Revolut Slashes AI Coding Costs 90% by Ditching Claude for Self-Hosted Open-Source Models

Revolut Achieves 90% Cost Reduction Through Self-Hosted AI Infrastructure

British fintech giant Revolut has dramatically reduced its AI coding costs by roughly 90% in less than a year by transitioning from commercial AI services to self-hosted models1

1

. Speaking at the AI Engineer conference in New York, David Abitbol, operations and infrastructure lead at Revolut, revealed that the company now pays approximately $0.30 per million input/output tokens by self-hosting the open-source GLM model from Chinese AI lab Z.ai2

2

. This represents about one-tenth the cost of using OpenAI's Codex or Anthropic's Claude Code1

1

.

The deployment has delivered significant improvements beyond cost savings. Revolut saw a 60% improvement in coding time and a 70% improvement in engineering lead times following the rollout of AI coding tools to roughly 2,000 engineers1

1

. The company, valued at $115 billion in July 2026, up from $75 billion in November and $45 billion in 2024, has reinforced its position as Europe's most valuable private technology company2

2

.

From Budget Constraints to Unlimited Tokens

Revolut's journey began with experimentation in early 2025 when it deployed Gemini, followed by Claude Code in early 20261

1

. However, the company remained highly skeptical about AI coding costs after witnessing engineers spend $1,000 to $10,000 monthly without clear return on investment. This prompted Revolut to launch AI tools with strict budget limits, giving engineers access to only a limited number of tokens1

1

.

In April 2026, Revolut began self-hosting open-source models, ultimately settling on the GLM family for coding workloads2

2

. The company initially adopted GLM-5.1 before moving to GLM-5.3, which it now serves using 56 NVIDIA B200 GPUs1

1

. When tokens became cheap, engineers no longer needed to save them for the end of the quarter or the most difficult tasks. Abitbol noted that the biggest gains came from letting AI models handle debugging and Jira and Confluence tickets rather than just code generation1

1

.

Technical Infrastructure and Performance Optimization

Developed by Z.ai, GLM 5.1 operates at roughly 754 billion total parameters with 40 billion activated for each token generated. The model requires approximately 380 GB of GPU memory just to hold the weights and another 6-26 GB per user to maintain context depending on precision1

1

. Revolut's infrastructure consists of 56 NVIDIA B200 GPUs, equivalent to seven eight-GPU DGX B200 systems, serving approximately 2,000 engineers1

1

2

2

.

The greatest performance gains came not from adding NVIDIA GPUs but from hyperparameter tuning, switching to NVFP4 quantization, implementing disaggregated prefill and decode, and optimizing context lengths1

1

. 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—an 88% performance improvement despite a 76% increase in demand2

2

.

Abitbol emphasized that supporting a one-million-token context window doesn't necessarily make such a large window useful or efficient. "Long context is a scam. You really need to be mindful of that," he said, noting that 256,000 tokens is what engineers actually need1

1

. The company targets a time to first token below three seconds, with latency exceeding five seconds causing engineers to complain. Revolut serves interactive workloads at around 200 tokens per second, though engineers can access throughput up to 800 tokens per second after 5 pm when GPU demand is lower1

1

2

2

.

Hybrid Approach Balances Cost and Capability

While self-hosted models have become Revolut's default for fintech AI infrastructure, GLM hasn't replaced commercial services entirely. Since the Chinese model is text-only, it isn't ideal for tasks like UI work. Revolut deployed a model router to direct vision queries to models like Gemini or GPT Luna while keeping code generation running on GLM1

1

. The company still uses frontier tools such as Claude Code, Codex, and Gemini for specialized workloads including image recognition2

2

.

"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 said1

1

2

2

. Abitbol acknowledged that Chinese open-weight models may raise privacy and security concerns, but emphasized that the major risk involves accessing them via APIs running out of China. When run on-premises or on regional cloud resources, there's minimal concern1

1

. Abitbol confirmed that Revolut downloads model weights and runs inference on US-based infrastructure rather than sending requests to Chinese-hosted endpoints2

2

. In October, CEO Nik Storonsky revealed that Revolut was developing proprietary AI models trained on customer transaction data, citing the company's 30 million to 40 million daily transactions as a training data source2

2

.