Chinese AI’s ‘Darkest Moment’: 5 Headwinds Checked in Data

A Chinese article titled “中国大模型的至暗时刻”, the darkest moment for China’s large models in our translation, lists five headwinds facing the country’s AI industry: compute, distillation, robot software, revenue and optical parts. We checked each against facts already on record in our own reporting. Four have recent on-record facts behind them. One detail is out of date: the US remote-access rule released on 16 November names 5 countries, not the 7 the article cites from earlier reporting.

On usage the data cut the other way: a DeepSeek model processed 61.2 trillion tokens on OpenRouter in the week through 16 November, the most of any model. Share prices of two listed labs have fallen hard. MiniMax closed at HK$124.90 in Hong Kong on 16 November, 89.9% below its highest close and 24.3% below its HK$165 IPO price, and Zhipu’s HK$353.60 close was 85.3% below its peak, by our arithmetic on exchange data.

Business Pill · Open Weights Explained

An evergreen Business Engineer explainer on open weights: rent a model through an API or run it yourself. It teaches the general idea only and does not discuss this article.

The key insight: As we read the five factors against our own reporting, the article is strongest on cost and weakest on demand: a US rule, OpenAI‘s new limits and a planned optics ban all raise what Chinese labs pay, while a DeepSeek model processed 61.2 trillion tokens on OpenRouter in a week, the most of any model.

Where the Article Came From

The article ran on Huxiu, the Chinese business site, dated 15 November, and says it comes from the WeChat account 嬉笑创客, with the author given as CB. Rui Ma, who covers Chinese tech, summarised it in English on X at 01:24 UTC on 17 November, writing that it “lays out the bear case quite extensively”. Kyle Chan shared her summary at 03:31 UTC as a “Good list of factors causing headwinds for China’s AI industry”.

When we checked at about 05:00 UTC on 17 November, the Huxiu page returned a message that the content could not be found or had been deleted. We read the full text from a cached copy retrieved through a scraping service. Huxiu has not said why the page is unavailable. Ma noted “pushback in the comments”, and one comment on the cached page called the author too pessimistic, in our translation.

The framing is the article’s. Ma added a caveat of her own: “it feels like every six months it is a” darkest moment for Chinese AI, she wrote, before adding: “But I think these are real concerns”.

Zhipu and MiniMax daily closes in Hong Kong as a multiple of each IPO price, from listing to 16 November 2026.
Zhipu and MiniMax daily closes in Hong Kong as a multiple of each IPO price, from listing to 16 November 2026. Closes from the Yahoo Finance chart feed; IPO prices HK$116.20 and HK$165. The multiples and percentages are our arithmetic.

1. Compute: The Rule Names 5 Countries, Not 7

Ma’s summary of the first factor: “Getting compute is harder whichever way you go”. The article cites The Information’s 11 November report that seven countries, including Norway and South Africa, could face tighter rules on advanced chips and remote access. It also cites the Remote Access Security Act and the MATCH Act on chipmaking equipment.

The rule the Commerce Department’s Bureau of Industry and Security released on 16 November is narrower on countries and broader on access, as our piece on the remote-access rule reported. It requires a licence to export advanced chips to data centres, anywhere, that give companies headquartered in China, Macau or other Country Group D:5 destinations access to more than 12,000,000 aggregate TPP of computing power.

It adds a licence requirement for 5 countries: Indonesia, Malaysia, Singapore, Thailand and Brazil. The text names neither Norway nor South Africa, by our search of its 132 pages.

The two bills match the article’s account. The Remote Access Security Act passed the House by 369 votes to 22 on 12 January and awaits Senate action, as our piece on Chinese firms renting Nvidia capacity set out. The MATCH Act cleared the House Foreign Affairs Committee 36 to 8 on 22 April and has had no floor vote, per our MATCH Act explainer.

The compute squeeze is also visible at DeepSeek. Users told STAR Market Daily on 16 November that its app cited insufficient computing power when it capped single inputs at about 40,000 characters, as our DeepSeek rationing piece reported. Not all the data point one way: SemiAnalysis wrote on 12 November that more than 1.5GW of restricted Nvidia chips “still ultimately ended up in China”, as our SemiAnalysis piece covered.

2. Distillation: OpenAI’s New Limits Start 20 November

The second factor, in Ma’s words: “Catching up could get a lot more expensive” if learning from leading models’ outputs becomes harder. The article cites an industry discussion note of 11 October in which unnamed participants estimated that the cost of direct distillation could rise 5 to 10 times, and 20 times in extreme cases.

One of those limits is now dated. OpenAI’s updated Usage Policies take effect on 20 November. They bar using OpenAI models or outputs to build training material for models meant to replace OpenAI’s. Services that resell direct access to its models need prior review and approval, as our piece on the policy reported. The 5 to 10 times cost estimate comes from unnamed participants, and we found no public data to check it against. Fresh reporting points the same way: Reuters reported on 17 November, from a review of more than 300 Chinese documents, that distillation “has expanded beyond academic experiments into mainstream industrial use in China”, and noted that the practice is not illegal, as our piece on the Reuters review reported.

3. Robot Software: General Models Move Into Robotics

Third, in Ma’s words: “Robot-brain startups could find their advantage disappearing”. The article cites a 74.8% task success rate for GPT-6 Astra on a grasping benchmark in Nvidia researchers’ work and a 96% sorting rate for Google DeepMind’s Gemini Robotics 2.5 in a factory pilot.

Our own reporting points the same way. Anthropic says Claude Fable 5.2 scored 64.9% on RoboDojo, a simulated two-arm robot test, against 29.8% for GPT-6 Astra on Anthropic’s runs, and Fortune reported that OpenAI disbanded its stand-alone robotics team because its general models can be applied directly to robots, as our robotics benchmark piece set out. The benchmarks differ and the scores are not comparable. The article’s second point, that robots must compete with China’s relatively low labour costs, is not something our data covers.

4. Revenue: Usage Is High, Spend Is Not

Fourth, in Ma’s words: “More users don’t necessarily mean enough revenue”. The article cites Epoch AI’s estimate that six leading Chinese AI companies had about $11 billion of annualised revenue, one-tenth of OpenAI’s and Anthropic’s combined $105 billion. Those are the figures our Epoch piece reported from Nikkei Asia; Anthropic’s own filing puts its run-rate revenue at about $81 billion as of 31 October, about 7.4 times the six firms’ total, by our arithmetic in that piece.

The article also says Doubao has 419 million monthly users while its app made less than 1 million yuan a day in the first half, citing LatePost. QuestMobile put Doubao at 419.08 million monthly active users in September, as our Doubao piece reported; we have not seen the LatePost revenue figure. Prices are being cut too: Alibaba sold nearly 100,000 QwenWork monthly passes in 30 minutes at 0.01 yuan each, cut from 28 yuan, as our QwenWork piece reported.

Usage is where the data cuts the other way. DeepSeek V4.1 Flash 1017 processed 61.2 trillion tokens on OpenRouter in the week through 16 November, the most of any model, and DeepSeek had 31.8% of OpenRouter’s text requests in the week beginning 9 November, ahead of OpenAI at 19.0%, as our OpenRouter piece showed. On Databricks, Zhipu, DeepSeek and Moonshot together had 97.4% of open-weight tokens in the week of 2 November.

Yet that usage does not turn into spend at the same rate. DeepSeek had 28.2% of open-weight tokens on Databricks that week but 9.1% of open-weight list-price spend, by our sums in our Databricks AI Index piece. Revenue is growing from a small base: Zhipu’s first-half revenue rose 399.7% and MiniMax’s 283.1%, and both reported losses, as our Kai-Fu Lee piece showed.

Epoch AI annualized revenue estimates: six Chinese AI firms about $11 billion combined against about $105 billion for OpenAI and Anthropic
Epoch AI’s annualized revenue estimates, published on 6 October 2026 and reported by Nikkei Asia on 10 November, from months between June and September. The ByteDance and Alibaba figures are Model-as-a-Service revenue only, Epoch says.

5. Optical Modules: A Ban Under Consideration

Fifth, in Ma’s words: “Even the areas where China already leads face risks”. The article says Chinese companies hold about two-thirds of the global optical-module market, citing Reuters and Counterpoint Research, and that the US is considering restrictions on new Chinese models.

Reuters reported on 10 November that Microsoft has been pressing the administration to halt plans to bar imports of new models of Chinese optical transceivers, and that the Federal Communications Commission is working on the ban, as our optics piece reported. Counterpoint Research’s estimate cited in that piece is about 60% of global datacom transceiver revenue, a narrower measure than the article’s two-thirds. No ban had been announced in the sources we read.

What the Share Prices Show

The article puts the falls of Zhipu and MiniMax from their highs this year at about 86% and 88%. Exchange data give similar figures. MiniMax’s 16 November close of HK$124.90 was 89.9% below its highest close, HK$1,238 on 18 March, and 24.3% below its HK$165 IPO price, by our arithmetic. Zhipu’s HK$353.60 close was 85.3% below its HK$2,410 high of 22 June, and still about 3 times its HK$116.20 IPO price.

Our earlier piece tracked the $150 billion drop Bloomberg reported for the two stocks on 23 October. Since then MiniMax has fallen a further 22.3% and Zhipu 23.3%, from their 22 October closes of HK$160.70 and HK$460.80, by our arithmetic.

The Structural Read

As we read it, the gap the article describes is between use and payment: DeepSeek had 28.2% of open-weight tokens on Databricks but 9.1% of open-weight spend, by our sums, and Epoch AI estimates six leading Chinese firms at about $11 billion of annualized revenue.

As we read the dates, two of the five factors are already in force or dated: the remote-access rule released on 16 November and OpenAI’s usage policies from 20 November; the MATCH Act and the optics ban are still proposals.

Rui Ma, summarising the article on X, 17 November 2026

“The worry is that companies keep making better models while needing more money to stay competitive, without earning enough to fund the next round.”

Three Implications

FOR CHINESE LABS Our pieces show access to US compute and to OpenAI outputs narrowing on dated rules, while usage of DeepSeek, Zhipu and Moonshot models stays high on OpenRouter and Databricks.

FOR INVESTORS MiniMax closed 89.9% below its highest close and 24.3% below its IPO price on 16 November, by our arithmetic on exchange data; Zhipu was 85.3% below its peak.

FOR US SUPPLIERS Reuters reported that Microsoft has pressed the administration to halt a planned ban on new Chinese optical transceivers, a market where Counterpoint puts Chinese makers at about 60% of revenue.

The Business Engineer Lens

This story maps onto the Business Engineer framework AI’s Geopolitical Chokepoint.

The essay argues the AI supply chain is “creating geopolitical leverage that operates independently of who has the best models”.

As we read it, the article’s compute, distillation and optics factors are that kind of leverage: each acts on Chinese labs whatever the quality of their models. The essay predates the article and does not discuss it.

What Is Not Established

The darkest-moment framing is the article’s view, and its own conclusion says Chinese AI is far from a total rout, in our translation. The distillation cost estimates and the article’s forecast of a wider US-China model gap come from unnamed participants in an industry discussion note.

We have not checked the article’s Reuters figure for Chinese projects naming restricted chips, its Financial Times report on an Oracle and Tencent lease, the LatePost revenue figure or a Tencent Research Institute survey against the originals. We do not know why the Huxiu page became unavailable. The usage figures come from OpenRouter and Databricks and do not cover China’s domestic market.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

AI’s Geopolitical Chokepoint

A Business Engineer map of where geopolitical leverage sits in the AI stack, beyond who has the best models.

Read the Map of AI →

The Bottom Line

Against our own reporting, four of the article’s five headwinds have recent on-record facts behind them: a US rule on remote access, OpenAI’s limits on training from its outputs, general models moving into robotics, and a planned optics ban. Revenue is small against the US leaders. Usage is the exception, with a DeepSeek model the most used on OpenRouter last week. MiniMax’s close of HK$124.90 sat 89.9% below its peak, by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The article’s claims are its own, read in full in Chinese from a cached copy of the Huxiu page, which was unavailable when we checked on 17 November; translations are ours. Rui Ma’s and Kyle Chan’s wording is from their posts on X. Every fact we set against the article comes from our earlier pieces, linked in each section, which name their primary sources.

Share prices are Hong Kong closes from the Yahoo Finance chart feed; the percentages and multiples are our arithmetic. This piece reports a view and checks it; it makes no forecast.

Sources: Huxiu: 中国大模型的至暗时刻 (15 Nov 2026; page unavailable when checked 17 Nov) · Rui Ma on X (17 Nov 2026) · Kyle Chan on X (17 Nov 2026)

Scroll to Top
FourWeekMBA