Revolut moved 2,000 engineers off metered AI APIs onto self-hosted open models running on 56 Nvidia B200 GPUs, and says that cut its AI coding costs 90%. David Abitbol of Revolut gave the figures in a talk posted on the AI Engineer YouTube channel on 29 October, StartupHub.ai reported.
Abitbol said the bank serves GLM 4.5 at about 30 cents per million tokens, which StartupHub reports as one tenth of its prior Anthropic and OpenAI invoice. By our arithmetic, that puts the earlier cost near $3 per million tokens. For comparison, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens.
Business Pill · Lead Time Explained
A short explainer on lead time as a business concept. It teaches the general idea only and says nothing about this story.
The key insight: As we read the reports, Revolut and Decagon each kept frontier models for a narrow slice of work and moved routine, repeatable tasks to open models they run or tune themselves.
StartupHub reports that the system “was demoed as a production coding assistant for Revolut’s engineers, not a lab test”. Its performance target was “sub 3 seconds to first token and about 200 tokens per second for interactive use”.
The results came in stages. StartupHub reports that “Gemini alone delivered roughly 10% faster lead time and 20% faster coding time in early 2025, while early Claude Code with strict budget caps showed little impact”.
After self-hosting GLM in April, StartupHub says, “Revolut saw about 60% improvement in coding time and 70% in lead time on a 2,000 engineer sample”. Time to first token fell “from over 10 seconds to 1.6 seconds” while requests per second rose “from 3 to over 7”.
Abitbol attributed the larger lead-time gain to work outside coding. StartupHub’s account: “like inbox triage and meetings, that engineers now automate when tokens are not rationed”.
Improvement Revolut’s David Abitbol reported, % faster: Gemini alone in early 2025 (coding time 20%, lead time 10%) vs self-hosted GLM from April (coding time about 60%, lead time 70%), on a 2,000-engineer sample, as reported by StartupHub.ai on 29 October 2026. Self-reported figures.
What Was Posted on X
Interesting.
Revolut reportedly only started seeing meaningful AI productivity gains after moving away from expensive OpenAI and Anthropic models to self-hosted GLM.
Decagon reportedly runs ~90% of its workloads on open weight models.
Why Because when you’re running AI agents in production for e.g. customer service, cost and latency matter enormously. You don’t need a frontier model with PhD-level reasoning or even a model that knows the capital of Finland to handle every customer interaction.
We’re seeing similar dynamics across hypergrowth AI startups and increasingly the large enterprises we work with.
But most importantly, the future isn’t binary.
When use cases are new, you want the smartest general-purpose model available. But once workflows mature and become predictable, smaller, faster, specialized models increasingly win.
I wouldn’t be surprised if 80โ90% of routine enterprise AI workflows eventually run on open-weight and smaller models.
This doesn’t mean frontier labs lose.
It means the market is splitting into two massive opportunities: frontier intelligence and specialized execution at scale.
Oliver Molander pointed to Revolut and Decagon in this 1 November post. He wrote: “I wouldn’t be surprised if 80โ90% of routine enterprise AI workflows eventually run on open-weight and smaller models.” Revolut’s figures in this piece come from StartupHub.ai’s account of David Abitbol’s talk; Decagon’s split comes from Jesse Zhang’s own post.
What It Took
The savings came after tuning. StartupHub reports that the team “improved performance 88% for 76% more demand without adding GPUs by tuning context limits, quantization and parallelism”. It capped context at 256k, and NVFP4 on Blackwell “saves about 40% GPU footprint versus FP8”.
The 30-cent figure has limits. StartupHub notes that “the talk left cloud spend, staffing and observability costs outside the 30 cent token math”. Image tasks still go to Gemini or similar models, so the self-hosted path is text only.
By our arithmetic, 2,000 engineers on 56 GPUs is about 36 engineers per B200.
Decagon Says 90% Runs on Open Models
Revolut is not the only company to report this shift. Decagon co-founder Jesse Zhang wrote on X on 24 August that “Decagonโs workflows are now roughly split 90% open source, 10% closed source”. He added that “you do need a minimum amount of frontier tokens for new or novel projects”.
On The a16z Show, StartupHub reported, Zhang said frontier labs “do have small models, but you can’t really control them in the way that you want”. Co-founder Ashwin Sreenivas said of fine-tuned smaller models: “It is better at the task, it is cheaper, and it is faster.”
The Business Lever: Price per Token
Oliver Molander pointed to both companies in a post on X on 1 November. He wrote: “I wouldn’t be surprised if 80โ90% of routine enterprise AI workflows eventually run on open-weight and smaller models.” He added that frontier labs do not lose, because “the market is splitting into two massive opportunities: frontier intelligence and specialized execution at scale”.
As we read the two cases, both moved routine, repeatable work off frontier APIs and kept frontier models for the rest: Revolut for images, Decagon for new or novel projects. Our token price table lists Claude Opus 5.5 at $4 and Claude Sonnet 5.5 at $2 per million input tokens. Revolut’s reported 30 cents is its own serving cost, which excludes cloud spend, staffing and observability, while those are list prices, so the two are not like for like.
The Structural Read
As we read StartupHub’s account, Revolut’s cost cut and its speed gain arrived together: time to first token fell from over 10 seconds to 1.6 seconds as it moved to self-hosted GLM.
As we read StartupHub’s account, the gain depended on tuning: Revolut reports 88% more performance for 76% more demand without adding GPUs.
As we read Zhang’s post, Decagon frames the split as a balance, not a replacement: it says you need a minimum amount of frontier tokens for new or novel projects.
Jesse Zhang, Decagon co-founder, on X, 24 August 2026
“Decagonโs workflows are now roughly split 90% open source, 10% closed source.”
Three Implications
FOR ANTHROPIC AND OPENAI StartupHub reports Revolut’s prior Anthropic and OpenAI invoice was ten times its self-hosted cost; Decagon says frontier tokens remain needed for new or novel projects.
FOR GPU CLOUDS AND NVIDIA Revolut reports running 2,000 engineers on 56 Nvidia B200s, about 36 engineers per GPU by our arithmetic.
FOR ENTERPRISE AI BUYERS StartupHub notes the 30-cent figure left cloud spend, staffing and observability costs outside the math, and that portability to smaller teams is unproven.
The framework says the winning pattern is to “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Revolut and Decagon describe buyers treating routine model work as the abundant layer. The framework predates these reports and does not discuss either company.
We did not watch Abitbol’s talk; the figures here come from StartupHub’s account of it, and the GLM version (4.5) is as StartupHub reports. The 30-cent cost excludes cloud spend, staffing and observability, by StartupHub’s account. Decagon’s 90% is the founders’ description of its workflows, without a stated measure (tokens, requests or spend). Neither company’s figures have been independently audited. The $3 prior cost and 36 engineers per GPU are our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Revolut says self-hosting GLM on 56 Nvidia B200s cut its AI coding costs 90%, to about 30 cents per million tokens for 2,000 engineers, according to StartupHub’s account of a talk posted on 29 October. Decagon’s co-founder wrote in August that its workflows run roughly 90% on open-source models.
A note on sourcing. Revolut’s figures come from StartupHub.ai’s 29 October account of David Abitbol’s talk on the AI Engineer YouTubechannel; we did not watch the talk and did not contact Revolut. Decagon’s split comes from Jesse Zhang’s 24 August post on X and StartupHub’s report of the founders’ a16z interview. Token list prices are from our own price table. The prior-cost and per-GPU figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
