Revolut cut AI coding costs 90% by self-hosting GLM on 56 GPUs
Revolut’s ops and infrastructure lead David Abitbol spoke at the AI Engineer conference in New York last month and laid out exactly how the British fintech dropped its AI coding inference costs by roughly 90 percent in under a year. The short version: they stopped defaulting to Claude and OpenAI and started running their own models on their own hardware.
The Before: $1,000 to $10,000 per engineer per month
Revolut started experimenting with AI coding tools in earnest in early 2025 with Gemini, then added Claude Code in early 2026. The cost picture got uncomfortable fast.
“We were highly skeptical about the cost of AI. We were seeing engineers spend $1,000 a month, $10,000 a month with not so much ROI.”
The company responded by capping token budgets. Engineers rationed usage, saving tokens for the hardest tasks and skipping the lower-stakes work that also benefits from AI help.
The Switch: GLM on 56 Nvidia B200 GPUs
Revolut adopted GLM-5.1 first, then moved to GLM-5.3. Both are open weights models from Chinese developer Z.ai. At roughly 754 billion total parameters with 40 billion activated per token, GLM is not a lightweight model. It needs about 380 GB of GPU memory just to load the weights, plus 6 to 26 GB per user for context depending on precision.
To serve it at scale for around 2,000 engineers, Revolut deployed 56 Nvidia B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems.
Self-hosting GLM-5.1 cost $0.30 per million input/output tokens, roughly one-tenth the cost of OpenAI’s Codex or Anthropic’s Claude Code. With pricing that low, token rationing stopped. Engineers could use the model freely for debugging, Jira tickets, Confluence documentation, and anything else that would previously have eaten into a monthly budget.
The Tuning: Where the Real Gains Came From
Adding GPUs was not the main performance lever. Abitbol identified four optimizations that drove the biggest improvements:
- Hyperparameter tuning
- Switching to NVFP4 quantization
- Implementing disaggregated prefill and decode
- Tuning context lengths
On context length specifically, Abitbol was direct: “Long context is a scam.” GLM can technically serve a one-million-token context window. Revolut found that 256,000 tokens is the practical ceiling that actually matters.
With those optimizations in place, the team hit sub-3-second time to first token on 256K context windows and a generation speed of 200 tokens per second. Abitbol noted that once latency exceeds 5 seconds, his Slack inbox fills up. Engineers working after 5 pm get closer to 800 tokens per second due to lower system load.
The Results
- 90% reduction in AI coding inference costs
- 60% improvement in coding time
- 70% improvement in lead time
The biggest productivity gains did not come from writing code faster. They came from using AI to handle debugging, Jira tickets, and Confluence work that engineers previously either did manually or skipped entirely because tokens were too expensive to spend on it.
The Caveats
GLM is text only, so it does not handle UI or vision work. Revolut solved this with a model router that sends vision queries to Gemini or GPT Luna while keeping code generation on GLM. Commercial models remain the exception, not the default.
On the security question around Chinese open weights models, Abitbol acknowledged the concern but drew a clear line: accessing these models via APIs running out of China is the actual risk. Running the same model on-premises or on regional cloud infrastructure removes that concern.
Revolut’s billing is now flat. No monthly surprises, no per-engineer spend tracking.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
