The Great AI Irony
CapEx bottleneck companies are capturing stock appreciation, but in the long-run, the app layer will create the most value
“We’re in the vertical ascent phase” is a phrase you hear often in the Valley right now. But it’s not the vertical takeoff of capabilities people are talking about — it’s the vertical takeoff of stock prices!
The AI bottleneck trade has become the talk of the town this year, as surging demand for AI has run headfirst into a supply chain that had not sufficiently scaled to meet it. The “Year of Delays” is turning into the “Year of Bottlenecks.”
The mouth watering profits of commodity providers — from HVAC makers to memory manufacturers to engine makers — has investors salivating. And it has captured retail momentum in a way that I’ve never seen before in my career. Every dinner party people are talking about their personal profits owning Micron or SK Hynix. It’s mind blowing to think that Micron’s valuation today exceeds that of OpenAI, the company at the very heart of the generative AI boom.
These are signs of exuberance, of course, but there is a great irony behind it.
The great irony is that the history of technology is rife with “big systems” like mainframes, cell phones, or engines that over time become smaller and more efficient. The very success of AI guarantees that the returns to making systems more efficient will attract entrepreneurs and capital. As the systems get cheaper, they also become more ubiquitous, which creates compounding growth and radical change. This is a theme of Alasdair Nairn’s writings, including his forthcoming book “How AI Will Move Markets.”
Today’s AI revolution is real precisely because the fact that intelligence can emerge from sufficiently scaled computing, applied to gargantuan quantums of data, is a scientific breakthrough of world-historic importance. Like most breakthroughs in history, we barely understand how these systems work today — the leading edge researchers at the frontier are tinkerers with empirical, rather than scientific, understanding of AI systems. Like an engine mechanic in the early days of the Industrial Revolution, they are experts at working with AI systems, rather than theorists of the craft. But eventually, there will be a science of LLMs, and once the principles behind these systems are fully understood, we will make dramatic progress. Today’s AI systems will look like yesterday’s mainframes.
There is a famous quote from John Maynard Keynes that there is more money to be made predicting what other investors will do than there is from true fundamental investing. This is why market cycles get so extreme. It is natural that in today’s markets, the surging profits of bottleneck providers are attracting such profound attention. But for those of us investing with a ten or twenty year time frame — and for founders who are optimizing for changing the world rather than making quick money — the long-term still matters.
And that is the second irony: it is precisely what will win in the long-term — real AI applications with tangible economic value for the end-user, that investors seem to be writing off these days. This is a hard, grindy business — value compounds slowly but surely. With each year, you understand your customer better, and you can deliver them a better experience. With each underlying upgrade cycle of the big LLMs, you have a better product to deliver your customers. There is a very real risk that foundation models vertically integrate to compete against you — but that is manageable, as their waterfront is wide. Time and focus are on your side.
It is a sign of the times that even Meta would rather rent out their compute than figure out how to derive tangible end-user value from this resource.
There is a narrative today that today’s bottlenecks will exist forever, or at least for decades, because of the promise of AGI. I think that’s optimistic thinking — it’s backsolving for what will make stocks keep working, rather than for what history teaches us. The beauty of capitalism is that mouth watering profits are a great motivator of CapEx expansion and competition. The whole point of semiconductors is that they eventually become boring — they are a means to an end. For those of us looking out into the horizon and dreaming of a future with abundant intelligence, its the next phase — the value delivery phase — that should be the really energizing one.