Revolut Slashes AI Coding Costs by 90% After Swapping Claude Code for Self-Hosted B200 Cluster.

Revolut Slashes Coding AI Costs by 90% After Replacing Claude Code with Self-Hosted NVIDIA B200 Infrastructure

At the AI Engineer Summit, David Abitbol, a senior software engineer at UK-based digital banking giant Revolut, revealed how the fintech firm achieved a staggering 90% reduction in AI infrastructure costs. By transitioning its enterprise developer workflows away from third-party commercial APIs like Claude Code to a self-hosted hardware stack, Revolut unlocked significantly higher developer productivity while drastically cutting operational expenditures.

With a global engineering workforce of over 2,000 developers, the move highlights a growing industry shift among technology enterprises toward internal AI infrastructure ownership.

Evolutionary AI Deployment, Hardware Architecture, and Cost Breakdown

• Multi-Year AI Model Transition:

  • Early Exploration (2025): Revolut initially integrated Google's Gemini models across its software engineering teams in early 2025 to evaluate automated code assistance.

  • Claude Code Era (Early 2026): The bank subsequently migrated to Claude Code in early 2026. However, API usage costs surged dramatically frequently reaching up to $10,000 per engineer per month in heavy-usage scenarios delivering diminishing marginal returns relative to expenditure.

• On-Premise Hardware Stack and API Cost Metrics:

  • NVIDIA B200 Cluster Deployment: Revolut deployed an internal compute cluster powered by 56 NVIDIA B200 GPUs to host open-weights foundational models on-premise.

  • Open-Source LLM Migration: The infrastructure initially ran the GLM-5.1 open-source architecture before upgrading to the newer GLM-5.3 model.

  • Dramatic Token Cost Reduction: Self-hosting reduced token processing costs to just $0.30 per million tokens—approximately one-tenth of the price of commercial services like Claude Code or OpenAI Codex.

  • Developer Efficiency Gains: Alongside the 90% cost savings, Revolut registered a 60% reduction in overall time spent writing code across its engineering divisions.

• Behavioral Impact of Uncapped Token Budgets:

  • Eliminating Token Conservation Behavior: Previously, developers operated under strict monthly token quotas, forcing them to hoard API capacity for complex coding tasks or pace usage toward the end of financial quarters.

  • Expanded Enterprise Automation: With near-zero marginal token costs, engineers began applying AI models across broader lifecycle tasks, including automated code debugging, real-time ticket state management inside Jira, and automated documentation syncing across Confluence.

Optimization Techniques, Latency Tuning, and Multi-Model Routing

• Hardware Quantization and Context Window Refactoring:

  • NVFP4 Precision Quantization: Abitbol detailed key technical tuning strategies, emphasizing the implementation of NVFP4 (NVIDIA 4-bit Floating Point) quantization to maximize GPU throughput and memory bandwidth efficiency on the B200 hardware.

  • Refuting the Ultra-Long Context Myth: Abitbol challenged the prevailing industry assumption that larger context windows yield better results, labeling extreme context length a misnomer. Revolut capped its operational context window at an optimal 256K tokens, balancing deep code retrieval with high generation speed.

  • Throughput and Time-to-First-Token (TTFT): Following optimization, the B200 cluster sustained a throughput of 200 tokens per second during peak working hours, successfully reducing Time-to-First-Token latency down to 3 seconds. During off-peak evening hours, throughput peaked at nearly 800 tokens per second.

• Hybrid Multi-Model Routing Strategy:

  • Limitations of Text-Only Open Models: Abitbol noted that text-only open models like GLM cannot entirely replace proprietary frontier models like GPT or Claude across all software domain tasks, particularly when handling UI design, image parsing, or complex visual front-end engineering.

  • Automated Model Router: To address this, Revolut built an intelligent internal model router. The gateway automatically routes standard coding, debugging, and ticket-triage requests to the local B200 GLM cluster while dynamically dispatching complex multimodal or front-end tasks to commercial APIs as needed.

 

Source: The Register 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Revolut Slashes Coding AI Costs by 90% After Replacing Claude Code with Self-Hosted NVIDIA B200 Infrastructure

At the AI Engineer Summit, David Abitbol, a senior software engineer at UK-based digital banking giant Revolut, revealed how the fintech firm achieved a staggering 90% reduction in AI infrastructure costs. By transitioning its enterprise developer workflows away from third-party commercial APIs like Claude Code to a self-hosted hardware stack, Revolut unlocked significantly higher developer productivity while drastically cutting operational expenditures.

With a global engineering workforce of over 2,000 developers, the move highlights a growing industry shift among technology enterprises toward internal AI infrastructure ownership.

Evolutionary AI Deployment, Hardware Architecture, and Cost Breakdown

• Multi-Year AI Model Transition:

  • Early Exploration (2025): Revolut initially integrated Google's Gemini models across its software engineering teams in early 2025 to evaluate automated code assistance.

  • Claude Code Era (Early 2026): The bank subsequently migrated to Claude Code in early 2026. However, API usage costs surged dramatically frequently reaching up to $10,000 per engineer per month in heavy-usage scenarios delivering diminishing marginal returns relative to expenditure.

• On-Premise Hardware Stack and API Cost Metrics:

  • NVIDIA B200 Cluster Deployment: Revolut deployed an internal compute cluster powered by 56 NVIDIA B200 GPUs to host open-weights foundational models on-premise.

  • Open-Source LLM Migration: The infrastructure initially ran the GLM-5.1 open-source architecture before upgrading to the newer GLM-5.3 model.

  • Dramatic Token Cost Reduction: Self-hosting reduced token processing costs to just $0.30 per million tokens—approximately one-tenth of the price of commercial services like Claude Code or OpenAI Codex.

  • Developer Efficiency Gains: Alongside the 90% cost savings, Revolut registered a 60% reduction in overall time spent writing code across its engineering divisions.

• Behavioral Impact of Uncapped Token Budgets:

  • Eliminating Token Conservation Behavior: Previously, developers operated under strict monthly token quotas, forcing them to hoard API capacity for complex coding tasks or pace usage toward the end of financial quarters.

  • Expanded Enterprise Automation: With near-zero marginal token costs, engineers began applying AI models across broader lifecycle tasks, including automated code debugging, real-time ticket state management inside Jira, and automated documentation syncing across Confluence.

Optimization Techniques, Latency Tuning, and Multi-Model Routing

• Hardware Quantization and Context Window Refactoring:

  • NVFP4 Precision Quantization: Abitbol detailed key technical tuning strategies, emphasizing the implementation of NVFP4 (NVIDIA 4-bit Floating Point) quantization to maximize GPU throughput and memory bandwidth efficiency on the B200 hardware.

  • Refuting the Ultra-Long Context Myth: Abitbol challenged the prevailing industry assumption that larger context windows yield better results, labeling extreme context length a misnomer. Revolut capped its operational context window at an optimal 256K tokens, balancing deep code retrieval with high generation speed.

  • Throughput and Time-to-First-Token (TTFT): Following optimization, the B200 cluster sustained a throughput of 200 tokens per second during peak working hours, successfully reducing Time-to-First-Token latency down to 3 seconds. During off-peak evening hours, throughput peaked at nearly 800 tokens per second.

• Hybrid Multi-Model Routing Strategy:

  • Limitations of Text-Only Open Models: Abitbol noted that text-only open models like GLM cannot entirely replace proprietary frontier models like GPT or Claude across all software domain tasks, particularly when handling UI design, image parsing, or complex visual front-end engineering.

  • Automated Model Router: To address this, Revolut built an intelligent internal model router. The gateway automatically routes standard coding, debugging, and ticket-triage requests to the local B200 GLM cluster while dynamically dispatching complex multimodal or front-end tasks to commercial APIs as needed.

 

Source: The Register