx, ops overhead, compliance review, or something else? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Christian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.