Sept. 21 was a Monday, and Xiaomi used it to publish something that a number of well-funded American companies should read twice. MiMo-V2.6 arrived in two flavors — Pro and Flash — with open model weights, a technical report, the RL environments, and the training code. Artificial Analysis measured Pro at 46 on its Intelligence Index, up from 26 for April’s V2.5-Pro, and ranked it the top open-weights model in the world. It is a mixture-of-experts model with 1.02 trillion total parameters, 42 billion active, a million-token context window, and text, image, audio and video input.

A consumer electronics company did that. Xiaomi’s core business is phones, scooters and air purifiers. If your mental model of frontier AI is a capital-intensive industry with three or four survivors, the back half of September should have dented it.

Ninety-Nine Percent Off Is the Actual Headline

The parameters aren’t the interesting part either. The price sheet is. Xiaomi is charging $0.435 per million uncached input tokens and $0.87 per million output tokens for Pro. Flash runs $0.14 and $0.28. For scale, VentureBeat’s comparison table lists Gemini 3.7 Flash at $0.75 in and $3.75 out through the end of the year.

Everyone quoted those numbers on Monday. Almost nobody quoted the cache-hit line. Pro’s cached input price is $0.0036 per million tokens — a discount of roughly 99% against the uncached rate. Flash’s cached input is $0.0028.

That is the number that matters, because of what the industry actually does with models now. The work that pays is not one-shot question answering. It’s agents: a process that wakes up, re-reads the same file tree, the same ticket queue, the same contract stack, the same repository, and takes one more step. In a loop like that, the overwhelming majority of tokens are context that hasn’t changed since the last iteration. Cached input isn’t an edge case. It’s the workload.

At $0.0036 per million, re-reading your own context stops being a line item. It stops being a design constraint. You can build a workflow that would have been economically absurd eighteen months ago and never think about it again.

The Casualty Isn’t OpenAI

Here is where the standard story goes wrong. The standard story says cheap open weights squeeze the frontier labs. They don’t, really. Frontier labs sell enterprise contracts, compliance paperwork, model choice as a procurement checkbox, and a brand a general counsel can defend at a board meeting. That business has margin to lose and can lose it for years.

The firms with no margin to lose are the ones in the middle — American inference hosts, model routers, “sovereign AI” hosting outfits, gateway startups whose entire product is someone else’s open weights behind a nicer billing page. They raised money on a straightforward thesis: open weights would commoditize the model layer, and they would be the ones selling the commodity. Their cost basis is Xiaomi’s cost basis. The same weights, the same accelerators. The only thing they control is the markup.

Xiaomi just set that markup at approximately zero. You cannot undercut a company willing to charge a rounding error for the input side, because there is nothing left to undercut. And they can’t match it by being better engineers, because the input is identical.

A solutions engineer at a US inference host, in a Slack DM a little after midnight once the pricing page went live, put it more plainly than the analysts did: “There’s no version of our roadmap where we win a price fight with a phone company.”

You Can’t Regulate a Free Download

Which brings us to the part that should bother people who spent two years cheering open weights as the market’s answer to Big Tech concentration. I was among them, mostly. The theory was that open weights would spread capability outward — that a thousand small firms with a trillion-parameter model apiece would keep the giants honest.

What Xiaomi demonstrated Monday is that open weights don’t spread the business. They spread the product and concentrate the economics. The weights are free; the ability to serve them at $0.0036 per million tokens belongs to whoever has the cheapest capital, the cheapest power, and the vertical integration to eat the loss. That is a short list, and it does not include the startups that treated “open” as a business model rather than a distribution strategy.

The survivors are the ends of the barbell: companies with proprietary distribution, and companies with proprietary workflow lock-in. The middle gets squeezed, and its engineers get absorbed. That isn’t decentralization. It’s a different shape of concentration, and it’s harder to argue with because nobody did anything wrong. Xiaomi ran a price.

The policy reflex will be to reach for something — procurement rules, compute subsidies, another round of thinking about weights as an export-controlled artifact. Note that what did the damage this week was a number on a pricing page, and numbers don’t clear customs. The right has spent years insisting that the market would discipline the AI industry. It will. It just won’t discipline it in the direction anyone was promised.

Sources