Revolut Moves 2,000 Engineers to Self-Hosted GLM and Cuts AI Coding Costs 90%

TL;DR

Revolut told the AI Engineer conference in New York that it now serves AI coding for roughly 2,000 engineers from self-hosted GLM-5.3, the open-weight coding model from Z.ai, running on 56 Nvidia B200 GPUs. The blended cost is about $0.30 per million tokens, which David Abitbol, who leads AI operations at the bank, says works out to one-tenth of the invoice Revolut was paying Anthropic and OpenAI. The Register reported the talk on November 3, and the Hacker News thread passed 190 points the same day. The saving is the headline; the tuning numbers are the part you can copy.

How the bill got capped

Abitbol's timeline starts in January 2025, when the bank rolled out Gemini as its first sanctioned AI tool. His chart credits it with a 10% improvement in lead time and 20% in coding time, mostly because people stopped digging through documentation by hand. Claude Code came next: experiments from February 2025, a full rollout in January 2026, and immediate sticker shock.

"We were highly skeptical about the cost of AI," he said. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits."

The caps worked on the bill and failed on behavior. Engineers hoarded tokens for the hardest task of the quarter and kept doing the surrounding work by hand: the inbox digging, the Jira triage, the Confluence archaeology. A rationed model is a model nobody sprays on the boring 80% of the job, and the boring 80% is where a bank's lead time actually goes.

In April 2026 Revolut started self-hosting GLM-5.1. By October the default was GLM-5.3, with Claude Code and Codex kept for exceptions. "Open source is now a priority at Revolut, and our economics are favorable," Abitbol said. "The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception."

What 56 B200s actually serve

Fifty-six B200s is seven eight-GPU systems, and the model is not small. The Register puts GLM-5.1 at roughly 754 billion total parameters, 40 billion active per token, needing about 380 GB of GPU memory for the weights plus another 6 to 26 GB per concurrent user for context, depending on precision. GLM is text-only, so a router sends vision work to Gemini or GPT-6 Luna, and compatible APIs mean engineers keep whichever coding harness they already use.

The economics Abitbol claims are a blended $0.30 per million tokens, input and output averaged, against roughly a tenth of the previous Anthropic and OpenAI spend. Those numbers describe Revolut's cluster at Revolut's utilization, one bank's meter reading. What the GPUs cost to buy or rent, what the platform team costs, and how busy the cluster runs were all left unsaid.

Tuning beat buying

The best slide in the talk has no model on it. During the migration to GLM-5.3, average time to first token was 14.26 seconds. After the serving stack was tuned, on the same 56 GPUs, it was 1.68 seconds, an 88% drop, with 76% more requests per second. No new hardware entered the story.

The levers: NVFP4 quantization on the Blackwell cards, splitting prefill from decode so the two phases stop competing for the same batch, eight-way tensor parallelism instead of four-way (the weights fit on four cards, but then the KV cache has nowhere to live), plus cache-hit monitoring and speculative decoding. His advice: resist adding GPUs until the current stack is tuned.

His line on context windows: "Long context is a scam. You really need to be mindful of that. Even though you can serve one million tokens, 256 [thousand] is really what you need." Reserving memory for a million-token window that coding traffic never sends is renting the banquet hall for a table of two. A bank calling long context a scam is also the most on-brand sentence of the conference.

The latency budget engineers will feel

His tolerance numbers: time to first token under three seconds, because past five "my Slack inbox starts to blow up," and around 200 tokens per second of generation, which he says people barely register. After 5 p.m., when the cluster quiets down, the same hardware serves whoever is still logged in at close to 800 tokens per second, the only performance upgrade in the talk that costs a calendar invite.

The indifference to throughput has a reason. "We no longer look at the stream of thinking tokens," he said. "People run tasks in parallel and check in when the task is done." When nobody watches the stream, decode speed stops being the constraint and spare capacity becomes the product.

The 60/70% chart, read carefully

The slide everyone quotes claims a 60% improvement in coding time and 70% in lead time after the GLM-5.1 rollout, measured across the 2,000-engineer sample. These are Revolut's own numbers from an internal chart, and the comparison does not isolate the model switch from everything else that changed during 2026.

The detail worth more than the percentages is where Abitbol says the gains came from: debugging, ticket triage, documentation digging, the work around the code. "When you have unlimited tokens, you're free to do that," he said. "You no longer have to save your tokens for the end of the quarter or the most difficult task." Cheap tokens changed what engineers were willing to delegate, and most of the delegated work was never code writing.

Caveats

• Every cost and speed number here is Revolut's own, delivered in a conference talk with no published methodology and no independent audit.
• The 60% and 70% improvements come from an internal chart that cannot separate the model change from other changes over the same period.
• The $0.30 per million figure is Revolut's blended serving cost at its own utilization; it is not stated whether hardware amortization, power, or the platform team sits inside it.
• Quality is contested. Hacker News commenters who use both argue current Claude and Codex models still handle the hardest work better, and the fact that Revolut keeps them as exceptions is its own evidence.
• GLM is text-only, so the frontier vision models still get paid elsewhere in the stack.
• Self-hosting moved the compliance work in-house. Sandboxing, approval gates for restricted operations, and telemetry are now Revolut's problem, which Abitbol counts as a feature and your risk team should price as a cost.

What to steal

The architecture is the copyable part: an open-weight model as the default, frontier models as exceptions behind a router, and harness-compatible APIs so nobody changes tools. That shape starts paying once per-seat API pricing starts to hurt; one Hacker News commenter put the crossover around 150 engineers, where enterprise per-token billing begins and subsidized subscriptions end.

The second steal is operational. An 8.5x latency improvement and 76% more throughput on identical hardware came from quantization, prefill/decode separation, parallelism width, and cache behavior, all of it software. Most teams asking for a bigger GPU budget have not finished that list.

The third is behavioral. The biggest measured gains at Revolut showed up when tokens stopped being scarce, which means a hard budget on AI tools can cost more in unbilled drudgery than it saves on the invoice. If your engineers are hoarding tokens for quarter-end, the budget is saving money by disabling the tool.

Key Takeaways

• Revolut serves AI coding for roughly 2,000 engineers from self-hosted GLM-5.3 on 56 Nvidia B200 GPUs, at a blended cost it puts at $0.30 per million tokens, about one-tenth of its previous Anthropic and OpenAI bill.
• The move followed bill shock: engineers spending $1,000 to $10,000 a month on Claude Code with unclear ROI, then strict budgets that made engineers hoard tokens and skip AI for everyday work.
• Tuning, not hardware, delivered the serving win: time to first token fell from 14.26 to 1.68 seconds (88% lower) with 76% more requests per second on the same 56 GPUs, via NVFP4 quantization, prefill/decode separation, and eight-way tensor parallelism.
• His latency rules of thumb: sub-3-second time to first token (Slack complaints past 5 seconds), about 200 tokens per second interactive, close to 800 after hours, and 256K of context instead of a million.
• Revolut's chart credits the GLM rollout with 60% better coding time and 70% better lead time; the numbers are self-reported and not isolated, and the gains came mostly from delegating debugging and ticket work.
• GLM is text-only, so vision work routes to Gemini or GPT-6 Luna, and Claude Code and Codex stay available as exceptions.

Sources: The Register report, November 3, AI Engineer NYC talk page with full transcript, Hacker News discussion (191 points), Z.ai GLM-5.1 documentation, Z.ai GLM-5.3 documentation