HYPE OR FACT? · CHECKED 3 NOVEMBER 2026
Revolut moved ~2,000 engineers to self-hosted GLM at a tenth of the cost?
OUR VERDICT
Partly factOur original verdict: “Fact or hype: partly fact. Revolut's AI infrastructure lead really did say all of this in two October conference talks, and the post quotes him closely. But every number is Revolut's own and unaudited, the 70% gain isn't separated from lifting engineers' token budgets, Claude Code and Codex are still in use, it isn't "breaking" news, and the IPO line is the poster's opinion.”
THE CLAIM
“BREAKING: Revolut moved ~2,000 engineers to self-hosted GLM, with Claude Code and Codex on the exception list. David Abitbol: 56 B200s, a blended $0.30 per million tokens, about a tenth of what they spent on Anthropic and OpenAI, and lead time improved 70%. A pricing problem for every lab heading to IPO.”
What’s actually true
- The source is real. David Abitbol, who runs AI operations and infrastructure at Revolut, gave the talk "Shipping like a startup, regulated like a bank" at AI Engineer NYC on 13 Oct 2026 and a shorter version at NVIDIA GTC Berlin in October (a session presented by Together AI). There is no Revolut press release or engineering blog; this is one executive's account on stage. The post calls it "BREAKING", but the talk was three weeks old.
- ~2,000 engineers: he said it, but in two ways. The productivity chart is "a sample of 2,000 engineers", and he said that to serve the 2,000 engineers they have, you need around 56 GPUs. He didn't say everyone was "moved" off other tools.
- GLM: confirmed in both talks. Revolut started self-hosting GLM 5.1 in April 2026 and now runs GLM 5.3 (he said about 800B parameters; Hugging Face lists 753B). It uses only the open weights on its own GPUs, not Z.ai's China API, caps context at 256K, runs FP4 quantisation and sends images to a separate model because GLM-5.3 is text-only.
- Claude Code and Codex on the exception list: true in spirit, but frontier models are the exception, not banned tools. He said "the frontier is the exception" and "we keep using Codex, Claude Code" for the hardest problems with a small budget. Claude Code, Codex and OpenCode can also be pointed at the self-hosted GLM, so the apps aren't necessarily gone.
- 56 B200s: he said "roughly 56". He also said the GPU count stayed fixed from spring to August while tuning alone cut time to first token from more than 10 seconds to under 2, and that out of the box the performance was "terrible".
- $0.30 per million tokens: he said a blended cost of 30 cents per million tokens. "Blended" isn't defined, and he advised keeping cache hits in the high 90s, so most of those tokens are probably cheap cached input. Our rough estimate, not his: 56 B200s rented at US$3-6 an hour is about US$120k-245k a month, which at $0.30 per million means roughly 400-800 billion tokens a month, or 200-400 million per engineer. That is heavy but plausible for agentic coding. Z.ai's own API charges US$0.26 per million for cached GLM-5.3 input, so $0.30 blended is in line with market prices.
- "About a tenth" is said, but it is described two different ways. In New York: "our invoice is one-tenth of what we were spending on Anthropic and OpenAI". In Berlin: one-tenth of the per-token cost of "the equivalent frontier model, so Opus". No baseline amount was given. At Anthropic's list prices, a mostly cached workload on Opus comes out only about 2-3x dearer per token (our estimate), so the tenth depends heavily on how tokens are counted. It is unaudited either way.
- 70% lead time: he said 70% better lead time (ticket picked up to feature merged/shipped) and 60% better coding time after the April GLM rollout. That is Revolut's internal chart, and it coincides with removing token limits: earlier, Claude Code came with a budget of about US$300 per engineer a month, which he said wasn't enough and showed little impact. So the jump reflects unlimited AI use, not GLM being better than Claude. The conference organiser notes that the comparison "does not isolate the model change from other causes".
- Missing from the post: the cost and effort of running GPUs (the $0.30 may not include staff), months of tuning, sandbox and human-in-the-loop controls, and the fact that frontier models are still used. "A pricing problem for every lab heading to IPO" is the poster's opinion, not something Revolut said.
- We haven't tested GLM-5.3 hands-on, and we read the talk transcripts rather than watching the full video.
Sources
- AI Engineer NYC 2026: David Abitbol, "Shipping like a startup, regulated like a bank" (talk page)
- Talk transcript (auto-generated)
- Talk video (AI Engineer NYC livestream, from about 5:37:00)
- NVIDIA GTC Berlin 2026: How Revolut Runs Production Agents on Open Models (presented by Together AI)
- David Abitbol on LinkedIn
- Hugging Face: zai-org/GLM-5.3
- Z.ai docs: GLM-5.3
- Z.ai API pricing
- Anthropic API pricing
- The New Stack: GLM-5.3 goes open weight, with a new license
- Trending Topics: Revolut cuts AI coding costs 90% with open-weight GLM (3 Nov)
- Oliver Molander on X (1 Nov)
- The claim: Guillermo Flor on X (2 Nov)
Cost
Self-hosting needs your own GPUs (Revolut: about 56 B200s for 2,000 engineers) plus the people to tune and run them. Or use the GLM Coding Plan from US$18/month, or the API at US$1.40 input / US$4.40 output per million tokens for GLM-5.3.
USE IT IF
Engineering teams with AI spend in the tens of thousands a month and the infrastructure skills to run GPUs, who want flat costs and no per-engineer token budget.
SKIP IT IF
Small teams without GPU or MLOps skills, since self-hosting is hard to get right. Also anyone who needs images or vision in the same model, or whose compliance rules forbid models from Chinese labs even with self-hosted weights.
How we knowChecked on 3 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗