Why more companies will run their own AI models

The best AI models are closed, and they will stay that way for a while. But for most of what companies use AI for, the best model isn't needed. A model that was the best four months ago is enough. And that one you can now download and run yourself.

That is the whole argument of this post: open-weight models are close enough to the frontier, and cheap enough to run, that companies with real volume will bring their AI in-house. Regulation, not quality, is now the main thing that could stop them.

How far behind are open models?

About four months. Epoch AI tracks this with its Epoch Capabilities Index. In October 2025 it measured the gap between the best open-weight model and the best closed model at about 3.5 months. In May 2026 it put the gap at about four months, an 8-point difference on its index, roughly the distance between GPT-5 and GPT-5.5.

Put plainly: the best model you can download today is about as good as the best model you could rent this summer. For summarising contracts, answering support tickets, classifying documents or writing everyday code, that was already more than good enough this summer.

What does the price gap look like?

Here are list prices per million tokens, input and output, as published in November 2026:

The open model is cheaper even when someone else hosts it. Once you run it on your own GPUs at high load, the price per token drops further. The catch is that you pay for the GPUs whether you use them or not, which matters later.

What did Revolut actually do?

Revolut is the most concrete example I've found, and it is often described too broadly, so here is what was reported, and what wasn't.

In a talk published by AI Engineer in October 2026, Revolut described moving its internal AI coding assistant from metered Anthropic and OpenAI APIs to open-weight GLM models from Z.ai, running on 56 Nvidia B200 GPUs for roughly 2,000 engineers. According to reports of the talk:

• the blended cost came to about $0.30 per million tokens, about a tenth of what the earlier API bills had cost, which is where the "90% cheaper" headline comes from;
• time to the first token fell from over 10 seconds to 1.6 seconds between April and August 2026, on the same hardware;
• the $0.30 figure leaves out cloud spend, staff and observability, and Revolut hasn't said whether it owns or rents the GPUs;
• it covers text only: image tasks still go to Google's Gemini, and customer-facing banking systems are not part of it.

So this is not "Revolut runs all its AI on its own servers." It is one high-volume workload, coding assistance for thousands of engineers, moved in-house because the volume made the bill large enough to be worth the engineering. That is exactly the kind of workload I expect other companies to move next.

Why are the frontier labs so expensive?

Because they can be, and because they have to pay for the next model.

Anthropic's public IPO filing from October 2026 shows it. In the first half of 2026, Anthropic made $16.3 billion in revenue, and the cost of serving its models was about $2.4 billion of that, which works out to a gross margin of roughly 85%. That margin only covers running the models. Training sits in research and development, and the company still reported an operating loss of $693 million for the half year.

OpenAI hasn't published a filing. Leaked 2025 financials, reported in June 2026, showed about $13.1 billion in revenue against $7.5 billion in cost of revenue, a margin near 43%, and an operating loss of about $20.9 billion.

Both companies sell each token with a healthy margin and still lose money, because the research is so expensive. That gap between the cost of serving a token and the price you pay for it is the space open models compete in. Anthropic's filing itself lists open-weight models as a source of pricing pressure.

What could stop it? Regulation.

The awkward fact in Revolut's story is the model's origin. Many of the best open-weight models today come from Chinese labs, and GLM is made by Zhipu (Z.ai).

• Zhipu has been on the US Entity List since January 2025. That restricts US exports to Zhipu. It doesn't ban anyone from running its published weights.
• In October 2026, the US Department of Homeland Security published an order requiring federal civilian agencies and their contractors to exclude DeepSeek, including "third-party products and services developed using DeepSeek implementations." It names DeepSeek, not GLM or Qwen. But it shows how fast a model can become off-limits for a whole class of customers.
• A proposed "No Adversarial AI Act" to ban Chinese AI in US government use was introduced in June 2025 and hasn't moved since.
• Z.ai's own license for GLM-5.3 requires companies with more than $10 billion in annual revenue to pass a Z.ai security review.
• In the EU, the AI Act's rules for general-purpose models have applied since August 2025, with enforcement starting in August 2026. They mainly bind the companies that provide a model, not companies that run one internally.

There is a US-made alternative now: Meta released the weights of Muse Spark Open in October 2026, under a license with conditions of its own. For a bank in a regulated market, that may matter more than benchmark points.

The lesson I take from this: build so that you can swap the model. That is the real advantage of open weights. If a rule changes, a company running GLM can switch to another open model on the same GPUs within weeks. A company locked into one API can't do that so easily.

What do I expect?

• Hybrid stacks become normal. A router sends routine, high-volume requests to a self-hosted open model and hard ones to a frontier API.
• Regulated industries prefer US and European open weights, even if they score a little lower, because the legal risk is lower.
• Frontier prices come under pressure as the four-month gap holds, which is what Anthropic's own filing anticipates.

This connects to the other side of the open-source story. The labs keep their very best models closed, as I describe in why AI is making open source less open. But you don't need the very best model to save 90% on a workload.