After Nvidia’s White House ‘coup,’ China may not be buying

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. I’m Mark Sullivan, a senior writer at Fast Company,covering emerging tech, AI, and tech policy.

This week, I’m focusing on Nvidia’s up-and-down fortunes stemming from Jensen Huang’s close relationship with Trump. I also look at some reported infighting over AI at Meta, and at the reasons for data centers in space. 

Sign up to receive this newsletter every week via email here. And if you have comments on this issue and/or ideas for future ones, drop me a line at sullivan@fastcompany.com, and follow me on X (formerly Twitter) @thesullivan. 

China may not want (many) Nvidia H200 chips after all

Nvidia appeared to have scored a major coup when President Trump on Monday wrote on Truth Social that the U.S. government would allow the sale of its powerful H200 AI chips to China. Previously, the chip company lobbied its way to an approval to sell its older and weaker H20 chip in China—the world’s second-largest economy and a hotbed of AI and robotics research—but President Xi Jinping told Chinese firms not to buy them, citing security reasons. 

The administration’s favor to Nvidia came with some conditions. The U.S. would get a 25% cut of the Chinese sales, and the chips would undergo a “security review” before their export. And Nvidia’s most powerful chips, the Blackwell GPU, would remain banned from export to China. But Nvidia still stood to make a lot of money selling the H200s. 

Now reports say that the Chinese government plans to restrict the import of the H200s, allowing only a small set of trusted Chinese companies or research organizations to get them. Reuters reports that Alibaba and ByteDance want to order H200s but are waiting for a final decision from the Chinese government.

Xi wants Chinese companies to use chips from domestic companies such as Huawei, which could help the Chinese chip companies catch up with Nvidia in a technological sense. The Information reports that the Chinese government sees the H200s as a “stopgap” solution in the meantime. The Chinese also have serious concerns about the security of the H200s, amplified no doubt by the chance that agents of the U.S. government might install security backdoors or location tracking codes in the chips during the security review. 

Huang reportedly talks to Trump on the phone regularly and has written checks for things like Trump’s new ballroom at the White House. The downside of embracing Trump so openly and unconditionally may have eroded trust for Nvidia in China. In the past, China has mounted state-sponsored or grassroots boycotts against American companies, including Apple, McDonald’s, and the NBA.

And there are other ways of getting Nvidia chips into China. The Information reports that the Chinese AI lab DeepSeek has been using thousands of Nvidia’s Blackwell chips (the most powerful in the world for AI) to train its newest model. Chinese companies have been setting up fake data centers in neutral countries, outfitting them with Nvidia servers loaded with chips, then dismantling the servers and sending the chips off to China. Nvidia said Wednesday that it’s unaware of any such activity.

‘Friction’ between Zuckerberg’s new superintelligence and other parts of Meta?: report

After the disappointing performance of Meta’s latest Llama models, CEO Mark Zuckerberg hatched a plan to put his AI lab in the running to build artificial superintelligence. He badly wants Meta to compete for that holy grail against the likes of OpenAI, Anthropic, xAI, and Google DeepMind. So, he paid $14.3 billion to buy Scale AI with the idea of having that company’s young CEO Alexandr Wang lead a new superintelligence research group at Meta. Over the summer, Wang and Zuckerberg went on a poaching spree to hire top AI research talent away from those companies, offering salaries in the hundreds of millions of dollars. They were successful: The new group has about 100 researchers. 

But all is not well, the New York Times reports. Wang has clashed with some of Zuckerberg’s top lieutenants—Chris Cox, who manages the company’s social network products, and Andrew Bosworth, who runs Meta’s mixed reality (metaverse) business—on how Wang’s group’s research should be applied. From the report:

In one case, Mr. Cox and Mr. Bosworth wanted Mr. Wang’s team to concentrate on using Instagram and Facebook data to help train Meta’s new foundational A.I. model — known as a “frontier” model — to improve the company’s social media feeds and advertising business, they said. But Mr. Wang, who is developing the model, pushed back. He argued that the goal should be to catch up to rival A.I. models from OpenAI and Google before focusing on products, the people said.

In other words, Cox and Bosworth are more interested in using Wang’s AI models as a means to an end (a business end): to pump up social engagement and better target ads at users. But Wang may see the superintelligence group as something more like a “pure research” group that sets its own research agenda. 

Wang, Cox, and Bosworth may simply be the latest actors in a much older tension between pure research and applied AI. “It’s unclear if Mr. Wang, Mr. Cox and Mr. Bosworth have resolved their debate,” the Times reports. After all the money he spent to chase superintelligence, Zuckerberg is likely to side with Wang and insulate the group from short-term demands of product managers. 

Why Musk and Bezos are putting data centers in space

Why are Elon Musk and Jeff Bezos working on missions to launch AI data centers into space? It sounds exotic. But it makes sense. 

Tech companies and their partners are spending trillions to build new terrestrial data centers to produce enough computing power for AI. In some areas, electricity costs have increased after the local energy provider built new grid infrastructure to accommodate new data centers. Data centers need a lot of electricity to power the AI chips inside them, and a lot of electricity and water to keep the chips cool.

It’s very cold in space, so the cooling problem goes away. An orbiting data center could use solar panels to collect the energy needed to run the servers (the sun is 30% more intense in space). Troubles associated with terrestrial data centers—land-use permitting, local zoning, water rights, etc.—don’t apply in space.

The Wall Street Journal reports that Bezos’s Blue Origin has had a team working on orbital AI data centers for more than a year. Musk’s SpaceX has plans to mod one of its Starlink satellites to host AI servers. Google and Planet Labs have plans to launch two test satellites into orbit loaded with Google AI chips (called Tensor Processing Units). 

Other, smaller companies, such as Starcloud and Axiom AI, have sprung up to focus all their efforts on orbiting data centers. Those involved acknowledge that while the floating data centers are technically feasible, lots of work remains to bring the costs down to a point where they’re competitive with earth-based data centers. 

More AI coverage from Fast Company: 

• OpenAI appoints Slack CEO Denise Dresser as first Chief Revenue Officer
• Nvidia’s Washington charm offensive has paid off big
• Google faces a new antitrust probe in Europe over content it uses for AI
• Trump allows Nvidia to sell H200 AI chips to China

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.