I watched “How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models” — David Abitbol, who leads AI operations at Revolut, speaking at AI Engineer New York. His framing: “Shipping like a startup, regulated like a bank.” It is a dense technical talk, and almost every number in it is relevant to anyone running local models for agentic coding, so here are the points that matter.

The rollout, with receipts

Revolut (80M customers, 45 markets) tracked a sample of 2,000 engineers for about a year:

Period Tool Lead time Coding time
Jan 2025 Gemini +10% +20%
Feb 2025 – Jan 2026 Claude Code minimal at launch minimal at launch
Apr 2026 Self-hosted GLM 5.1 +70% +60%

Two details worth pausing on. First, Claude Code showed little measurable impact at first while engineers were spending $1,000–$10,000/month, which is why the rollout came with strict budget caps. Second, the asymmetry of the GLM numbers: the bigger gain landed on lead time, not coding time. His explanation is that the bottleneck in engineering is not typing code — it is the surrounding work (inbox triage, Jira, Confluence). When tokens are effectively unlimited, engineers stop rationing them for the hardest task or the end of the quarter.

The current posture: self-hosted open source is the default, billing is flat with no month-to-month bill shock, and frontier models (Codex, Claude Code, Gemini) are the exception for the hardest tasks.

The economics

Metrics that matter for happy engineers: 2,000 engineers, <3s TTFT, >200 TPS, $0.30 per million tokens

The capacity math: serving 2,000 engineers on GLM 5.3 — an 800-billion-parameter model — takes roughly 56 B200 GPUs (dedicated, hosted with Together AI). The headline numbers:

  • $0.30 per million tokens blended (in/out, cache is free) — the invoice is one tenth of what they were paying Anthropic and OpenAI.
  • Latency targets are empirical, not theoretical: TTFT under 3 seconds, because past 5 seconds “my Slack inbox starts to blow up.”
  • Throughput barely matters for interactivity: 200 TPS is fine, because engineers now run tasks in parallel and check back when they finish instead of watching tokens stream. After 5 p.m., when the GPUs are idle, they get 800 TPS.

Resist scaling, tune first

This is the slide I would tattoo on the ops team’s wall:

Resist scaling, tune first: 76% higher RPS and 88% lower TTFT on the same 56 GPUs

80% of the performance gain came from adjusting hyperparameters, not more GPUs. Same 56 B200s from April to August: requests per second up 76% (3 → 7+) and time-to-first-token down 88% (14.26s → 1.68s stable). Dropping weights onto GPUs and calling it a day produces terrible performance; the configuration is the product.

His tuning checklist, roughly in order of contrarian value:

  • Long context is a scam. Even if the model advertises 1M tokens, the workload needs 256K — and serving more costs KV cache for little gain.
  • FP4 is the floor on Blackwell. Going from FP4 to FP8 costs about 40% more GPU footprint for zero precision gain.
  • Tensor parallelism: GLM fits on TP4, but the KV cache runs short; TP8 may serve better depending on workload.
  • Cache hit rate above 90% for coding tasks — this is why sticky sessions matter.
  • Plus speculative decoding and prefill/decode (PD) disaggregation.

Take the smarter model, route vision elsewhere

Production setup: text models, vision models, harness

GLM is text-only, and he treats that as an acceptable trade-off rather than a blocker: “we would make [that trade-off] every time” — get the smarter model and handle image recognition another way. A local router (they use OpenCode) sends image tasks to Gemini while GLM handles the main coding work. For teams that need vision directly, GLM’s Flash variant works well; their UI designer uses it and UI product turnaround improved substantially. All endpoints are OpenAI-compatible, which is what plugs into OpenCode, Codex and friends — compatibility is the feature he recommends optimizing for.

The compliance answer for Chinese open weights

For the CIO in the room: sandbox, data privacy, analytics

Revolut is an EU-regulated bank running Chinese open-weight models, and the argument to regulators is clean:

  • Weights only, no inference to China. They download the weights and run them on US-based infrastructure — “for data sharing side, this is totally fine.”
  • Coding agents are not autonomous agents. A human is always in the loop, and the sandbox is built to guarantee it: allow/deny rules per operation (allow grep, deny gradle, for instance). That classification — “productivity use case with humans in the loop” — is what keeps it out of the heaviest regulatory audit.
  • The sandbox does double duty: it also shields against rogue internal use (“you’d be surprised what you find in a large company”) and prompt injection, which he considers less of a threat now that models are smarter, but still a live one.
  • Data privacy: GPUs can sit in any country in the data equivalency framework (US, EU, UK, Israel); GPU and inference providers have zero data retention by default.
  • Strong OpenTelemetry so any incident can be traced through deep logs.

The through-line

The cost win and the compliance win come from the same discipline: constrain the system tightly — cap context, cap precision, cap permissions — instead of buying more capacity or a bigger model. The open question the talk leaves hanging is durability: these are one company’s internal numbers, the 1/10 cost figure assumes frontier API pricing holds, and nobody knows yet whether regulators will accept the human-in-the-loop sandbox argument as a general precedent or a Revolut-specific arrangement.

Talk: How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models — David Abitbol, Revolut @ AI Engineer. Slides captured from the video.