The U.S. is being played.
Corporatism always comes with a “for sale” sign. Sometimes it’s straightforward, good old graft, an exchange of favors, that sort of thing.
And then there is the arrangement under which Nvidia and AMD (Advanced Micro Devices) have agreed to pay Uncle Sam 15 percent of the revenues they receive from sales into China of, respectively, their sales of H20 and MI308 chips in exchange for the grant of the necessary export licenses.
Reporting on this story for the Financial Times, Demetri Sevastopulo and Michael Acton described this agreement as “unusual,” which, I suppose, is one way of spelling “nuts.”
As they explain:
In April, the Trump administration said it would ban H20 exports to China. However, Trump reversed course in June after meeting Huang at the White House. Over the following weeks, Nvidia became concerned because the Bureau of Industry and Security [BIS], the arm of the commerce department that runs export controls, had not issued any licenses.
Huang raised the issue with Trump on Wednesday, according to people familiar with the exchange, and BIS started issuing licenses on Friday.
The pressure on Fed Chairman Powell to cut interest rates might be taken as a warning sign that the heavily indebted U.S. is closing in on an era of fiscal dominance (an unpleasant state of affairs when a country’s need to keep financing its debt overwhelms monetary policy), but the Treasury’s need for extra dollars is not meant to seep into national security policy — nor is madness.
The Financial Times:
In a recent letter to commerce secretary Howard Lutnick, Matt Pottinger, a China expert who was deputy national security adviser in Trump’s first term, and 19 other security experts urged the US not to grant H20 licences. They said the H20 was a “potent accelerator of China’s frontier AI capabilities” and would ultimately be used by the Chinese military.
Nvidia said the claims were “misguided” and rejected the notion that China could use the H20 for military purposes.
Oh come on.
The FT:
On Saturday, Nvidia said: “While we haven’t shipped H20 to China for months, we hope export control rules will let America compete in China and worldwide. America cannot repeat 5G and lose telecommunication leadership. America’s AI tech stack can be the world’s standard if we race.”
I don’t blame Nvidia for making that argument, even if it is one more suited to the (imagined) world of 2000 than the real, dangerous world of 2025. The company has its shareholders to think of.
If the price of American AI becoming the “world’s standard” means giving China a helping hand, it is not worth it. Moreover, and this is not only the case with Nvidia or AMD, the less dependent that American companies are on Chinese demand (or supply) in anything that matters the better. Given the regrettable depth of the Sino-American commercial entanglement, a full “divorce” is not feasible yet, but, regardless of the direct security implications, granting these export licenses is the reverse of the Paltrovian “conscious uncoupling” that is clearly called for.
In an article written a couple of weeks ago, The Economist considered the wisdom of selling U.S. “world-beating” AI chips to China and found none:
That is because as impressive as Chinese models have been, America’s chip controls were clearly working. When Nvidia devised the H20 to comply with an earlier set of rules, it inadvertently created a chip that was hobbled for training new AI models, but perfect for running them—a process called inference. Since exports of the H20 were banned in April, even the Chinese labs that had overcome the shortage of training chips to produce world-class AI models have been unable to access enough computing capacity to offer those models to paying customers. They have had to resort to relying on outsourced hosting, and making the most of the limited quantity of AI chips produced by Huawei and other Chinese hardware firms. But the trend seems clear: without H20s, Chinese companies cannot keep up with demand.
And as AI adoption increases, having enough capacity for inference will become ever more important, making export controls even more potent. America’s ban on the export of H20s, in short, has impeded China’s progress in AI. It seems perverse for America, engaged in an arms race with China, to give up this advantage.
The argument that selling H20s and their kin will make China “dependent” on the U.S. in this area can only be made by those who do not understand how China operates and, yet, incredibly it is being made within the administration.
A stronger argument is that depriving China of the American chips it needs will force it to turn to its own formidable resources. That’s true, but as The Economist’s writers warn:
if there is even a small chance that the time-frame for AI development suggested by America’s AI leaders is correct, the race for superintelligence may have been won by 2030. Accordingly, America should do everything it can to win that race in the short term, even if that means it fails to hamper the development of China’s hardware industry in the longer term.
Meanwhile (via the Wall Street Journal, August 3):
China is limiting the flow of critical minerals to Western defense manufacturers, delaying production and forcing companies to scour the world for stockpiles of the minerals needed to make everything from bullets to jet fighters.
Earlier this year, as U.S.-China trade tensions soared, Beijing tightened the controls it places on the export of rare earths. While Beijing allowed them to start flowing after the Trump administration agreed in June to a series of trade concessions, China has maintained a lock on critical minerals for defense purposes. China supplies around 90% of the world’s rare earths and dominates the production of many other critical minerals.
As a result, one drone-parts manufacturer that supplies the U.S. military was forced to delay orders by up to two months while it searched for a non-Chinese source of magnets, which are assembled from rare earths.
The U.S. is being played.