Manhattan Projects Everywhere
What we need is a new TVA.
Kevin Xu recently said the process of America and China becoming more alike, the Chimerica, is complete. Kevin was referencing the “golden share” and America’s investment in Intel. In that mirror he saw reflected China’s state owned enterprises.
It is not just Intel and it is not just the Trump Administration that has advanced the idea of U.S. public ownership of private corporations. Senators Warren and Sanders supported a measure in the CHIPS Act that would have seen the U.S. control a portion of countries that accepted money. That failed to make the final bill. There is value in a discussion of how much money the government should give a company and what that should mean for the country. Reasonable people can disagree. That said, I think those demanding government equity are willfully denying why the government chooses to operate this way.
Basic macro says that when the government increases spending in an economy, all things equal, companies and firms decrease their investment. This is called crowding out. One way the Government can offset the drop in investment is to tilt in some investment. They do this via a combination of grants and tax credits on investment. That is exactly what the Advanced Manufacturing Investment Credit (AMIC) in the CHIPS Act did. Companies got 25% back on every dollar of investment in semiconductor manufacturing in the States.
Semiconductors are the backbones of the modern economy and modern defense industry. I would have said they were the bedrock, but that is the rare earth minerals that we will get to at a later date.
Leading edge chips are hard to produce. They require deep process knowledge combined with the ability to replicate at scale. TSMC did not wake up with a production lead. They spent decades producing trail edge chips before they caught up with Samsung and Intel. That long tail allowed them to build and inculcate process knowledge that allows then to scale quickly with new chips.
Fabs are expensive to build. One ASML EUV cost $250-$350 million. It has to be put in a building that is 4 stories tall (up or down does not really matter.) It needs filtered water, assured power, reliable HVAC, clean rooms, the list continues. All of this is before wafer one is touched. This is not a garage project.
Why does any of this matter? Because of Intel’s golden share. Part of me wants to scream “that’s not the American way!” Equally part of me wants to say, this is a national security issue. The U.S. has to have some native ability to produce chips at scale.
What kind of chips? Intel’s 14A sounds like it will be great, higher perf-per-watt than the 18A. (That is geek speak and means it’s more efficient.) Intel only has one issue. No one wanted to be their customer because, they are Intel. The USG might have solved that problem for them. Is the 14A what the military needs? No, it’s not. But, the U.S. could mandate Nvidia or other tech companies use Intel chips, at least for domestic sales. Chimerica indeed.
What does the military need? A metric boat load of legacy chips. The main reason for this is you don’t need bleeding edge chips to make a tank turret spin. In most military applications, a 90nm and above chip works just fine. The cost savings alone justify it.
How do we know legacy chips matter? In 2022, we had reports of Russian troops pulling chips out of washing machines. They had “run out of semiconductors” as a part of US and European sanctions. Damn, good thing we have massive capacity for legacy chip production in the U.S….
What are people focusing on instead? An “AI Manhattan Project.” Again, it’s a bipartisan issue. Senator Slotkin (D) and the Trump Administration, along with others across the spectrum, have expressed the need for a Manhattan Project 2.0.
Does a Manhattan Project for AI mean that the development of AI will move from private companies and into the hands of the Government? God, I hope not. There are plenty of examples where the Government developed cutting edge technology and gave it to civilian markets, ARPAnet and the space program’s myriad benefits jump to mind. Less come to mind where the government took over private companies and “finished” their research.
Hopefully, the rhetoric behind “an AI Manhattan Project” is aimed at spurring investment in infrastructure, expanding STRM education at all levels, and incentivizing companies to build. This would be much less of a Manhattan project and much more of a CHIPS Act without the baggage.
The baggage around the CHIPS Act is important. What can’t happen, or at least should not happen, is the same discussion of “how much of these companies should the U.S. own” that accompanied the CHIPS Act. None is the answer. On its face the reason is clear, xAI, Google, Meta, and OpenAI (backed by Microsoft) have no incentive to take money that results in losing control of a portion of their company. This version of the “AI Manhattan Project” would not even solve the actual problem.
The bottleneck we currently face is power generation and distribution. Power constrains compute, which limits AI development. Instead of a Manhattan Project of AI, we need a Tennessee Valley Authority (TVA) and the Civilian Conservation Corps (CCC) 2.0. The U.S. has the ability to build cheap abundant energy for the country. It could also be overwhelmingly clean with today’s technology.
What would it require? (This is my chance to channel Derek Thompson.) It requires permitting reform to speed up building nuclear and non-nuclear power production and transmission corridors across the country. This needs to be coupled with incentives to companies that collocate their data centers near power production to reduce stress on the grid. Lastly, it means a complete change in how power is bought and sold. No more kilowatts per hour, but priced on reliability and capacity.
This would hurt utility companies. The cost of abundant energy is that Duke Energy and Southern Company would fundamentally change. Their model is, to some degree, based on scarcity we can actually remove. The benefits greatly outweigh the costs.
What does any of this have to do with defense? In my mind, everything. Right now the U.S. still enjoys an advantage in AI and compute because SMIC and Huawei can’t produce leading edge chips at scale (or at all depending on who you ask but, I think 7nm is pretty good.) Where the U.S. is falling behind is power generation, which limits future progress. Call in an abundance mindset, call it energy independence, I don’t care what you call it, but fixing this issue will actually address the bottleneck the U.S. faces and it is a national security concern.