**Revolut says it has cut its AI coding costs by about 90% by dropping per-token payments to commercial model providers such as Anthropic's Claude and OpenAI and running an open model on its own GPU infrastructure.
AI infrastructure · Revolut
Revolut says self-hosting an open model cut its AI coding costs by about 90%
Instead of paying per token to commercial providers, the company reportedly runs GLM 4.5 on its own GPUs, serving about 2,000 engineers at an effective cost of roughly $0.30 per million tokens.
56
NVIDIA B200 GPUs running the model in-house
~2,000
engineers using the setup
~$0.30
effective cost per million tokens (reported)
~90%
claimed reduction in AI coding spend
Relative cost of the same coding work
Previous spend shown as 10 units, matching the reported "about a tenth" figure. Each block is one unit.
~10×
Previous per-token API spend
1×
Self-hosted open model (about 90% lower)
The trade-off
Owning the inference stack changes the cost math, but it moves cost into new places.
Hardware
Dedicated GPUs must be bought or leased and kept running.
Staff
Engineers must operate, monitor, and maintain the system.
Capacity planning
Peak usage must be covered so developers stay productive.
Why it matters
Heavy code-assist traffic can make per-token API billing one of the largest AI budget lines.
Open-weight models are being judged as production options for internal developer tools.
Hosted providers may have to justify their pricing more directly.
What is still unproven
The savings are company-level claims and have not been independently verified.
The breakdown between hardware amortization, utilization, and the model itself is not disclosed.
Commercial frontier models may still lead on capability, reliability, and ease of use.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...