So the fintech did something the AI industrial complex didn’t expect. It stopped buying tokens and started running its own models.
The results are embarrassing for the frontier labs: a 90% reduction in AI coding inference costs in under a year, a 60% improvement in coding time, and a 70% cut in lead time across roughly 2,000 engineers. All powered by GLM-5.3, an open-weights model from Chinese lab Z.ai, served on 56 Nvidia B200 GPUs.
Let’s break down exactly what Revolut did, what it means for the “frontier models are the only way” narrative, and why your engineering org might want to run the same math.
The Path to GLM: A Study in Token Rationing
Revolut didn’t start hostile to commercial APIs. It started with Gemini in early 2025, which delivered modest gains: a 20% reduction in coding time and a 10% improvement in lead time. Then came Claude Code in early 2026, which forced the company into a position most engineering leaders will recognize: strict budget caps because the costs were spiraling.
Speaking at the AI Engineer conference in New York, David Abitbol, Revolut’s ops and infrastructure lead, was blunt about what was happening: “We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI.”
The token limits created a perverse constraint. Engineers were hoarding their allocations for the end of the quarter or the most difficult tasks, which meant the AI wasn’t being used where it actually helped, on the boring, time-consuming work that pads every sprint.
The self-hosted pivot happened in April 2026 with GLM-5.1, later upgraded to GLM-5.3. At $0.30 per million input/output tokens, the cost was roughly a tenth of what OpenAI’s Codex or Anthropic’s Claude Code was charging. When tokens stop being scarce, the economics of AI-assisted development shift fundamentally.
The Hardware Math That Makes It Work
Here’s where the “open weights aren’t viable” crowd starts flinching. GLM-5.3 is not a small model. At roughly 754 billion total parameters (with 40 billion activated per token), it requires about 380 GB of GPU memory just for weights, plus another 6-26 GB per user for context, depending on precision.
Revolut serves this with 56 Nvidia B200 GPUs, equivalent to seven eight-GPU DGX B200 systems. At roughly $30,000-$40,000 per chip, that’s a $1.7-2.2 million hardware investment before you account for servers, networking, power, and the engineering talent to operate it all. Not pocket change.
But here’s the trick: when you’re running 2,000 engineers through that infrastructure every day, the fixed cost amortizes faster than the monthly variable bill from Anthropic ever did. The per-token cost drops by an order of magnitude, and the bill becomes predictable. No more invoice anxiety at month-end.
The Performance Tuning Nobody Talks About
You don’t just buy GPUs and get instant 90% savings. Abitbol was explicit that out of the box, inference performance was “terrible.” The actual performance gains came from a stack of optimizations that would make most platform teams sweat:
- NVFP4 quantization to reduce memory overhead
- Disaggregated prefill and decode for better resource utilization
- Hyperparameter tuning that’s apparently more art than science
- Aggressive context-length limits capped at 256K tokens
That last point deserves attention. “Long context is a scam”, Abitbol said, directly contradicting the marketing narrative from every frontier lab selling million-token context windows. Revolut’s experience: 256K tokens is what engineers actually need, and trying to serve more just degrades latency and inflates costs.
The results after tuning: sub-3-second time to first token on 256K context windows (down from over 10 seconds), generation speeds of ~200 tokens per second during business hours, and 800 tokens per second after 5 pm when fewer engineers are online. Abitbol set a hard SLO: “As soon as you exceed 5 seconds, my Slack inbox starts to blow up.”
That’s the real engineering story here. The open-weights model was good enough, the systems engineering is what made it viable.
The Surprising Productivity Wins
Here’s what challenges the “AI writes code faster” narrative: Revolut saw a 60% improvement in coding time, but a 70% improvement in lead time, the metric from task pickup to merge.
The bottleneck in software development was never actually the code. When tokens become unlimited, engineers let AI handle the drudgery: debugging, Jira tickets, Confluence searches, navigating the sprawl of internal tooling. Abitbol’s observation cuts against the “10x engineer” hype: the AI gains come from removing overhead, not accelerating typing.
The other shift was behavioral. With flat pricing instead of metered tokens, engineers stopped rationing their AI usage. No more saving prompts for the hardest tasks. No more anxiety about burning the monthly budget on routine questions.
Why This Matters Beyond Revolut
Revolut is not a marginal startup. With over 80 million customers, $6 billion in 2025 revenue, and a valuation around $115 billion, it runs engineering at a scale where small per-engineer AI costs compound into real money fast. This is the most detailed public breakdown yet of what the AI pricing disruption looks like in practice.
The backlash against frontier API pricing has been building. When a cloud financial analyst says costs moved “from an afterthought to a product constraint”, this is what they mean. Organizations are realizing that the frontier labs’ margins are their own budget lines.
And it’s not just Revolut. The trend toward efficiency breakthroughs in small local LLMs and hybrid reasoning agents that drastically cut token consumption suggests the industry is converging on cost optimization as the primary constraint, not raw capability.
The Honest Caveats
Before you run out and buy a stack of B200s, let’s be clear about what Revolut’s numbers don’t include. The 90% savings figure covers inference costs only. It excludes the cloud infrastructure, the platform team’s salaries, and the monitoring/observability stack required to operate this reliably. StartupHub.ai’s analysis is worth reading on this point.
The hardware also has an expiration date. Nvidia’s Rubin architecture is already on the horizon, and DIY datacenter cooling and GPU efficiency remains a persistent challenge. If Revolut bought those B200s outright, it’s betting that a two-to-three year useful life gives enough ROI before the next generation makes them obsolete.
GLM doesn’t handle vision tasks either. Revolut uses a model router to send image-related queries to Gemini or other hosted models. The open-weights win is specifically for text and code generation, the high-volume, internally-facing workload that makes up the bulk of AI coding tools.
And the privacy angle: Abitbol acknowledged the “stigma” of using a Chinese open-weights model at a regulated bank. Revolut runs GLM on US infrastructure rather than Z.ai’s Chinese inference endpoints, keeps data in jurisdictions with equivalent data protection, and maintains human-in-the-loop sandboxing with approval rules for commands. The model weights might be Chinese, but the data flows stay within regulatory boundaries.
The Real Takeaway
The “10x engineer” promise was always marketing. What Revolut actually achieved was a 10x reduction in the cost of AI-assisted development, and the productivity gains followed.
The frontier labs will survive this. Anthropic’s Q3 revenue is approaching $17 billion, and there will always be workloads that demand the best model available, not the cheapest adequate one. But the era of treating frontier API spend as an unavoidable cost of doing AI is over. The muscle memory of running your own infrastructure is still strong in fintech, and Revolut’s numbers just gave every CFO a benchmark.
For engineering leaders, the math is now on the table. If you have 500+ engineers using AI coding tools, run the numbers on self-hosting an open-weights model. The extreme quantization techniques and code-as-middleware approaches emerging in the open-source ecosystem are making this more accessible by the quarter.
Revolut’s real innovation wasn’t the model. It was recognizing that at a certain scale, you stop being a customer and start being a competitor. The question every large engineering organization now faces: are you big enough to build, or are you still paying the tax?
