Back to Home

Revolut's $0.30 Bet: Inside the 90% AI Bill Cut

The slide behind David Abitbol did not look like a victory lap. It was a cost curve, falling from left to right, shown to a room of builders at the AI Engineer conference in New York in October 2026.

Abitbol runs operations and infrastructure at Revolut, the British fintech that treats engineering headcount as a growth metric. What he described was not a pilot or a discount code. It was the moment his company stopped buying AI coding from the frontier labs and started serving Chinese open-weight models on its own GPUs.

The result: roughly 90 percent off its AI coding inference bill in under a year, with about 2,000 engineers now working on a model most of their peers have never run in production.

It did not start that way. Early 2025 brought Gemini. Early 2026 brought Claude Code, rolled out at scale. And with it came a spreadsheet that refused to behave. "We were highly skeptical about the cost of AI," Abitbol said. "We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI."

So the company did the obvious enterprise thing. It put a meter on the tool: strict budget limits, a capped pool of tokens, a finance department that could sleep at night.

That works until it does not. Teams stopped asking hard questions because hard questions burn context. Debugging an unfamiliar service is expensive. Reading a Jira trail is expensive. The assistant was there, but everyone was saving it for a rainy day that never came.

Meanwhile, the bill kept scaling with curiosity.

The Turn: Thirty Cents a Million

Revolut began experimenting with self-hosting, first on GLM-5.1 and now on GLM-5.3, both built by Z.ai. At roughly $0.30 per million input and output tokens, running the model in-house came out about a tenth of what Codex or Claude Code cost at the time.

The hardware is not casual. GLM-5.1 carries around 754 billion total parameters with 40 billion active for each token generated, which means roughly 380 GB of GPU memory just to hold the weights, plus 6 to 26 GB per user to maintain context. Revolut serves GLM-5.3 across 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems.

Here is the part every infrastructure lead should read twice. The gains did not come from stacking more silicon. Abitbol credited hyperparameter tuning, a shift to NVFP4 quantization, disaggregated prefill and decode, and a strict discipline around context length.

  • Sub-3 second time to first token at contexts of roughly 256,000 tokens
  • About 200 tokens per second for interactive work, and closer to 800 for engineers signing in after 5 pm
  • 60 percent faster coding time and 70 percent better lead time across the rollout
  • Flat monthly billing, with no end-of-quarter bill shock

Latency is where engineers cast their votes. "As soon as you exceed 5 seconds, my Slack inbox starts to blow up," he said. Throughput, he argued, barely matters anymore, because people no longer sit and watch the thinking tokens scroll. They launch tasks in parallel and check back when the work is done.

What Unlimited Tokens Do to a Team

The bill is the headline. Behavior is the story. Once tokens stopped being scarce, engineers started handing the model the unglamorous work: long debugging sessions, Jira tickets, Confluence archaeology, the tasks that always lost the priority argument. "When you have unlimited tokens, you're free to do that," Abitbol said. "You no longer have to save your tokens for the end of the quarter or the most difficult task."

He was equally blunt about one sacred cow of the model marketing era. "Long context is a scam," he said, arguing that a million-token window is mostly a spec sheet flex. For Revolut's engineers, 256,000 tokens is what the job actually needs, and serving it well beats advertising more.

GLM has not replaced everything, either. The model is text only, so UI and vision work still goes elsewhere. Revolut runs a router that sends visual queries to Gemini or GPT Luna while code generation stays on GLM, and Claude Code and Codex survive as exceptions rather than defaults. "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."

He also answered the question every security team asks about Chinese open weights. The risk, in his view, sits in calling APIs hosted in China, not in running the weights on hardware you control.

So the 90 percent is the number, but it is not the lesson. When every token has a meter on it, engineers budget their attention and save the tool for emergencies. When the meter goes flat, they spend it on the messy work nobody wanted to fund. Revolut did not just find cheaper tokens. It found the work its teams had been quietly skipping.

Comments

No comments yet. Be the first to share your thoughts!