Geopolitics is increasingly defined by the race for advanced chips. China is improving its homegrown semiconductor industry while the United States preserves its access to cutting-edge silicon. The artificial intelligence boom has made each gain and setback more consequential. To understand all this and more, I spoke with Chris Miller, author of the 2022 bestseller “Chip War.”
This interview has been condensed and edited for clarity.
***
Damir Marusic: Chris, you wrote “Chip War,” one of the most prescient books about the U.S. and China vying over microchip technology. Four years later, how’s that competition going?
Chris Miller: Well, the reason there’s a chip war is that AI is computationally intensive. It requires a ton of chips to train AI systems and a ton more to deploy them. The good news is that it’s predominantly U.S. companies, plus companies in allied countries, that produce most of the chips we need to train and deploy those systems. We started in an advantageous position; we weren’t dependent on China for the critical, high-value, technologically advanced inputs for AI.
It’s no surprise that China is trying to catch up. For more than a decade, it’s been pouring tens of billions of dollars a year into building its own semiconductor design and manufacturing capability. It’s a capable country, and that’s a lot of money. Even today, though, most of its leading AI systems are trained and deployed on Western-designed and manufactured chips, which shows that the country still has some way to go.
Does instituting an export control on chips incentivize an adversary to work harder? The argument you hear — especially from Nvidia and its CEO Jensen Huang — is that the real fight is over the infrastructure that will run the global economy. Is there an advantage to preserving China’s dependency on U.S. chips?
There’s more than a grain of truth to the argument that chip-export controls incentivize some Chinese players to build domestic capacity. But it’s important not to think of the country as a monolith. It’s clear that the government has long wanted its own technology stack and has spent a lot of money trying to build it. I don’t think it’s plausible to say we’ve changed its incentives: Beijing was talking about self-sufficiency in semiconductors before anyone in Washington knew what a semiconductor was.
The Chinese tech sector is varied. Companies like Huawei have every incentive to want a self-sufficient tech stack. If Nvidia is locked out of China, Huawei benefits because it won’t have tough competition. China has been restrictive on letting Nvidia chips into the country. One interpretation is that it’s because Huawei has been lobbying to keep them out.
Then there’s the Chinese cloud computing and AI companies, which would prefer imported chips. They still do use a fair number of them. There is some risk that if we don’t give them any access to Nvidia, they’ll have no choice but to help Huawei and other local players build the domestic chip ecosystem.
But when you read the various reports and listen to U.S. government officials talk about Chinese models being trained on Nvidia chips in places like Thailand, it doesn’t seem like the Chinese AI companies have had that much trouble accessing U.S.-designed chips, either by smuggling them into China or by using cloud computing services outside China.
Is that a loophole we need to patch up?
On smuggling: The combination of a lot of regulatory back-and-forth between the Biden and Trump administrations, a limited budget and confusion as to how aggressively to enforce certain rules has meant that enforcement has been surprisingly bad. And that’s a problem, because the worst outcome is an export control that doesn’t even get enforced.
There’s a different debate about remote access to cloud computing. There are those who want to shut it down to keep China from training its models versus those who point out that it’s good to have China’s tech ecosystem dependent on computing outside the country. But I wouldn’t put it in the category of failed export-control enforcement, because the U.S. has thus far made a choice not to close it off.
One of your colleagues at the American Enterprise Institute recently published a report in which he estimated that, by 2028, Chinese chip output might be able to cover as much as 87 percent of domestic demand. What do you think?
Predictions are hard, especially about the future. A lot depends on what you think AI demand is going to be and whether Chinese companies will be allowed to access the foreign cloud.
I’m probably more pessimistic about China’s ability to increase production. It’s been a struggle for companies like Semiconductor Manufacturing International Corp., or SMIC — the leading producer of logic chips in China and the key partner for Huawei — to ramp up output. For much of 2024 and 2025, Huawei’s primary strategy for acquiring high-end chips was to set up a shell company and smuggle them in from Taiwan. You would do that only if you were really struggling domestically.
Will SMIC improve? No doubt. It’s impossible to know how quickly. But in some ways, the surprise for most people has been how much China has struggled to ramp up domestic production capacity for high-end chips.
There have been reports that China has made progress on developing its own advanced lithography machines needed to make chips — in part through industrial espionage against ASML, the Dutch company that has a near monopoly on this technology. Is this a Sputnik moment?
Let’s put the story in context. The most recent one was about a Chinese firm preparing to ship about five “immersion DUV” lithography machines to customers: a type of tool ASML first produced in the early 2000s — almost a quarter of a century ago. Five of them are supposed to be shipped to customers this year, which is a couple percent of ASML’s global production of DUV machines. If these machines work at the requisite level of throughput, we’re talking about a small competitor in an old technology.
For ASML, it took a long time to get these machines working well at the highest throughput. A reasonable assumption would be that the first generation of Chinese manufacturers are going to have a steep learning curve as they gather data, figure out what breaks more often than it should and try to fix it.
I don’t think we should be surprised that a quarter century after ASML figured it out, China has begun manufacturing its own tools. The surprise, in some ways, is that it’s taken so long. So we shouldn’t overreact to any one headline. Most of them have been more noise than signal.
Is Taiwan still as important, strategically, as it was four years ago?
Yes. Ninety-plus percent of AI accelerators are going to be produced there this year, and almost certainly next year, too. But we are starting to see increased diversification.
Taiwan Semiconductor Manufacturing Co. is building plants in Arizona. The first one is up and running, and it has almost a dozen more coming online over the next decade. The company talks about maybe a quarter of its leading-edge chips being produced in the state by the mid-2030s.
In part because there has been such extraordinary demand for silicon, we have also seen companies like Samsung and Intel win major orders from new customers. The health of Intel Foundry and Samsung Foundry today looks much better than it did even six months ago. Intel’s plants are largely in the U.S., and Samsung’s are in South Korea and the U.S.
Finally: Elon Musk is entering the chip manufacturing business with his new venture, Terafab. Between space, AI models and data centers, he’s a busy guy. But everything I hear from those in the semiconductor supply chain suggests he’s taking this seriously. When you put all these pieces together, I think we’re guaranteed to see a more geographically diverse chip production base by the middle of next decade.
Taiwan, however, will be an irreplaceable supplier of silicon. U.S. companies account for 70 percent or 80 percent of TSMC’s revenue. Even if 25 percent of TSMC’s advanced production were in the U.S. and 70-plus percent were in Taiwan, U.S. firms would still largely be buying that remaining capacity. There’s no world in which the U.S. tech sector is independent of Taiwan for the next decade.
We therefore need to ensure we do what’s necessary to help secure the island’s autonomy and our ability to source the products we need from it.
The post China is still losing the chip war appeared first on Washington Post.
