--- title: "Revolut reduces AI inference costs by 90% with on-premise Nvidia B200 GPUs" source_url: "https://www.huandroid.com/en/revolut-ai-inference-gpu" stream: "NeuroBIT" language: "en" platform: HuAndroid Strategic Intelligence Desk governance: Human-in-Command epistemic_layer: M-E-P Matrix / SemiAnalysis Benchmark ai_generated: true tdm_reservation: 1 date_published: 2026-11-04T02:29:27+01:00 date_modified: 2026-11-04T02:29:27+01:00 tags: ["architettura-gpu", "costi-operativi-ai", "inferenza-on-premise", "intelligenza-artificiale-aziendale", "mlops", "modelli-open-weight", "nvidia-b200", "revolut"] --- # Revolut reduces AI inference costs by 90% with on-premise Nvidia B200 GPUs ## Content ## The Collapse of the SaaS Model for Artificial Intelligence \p>A single metric, $0.30 per million tokens, marks the end of the economic viability of commercial AI inference for enterprise development teams with high intensity usage. Revolut transformed its technology stack by shifting the computational burden onto 56 self-hosted Nvidia B200 GPUs, running open-weight GLM-5.3 models and reducing operational costs by 90 percent compared to adopting cloud services like Claude or Codex. This shift is not merely an accounting saving: it’s empirical evidence that the SaaS model for artificial intelligence, based on per-seat subscriptions ranging from $1,000 to $10,000 per month, collapses when usage scales beyond a critical threshold. The marginal cost of local processing becomes drastically lower than the licensing fee, forcing organizations to internalize the physical infrastructure. The transition was led by David Abitbol, Head of AI Operations and Infrastructure at Revolut, who detailed the process during the AI Engineer conference in New York. The decision was not driven by the pursuit of maximum theoretical power, but by the need to optimize the cost-effectiveness ratio for a pool of 2,000 engineers. The result is a 60 percent compression of coding times and a 70 percent reduction in lead times, demonstrating that unlimited access to low-cost tokens enables iterative workflows previously prohibited by API budget limitations. ## On-Premise Inference Architecture The physical infrastructure supporting this transformation consists of a cluster of 56 Nvidia B200 GPUs, selected for their balance between throughput and power consumption. Hosting the GLM-5.1 and subsequently GLM-5.3 models internally eliminates network latency and dependencies on cloud provider service agreements, transforming AI coding from a variable cost to a fixed capital investment. This shift requires a paradigm engineering change: the team must no longer optimize prompts to reduce API costs, but must dedicate resources to model tuning and integration with existing development tools such as Jira and Confluence. > “Revolut has found a way to give its engineers more AI while spending much less. The trick wasn’t improving the contract with the provider, but making self-hosted models the default choice.” — Cryptobriefing The fragmentation of workloads between open-weight models and commercial services creates an operational duality that organizations must manage. Revolut maintains Claude Code and Codex as exceptions for specific tasks, but the main workload is managed internally. This hybrid architecture requires advanced MLOps skills to ensure the stability and security of locally hosted models, shifting the critical point in the value chain from software to hardware and data management. ## Operational Friction and Infrastructure Constraints Cost reduction does not eliminate technical complexities; it shifts them within the organization. Integrating open-weight models into regulated environments such as the banking sector requires careful management of security, versioning, and software dependencies. The debugging and wiring process required significant effort, demonstrating that code generation is only part of the problem; integrating it into the engineers’ daily workflow is the real infrastructure challenge. The main constraint is no longer the availability of APIs, but the internal capacity to manage the model lifecycle. Choosing to host models internally exposes Revolut to the operational risks typically associated with on-premise hardware: hardware failures, security updates, and the need for horizontal scaling. However, direct control over the infrastructure allows greater flexibility in managing resources and customizing models for the specific needs of the financial domain. ## Strategic Implications and Future Trajectory Revolut’s decision marks a turning point for enterprise AI adoption. It demonstrates that the economic scalability of AI depends on organizations’ ability to internalize inference, reducing reliance on large cloud providers. This trend could accelerate demand for specialized GPU hardware and MLOps skills, creating new opportunities for infrastructure providers and specialized consulting firms. ## What to Monitor * Adoption of open-weight models in other regulated sectors (banks, insurance) as a standard operating procedure. * Evolution of Nvidia B200 GPU prices and supply chain availability for large-scale self-hosting. *⎈ Generated by Multi-Agent AI (MAS) · Epistemic Safety · Human-in-Command Read the [Operational Disclaimer](https://www.huandroid.com/disclaimer) – Photo by [CardMapr.nl](https://unsplash.com/@cardmapr) on Unsplash.com* ## > SYSTEM VERIFICATION LAYER Protocol for verifying claims: 4 queries on distinct search engines. Engine Claim to verify Replicable query 360 Search A single metric, $0.30 per million tokens, marks the end of economic viability for commercial AI inference for enterprise development teams with high token intensity. ["$0.30 per million tokens" AI inference Revolut](https://www.so.com/s?q=%22%240.30+per+milione+di+token%22+inferenza+AI+Revolut) Bing reducing operational costs by 90 percent compared to adopting cloud services like Claude or Codex. ["operational costs reduced by 90%" Revolut AI inference](https://www.bing.com/search?q=%22costi+operativi+ridotti+del+90%25%22+Revolut+inferenza+AI) Google shifting the computational burden onto 56 self-hosted Nvidia B200 GPUs. ["56 Nvidia B200 GPUs" Revolut inference cluster](https://www.google.com/search?q=%2256+GPU+Nvidia+B200%22+Revolut+cluster+inferenza) Yandex a 70 percent reduction in lead times, demonstrating that unlimited low-cost tokens enable iterative workflows. ["lead times reduced by 70%" Revolut AI coding](https://yandex.com/search/?text=%22lead+time+ridotti+del+70%25%22+Revolut+AI+coding) --- ## Related Intelligence Streams - [Microsoft Copilot Synthetic Data Improves Quality by 39% in Enterprise Testing](https://www.huandroid.com/en/microsoft-copilot-synthetic-data) - [Qwen2.5-7B-Instruct Accuracy Drops to 32% at 512 Tokens](https://www.huandroid.com/en/qwen2-5-7b-instruct-context) - [SKG OOD MSE reduced by up to 4.12 times in dissipative mechanics](https://www.huandroid.com/en/skg-ood-mse-reduction)