Revolut Cut AI Coding Costs 10x With Self-Hosted GLM

An EU-regulated bank just told a New York conference room that it cut its AI coding inference costs by roughly 90 percent. The way it did it: stop paying for the tools everyone else rents. Revolut, the digital bank serving 80 million customers across 45 markets, watched its engineers rack up $1,000 to $10,000 a month each on Claude Code with, in its own words, "not so much ROI." Its answer was not a better enterprise discount. It was to self-host open-weight GLM models from Chinese lab Z.ai on 56 Nvidia B200 GPUs and make tokens so cheap that engineers stopped rationing them. Here is the story behind the 10x claim: what Revolut actually measured, what it spent to get there, and what the numbers leave out.

The story in 90 seconds

The timeline comes from David Abitbol, Revolut's ops and infrastructure lead, speaking at the AI Engineer conference in New York in October 2026. Revolut first introduced AI to engineers in January 2025 with Google's Gemini, which delivered a 10 percent improvement in lead time and a 20 percent improvement in coding time, mostly by replacing Stack Overflow lookups and documentation reading.

Claude Code came next: trialled from February 2025 and launched at scale in January 2026. This is where the trouble started. "We were highly skeptical about the cost of AI. We were seeing engineers spend a thousand a month, $10,000 a month with not so much ROI, and so when we launched it we launched it with strict budget limits," Abitbol said. Capped tokens meant engineers hoarded them for the hardest tasks instead of folding AI into everyday work.

Around April 2026, Revolut started self-hosting open-source models ("we're a big fan of GLM for coding"), beginning with GLM 5.1 and later moving to GLM 5.3. The result, per the company's own chart tracking 2,000 engineers: a 60 percent improvement in coding time and a 70 percent improvement in lead time, at a blended cost of $0.30 per million tokens. "Our invoice is one-tenth of what we were spending on Anthropic and OpenAI," Abitbol said.

The numbers behind the 10x claim

All figures below are Revolut's own, as presented in the talk. No independent audit or invoice has been published.

Currency conversions use November 4, 2026 market rates (1 USD ≈ £0.74 / C$1.40) and are approximate.

Why tokens got cheap: the hardware math

To serve 2,000 engineers on GLM 5.3, Revolut runs 56 Nvidia B200 GPUs, equivalent in GPU count to seven eight-GPU DGX B200 systems. GLM 5.1 is not a small model: roughly 754 billion total parameters with 40 billion activated per generated token, needing about 380 GB of GPU memory just to hold the weights plus 6–26 GB per user for context, according to The Register's reporting.

But the striking part of the talk is that adding GPUs was not the lever. Between April and August 2026 the GPU count stayed flat while time to first token fell from over 10 seconds to about 1.6 and requests per second more than doubled, an "88% improvement in performance for 76% increased demand." The gains came from tuning:

• NVFP4 quantization: moving from FP8 to FP4 on Blackwell GPUs cut the GPU memory footprint by 40% with, Abitbol said, zero loss in precision.
• Tensor parallelism sized for the KV cache: GLM fits on TP4, but TP8 works better once context memory is accounted for.
• Cache-hit discipline: the coding-test cache hit rate has to stay above 90%.
• Inference plumbing: speculative decoding, sequential prefill, and disaggregated prefill/decode.

The lesson Abitbol drew is blunt: "there's a lot of parameters that you can improve on, and you really need to resist scaling if you're looking to self-host a model."

"Long context is a scam"

The talk's most quoted line is about context windows. GLM can serve a one-million-token context, but Abitbol's verdict: "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."

His performance targets are framed around human patience, not benchmarks. Aim for sub-3-second time to first token: "as soon as you exceed 5 seconds, my Slack inbox starts to blow up." Serve around 200 tokens per second for interactive work — "people are fine [with that]. You don't need to go higher." The workflow changed with the economics: engineers run tasks in parallel and check back when they finish instead of watching thinking tokens stream. After 5 pm, when demand drops, throughput climbs toward 800 tokens a second.

The security playbook for a Chinese open-weights model

Abitbol addressed the obvious objection head-on. The concern with Chinese models, he argued, is concentrated in one place: querying them through inference APIs that run in China. "You're not using inference, dedicated endpoint that goes to China. You're just getting the weights and putting them on a US-based service. So for data sharing side, this is totally fine."

Revolut deploys the openly available weights on regional cloud or on-premise infrastructure in the United States, the European Union, Britain, and Israel, with zero data retention by default, according to the talk coverage. As an EU-regulated bank under what Abitbol called "the most strict regime in terms of compliance," the company treats coding tasks as human-in-the-loop work, never autonomous agents. The programming agents run in a sandbox with strict approval rules, the same "approve for me" gating found in Codex and Claude Code — which also mitigates prompt-injection risk.

What GLM didn't replace

GLM is a text-only model, so it was never going to cover everything. Revolut uses a local model router: image-recognition and UI queries go to Gemini or GPT Luna, while code generation stays on GLM. A smaller GLM 5 Flash variant handles vision needs: "we have our UI designer making great use of it," Abitbol said.

The self-hosted endpoints are OpenAI-compatible, including the Responses API, so engineers can point their existing tools at them. Abitbol said the setup works inside OpenCode, Claude Code, and the Codex app. And the frontier models were not fired: "We keep using Codex, Claude Code. The latest models are fantastic... But we keep them as exception." For background on the tool Revolut kept as the exception, see our Claude Code explainer.

The fine print

The honest version of this story has caveats, and they matter:

• Every headline number is Revolut's own. The 90%, the $0.30, the 60/70% improvements all come from one conference talk. No invoices, no methodology, no independent audit have been published.
• $0.30 per million is a blended inference cost. The 56 B200 GPUs, the power, the engineers tuning quantization. None of that capex or ops cost is itemized in the talk. Self-hosting trades a variable bill for a fixed one; the fixed one is not zero.
• The 10x compares against Revolut's own enterprise usage billing. This is the crucial context: on Anthropic's enterprise plan, seats cost $20 per user per month but usage is billed at API rates. The seat is cheap, the tokens are the bill. That is exactly how engineers reached $1,000–$10,000 a month. Revolut's one-tenth is measured against that metered spend, not against seat prices.
• The productivity numbers come from an internal chart. A 2,000-engineer sample over roughly a year is substantial, but the chart's methodology is unpublished, and the Gemini and Claude Code eras are compared against different baselines.
• The timeline is per Abitbol's own account: Gemini in January 2025, Claude Code trials from February 2025 and at scale from January 2026, self-hosted GLM from April 2026.

Why it matters beyond one bank

Revolut is not an outlier in rethinking AI coding economics. It is just the first to show its math in public. In October, Santander said it had moved from large general-purpose models to smaller specialist ones, cutting costs while raising output quality. On November 4, Goldman Sachs reportedly restricted most traders and bankers from using advanced AI coding tools after a review raised productivity and risk-management concerns, per FStech's reporting on the Revolut story.

The pattern is the same in each case: metered AI pricing makes experimentation expensive at scale, and the response is control: smaller models, tighter policies, or, in Revolut's case, owning the meter. "Open source is now a priority at Revolut, and our economics are favorable. The billing is flat, we don't have any bill shock at the end of the month, and the frontier is the exception," Abitbol said. For teams running coding-agent tooling of their own, our ECC toolkit rundown covers the workflow side of the same stack.

How much did Revolut actually save on AI coding?

Revolut says its blended inference cost is $0.30 per million tokens (about £0.22 or C$0.42) and that its total AI invoice is roughly one-tenth of what it spent on Anthropic and OpenAI. That is the company's own arithmetic from a conference talk; no invoices have been published, so treat the 10x as directionally credible rather than audited.

What hardware does Revolut use for self-hosted AI coding?

56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems, serving GLM 5.3 for about 2,000 engineers. The company says the biggest performance gains came from tuning (NVFP4 quantization, tensor parallelism, cache-hit discipline) rather than adding hardware.

Is GLM free to use?

The model weights are openly available, which removes per-token API fees, but self-hosting is not free: 56 B200 GPUs plus power, networking, and the engineering time to tune inference all cost real money. Revolut's pitch is that this fixed cost beats an unpredictable metered bill at its scale. The breakeven math looks very different for a 20-person team.

Why does Revolut say 256K context beats 1M?

Serving a one-million-token context window is possible but, in Abitbol's words, "long context is a scam" for engineering work: the 256K-token range is the practical sweet spot where latency stays under the sub-3-second threshold engineers tolerate. Longer contexts mostly buy slower responses.

Isn't it risky for a bank to use a Chinese AI model?

Revolut's answer is that the risk it worries about is inference APIs running in China, not the weights themselves. It runs the open weights on infrastructure in the US, EU, Britain, and Israel with zero data retention by default, keeps a human in the loop, and sandboxes the coding agents. Whether that satisfies every regulator is a separate question the talk does not fully answer.

Could a smaller company copy this?

The tuning playbook transfers. Quantization, cache-hit discipline, and context sizing work at any scale. The 56-GPU fleet does not. For smaller teams, the practical takeaway is narrower: measure your per-token spend, know your latency threshold, and check whether your cloud AI bill has quietly passed the cost of dedicated inference before assuming the meter is the only option.

Revolut's story is less "open source beats Big AI" than "flat infrastructure beats metered tokens at 2,000-engineer scale." The 90% figure is real arithmetic on real invoices, just Revolut's invoices, not yours. If you run an AI coding budget, the question the talk actually answers is simpler: at what token volume does your usage bill exceed the cost of tuning your own GPUs?

This is general information, not procurement or legal advice.