And: Qwen4-Max is just ok Monday. Tech stocks are down 1.3% this morning, the first day of Anthropic's IPO week. The American AI lab is expected to price on Wednesday and trade on Thursday. The company has yet to file a new S-1/A filing, so its previously set $70 to $78 per-share price range target remains in force.
Tech Twitter spent the weekend agitating for Jensen Huang's return to Nvidia. We're taking a look at the mess today, as well as new changes to American laws that (should) limit Chinese companies' ability to access Nvidia's chips located in other countries.
Looking ahead, Baidu and Xpeng are reporting earnings results today; Nvidia, JD.com, Karna, and Weibo report tomorrow; Palo Alto Networks, Bullish, and Pony.AI on Wednesday; and Thursday will bring data from Alibaba, Elastic, and NetEase. On Friday, everyone's favorite chop-shop BendingSpoons reports.
Ready for another wild week? We're in the technology news tornado, so strap in! — Your friend, Alex.
- 📈 Trending Up: *** AI risk control in China*** ... *** Starlink competition***? ... *** picking winners and losers already***
- 📉 Trending Down: *** Starlink in Ukraine*** ... *** the right and the tech-right*** ... *** Russian refineries*** ... *** Starlink competition***
Chinese AI fails to breach the frontier
The Alibaba AI team just dropped Qwen4-Max, the first model in the new Qwen4 family, with a 2.4 trillion-parameter base and a 1 million token context window. Compared to its predecessor, Qwen3.8-Max (version 0902), Qwen4 is said to be dramatically more capable, and cost less to boot.
While New Model Day is always fun, Qwen4-Max scores a somewhat modest 43 points on Artifical Analysis' Intelligence Index (the current leading AI benchmark). That's lower than Kimi K3.1 Pro, which is still the leading Chinese AI model. What's more, Qwen4-Max isn't open-weight; you can't fine-tune it or host it, which makes it similar to offerings from OpenAI and Anthropic in terms of customizability.
But we have several commercially open, advanced AI models in the market today:
So Qwen4-Max doesn't get bonus points for openness, and we're judging it merely on its merits. And on that front, it underperforms GPT-6.1 Terra, the current mid-sized model from OpenAI, which also costs a similar amount.
Qwen4-Max also costs more than Terra per Intelligence Index task, further underscoring how far China is at present from the frontier: The latest and greatest model from the mighty Alibaba is dumber and costlier than the middle-grade model from the No. 2 American AI lab.
- Partly, CO thinks a crackdown on distillation (using an existing AI model to train another) is hampering Chinese AI labs' ability to fast-follow AI advances.
- And there's the****bitter lesson: Here in the States, we have more installed compute than China, and are spending a multiple more on new construction.
Which brings us to:
Nvidia's growth question
Rewind the clock to July 2025. Nvidia has just reported its Q2 F2026 results: $10.16 billion worth of sales to customers in Singapore, good for 22% of its total revenue. At the time, Nvidia told investors that its customers "use Singapore to centralize invoicing, while our products are almost always shipped elsewhere," and that over "99% of controlled Data Center compute revenue billed to Singapore was for orders from U.S.-based customers" in the second quarter.
In the next quarter, Nvidia started attributing revenue to geographies that were grouped by where its customers' headquarters were instead of its billing location. And poof, Singapore left the rankings.
Normally, accounting changes listed in the weeds of 10-Q filings are beyond our remit (I don't want to torture you). But given the following piece of reporting from the Wall Street Journal, you'll understand why we care:
Records seen by The Wall Street Journal show Megaspeed—the company of the former Nvidia CEO’s onetime mistress, Alice Huang—arranged to send nearly $12 billion worth of top-of-the-line Nvidia chips to companies in China between March 2024 and June 2025.
Megaspeed is based in Singapore. Nvidia says it cut Megaspeed off in June 2025. Huang, the Megaspeed executive and Jensen's alleged romantic partner, left her role at the company that August. Hmm.
Someone will eventually get to the bottom of how Megaspeed managed to secure and export so many Nvidia chips to China. But right now, it's worth noting just how much Nvidia gear was historically purchased and funneled to China.
But July 2025 feels like decades ago given how much has changed since, and I doubt Nvidia is in serious danger of being punished for illicit chip transfers to China in a way that would crush its revenue base or growth. But there is an interesting way that companies have offered Nvidia chip capacity to Chinese companies, without crossing the border, that still matters today:
Data centers outside China can freely import Nvidia chips and rent the computing power to Chinese customers. In May 2025, the Trump administration opened the floodgates to the tactic by revoking a Biden administration plan to limit how many chips countries in Southeast Asia and elsewhere could import. The administration said it would put in place alternative curbs—but 18 months later, it has yet to do so. [...] Chai, the Singapore-based researcher, said in a recent report that Malaysia has attracted more than $34 billion in data-center investments, mostly from U.S. and Chinese firms.
The same WSJ piece cites Alibaba and ByteDance as two customers of compute run on Nvidia chips just over the border.
But the United States is now cracking down on such tactics: The Commerce Department on Monday said it is considering new rules that would limit Chinese companies' ability to access American chips in data centers located elsewhere. (Naturally, there's a corruption angle here, too.)
Breaking news: While writing this, the Bureau of Industry & Security (part of the Commerce Department) announced new interim rules to "address remote access to advanced [AI] chips and servers [...] designed to ensure that existing restrictions on access to sensitive American technology cannot be circumvented through remote computing arrangements involving data centers in third countries."
A license will now be required for advanced AI chips "destined for use in data centers, wherever located, that will provide direct or indirect access to a significant volume of computing power to or for an entity headquartered in, or whose ultimate parent company is headquartered in, Macau or destination specified in Country Group D:5," which includes China.
What's more, companies that currently provide "significant volume of compute power to companies headquartered, or with an ultimate parent headquartered, in Macau or Country Group D:5 prior to the publication date of this rule that wishes to continue providing these services must obtain a license with[in] 90 days of the effective date of this IFR or face future export restrictions."
The rule also adds new licensing requirements for chip exports of advanced semiconductors to Malaysia, Singapore, Indonesia, Thailand and Brazil, as the Commerce Department thinks they “present a higher-than-normal risk of diversion.”
Summarizing: The United States government is cracking down on building new data centers in nations near China that intend to rent their capacity to Chinese customers. Now, existing facilities built for that purpose must apply for permission to continue doing business.
Ouch. That could cut off much of China's access to American AI chips. Should we have had tighter rules? Yes. Do we have them now? Yes.
- Is Nvidia's revenue base in trouble? Not so far as we can tell.
- Some 62.5% of the company's revenue in its July 2026 quarter went to U.S. customers, and another 28% to Taiwanese customers. I would classify sales to cloud setups to serve Chinese customers as accelerants rather than the full booster. We're talking more about a growth slowdown than aggregate backward motion.
In the previous section, we noted that Chinese AI models are struggling to catch up to their American counterparts. Further crimping of available compute to Chinese AI labs will further bolster American AI shops, and give allied nations with easy access to Nvidia gear a leg-up. Europe, this is your chance.
Folks clamoring for Jensen's return to power cannot undo market changes already put into effect. And reporting indicates that Nvidia was already pulling back from its AI-industry financing push. So that's over, too. Bringing Jensen back now won't rekindle the old magic. The free-wheeling days are over. Welcome to playing by the rules. Hopefully.
I have a piece bubbling on the Nvidia-Jensen-founder-ethics mess that everyone is talking about. If I get it done today, I'll ship it and share a link in the newsletter tomorrow, I reckon.