Picks and Shovels

It's Tuesday, March thirty first, and welcome to HigherIQ. On March twenty fifth, Google published a blog post about a math paper. By the end of that day, memory chip stocks in the US were falling. By the next morning, the sell-off had spread to South Korea and Japan. Over the last week, Micron, one of the biggest AI hardware winners of the last two years, dropped more than thirty percent. If you own an index fund, a target date retirement fund, or frankly any diversified portfolio, this probably affected you, and you may not have noticed. The question it raised is one worth understanding: is AI a hardware story or a software story? Because the answer just changed. To understand what happened, we need a quick analogy. Think of an AI model like a person reading a very long book while answering questions about it. As the person reads, they take notes so they don't have to reread earlier chapters every time a new question comes up. Those notes are called the key-value cache, and they're stored in memory chips. The longer the book, the more notes, the more memory you need. For the last three years, the AI industry has been buying enormous quantities of memory chips because the books keep getting longer and the questions keep getting harder. Google's paper introduced an algorithm called TurboQuant. What it does, stripped of the technical language, is make those notes dramatically smaller without losing any of the information. Specifically, it compresses each note from sixteen bits of data down to three. That's a six-times reduction in how much memory the system needs to do the same job, with zero loss in accuracy and no retraining required. A mathematical trick that makes the existing notes take up less space. The sell-off started on March twenty fifth, the day Google published. Micron fell four percent. SanDisk fell nearly six percent. Western Digital dropped almost five percent. The next morning in Seoul, SK Hynix, the company that had become one of the hottest semiconductor stocks in the world because of AI demand, fell six percent. Samsung dropped five percent. In Tokyo, Kioxia lost six percent. And it didn't stop there. Over the next eight trading sessions, Micron fell more than thirty percent from its mid-March highs, erasing months of gains. The sell-off wasn't a one-day panic. It was the market systematically repricing how much hardware AI actually needs. Here's why this matters beyond the stock tickers. For the last three years, the dominant investment thesis in AI has been simple: AI needs chips, chips need memory, and the companies that make those things are the safest way to profit from the AI boom. You don't have to pick which AI company wins. You just invest in the picks and shovels. Nvidia makes the GPUs that train the models. Micron and SK Hynix make the memory that runs them. Taiwan Semiconductor manufactures the chips. This thesis drove some of the largest stock gains in market history. Nvidia went from a two hundred billion dollar company to a three trillion dollar company in roughly two years. Micron's stock tripled. SK Hynix became South Korea's most valuable company. The "picks and shovels" thesis works as long as one assumption holds: that more AI means more hardware. TurboQuant is the first credible challenge to that assumption. If a software optimization can reduce memory needs by six times, the math changes. Not all of the hardware demand disappears, but the growth rate investors were pricing in might be too high. Now, two important caveats. First, TurboQuant only works during inference, which is when the AI model is running and generating answers. It doesn't help with training, which is the expensive, months-long process of teaching the model in the first place. Training is what drives the biggest chip orders, and that demand is untouched. Micron's entire high-bandwidth memory production for 2026 is already sold out under binding contracts. The structural shortage in the memory market is expected to persist through the end of the decade. Second, there's a concept in economics called the Jevons Paradox that applies here, and it's one of the most counterintuitive ideas in the history of economics. In 1865, the English economist William Stanley Jevons observed that when James Watt's steam engine became dramatically more fuel-efficient than earlier models, total coal consumption in England didn't fall. It rose. It rose a lot. Why? Because cheaper energy made it economical to use steam engines in factories, mines, ships, and mills that couldn't afford the older, less efficient versions. The efficiency gains expanded the market faster than they reduced per-unit consumption. Jevons was worried about coal depletion, but his insight applies to every technology cycle since. When something gets cheaper to operate, people use more of it, not less. Several analysts made exactly this argument about TurboQuant. If running AI models suddenly costs six times less memory, companies will run bigger models, serve more users, process longer documents, and deploy AI in use cases that were previously too expensive. Bank of America analyst Vivek Arya argued that the six-times improvement likely leads to six times more capability, not six times less spending. Total memory demand could actually increase. This is the same debate that played out in January 2025 when a Chinese AI lab called DeepSeek released a model that performed comparably to American frontier models at a fraction of the cost. The market panicked. Nvidia lost nearly six hundred billion dollars in market value in a single day, the largest one-day loss for any company in stock market history. Headlines declared the AI hardware trade over. Within weeks, Nvidia's stock recovered. The logic was the same: cheaper AI doesn't mean less AI. It means more AI. More companies building with it, more applications that become economically viable, more compute consumed in aggregate even as the cost per unit falls. One analyst at Citrini Research compared the TurboQuant panic to saying "Aramco should crash because Toyota came out with a next-generation hybrid engine." The efficiency of the engine doesn't change the demand for fuel if it puts more cars on the road. So how should someone who isn't a chip analyst or an AI researcher think about this? If you own a broad index fund like the S&P 500, roughly thirty percent of your money is in technology stocks, and a meaningful chunk of that is in companies whose valuations depend on the AI hardware thesis. Nvidia alone is one of the largest weightings in the index. You are, whether you chose to be or not, making a bet on how much hardware AI will need. There are two types of AI investments. The first type bets on the infrastructure: the chip makers, the memory companies, the data center builders, the cloud providers. These are the picks and shovels. They do well when AI spending goes up, regardless of who wins the AI product race. The second type bets on the applications: the companies that use AI to do something useful for customers. Software companies, healthcare companies, financial services firms, anyone building products on top of the models. What TurboQuant and DeepSeek both suggest is that the infrastructure story has more risk than people assumed. Software breakthroughs can change the hardware equation overnight. A blog post can erase tens of billions in value before the market closes. The application story, by contrast, actually gets better when AI gets cheaper. If it costs less to run a model, more companies can afford to build with it, and the products get better and cheaper for consumers. None of this means memory stocks are bad investments. Micron just reported revenue growth of nearly two hundred percent and its entire high-bandwidth memory production for 2026 is sold out. The structural demand for AI hardware is real. But the episode illustrates something the market is still learning: in technology, the thing you're sure about is the thing that changes fastest. Three years ago, nobody thought AI would need this much hardware. Last week, a math paper suggested it might need less. Both of those surprises moved billions of dollars. The next one will too. So if this comes up in conversation, here's how to think about it. Google published a research paper that compresses AI memory needs by six times. Memory chip stocks dropped across the US, South Korea, and Japan within hours. The deeper question isn't about one algorithm. It's about whether AI is a hardware story or a software story. For three years, the market has treated it as hardware. Picks and shovels. Buy the chip makers. But software breakthroughs can change the hardware math overnight, and they keep happening. If you're investing in AI through an index fund, you're probably more exposed to the hardware bet than you realize. The companies that benefit most from cheaper AI aren't the ones selling the chips. They're the ones building with them. Stay informed, stay curious, and we'll see you tomorrow.