A Chinese AI lab gives its best models away for free. This morning, the market briefly decided that company was worth $128 billion. Both of those things are true at once, and the gap between them is one of the most interesting puzzles in AI right now.

The company is Zhipu — the Beijing lab behind the open-weight GLM models. On Monday its Hong Kong shares (as we covered in today's Daily Update) surged as much as 42%, briefly pushing its market value above one trillion Hong Kong dollars, roughly $128bn. So how is a company that doesn't charge for its core product worth a quarter of what Wall Street thinks Anthropic is? The answer comes in two parts: how Zhipu actually makes money, and what the market is really buying.

The model is the funnel, not the product

Start with the business, because it's more conventional than the "free AI" framing suggests. The open GLM models aren't what Zhipu sells — they're how it gets in the door. What it charges for is the service layer wrapped around them.

According to Zhipu's first audited annual results since its January listing, the company brought in RMB 724m (about $105m) in 2025 — up a striking 132% on the year. The biggest chunk, around three-quarters, comes from on-premise private deployments: bespoke GLM installations for banks, telecoms, big enterprises and the Chinese state sector — buyers who, for data-sovereignty or security reasons, can't or won't send their data to a public-cloud AI. It's a project-based, services-heavy business, structurally closer to a systems-integrator like SenseTime than to a subscription app. Two faster-growing lines fill out the rest: a paid API platform (up nearly four-fold) and "enterprise agents" (up almost 250%).

So Zhipu does have real revenue, and it's growing fast. But here's the part the trillion-HKD headline skips: it is deeply unprofitable. On an adjusted basis the company lost around RMB 3.2bn last year — roughly $440m, or more than four times its revenue — with research and development alone running at over four times what it takes in. By analysts' rough shorthand, labs like Zhipu burn about $10 for every $1 they earn. It has perhaps twelve to eighteen months of cash to change that.

Why give the model away at all?

If the deployments are the business, why open-source the models that drive them? Because in this market, openness is itself a strategy — and a well-theorised one.

The cleanest version is an old strategy: commoditising your complement. If frontier models are becoming a cheap, interchangeable commodity, you don't try to own the model — you make it free, drive mass adoption, and capture the value in the layer above. Open weights become a distribution flywheel: downloads lead to developers, developers lead to usage, usage feeds back into a better model. A US government commission recently noted that this loop has made China's Qwen, not Meta's Llama, the most-forked model family on Hugging Face, with over 100,000 derivatives.

There's a cultural and strategic dimension too. DeepSeek's founder Liang Wenfeng — the most quotable voice of China's open camp — has argued that "in disruptive technology, a closed-source moat is always temporary," and that giving models away "helps attract top-tier talent." And there's geopolitics: open weights are soft power. A model anyone can download and run on their own hardware is one no government can switch off — which is precisely the appeal in a world where the US has been restricting access to American models. It's worth being honest that, as that same US commission put it, whether China's open strategy is genuine preference or partly necessity "is an open question."

The export-control catalyst

That last point is not abstract this week. The specific trigger for Monday's surge was last week's release of GLM-5.2, which ranked second in the world on the Code Arena leaderboard for front-end web development — behind only Anthropic's Claude Fable 5.

But Fable 5 was ordered blocked for foreign nationals by the US government on 12 June. Zhipu released GLM-5.2 the next day, with its open MIT-licensed weights following on 16 June. Put those together and GLM-5.2 isn't just "the number-two model" — for most of the world, it's the best model still freely usable. American export controls, designed to slow China down, handed China's leading open lab an unusually clean marketing moment.

What the market is actually buying

None of that, though, explains a $128bn valuation. At that price Zhipu trades at more than 1,000 times its trailing sales — an extraordinary multiple for any company, let alone a loss-making one. So the honest read is that the market isn't pricing the $105m revenue stream. It's pricing a story, amplified by some unusual mechanics.

The story is scarcity and optionality: Zhipu is one of the first pure-play large-language-model labs that public investors can actually buy, and the bet is that today's developer adoption eventually converts into tomorrow's enterprise and sovereign-AI revenue. The mechanics are what make the move so violent. Only about 4% of Zhipu's shares trade freely right now — a cornerstone tranche unlocks in July, and the big employee unlock doesn't come until January 2027 — so a thin float meets heavy demand. The stock was freshly added to the Hang Seng Tech Index and to Stock Connect this month, opening the floodgates to mainland buyers. And it is hard to borrow and therefore hard to short, which turns momentum into something closer to a squeeze. Its sister lab MiniMax saw its small retail share tranche oversubscribed more than a thousand times.

The bull case is real but should be read with one fact in mind: even the Wall Street banks that upgraded the stock — JPMorgan and Bank of America among them — set price targets that sit below where it traded on Monday. When the optimists' targets are beneath the market price, the market is running on sentiment, not arithmetic.

The bet underneath it all

Step back from the ticker and the genuinely important question is whether the open-weight model as a business works at all. Here the evidence cuts both ways, and sharply.

On adoption, the open camp is winning. Chinese open models have overtaken Meta's Llama on cumulative downloads, and now lead it by a wide margin. On the raw measure of who builds on what, open and Chinese models lead.

On money, it isn't close. According to enterprise surveys, the closed Western labs — Anthropic, OpenAI and Google — together account for around 88% of enterprise large-language-model usage. Chinese open models account for barely 1% of enterprise API usage, held back by exactly the security and data-residency caution that, ironically, drives Zhipu's on-premise business at home. Open weights are winning the developer world and losing the enterprise budget.

That is the bet a $128bn valuation makes: that "give the model away, sell the service" converts adoption into durable revenue before the cash runs out — and that the moat really has moved from the model to the layer around it. It might. Closed labs are still comfortably ahead at the frontier and on revenue, and the smartest analysts think the open camp is catching up in capability but not closing the gap. What Monday proved is narrower, and still remarkable: a Chinese lab that charges nothing for its headline product is now, on paper, one of the most valuable AI companies on Earth — and the open-versus-closed argument just acquired a very large number on the open side of the ledger.