On the morning of November 17, the order that mattered was not the biggest one of the session. It was the sixth consecutive one. Z.AI Co., the Hong Kong-listed developer most of the world still calls Zhipu, opened weak, slid as much as 6.8%, and closed out a decline of more than 85% from its June peak. Five months earlier the same stock had been the cleanest way to own a national bet: Chinese frontier AI, listed and tradable, up almost 2,000% in under six months after its January debut.
The reversal is now the sharpest illustration of a problem the sector spent 2026 pretending it had solved. China's labs got very good at closing the capability gap. They did not close the compute gap. Every valuation built on the first assumption carried an unhedged exposure to the second, and Washington moved last week to make that exposure explicit.
The trigger was a rule, not a model release. The US will require companies in five countries, including Singapore and Malaysia, to obtain export licences before giving Chinese firms remote access to advanced AI chips and servers. The measure takes effect on November 19 and lands after a controversy involving Nvidia that has put export controls back under scrutiny in Washington.
That distinction matters more than it sounds. Remote access to overseas capacity had become the quiet workaround for labs facing domestic supply limits: rent the compute abroad, keep training at home. Billy Leung, an investment strategist at Global X Management, described the consequence without decoration. The restriction, he said, closes another workaround for access to Nvidia compute, which could mean higher costs and slower model development.
The market did not wait for the cost curve to move.
The Numbers Behind the Slide
- Down more than 85% from the June peak, after a rally of roughly 2,000% in under six months.
- A 6.8% intraday drop on Tuesday, the first session after the licence rule was reported.
- About a third of market value erased since the previous Wednesday's close, a fall that matches rival MiniMax almost exactly.
- Short interest at a record 9.9% of free float, according to S&P Global data.
- Still up around 200% from its listing price, the number bulls keep repeating.
The company's answer had arrived a week earlier, and it read like a management team watching the tape. On November 10, Z.AI proposed buying back up to HK$4 billion of its H shares, capped at 1% of issued shares, with chairman Liu Debing behind the plan and the June 22 annual general meeting mandate as its legal cover. The board framed the move as confidence in the long-term business.
It has not worked yet. Shares were already about 58% below where they traded at interim results when the buyback was announced, and within days the slide extended past 85%. That gap is instructive. A company buying its own stock is a signal about price. A market repricing a cost structure is a verdict on the business model, and HK$4 billion of demand looks small against a float that was valued on the assumption of unlimited compute.
There is a second signal in the same week, easier to miss. Chinese labs including Zhipu, Moonshot and Tencent have started trimming offensive cyber capabilities from new models, a retreat that follows both the export-control fight and the earlier scrutiny of GLM-5.3's hacking scores. Capability roadmaps are now being written with one eye on a licence application, which is what happens when a research agenda turns into a compliance document.
Tencent told investors the new curbs would not hurt cloud revenue but might slow model training. That sentence deserves a second read. The revenue line holds up because inference on existing models still sells. The research line is where the damage sits, and research is precisely what the multiple was paying for.
So what is actually left? A lab that quadrupled revenue while still losing billions, that ships a genuinely competitive open-weight family in GLM-5.3, and that now trades on assumptions it can no longer underwrite. That is not a collapse story. It is a de-rating, and the two get confused constantly.
What to watch next. The interesting variables are no longer benchmark scores.
- The November 19 effective date, and whether licence applications for Singapore and Malaysia capacity are granted at all.
- Whether the HK$4 billion buyback is executed aggressively or left as a shelf announcement.
- Short interest above 9.9% of free float, and whether it keeps climbing into any rebound.
- Whether GLM-6.0 arrives on schedule, or slips the way GLM-5.5 never did.
The uncomfortable lesson is for everyone who spent the year measuring progress in benchmark deltas. Z.AI did not lose 85% because its models got worse. It lost 85% because the market finally priced the one input the company cannot control, and no amount of coding leadership hedges a licence requirement. The frontier is still a physical place, and passports still get checked at the door.
Comments