---
title: "Revolut Cut AI Coding Costs About 90% by Self-Hosting Open Models. What the Numbers Actually Say"
slug: "revolut-self-hosted-open-models-ai-coding-costs"
author: "ibl.ai"
date: "2026-11-04 12:30:00"
category: "Premium"
topics: "Revolut AI, self-hosted LLM, open-weight models, GLM, AI coding costs, LLM cost per million tokens, Nvidia B200, Decagon open-source models, frontier API pricing, model-agnostic AI, enterprise AI cost, self-hosting inflection point"
summary: "Revolut says self-hosting open GLM models on 56 Nvidia B200 GPUs cut its AI coding costs to roughly a tenth of what it paid frontier labs, about $0.30 per million tokens. What the talk supports, what the viral version got wrong, and what it means for organizations without a GPU cluster."
banner: ""
thumbnail: ""
linkedin: |
  Revolut says it cut AI coding costs to roughly a tenth by moving off frontier-lab APIs onto open GLM models it hosts itself, about $0.30 per million tokens on 56 Nvidia B200 GPUs.

  The viral version of this story has three details wrong. The write-up of the talk names GLM 4.5, not GLM-5.3, and says self-hosting began in April, before GLM-5.3 existed. The "$3 per million before" figure appears in no source we could find. And Revolut's last reported valuation is about $115 billion, not $200 billion.

  The core result stands, and it is a big one. Most production work does not need the most expensive model. Decagon's CEO has said about 90% of its workloads run on open-source models.

  The lesson for everyone without 56 B200s: the saving comes from being able to choose the model per task. That only works if your agents are not wired to one vendor.

  #iblai #EnterpriseAI #OpenSource #LLM #AIInfrastructure
---

## The Short Answer

**Revolut says self-hosting open GLM models cut its AI coding costs to roughly a tenth of frontier-lab API prices, about $0.30 per million tokens. The saving comes from choosing the model per task. On ibl.ai you own all the code and the data and run agents model-agnostic, so that choice stays yours.**

The account comes from a talk by David Abitbol, who leads AI operations at Revolut, published by the [AI Engineer channel](https://www.youtube.com/watch?v=Ok6QdZHCcI8) on October 29, 2026, and a [StartupHub write-up](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/revolut-says-self-hosted-ai-cut-coding-costs-90) the same day.

## What did Revolut actually change in its AI coding setup?

Revolut moved its AI coding work off frontier-lab APIs, which StartupHub names as Anthropic and OpenAI, onto open GLM models it serves itself. Per that write-up, it runs them on 56 Nvidia B200 GPUs and began self-hosting in April.

The talk's title states the headline: AI coding costs cut 90%. StartupHub reports GLM 4.5 served at about 30 cents per million tokens, roughly a tenth of what Revolut had been paying.

Revolut also reported about 60 to 70% faster delivery, and says tuning made the same GPUs 88% faster. The delivery figures come from a sample of about 2,000 engineers.

## What did the viral version of the Revolut story get wrong?

The viral version got three details wrong, and none of them change the conclusion. First, the model: posts said GLM-5.3, but the written account names GLM 4.5. [Z.ai released GLM-5.3](https://technode.com/2026/08/17/z-ai-launches-glm-5-3-with-claimed-50-gain-on-coding-benchmark/) in August 2026, after Revolut started.

Second, the "$3 per million tokens before" figure. We found it in no source; the talk supports "roughly a tenth", not a specific prior price.

Third, the valuation. Revolut's most recent reported mark is about $115 billion, from a [July 2026 secondary share sale](https://tech.eu/2026/07/22/revolut-confirms-fresh-secondary-share-sale-at-reported-115bn-valuation/). $200 billion is a target, not a valuation.

## Is Revolut the only company moving production AI to open models?

Revolut is not alone. Decagon CEO Jesse Zhang has said the company runs about 90% of its workloads on open-source models rather than OpenAI's or Anthropic's, as [36Kr reported](https://www.36kr.com/p/3889567082265352) in July 2026.

The pattern is the same in both: most production tasks, code completion, classification, extraction, routine drafting, run well on models that cost a fraction of the frontier.

Even hosted list prices show the spread. [Z.ai's pricing page](https://docs.z.ai/guides/overview/pricing) lists GLM-4.5 at $0.60 per million input tokens and $2.20 output, and GLM-5.3 at $1.40 and $4.40.

## Does self-hosting open models make sense without a GPU cluster?

Self-hosting at Revolut's scale needs a GPU cluster and a team to tune it. Most organizations have neither, and should not pretend to.

The transferable part is not the hardware. It is the right to pick the model per task: a cheap open model for the bulk of the work, a frontier model where the task earns it, and the ability to move either when prices change.

For the serving side, our comparison of [vLLM, Ollama, TGI and LocalAI](/blog/self-hosted-llm-serving-stack-ollama-vllm-tgi) covers which runtime fits which job and what hardware each one needs.

## Why does model-agnostic matter more than the model you pick?

Model-agnostic matters more because the cheapest good model changes every few months. GLM went from 4.5 to 5.3 in about a year, and pricing shifted each time.

An agent wired to one vendor's API cannot follow those prices. An agent that names its model as a setting can. On ibl.ai, each agent's [LLM is chosen per agent](/docs/os/agent-settings/llm-selection), and [Costs](/docs/os/analytics/costs) shows what each one spends.

That is also the security half of the story. A model you host keeps prompts and code inside your network, as Revolut's does.

## Who should own the AI stack that makes this saving possible?

The organization paying for the tokens should own the AI stack. Revolut could switch because it controlled where inference ran. Most companies renting AI per seat cannot.

On ibl.ai you own all the code and the data, the platform runs model-agnostic, there is no per-seat pricing, and you can deploy anywhere, from your own cloud to a fully air-gapped network. That is the same choice Revolut made, without building the platform yourself.

Revolut's earlier move, [its own foundation model, PRAGMA](/blog/banks-building-own-ai-foundation-models-revolut-pragma), came from the same instinct. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.

*Sources: the talk "How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models" by David Abitbol on the [AI Engineer channel](https://www.youtube.com/watch?v=Ok6QdZHCcI8), October 29, 2026; model, price, GPU count, start month and delivery figures from [StartupHub's account](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/revolut-says-self-hosted-ai-cut-coding-costs-90); GLM-5.3 release from [TechNode](https://technode.com/2026/08/17/z-ai-launches-glm-5-3-with-claimed-50-gain-on-coding-benchmark/); list prices from [Z.ai](https://docs.z.ai/guides/overview/pricing); valuation from [Tech.eu](https://tech.eu/2026/07/22/revolut-confirms-fresh-secondary-share-sale-at-reported-115bn-valuation/); Decagon from [36Kr](https://www.36kr.com/p/3889567082265352). The briefs said GLM-5.3, a $3 prior price, a $200 billion valuation and an AI Engineering NYC venue; the sources support GLM 4.5, no prior price, about $115 billion, and an AI Engineer talk with the venue unconfirmed.*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.