Model

Z.ai falls to HK$335, more than 85% below its June peak, as Alibaba cuts token prices

The Beijing lab still carries about HK$163 billion of market value on HK$1.49 billion of trailing sales, with a HK$12.4 billion annual cash burn underneath. A new US rule on rented compute and Alibaba's cut to token prices are squeezing what is left for the funds holding Hong Kong's AI listings.

Vincent JiangVincent Jiangthree-minute read
Share
An Nvidia H100 accelerator card standing upright against a black background, showing its gold backplate, silver bracket and PCIe connector.
An Nvidia H100 accelerator. Chips like this are what the new US rule keeps from reaching China-headquartered labs through third-country data centers, and the hardware Nvidia's self-hosting recipe for Z.ai's GLM-5.3 runs on.

A buyback that could not hold the line

Z.ai−5.26% — Z.ai, down 5.26 percent today, the Beijing lab behind the GLM models and formerly known as Zhipu, closed at HK$335.00 on Tuesday, down 5.3%, after a 13.9% slide on Monday 12. The shares sit more than 85% below their June peak, unwinding a run that lifted them almost 2,000% in under six months from the January listing 1.

The chair's $510 million buyback proposal, reported on November 13, has not held the line: from that day's close of HK$408.40 the stock has fallen a further 18% 34. Short interest has reached a record 9.9% of the free float, on S&P Global data 5.

HK$163 billion riding on HK$1.49 billion of sales

Tuesday's close values Z.ai at about HK$163.3 billion, roughly 110 times its HK$1.49 billion of trailing sales 2. Underneath sit a HK$2.95 billion trailing net loss, HK$12.38 billion of levered free cash flow burned over the past year, and HK$4.58 billion of cash 2. Two sessions earlier, at HK$408.40, the same shares carried 118.6 times enterprise value to sales 4. Zhipu has also sat on the Entity List, the US export-control blacklist, since January 2025 6.

HK$163B of market value on HK$1.49B of trailing sales

The crash still leaves about 110 times trailing sales

-HK$50BHK$0BHK$50BHK$100BHK$150BHK$200BMarket valueHK$163.34Babout 110 times trailing salesCash on handHK$4.58BTrailing 12m revenueHK$1.49BTrailing 12m free cash flow-HK$12.38B
Data
Value
Market valueHK$163.34B
Cash on handHK$4.58B
Trailing 12m revenueHK$1.49B
Trailing 12m free cash flow-HK$12.38B
Hong Kong dollars, billions. Market value at the November 17, 2026, close; revenue, cash and levered free cash flow are trailing twelve months.2

The new rule licenses rented compute

BIS, the Commerce Department's export-control bureau, released the interim final rule on November 15, written at the direction of the Treasury and the National Security Council so existing chip curbs "cannot be circumvented through remote computing arrangements involving data centers in third countries" 7.

The mechanics reach past shipments. A US license is required for computing equipment that exporters know will serve China-headquartered firms from data centers above 12 million units of total processing performance, the export-control measure of compute 8. Advanced chips bound for Singapore, Malaysia, Thailand, Indonesia and Brazil now need licenses too, and existing arrangements have 90 days to comply once the rule publishes on November 19 8.

The rule is the smaller threat to Z.ai itself

Two threats squeeze Z.ai's token business. The US rule is the smaller one for the company itself: Z.ai trains on its own data center of Chinese-made accelerators 4, and the license requirement lands on the labs that rent US-chip compute through those five countries instead. Alibaba's−3.40% — Alibaba, down 3.40 percent today price cut has no shield.

The price cut lands where Z.ai actually earns

Qwen4-Max lists at $1.50 per million input tokens and $4.50 per million output, down from $2 and $6, with a claimed 62% lower cost per task on Alibaba's own benchmark 9. Z.ai's flagship GLM-5.3 lists at $1.4 and $4.4, and its open weights, the downloadable model files anyone can run, are the ones Nvidia−1.33% — Nvidia, down 1.33 percent today hands out in a $190,000 self-hosting recipe that pays Z.ai nothing 4.

Alibaba's cut leaves Qwen4-Max within $0.10 of GLM-5.3 on both input and output

  • Input, $ per million tokens
  • Output, $ per million tokens
$0$2$4$6Qwen4-Max (Alibaba)GLM-5.3 (Z.ai)$1.4$4.4
Data
Input, $ per million tokensOutput, $ per million tokens
Qwen4-Max (Alibaba)$1.5$4.5
GLM-5.3 (Z.ai)$1.4$4.4
US dollars per million tokens, list prices. Qwen4-Max is shown after Alibaba's cut from $2 per million input and $6 per million output, which closed the gap to Z.ai's flagship to ten cents on both measures.9,4

Amazon's+0.42% — Amazon, up 0.42 percent today Bedrock, a pay-per-token model service, rents the same weights, and AWS says prompts are never shared with the model provider 6. Analysts at the Japanese bank Nomura expect tighter compute access to constrain model training and slow cloud revenue growth 10.

The money moved to the chipmakers

One session split China's AI chain. SMIC and Hua Hong, China's two big contract chipmakers, rose as much as 6.4% and 12% on Tuesday, while Alibaba fell 5.6%, Baidu−2.39% — Baidu, down 2.39 percent today 6.6%, robot maker Yushu Technology 11%, and Z.ai and MiniMax, a rival AI lab listed in Hong Kong in January, more than 6% each 10. MiniMax shows where this ends: HK$40.5 billion of market value against a March peak above HK$410 billion, on $116.6 million of first-half revenue and a $293 million adjusted net loss 211.

Nvidia's results and a January lockup decide what comes next

Three dates decide the next leg. Nvidia reports after the US close tonight, its first results under interim chief executive Colette Kress 12. The rule reaches the Federal Register on Thursday 8. A lockup on roughly 178 million pre-IPO Z.ai shares expires in January 5. Then the next income statement shows whether giving the weights away earns Z.ai any revenue at all.

One session, two directions: chipmakers up, model makers down

The rule's first session moved money to China's chipmakers

-15%-10%-5%0%5%10%15%Hua Hong12%SMIC6.4%Alibaba-5.6%Z.ai-6%+Baidu-6.6%Yushu-11%
Data
Value
Hua Hong12%
SMIC6.4%
Alibaba-5.6%
Z.ai-6%+
Baidu-6.6%
Yushu-11%
Peak intraday share-price moves reported on November 17, 2026, Hong Kong and Shanghai listings; the Z.ai bar marks the reported floor of more than 6%.10

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading BABA

Reader comments

0 comments

    Model

    Anthropic's Thursday Debut Puts a Daily Price on Alphabet's Stake

    Alphabet holds the equity, rents Anthropic its compute and undercuts its prices, and from November 19 the market marks that bundle daily. The consensus says weaker American open source restores Anthropic's pricing power; the models actually holding the floor are Chinese.

    Model

    Meta's Q3 earnings fell to $3.83 against $5.46 expected, on a $9.89 billion teen-harm legal charge

    Excluding the charge Meta beat, and Businessweek's settlement pricing implies only about half the pending teen-harm claims are covered, with 2,000 school districts unpriced. The reserve note in January's filing decides.

    Model

    Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI

    The tutorial walks enterprises through serving GLM-5.3 on two DGX Stations, roughly $190,000 at listed reseller prices, with no per-token bill to Z.AI. For a stock carrying 118.6 times sales, that reads as a price list for everything Z.AI sells.

    Model

    Anthropic's Claude models tie at 58 on a coding index Cognition partly built, days before its Nasdaq IPO

    A third of the refreshed index runs on FrontierCode, a benchmark built by Cognition, whose Devin agent announced one of the two tied winners, Claude Fable 5.2, on the day of its release. Anthropic's model lists on Nasdaq on November 19; Cognition's Devin has offers from $70 billion on the table.

    Model

    Anthropic's Fable 5.2 passes 41.3% of real freelance jobs before its Nasdaq debut

    The one independent test of finished freelance work puts Anthropic's Fable 5.2 first at 41.3%, and the model still fails most of it, days before a Nasdaq debut priced at up to $1.4 trillion.

    Model

    Hedra cut Krea 3's edit time from 22.9 to 3.3 seconds on AMD, and the app doesn't own the speed

    The eight-second edit Krea advertises is manufactured in the serving stack, and a partner's measurement shows the app does not own the speed. Inference efficiency accrues to the optimizer and the chip vendor, which caps a per-image app's margin and hands AMD a demand channel beyond hyperscalers.

    Subscribe to TheInference for free