Europe's Biggest Tech Startup Cuts AI Coding Costs by 90% Using Chinese Model

British fintech giant Revolut has reduced its inference costs for AI coding tasks to roughly one-tenth of its previous spending on Anthropic and OpenAI.

The company achieved this by self-hosting the open-weight GLM model from the Chinese AI lab Z.ai on dedicated NVIDIA GPUs.

The European digital bank said the deployment has also improved engineering lead times by 70% and coding time by 60% across a sample of 2,000 engineers.

David Abitbol, who leads AI operations at Revolut, revealed this at a technical session at the AI Engineering conference. 

"Self-hosting GLM brings our blended cost of tokens to 30 cents per million. Our invoice is now one-tenth of what we were spending on Anthropic and OpenAI,” he said. 

Revolut was valued at $115 billion in a secondary share sale in July, up from $75 billion last November and $45 billion in 2024. 

The transaction reinforced its position as Europe's most valuable private technology company.

The company began experimenting with AI tools in January 2025, initially introducing Gemini for internal data access before testing Anthropic's Claude Code.

Concerned about commercial AI costs, with some engineers spending $1,000–$10,000 monthly without clear ROI, Revolut imposed strict usage budgets.

Abitbol said the company then began looking for ways to give engineers unrestricted access to AI coding tools without incurring comparable commercial inference costs.

In April 2026, Revolut began self-hosting open-source models, settling on the GLM family for coding workloads. 

Abitbol said Revolut downloads model weights and runs inference on US-based infrastructure rather than sending requests to Chinese-hosted inference endpoints. 

Revolut found that removing token constraints also let engineers spend less time on activities such as searching documentation, managing inboxes and coordinating development tasks.

The infrastructure supporting the deployment consists of 56 NVIDIA B200 GPUs serving approximately 2,000 engineers running GLM-5.3.

Revolut targets a time to first token below three seconds, with Abitbol saying latency exceeding five seconds causes engineers to complain. 

The company serves interactive workloads at around 200 tokens per second, although engineers can access throughput of up to 800 tokens per second after 5 pm, when GPU demand is lower.

“Open source is now a priority at Revolut. The economics are favourable, the billing is flat, and we don't have any bill shock at the end of the month,” he said. 

The company also improved inference performance without adding GPUs. 

Between April and August, requests per second increased from approximately three to more than seven, while time to first token fell from over 10 seconds to around 1.6 seconds. 

This represented an 88% performance improvement despite a 76% increase in demand.

Abitbol attributed the gains to inference optimisation rather than simply adding computing capacity. 

Revolut has not entirely replaced commercial models. It still uses frontier tools such as Claude Code, Codex, and Gemini for specialised workloads, including image recognition. Its internally hosted GLM models handle most coding tasks.

In October, CEO Nik Storonsky said Revolut was developing proprietary AI models trained on customer transaction data. 

He cited the company's 30 million to 40 million daily transactions as a source of training data.