Deep Dive — Zhipu and MiniMax just broke the "sell the unlock" rule in Hong Kong

For eight months, the story of China's two publicly listed large-language-model companies was a story about gravity. Zhipu and MiniMax went public in Hong Kong in January, tripled and quadrupled, then spent the summer giving it back — Zhipu's stock halving from its peak, MiniMax losing nearly half its value in a single session when its own shares unlocked in July. On Monday, gravity broke.

Zhipu hit its first post-IPO lockup expiry — the moment when early investors are finally free to sell — and instead of a stampede, the stock jumped 19 percent. MiniMax rose roughly 17 percent alongside it. The Hang Seng Tech Index climbed nearly 5 percent on the day, with Alibaba, SMIC, Hua Hong Semiconductor and Kuaishou all posting gains between 8 and 12 percent. It was the clearest single-day signal yet that the market has stopped treating Chinese AI model labs as speculative tickers and started treating them as operating businesses.

Why an unlock normally means a crash

Lockup expiries are the closest thing public markets have to a scheduled heart attack. Early investors and employees hold stock they couldn't sell since before the IPO; when the restriction lapses, supply hits a market that may not have the depth to absorb it. The conventional wisdom in both mainland and Hong Kong markets is brutal and usually correct: unlocks mean selling.

MiniMax lived that script in July. About 150 million shares — roughly 48.9 percent of total share capital, amounting to nearly half the company — came free on July 9. The float went from under 3 percent to close to 50 percent in a single day. The stock fell 17.98 percent and closed at HK$297.40. It kept sliding to a record low of HK$209.20 intraday on July 14. JPMorgan cut its target twice in a week, from HK$300 to HK$240. By mid-July, MiniMax's market cap had fallen from a March peak above HK$410 billion to under HK$70 billion.

Zhipu ran the same gauntlet on July 8 and did something almost nobody expected: it rose 13.85 percent on the day. The difference was structural. Only about 5.76 percent of Zhipu's shares unlocked, and the holders were mostly industrial capital and long-duration institutions rather than financial investors looking for an exit. Zhipu also priced a placement at HK$1,588 that was fully subscribed by institutions, while MiniMax's raise combined shares and convertible bonds and carried a dilution overhang. Same event, opposite outcome — the unlock punished whoever had the weaker holder base.

What actually changed between July and now

Two things, and they are connected.

The first is that the fundamentals arrived. MiniMax reported first-half results on August 26, and the numbers were not a story about a company running out of narrative. Revenue for the six months ended June 30 rose 283.1 percent year over year to $116.6 million. The part that matters for anyone trying to value a model lab is the mix: revenue from its Open Platform and AI-based enterprise services jumped 703.1 percent to $73.9 million, lifting that segment from 30.3 percent of total revenue a year earlier to 63.4 percent. Revenue from consumer AI products roughly doubled, up 100.9 percent to $42.6 million. The half-year loss attributable narrowed to $358 million from $402.2 million.

That shift in mix is the whole argument. Consumer apps like Talkie and Hailuo generate users and headlines; enterprise API and platform revenue generates the kind of contracted, recurring, high-margin income that supports a multiple. MiniMax crossed from being mostly a consumer company to being mostly a platform company in twelve months.

The second change is the sell side turning. JPMorgan raised its Zhipu target from HK$1,800 to HK$2,000 with an Overweight rating. Goldman Sachs, Bank of America and Citigroup simultaneously put Buy ratings on MiniMax. JPMorgan's framing is worth sitting with: at current valuations, the market has already priced in Zhipu's guidance of $1 billion in annual recurring revenue by year end, so further upside depends on whether its open-weight models can scale through outside infrastructure and distribution. That is a much more sober thesis than "Chinese AI goes up" — it says the next leg has to be earned.

Zhipu's ARR history is why the bulls have something to point at. Reporting in July put its ARR at $1 billion as of that month, up roughly 15-fold between January and July. Anthropic took 15 months to go from $100 million to $1 billion in ARR; Zhipu did it in about five, on API pricing that rose roughly 83 percent cumulatively in Q1 while volumes still grew about 400 percent. Any company that can raise prices that much and still grow usage that fast has something investors usually call pricing power — a phrase almost never applied to Chinese AI until this year.

The skepticism that hasn't gone away

The bear case did not evaporate on Monday, and it deserves a straight hearing.

Start with the shorts. Bearish bets on both names hit records before the latest earnings, with short interest reaching about 20 percent of MiniMax's free float and 6 percent of Zhipu's, per S&P Global data. That is a large, well-informed group of investors who looked at the same H1 numbers and concluded the risk is to the downside. Competition is the reason they give: Moonshot's Kimi K3 landed as a direct threat to Zhipu's GLM line, and MiniMax's own M3 price cuts in June are the kind of move that reads as defensive. Jefferies has argued Zhipu's GLM-5.3 delivers performance comparable to Kimi K3 at roughly 19 percent lower cost per task, which is a real competitive answer — but it is also an admission that the comparison is now the relevant one.

There is also a difference between a lockup that goes well and a lockup that is over. MiniMax's founding team voluntarily accepted a 12-month lockup rather than the standard six, and this unlocking round did not involve founder or employee holdings. The dilution test has been deferred, not passed. The same is true of the float expansion: both stocks have more than halved from their peaks, and Zhipu is still up more than 800 percent since its January IPO while MiniMax is up more than 80 percent. Anyone who bought at the March top is still badly underwater on a 19 percent day.

And there is the broader question about what a model company is worth. The valuation anchor has visibly moved during 2026 — away from parameter counts, user numbers and listing scarcity, toward API revenue growth, agent ecosystem depth and enterprise service revenue. That is a healthier framework, but it is also a harsher one. It means these companies now get marked against their ability to convert token volume into margin, quarter after quarter, with DeepSeek and Moonshot and Alibaba competing for the same workloads. Analysts quoted in Chinese coverage have predicted consolidation is coming, with only a handful of independent model labs surviving past 2027.

What this means outside Hong Kong

The read that matters for everyone else is that China now has working price discovery for frontier AI. Hong Kong is the only market where pure-play model companies trade publicly — Zhipu listed on January 8 as arguably the world's first publicly listed LLM firm, raising about $558 million, with MiniMax following a day later and raising about $615 million. Every other frontier lab on earth is valued by private marks set in funding rounds.

That makes these two tickers the closest thing the industry has to a live index of whether model-building is a good business. And the answer they gave this week is more interesting than the share price: it is that the market will now pay for demonstrated enterprise revenue and credible cost curves, and will punish consumer-user stories and lockup overhangs without mercy. The same discipline that MiniMax's 703 percent enterprise growth was rewarded for is what its consumer business gets discounted for.

It also reframes the cost story that has run through this year's Chinese releases. Our own coverage of GLM-5.3-Flash traced how Z.ai served a week of OpenRouter-topping traffic on a large cluster of domestically developed accelerators rather than Nvidia silicon, claiming per-token cost roughly comparable to mainstream GPUs. The market move on Monday is, in part, investors pricing in that a lab which can serve frontier-class models on constrained hardware has a margin story that survives export controls. The DeepSeek P&L deep dive made the same point from the inside; public markets are now underwriting it.

What to watch next

Three things will tell us whether Monday was a re-rating or a rally.

Zhipu's earnings, due the Monday after MiniMax's — the companies reported days apart, and Zhipu's numbers are the ones that have to confirm the $1 billion ARR figure that reporting in July sourced to people familiar with the matter rather than to the company. A disclosed, audited number is a different animal from a well-sourced one.

The float. MiniMax's founders are locked for twelve months and this round didn't include them, so the real supply test is still ahead. Watch whether the institutions now publicly committing to hold actually do when their own windows open — and whether MiniMax's Shanghai STAR Market listing, which gives it an A-plus-H structure, pulls mainland demand in or simply adds more shares.

And the competitive response. Moonshot's Kimi K3 and DeepSeek's own pricing moves are the variables nobody controls. If Zhipu and MiniMax can hold pricing while volumes grow, the enterprise-revenue thesis holds. If the next round of model releases forces another round of cuts, the "pricing power" argument that justified Monday's move evaporates fast — and the record short interest will have been right.

Zhipu and MiniMax just passed a test most newly listed AI companies fail. Is enterprise revenue the durable moat here, or is a price war one release away? Tell us in the comments.

Sources: Reuters — MiniMax H1 revenue · Odaily — lockup reversal and analyst ratings · TechFlow — MiniMax 17%, Zhipu 5% · Bloomberg via NDTV Profit — record short interest · 36Kr — Zhipu ARR hits $1 billion · SCMP — MiniMax and Zhipu listings