Can self-hosted open models really cut your AI coding costs by 90%?

Last Updated: 2026-11-06

Key takeaways

• ▪ Revolut moved about 2,000 engineers from Claude Code and Codex to a self-hosted GLM deployment and reported roughly 90% lower AI coding inference costs in under a year.
• ▪ The saving came from flat, cheap tokens, not from a smaller model bill alone. Once tokens stopped being rationed, engineers used AI for debugging, Jira tickets and Confluence work.
• ▪ Revolut serves GLM-5.3 on 56 Nvidia B200 GPUs. A 754 billion parameter model needs roughly 380 GB of GPU memory for weights before any context is added.
• ▪ A smaller team can run the same play on two DGX Sparks with an open-weight model. We measured 35 tokens per second on one stream and 62 tokens per second at four concurrent streams.
• ▪ Our illustrative break-even for 10 developers is 2.5 months, assuming 5 million tokens each per working day and a A$12,000 hardware cost. The model is a planning tool, not a quote.
• ▪ Self-hosting still loses on vision work, frontier edge cases and peak concurrency, so the winning pattern is a router, not a wholesale replacement.

Revolut says it cut AI coding inference costs by roughly 90% in under a year by making self-hosted open models the default for about 2,000 engineers. The number is real enough to plan against, but the lesson is not "cancel Claude". It is that cheap tokens change how people work. Revolut's infrastructure lead told the AI Engineer conference in New York that the biggest gains came from letting engineers spend tokens freely on debugging and ticket work, not from writing code faster. The playbook is a hybrid: a self-hosted model for the bulk of code work, a router for vision and frontier cases, and a harness good enough that the cheaper model can finish the job.

What did Revolut actually change?

Revolut replaced its primary coding tools, Claude Code and Codex, with a self-hosted deployment of Z.ai's GLM model. It started on GLM-5.1 and moved to GLM-5.3, served on 56 Nvidia B200 GPUs. The company reported about 90% lower inference cost, a 60% improvement in coding time and a 70% improvement in lead time after rolling AI tools out to roughly 2,000 engineers, according to The Register's coverage of the AI Engineer talk. Those performance figures are company-reported and should be treated as directional. The cost figure is easier to reason about because it comes with a unit price.

At roughly $0.30 per million input and output tokens, self-hosted GLM cost about a tenth of Claude Code or Codex. Before the switch, Revolut had engineers spending between $1,000 and $10,000 a month with what its infrastructure lead described as not much return. The company had imposed strict budget limits, which meant engineers rationed tokens. Cheap tokens removed the rationing, and usage spread into the unglamorous work that actually consumes a developer's week: debugging, Jira tickets, Confluence pages, test failures and small refactors.

Why do cheap tokens change behaviour more than bills?

Rationing changes behaviour before it changes spend. When every prompt has a visible price, engineers save tokens for the task they think matters most and skip the small investigations that would have paid for themselves. The result is a low, tidy bill and a low, tidy return. Revolut's experience suggests the opposite curve: once the marginal token is nearly free, people run the extra check, ask the model to read the ticket, and let it try the boring fix. The Hacker News thread on the move captured this well. One highly rated comment argued that for a large number of tasks you do not need a frontier model, you need a good harness around it. Another pushed back hard, saying that swapping Opus for GLM would be a huge downgrade for their work. Both can be true at the same time, which is why the router matters.

What hardware does a 754 billion parameter model need?

GLM-5.3 is not a small model. It has roughly 754 billion total parameters with about 40 billion activated per token, so it needs around 380 GB of GPU memory just to hold the weights, plus another 6 to 26 GB per user for context depending on precision. That is why Revolut runs 56 B200 GPUs, the equivalent of seven eight-GPU DGX B200 systems. The interesting part is that the biggest performance gains came from tuning rather than adding GPUs: NVFP4 quantization, disaggregated prefill and decode, and deliberate context limits. Revolut's lead called long context a scam and said 256,000 tokens is what engineers actually need, even when the model can serve a million. With that tuning, Revolut reported sub-3-second time to first token at 256,000 tokens and about 200 tokens per second of generation, rising toward 800 tokens per second for engineers working after 5pm.

GLM is also text only, so Revolut routes vision queries to Gemini or GPT Luna and keeps code generation on the local cluster. That detail matters more than the headline saving. The architecture is not "open model replaces everything". It is a default path plus an exception path, with the frontier reserved for the work that needs it.

What does the same play look like at 10 developers?

Most Australian teams cannot buy 56 B200s, and most do not need to. A 284 billion parameter class open-weight model such as DeepSeek V4.1-Flash fits on two DGX Sparks, which is the setup we run. On our dual Spark cluster we measured 35 tokens per second on a single stream and 62 tokens per second at four concurrent streams. That is slower than a B200 cluster and it will queue under heavy parallel use, but it is comfortably interactive for a small team that runs tasks in parallel and checks in when they finish, which is how Revolut says its engineers now work.

Here is the break-even arithmetic for a 10-person team, using round assumptions. Five million tokens per developer per working day, 22 working days a month, so 1.1 billion tokens a month. At a blended hosted price of A$4.50 per million tokens, roughly the Australian equivalent of US$3, that is A$4,950 a month. Two DGX Sparks draw about 480 watts together, which is roughly 350 kWh a month, or about A$105 at A$0.30 per kWh. The hardware is about A$12,000 all in, an illustrative figure rather than a quote. Net saving is about A$4,845 a month, so the hardware pays back in about 2.5 months. After that the marginal cost of a token is electricity.

Where does self-hosting still lose?

Self-hosting loses on three fronts. First, capability: a 284B open-weight model is good at code, tickets and refactors, but it is not a frontier model on hard reasoning, large refactors across unfamiliar codebases, or UI work that needs vision. Second, concurrency: two Sparks serve a handful of parallel sessions well, not twenty. Third, operations: someone has to own model upgrades, quantization, evaluation, security patches and the harness that gives the model repo context, test runs and tool access. A cheap token on a badly wired harness is just a cheaper bad answer.

The harness is the part most teams underestimate. Revolut did not simply swap a model. It tuned serving, capped context, routed by modality and gave engineers tools that let the model act on the repo and the ticket system. If your setup is a chat box with a copy-pasted file, moving to a local model will save money and lose quality. If your setup can read the repo, run the tests and read the ticket, a local model becomes a credible default.

What should an Australian team do this quarter?

Start with a measured pilot, not a procurement project. Pick one team of five to ten developers, give them an open-weight coding model on owned hardware, and keep a hosted frontier model behind a router for vision and hard cases. Instrument three things: tokens per developer per day, time to first token, and the share of tasks the local model completes without escalation. Compare that with the same team's hosted baseline for a month. If the local path completes most routine work and the escalation rate is tolerable, the economics will look like Revolut's at a smaller scale. If it does not, you have spent a month and a modest hardware bill to learn where your frontier dependency actually sits.

For the inference side of the maths, see our earlier post on what actually makes a local LLM fast. For the harness side, see how an open agent harness changes the build-versus-buy decision. The two together are the real playbook: cheap inference plus a harness that can use it.

Frequently asked questions

Can self-hosted open models really cut AI coding costs by 90%?

Yes, for large engineering organisations with the right workload mix. Revolut reported roughly 90% lower inference cost after moving about 2,000 engineers to self-hosted GLM on 56 B200 GPUs. The saving depends on volume, utilisation and how much work stays on the local model. A small team with low token volume will see a smaller percentage saving because hardware and operations costs are spread over fewer developers.

What hardware do you need to self-host a coding model?

It depends on model size. A 754 billion parameter model such as GLM-5.3 needs roughly 380 GB of GPU memory for weights alone, which is why Revolut uses 56 B200 GPUs. A 284 billion parameter class model can run on two DGX Sparks with 128 GB of unified memory each, at lower throughput. Plan for 35 to 60 tokens per second on a dual Spark setup, not data-centre throughput.

Is GLM-5.3 as good as Claude Code or Codex?

Not across every task. GLM-5.3 is text only, so it is weaker on UI and vision work, and frontier models still lead on hard reasoning and unfamiliar codebases. Revolut routes vision queries to Gemini or GPT Luna and treats the frontier as the exception. With a good harness, the local model handles the bulk of routine coding, debugging and ticket work.

What is the break-even point for a small team?

Using our illustrative assumptions, about 2.5 months. Ten developers at 5 million tokens each per working day consume 1.1 billion tokens a month, which costs about A$4,950 at a blended hosted price of A$4.50 per million tokens. Two DGX Sparks cost about A$12,000 and draw about A$105 of electricity a month, so the net saving is about A$4,845 a month. Change any assumption and the answer moves.

Does self-hosting keep client code onshore?

Yes, if the hardware and the model weights sit in your environment and the router does not send code to a hosted provider. That is a genuine advantage for Australian consultancies and regulated clients with data residency obligations. You still need egress controls and audit logs to prove that code stayed local, and you should document which requests are allowed to leave for vision or frontier work.

What should stay on hosted APIs?

Vision and UI work, hard reasoning and large refactors across unfamiliar codebases, peak concurrency bursts, and any task where the frontier model's quality difference is worth the price. The practical pattern is a router with a local default and a hosted exception path, plus a budget and an audit trail for the exception path.