Our 50-page report contains 30+ charts and graphs, but this is probably my favorite: The amount of compute we project U.S. and Chinese chipmakers to manufacture over the next few years, put in terms of the chips' continuous power-draw. We call it "Annual Energizable Compute."

China is sprinting to chip independence. In a new report for @AEI, @Satvik_Pen and I estimate Huawei could meet half of China’s compute demand as soon as 2028—permanently transforming the global race to deploy AI. The time to prepare is now. 🧵 1/16 aei.org/research-products/wo…

Here are some observations on the current state of things and low-confidence predictions about the future: - The AI market is decentralizing even as frontier capabilities become more concentrated and potentially dangerous. - We are witnessing a civil war in American tech between incumbent frontier AI labs on the one hand and “little tech” + infrastructure providers on the other. - Google and Amazon are in an interesting position, being able to hedge. - The frontier labs are, in relative terms, likely to lose sole custody over AI’s development. Their best strategy is to align heavily with the U.S. government and capture the defense enterprise. - As we’ve seen over the past few days—and as we witnessed in the encryption wars of the 1990s—a government strategy to crack down on open source software is unlikely to succeed. Over the long term: - Market pressure will fragment AI software development across many smaller labs (as we are already seeing with Thinking Machines and Chinese competitors) and reward infrastructure providers capable of serving everyone. - Surging compute costs (and finite production capacity) will in turn push consumers toward “small” AI models capable of being run on edge devices and applications. - GPUs will also probably fragment as the dominant chip architecture. Tighter integration between software and hardware development will reward specialization into ASICs and low-power inference (hello, Etched). - Such fragmentation at every layer of the value chain will make it hard to sustain—in relative terms—the kind of AI capex spend we have seen over the past few years. In the short term: - A handful of hyperscalers and large labs have already pot-committed to building massive compute clusters—the question is what technical capabilities will emerge from these investments, and whether the physical demands of inference computing at scale will continue to serve as a soft moat around them. - Blackwell clusters might be good enough to squeeze out another order-of-magnitude improvement in frontier AI’s capabilities through the automation of AI research, even before the introduction of Vera Rubin. - That could mean we will soon start seeing even more really disruptive effects from frontier AI—such as the autonomous hacking of HuggingFace last week by an unreleased OpenAI model. - Frontier labs will continue emphasizing catastrophic risks to justify AI’s careful and centralized development, while their competitors will continue downplaying AI’s rapidly improving capabilities and the legitimate risks it might bring. - Near-term capabilities could give rise to an autonomously replicating “AI worm.” - No frontier lab in the United States or China will willfully release truly-destructive superintelligence.

Tectonic shifts in American tech policy have unleashed a new form of geopolitical power: granting and revoking access to frontier AI. Now, to wield a durable “AI umbrella,” we must create a system of rules that provide for its international access. choosingvictory.com/p/the-ca…

Great to join @CNBCi’s Squawk Box Asia to talk about the U.S.-China AI race. China will remain compute-constrained for at least the next 2 years—there’s no way around this. But labs like Z.ai are chasing closely. It’s a mistake to close off the 🇺🇸 frontier.

Just before the White House is set to unveil its new Executive Order on AI safety, it was great to join @CNBCi Squawk Box Asia to discuss U.S.-China competition in AI services. Thanks to hosts @cherykang and @Manisha3005 for having me on the program.

🦁 Into the lion's den! In a guest essay for @AsiaPolicy's Center for China Analysis, I lay out a defense of the U.S. "small yard, high fence" approach to U.S.-China tech competition. There's never been a more important time to restrict China's compute. centerforchinaanalysis.asias…

What the hell is going on with H200 chip sales to China? A bizarre confluence of internal fights within the U.S. and Chinese governments has led to an extremely confusing situation—whereby China may graciously “agree” to import the Nvidia chips its AI labs desperately need, while also framing this as some sort of “concession” the United States should be thankful for. Here’s what’s actually going on: Chinese AI labs desperately need American compute, which is why they’ve been renting or outright smuggling huge numbers of Nvidia-designed chips from third countries. A lack of computational power continues to materially constrain Chinese AI labs’ ability to train frontier models and serve them to global publics. But not every part of the Chinese Communist Party cares equally about the plight of the country’s AI labs. In fact, different segments of the CCP are optimizing for different security and development objectives: While economy-promoting organs like the Ministry of Commerce (MOFCOM) and the National Development and Reform Commission (NDRC) generally want labs like DeepSeek and Alibaba to succeed in building and selling more capable AI services, cadres in the Cyberspace Affairs Commission (CAC) and across the Chinese security services are deeply wary of importing U.S.-designed hardware, fearing U.S. chips might contain back-doors or location-tracking features that could jeopardize China’s national security. Still some other offices—and, I believe, Xi Jinping personally—worry primarily about accelerating China’s indigenization of every part of the semiconductor supply chain—and would rather see Chinese AI labs eat the bitter medicine (吃苦) of temporary chip scarcity, so long as it helps to create a captive market for Chinese chipmakers in the medium-term. The U.S. government, for its part, has had its own disagreements about the wisdom of export control. While it would seem truly idiotic to sell Chinese AI labs the single resource they need to build AI systems that threaten the United States or compete with U.S. companies like @OpenAI, @Anthropic, and @Google in international markets, some parts of the Trump administration are—like parts of the CCP—optimizing for a different objective: A prevailing faction within the Trump administration believes it can disrupt Huawei’s position within the Chinese chip market by keeping China “dependent” on American hardware. Under this worldview, successfully persuading the Chinese government to approve the import of Nvidia’s H200s would be a “win” for American industry. This has led to a bizarre situation where different sets of officials across both governments simultaneously view the sale of H200s to China as a “concession” to be fought for and/or guarded against. USTR Greer’s comment that the import of H200s would be China’s own “sovereign decision” suggests to me that the United States probably asked China to approve their import, CCP security services refused, and the United States opted not to press the issue further. At the end of the day, China’s AI labs remain compute-constrained and will continue desperately trying to gain access to American chips, regardless of whether they may be imported legally—and this will be broadly tolerated by the CCP. Though it probably did not mean to, China has managed to accidentally achieve the best of both worlds: a captive hardware market protected from competitive U.S. exports, and a software industry quietly empowered to rent or smuggle whatever resources it needs to chase the American frontier.

.@SecScottBessent is correct that now is the right moment to reopen conversation with Beijing on AI—Chinese models are only becoming more capable. I fully expect the U.S. and China to establish a formal working group to coordinate on AI security. The hard part will be ensuring whichever interlocutors are selected to maintain that channel actually have some sway over setting their country’s AI policy. At least in the United States, we’re learning in real time that AI is a team sport—it touches every part of U.S. government. That makes it hard to have a meaningful and productive conversation with whoever is supposed to speak for China.

Here is my best assessment of what the U.S. and Chinese sides will both ask for when Trump and Xi inevitably talk about “AI” at next week’s summit. There seems to be a narrow path to “coordination” that is both possible and desirable. choosingvictory.com/p/ai-coo…

In super simple terms: I want DeepSeek, Qwen, Moonshot, and Zhipu to coordinate their model releases with China’s CERT, just like our labs are doing with the U.S. CAISI. This is in both America’s and China’s interest. China's probably going to end up doing some version of this, anyway. Let's make sure we are on the same page about it — and actually call them out when they release easily-jailbroken exploit-generators.

Amid the frenzied food fight in Washington to define which AI risks are real and how seriously to weigh them, the U.S. interagency is breaking—quickly and visibly— toward treating AI as an abnormal technology. Next week, it will ask China to do the same. choosingvictory.com/p/ai-coo…

Every administration that has sought to cooperate with Beijing has faced that bizarre-but-familiar confluence of pressures: an incentive to claim success even when there is little to show for it, and an inability to acknowledge when Beijing is cheating. Under these ordinary circumstances, China can typically get away with just going through the motions or even defecting from the letter of a bilateral agreement with the United States. But pressing that luck now could prove fatal to the entire project. The reality is that the current space for negotiation is fragile enough that even the appearance of China’s non-compliance could easily shatter it. Ahead of next week's Trump-Xi summit — my analysis from January: choosingvictory.com/p/americ…

Congratulations to @RepBaumgartner and all ten bipartisan cosponsors on the MATCH Act's movement through @HouseForeign, by a wide margin — 36-8. It's so important that 🇺🇸 move to restrict the import and servicing of 🇨🇳's advanced lithography machines. aei.org/research-products/re…

Great to speak with @PunchbowlNews' @Dareasmunhoz about the MATCH Act—one of several pieces of legislation due for markup by @HouseForeign on Wednesday—which would require companies to apply for a license to provide maintenance service to their lithography machines in China. I told Punchbowl about our recent research at @AEIfdp diving into the SME supply chain. Julia Torres and I estimate that, absent U.S. action, the hundreds of deep ultraviolet lithography machines imported by Chinese companies will be more than sufficient to manufacture huge numbers of cutting-edge AI chips in the years ahead. On the other hand: “If you take away the parts and prevent them from being serviced, then Huawei is going to be sitting on tens of billions of dollars of wasted investment—unable to make chips at a really critical moment in the trajectory of the global AI industry,” Ryan Fedasiuk, a fellow at @AEI, said. Grateful to @RepBaumgartner and @BrianMastFL for their leadership on this important issue. punchbowl.news/article/tech/…

Look, a lot is being written about export controls on semiconductor manufacturing equipment (SME) by people who haven't gone deep on the science of immersion lithography or chip supply chains. I'm out with a new report today with @AEI's Julia Torres on this exact question. Here's the deal: Huawei and other Chinese chipmakers spent 2023-2024 importing hundreds of older-generation "deep" ultraviolet lithography machines from the Netherlands that fell below the level of American and Dutch export controls on "extreme" UV. These are the machines needed to "print" circuit patterns onto silicon wafers—the basis of a modern computer chip. We thought controlling high-end printers would be enough to stop China from printing high-end chips. But China has steadily upgraded these "older" lithography machines, effectively jailbreaking them to print more sophisticated logic chips than previously thought possible—by having them run over the same patch of silicon multiple times, a process called "multipatterning." It's a super inefficient process that is causing Huawei to struggle with yield rates—they can't make as many chips as efficiently as TSMC. But the bottom line is this: The hundreds of DUV machines already imported by China are more than capable of (expensively) printing the millions of logic dies Huawei needs to build huge quantities of cutting-edge AI chips. These are the chips that will power China's frontier AI systems in 2027-2028. And, unless we do something about it, these old machines will allow Huawei to print—by our estimate—*millions* of high-end (7nm and better) logic chips in the coming years. That would allow Huawei to actually start competing with U.S. chip giants like NVIDIA and AMD in global markets—threatening our goal of exporting the American AI Stack. Thankfully, lithography machines are hundred-million-dollar pieces of equipment—so large and complex, they need to be shipped in 40+ freight containers and serviced constantly by engineers from the single Dutch company that makes them, because their scanner optics degrade over time, and Chinese engineers are working them round-the-clock. Obviously, Chinese chipmakers know they are vulnerable, so they are working on their own optics and lithography equipment to break free of their dependence on ASML. That's why everyone freaked out in December when @f_potkin reported a lab in Shenzhen made its own DUV prototype. China's eventually going to succeed at building its own lithography industry. But a really easy way to hamstring Huawei and frustrate its logic chip production for several years (right now, during the heat of the AI race) would be to stop these older "DUV" machines from being serviced in China, and to stop selling them new ones. We just need to get the Netherlands on board with blocking older equipment sales and servicing. (Het spijt me, Dutch friends). That's what @RepBaumgartner's MATCH Act does. The @HouseForeign will vote on it Wednesday. aei.org/research-products/wo…

Even if Mythos-class AI capabilities remain gated behind large amounts of compute, we're still in for a much less secure online world. In my first-ever piece for @TheFP, I lay out what you can do to stay safe from online threats—AI-enabled or otherwise. thefp.com/p/the-ai-cyber-thr…