Model

MiniMax hit a record low, about 90% below its January peak, as a US rule gates its rented Nvidia compute

Washington's new remote-access rule puts a US license between China's listed model makers and the offshore Nvidia capacity they rent. MiniMax, the purest listed bet on Chinese foundation models, holds the price tag.

Vincent JiangVincent Jiangtwo-minute read
Share
An Nvidia data center GPU server in a rack, blue status lights lit behind its mesh front door
An Nvidia data center GPU system in a rack. Chinese model developers rent Blackwell hardware like this in Singapore, Malaysia and elsewhere instead of buying it at home.

The US Bureau of Industry and Security published an interim final rule on November 16: data center operators in Singapore, Malaysia, Thailand, Indonesia and Brazil now need an American license to serve advanced chips to customers, with 90 days to apply 3. The rule targets the exact route Chinese model developers have used to reach Nvidia−2.79% — Nvidia, down 2.79 percent today hardware their own country cannot buy: rented servers in Southeast Asian data centers 34.

The worst of it lands on the two listed pure plays

That lands on MiniMax harder than on anyone else listed. The Hong Kong company closed November 16 at HK$124.9, down 18.04%, while the Hang Seng Index actually rose 4. It fell more than 6% again on November 17 alongside Z.ai, the only other listed pure bet on Chinese foundation models, and hit a record low on November 18 per Reuters 37. Neither company owns an advertising, e-commerce or games business to lean on; their whole model is building models, and their training compute is rented abroad 34.

MiniMax trades near its 52-week low, about 90% below its debut-year peak

MiniMax is down about 90% from its peak and trades at a third of its first-day close

HK$0HK$500HK$1,000HK$1,500First-day close, Jan 92026 peakClose, Nov 1652-week low, Nov 18HK$114HK$124.9
Data
MiniMax share price
First-day close, Jan 9HK$345
2026 peakHK$1,330
Close, Nov 16HK$124.9
52-week low, Nov 18HK$114
Hong Kong listing prices in HK dollars: first-day close and peak from China Securities Journal via EastFrontier; Nov 16 close down 18.04%; 52-week low per Investing.com. Sources 3, 4, 5.3,4,5

The fingerprints are on the model itself

Unemployment Labs, a US nonprofit led by former OpenAI$1.18T — OpenAI, private, latest valuation $1.18T researcher Will DePue, reported on November 15 that numerical patterns in the published weights of MiniMax's M3 and five other Chinese models match Nvidia's Blackwell architecture, and estimates roughly 12,000 Blackwell GPUs in a cluster costing more than $500 million across the six models 12. The group is careful: the analysis is technical, and where the chips sit and whether rules were broken is unconfirmed 1.

The cheap escape runs on the wrong silicon

Inside MiniMax the arithmetic pulls both ways: gross margin more than doubled, but the net loss quadrupled to $1.87 billion from $465 million a year earlier, and the stock sits about 90% below its January peak of HK$1,330, below its IPO price 54. The cheaper escape is onshore: an A100 80G server rents for about 34,500 yuan a month in east China and Ascend 910B4 machines for about 15,500 yuan, but those are Hopper and Huawei silicon, not Blackwell 8.

Onshore compute is getting cheaper, but it is not Blackwell

Onshore rental is cheap, but it is Hopper and Ascend, not Blackwell

¥0¥10,000¥20,000¥30,000¥40,000A100 80G full server, east China¥34,500A100 floor, Yangtze River Delta¥33,750Ascend 910B4 machine¥15,500
Data
Value
A100 80G full server, east China¥34,500
A100 floor, Yangtze River Delta¥33,750
Ascend 910B4 machine¥15,500
Monthly rental quotes in yuan per unit, November 18, 2026, including VAT where stated; the A100 floor fell 3.6% in one session on cheaper DDR5 memory. Source 8.8

The disclosure to demand at the next results: how much of MiniMax's training runs on rented capacity outside China, and what a licensed Blackwell cluster costs. Until then, shareholders hold the only listed position where the compute bill is the whole story.

Get Model by email, free

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

Loading NVDA

More about NVIDIA

NVDAFiscal Q2 2027
Fiscal Q2 2027 · quarter to 26 Jul 2026 · changes vs a year earlier

Revenue rose 106% to $96.2B, 92.5% of it from Data Center, “driven by the ramp of our Blackwell Ultra infrastructure”. Gains on equity stakes of $7.8B lifted net income to $59.7B.

Revenue$96.2B▲ 106%
Gross margin75.0%▲ 2.6 pts
Operating margin66.2%▲ 5.4 pts
Net income$59.7B▲ 126%
Hyperscale$48.7B▲ 102%AI clouds & enterprise$40.3B▲ 138%Data Center$89.0B▲ 117%Edge Computing$7.2B▲ 27%Revenue$96.2B▲ 106%Gross profit$72.1B75.0% marginCost of revenue$24.1BOther income$7.8Bequity-stake gainsOperating income$63.7B66.2% marginOperating expenses$8.4BNet income$59.7B▲ 126%Tax$11.8B16.5% rateR&D$7.1BSG&A$1.4B
Show as a table
LineValue
Revenue$96.2B
Gross margin75.0%
Operating margin66.2%
Net income$59.7B
Hyperscale$48.7B
AI clouds & enterprise$40.3B
Data Center$89.0B
Edge Computing$7.2B
Gross profit$72.1B
Cost of revenue$24.1B
Other income$7.8B
Operating income$63.7B
Operating expenses$8.4B
Tax$11.8B
R&D$7.1B
SG&A$1.4B

Reader comments

0 comments

    Model

    HeyGen's Avatar 6 scored 84% on realism, but its 42.2% pick rate was below chance

    Avatar 6 launched November 17, 2026, with an 84% realism figure and first place in a blind arena; the score doubles a below-chance pick rate, and the arena capped 97% of clips at 35 seconds. That is the evidence base under a company claiming $200 million of annual revenue.

    Model

    OpenAI cuts flagship pricing 20% to match Anthropic as heavy users call GPT-6.2 Astra unusable

    OpenAI cut its flagship's price to match Claude Fable 5.2's rate, and the users who run models hardest rejected it, 24 hours before Anthropic priced a $1.4 trillion IPO on the assumption the quality gap holds. The independent labs say the gap is real, for now.

    Model

    Microsoft books Azure revenue on Chinese open models running on Baseten hardware it does not own

    Microsoft now bills customers' committed Azure dollars for open models from Chinese labs running on Baseten's hardware, a toll booth earning rent on compute it does not own.

    Model

    Alibaba stops giving away its best AI model, three days before earnings

    Alibaba launched Qwen4-Max on November 16 and kept the 2.4-trillion-parameter flagship behind its own paid API, promising open weights only for smaller models. Thursday's report is the first test of whether Qwen's three billion downloads become cloud revenue, and it lands two days after a US rule put the chips behind the model under license.

    Model

    Perplexity claims $1 billion of revenue on five times less spend than Anthropic's $1.4 trillion IPO

    Perplexity says it hit $1 billion in recurring revenue on a fraction of Anthropic's spend, three days before Anthropic prices a $1.4 trillion IPO on the opposite assumption. Whose economics hold will only be settled by a filing Srinivas now says could come in early 2027.

    Model

    Anthropic prices a $1.4 trillion IPO as Databricks logs show its costliest model at 0.6% of enterprise tokens

    Databricks' gateway logs show the costliest Claude tier at 0.6% of enterprise tokens and open-weight models near a third, hours before Anthropic prices a $1.4 trillion debut on premium-model revenue.

    Subscribe to TheInference for free