Revolut did not publish a glossy product blog. It sent its AI operations lead on stage. At AI Engineer NYC on October 13, 2026, David Abitbol described how the fintech made self-hosted open-weight GLM models the default coding stack for roughly 2,000 engineers, running on about 56 Nvidia B200 GPUs. The talk, later covered by The Register and transcribed by the conference, is the source behind this week’s X trend — and several viral numbers need their footnotes attached.
Quick Take
- What: Revolut self-hosts GLM (started on GLM 5.1 in April 2026, later GLM 5.3) as the default for ~2,000 engineers.
- Fleet: About 56 Nvidia B200 GPUs; blended cost cited near $0.30 per million tokens; roughly one-tenth the prior Anthropic/OpenAI invoice.
- Latency targets: Sub-3s time to first token (complaint threshold ~5s); ~200 tokens/s adequate for interactive use after tuning.
- Caveat: 60% coding-time and 70% lead-time improvements coincide with the switch but, per the conference write-up, do not isolate the model change from other causes. Claude/Codex remain for exceptions; vision routes elsewhere.
What happened
Abitbol’s session, “Shipping like a startup, regulated like a bank,” walks through why Revolut moved off a primarily commercial coding-agent bill. Token scarcity made engineers ration AI to hard coding tasks and skip surrounding work (inbox, Jira, docs). Self-hosting GLM on dedicated GPUs flattened the bill and removed the fear of month-end spike. The Register reports inference costs down roughly 90% versus the prior commercial mix. Open-weight GLM comes from Zhipu / Z.ai; Abitbol describes starting self-hosting around April 26 with GLM 5.1 and later serving GLM 5.3.
Background: why a bank self-hosts
Revolut is an EU-regulated digital bank. Abitbol’s compliance sketch, as summarized by conference materials, includes human-in-the-loop coding in a sandbox with per-operation rules, data in equivalency-framework countries, zero data retention by default, and strong OpenTelemetry. GPUs can sit in the US, EU, UK or Israel on “a US-based service” — the talk does not name a cloud vendor in the sources we read. Frontier tools stay available: Claude Code, Codex and others are exceptions, not the default. Because GLM is text-only in their setup, vision and UI work still route to models such as Gemini or GPT Luna via a router.
How they made 56 GPUs feel fast
Loading weights was not enough. Abitbol shows the GPU count held at 56 from April to August while requests per second rose about 76% and time to first token fell from 14.26 seconds during migration to 1.68 seconds in the stable setup — an 88% latency cut. Levers cited across The Register and the conference write-up include NVFP4 quantization on Blackwell, context length around 256K, tensor parallelism trade-offs (TP4 vs TP8), cache-hit rates above 90% for coding, speculative decoding, and disaggregated prefill/decode. Endpoints speak an OpenAI-compatible API (including Responses API) so engineers keep harnesses such as OpenCode, Codex and the Codex app.
| Metric (as attributed to Abitbol / coverage) | Figure |
|---|---|
| Engineers on default GLM | ~2,000 |
| GPUs | ~56× Nvidia B200 |
| Blended token cost | ~$0.30 per million |
| vs prior Anthropic/OpenAI invoice | ~1/10 (≈90% less) |
| Coding time / lead time (correlational) | ~60% / ~70% better after April switch |
| TTFT migration → stable | 14.26s → 1.68s |
| Target TTFT / complaint threshold | <3s / ~5s |
Pros and cons
For users
- Pro: If more fintechs copy the pattern, consumer apps may ship features faster when internal engineering throughput rises.
- Con: None direct; this is an internal platform story.
For developers
- Pro: A concrete playbook: measure TTFT, tune before buying more GPUs, keep a router to frontier models for hard or multimodal tasks.
- Con: Revolut’s utilization and regulated context do not transfer to a 20-person startup. Premature self-hosting wastes capital.
For companies
- Pro: At sufficient steady token volume, open weights plus owned GPUs can crush API unit economics and keep code off third-party trainers by default.
- Con: CapEx/OpEx, MLOps talent, model refresh cadence, and regulatory review are real. The 60/70% productivity figures are correlational.
What X is saying
Developers mostly debated when self-hosting wins, not whether Revolut’s slides exist.
Revolut cut its AI coding inference costs ~90% by making a self-hosted open model its default. GLM-5.3 on 56 Nvidia B200s. ~2,000 engineers. $0.30 per 1M tokens. <3s to first token at 256K context. Vision still goes to Gemini/GPT.
— Neyvorn (@Neyvorn) November 4, 2026 · View on X
Revolut moved about 2,000 engineers to self-hosted GLM and says its AI coding bill fell roughly 90%. The bigger lesson is what engineers did once tokens stopped being rationed.
— Josh Builds (@Joshua_WD) November 4, 2026 · View on X
Same week, two opposite answers to “should we self-host our coding model?” Both are right. Revolut reportedly moved ~2,000 engineers to self-hosted GLM 5.3. Others say a GPU can’t pay its rent. The difference is one number: utilization.
— Dharmendra Valiya (@dvvaliya) November 4, 2026 · View on X
Revolut cut its AI coding costs ~90% by switching from Claude Code to self-hosted GLM models. The lesson isn’t ‘self-host everything’ — it’s that the default tool is the most expensive line in your stack.
— Omer (@OmerYounus1) November 5, 2026 · View on X
What Grok says
Grok, answering a sourcing question, pointed to Abitbol’s AI Engineer NYC talk as the primary source and summarized the headline figures (GLM-5.3, ~56 B200s, ~$0.30/M tokens, ~90% less spend) as confirmed by The Register and others — while separate replies stressed that TCO, frontier exceptions and scale thresholds still decide whether self-hosting pays for smaller firms.
Critical analysis
The useful story is utilization and tuning, not “GLM beat Claude.” Revolut kept Claude and Codex for complex tasks and routed vision elsewhere. The $0.30/M figure is a blended cost at Revolut’s load, not a portable price list. The productivity percentages lack a clean causal design. For regulated enterprises with thousands of engineers and sustained coding-agent traffic, the talk is a serious data point. For everyone else, it is a warning against cargo-culting 56 B200s.
What’s still open
- Will Revolut publish a written engineering blog with reproducible configs?
- How often do they refresh base models as Z.ai ships new GLM versions?
- What share of tasks still escalate to Claude/Codex?
- At what monthly token volume do mid-size teams break even?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis is based on The Register’s report, the AI Engineer NYC talk page and transcript excerpts, conference write-ups, and posts on X from November 3–5, 2026. Figures are as attributed to David Abitbol’s talk unless noted. Revolut and Zhipu/Z.ai have no dedicated company pages in the i10x taxonomy; filed under Enterprise & Work.
i10x Perspective
Agent coding made tokens a first-class line item. Revolut’s answer — default to open weights, reserve frontier APIs, tune ruthlessly before scaling GPUs — will be copied by other high-volume engineering orgs. The winners will be teams that measure utilization honestly; the losers will buy clusters to recreate a bill they did not understand.
