Trump's Tariff Gun

Fab Diplomacy

The conventional wisdom on President Donald Trump’s trade policy is that it’s mostly just chaos and bluster, a series of threats made on social media that are often walked back a day later. This week, however, the bluster got very real for a factory in Japan.

Here is the Wall Street Journal:

Taiwan Semiconductor Manufacturing is delaying construction of a second plant in Japan partly because it is pouring funds more quickly into U.S. expansion ahead of potential Trump administration tariffs, people familiar with the plans said.

The revised schedule is the latest example of how President Trump’s aggressive stance on trade is pulling some investment toward the U.S. at the expense of allies.

This is, simply put, a huge deal. TSMC is arguably the most important manufacturing company in the world; it makes the advanced chips that power everything from Apple iPhones to Nvidia AI servers. For years, the US has tried to lure TSMC to build more factories on American soil, offering billions in CHIPS Act subsidies as an incentive. Japan, also desperate for domestic chip production, offered TSMC billions of its own.

But it seems the stick is proving more effective than the carrot. The threat of a US tariff on imported semiconductors has apparently forced TSMC to re-evaluate its "meticulous approach to capital spending." The company, famously wary of building too much capacity, is now prioritizing its multi-billion-dollar investments in Arizona to ensure it can supply its American customers without getting hit by a tariff. The plant in Japan, a key US ally, gets pushed down the priority list.

This is the "America First" supply chain taking shape in real time. It’s not just about asking nicely or offering subsidies. The implicit threat is that if you want to sell your critical technology into the world's largest market, you had better make it here. And this logic doesn't just apply to TSMC. We saw Nvidia recently announce plans for up to $500 billion in US-based AI infrastructure production over the next four years.

The move is a blow to Japan, which is already struggling with 25% US tariffs on its auto and steel industries and had hoped for a quick trade resolution. It shows that in this new era of geoeconomics, even close alliances take a backseat to domestic industrial policy. The US government wants advanced semiconductor manufacturing to happen in the US, and it has found a very effective, if blunt, instrument to make its wishes known. You don’t have to like the tariff gun, but it’s hard to argue that it isn’t working.

Recursive Self-Improvement

There is a research group called Model Evaluation & Threat Research, or METR, that studies how good AI models are getting. This is, you know, a thing. People get paid to benchmark large language models, and METR’s latest paper has a startling conclusion: AI capabilities are doubling every seven months. If that trend holds, they project that by 2030 an AI could complete a task that would take a human a full month of work with a 50% success rate.

But the really mind-bending part of METR's research isn't what the AI can do for us. It's what it can do for itself. As IEEE Spectrum reports:

Such tasks might include starting up a company, writing a novel, or greatly improving an existing LLM. The availability of LLMs with that kind of capability “would come with enormous stakes, both in terms of potential benefits and potential risks,” AI researcher Zach Stein-Perlman wrote in a blog post.

The ability to "greatly improve an existing LLM" is the really mind-bending part. It’s a bit recursive. We are in the midst of a technological revolution where the goal is to build a machine whose primary job is to design a better version of itself. It is like hiring a brilliant intern and, on their first day, telling them their only task is to find and train their replacement, who will be twice as smart and half as expensive. And then that replacement's only job is to hire their replacement, and so on.

This is a strange business to be in. How do you value a company whose core mission is to make its own products obsolete as quickly as possible? And yet, this is essentially the bull case for the entire AI infrastructure sector. The reason companies are spending hundreds of billions of dollars on Nvidia chips and data center capacity is not just to serve the current demand for chatbots. It is a bet on this recursive loop of self-improvement. The better the models get, the more compute they need to get even better.

For now, the ones doing the improving are still human. The "AI talent war" is really a war for the handful of researchers who can coax these improvements out of the models. We’ve seen Meta reportedly offering nine-figure compensation packages to poach people from OpenAI. But the premise of the METR paper is that eventually the AI will take over that job too.

This self-improvement cycle is the "world-changing" potential that people like former Google CEO Eric Schmidt talk about. It’s also what creates the immense, almost unquantifiable demand for compute that has made Nvidia a multi-trillion-dollar company. The bet is that the machines will get so good at improving themselves that the demand for the hardware they run on will be functionally infinite.

Of course, as the METR researchers note, there are potential bottlenecks. Progress could be limited by the availability of hardware and robotics. You still need to physically build the data centers and supply them with power, a process that is running into very real-world constraints. But the underlying logic of the market is now clear. The AI race is not just about building a single, static product. It’s about investing in a process of accelerating intelligence, a process that, if the trend holds, will soon be running itself.

The Scoreboard

• AI: US Plans AI Chip Curbs on Malaysia, Thailand Over China Concerns (Bloomberg)
• AI: Nvidia Partner Hon Hai Meets Sales Estimates on Strong AI Demand (Bloomberg)
• Search: The Companies Betting They Can Profit From Google Search’s Demise (WSJ)
• Social Media: Trump Says US Will Start Talks With China on TikTok Deal This Week (Reuters)