The robot in the Micron call nobody is talking about

In part one I made the case that Micron Technology ($MU) may have quietly turned memory from a cyclical business into a structural one, using sixteen take-or-pay contracts with price floors and billions in customer cash put down up front. If you missed it, start there, because this issue builds directly on it.

A floor under pricing only matters if demand keeps showing up to sit on top of it. So today is about the engines underneath the shortage, four of them, running on different clocks. Three are already here. The fourth is the one I have been waiting years to see a memory company put a hard number on, and it is about robots.

The clearest driver is high-bandwidth memory, or HBM, the tall stacks of memory chips that sit shoulder to shoulder with an AI accelerator and feed it data fast enough to keep up. Picture stacking a dozen memory chips vertically and wiring them straight through with thousands of microscopic copper channels, instead of laying them flat on a board. That is HBM, and it is where the money is.

What changed in the most recent earnings release is the timeline. Micron now sees the HBM market crossing $100 billion as soon as calendar 2027, pulled forward a full year from its prior call of 2028, against a 2025 market its own projections put around $35 billion. Micron says its newest generation, HBM4, is ramping twice as fast as the prior one, and that it has already shipped over $1 billion of it. When a market gets dragged a year closer in a single quarter, that is demand arriving faster than the people closest to it expected.

The second engine is less obvious and, to me, more interesting, because it widens the story beyond the AI accelerators everyone already watches.

Agentic AI, the systems that take actions rather than just answer questions, does not run on accelerators alone. It needs racks of ordinary CPUs to run what Micron calls the "control plane," the part that directs the AI agents and executes their programs, plus storage racks to hold the growing pile of context these agents carry around. That pulls in conventional server memory and a low-power variant Micron pioneered called SOCAMM, designed to cut the power and space that memory takes up inside a server. On the strength of this, Micron lifted its 2026 server-unit growth outlook to the high teens, up from low double digits.

The reason this matters for the structural thesis: it means the demand is not riding on a single product in a single kind of rack. It is broadening across the whole data center, which is exactly what you want to see if you believe the shortage has staying power rather than being one hot product cycle.

The third engine is storage. Enterprise SSDs, the high-capacity flash drives that feed AI servers, hit a $5 billion quarter for Micron and more than doubled from the prior quarter. AI does not only compute, it remembers, and that growing "context store," everything a model needs to keep on hand, is becoming a structural buyer of flash. Management also pointed to AI storage displacing traditional hard drives as a second leg of that demand. This is the part of the AI build that gets the least attention and is quietly inflecting.

Now the one I have been waiting for. I have been telling you for a while that physical AI, robots and humanoids that act in the real world, is the next five-year growth theme, the natural next step after AI learned to think. Micron's most recent earnings release was the first time I have seen a memory leader put a hard number on it, and the number is worth your attention.

Here is how management framed it. An advanced driver-assist car, the kind that handles more of the driving itself, already carries over five times the memory and storage of an average car. A humanoid robot, they said, carries ten times the memory of one of those advanced cars. Stack those together and a humanoid is roughly fifty times the memory content of the car in your driveway today. Think of a humanoid as a walking data center that has to see its surroundings, decide what to do, and move, all in real time, all on board. Micron expects this to begin "a sustained, substantial multi-decade memory demand cycle" in the latter part of this decade.

The automotive side of this is already moving, not speculative. Micron said the share of these advanced driver-assist vehicles is more than doubling this year to over 20% of the market, and is expected to pass 40% by 2030, with memory content per car climbing as vehicles take on more autonomy. So the car part of the curve is underway today, and the humanoid part is the same idea taken to an extreme, just arriving later. That is the bridge I find convincing: you do not have to believe in walking robots next year to see the memory content per machine ratcheting up across an entire category of devices.

I want to be careful here, because this is the easiest part of the story to get carried away with. The robotics demand is late-decade, not next quarter, and management framed it as a long arc rather than a current driver. But it tells you why five-year contracts exist at all. When your customers can see autonomous vehicles and humanoids coming, each carrying orders of magnitude more memory than today's devices, locking up supply through 2030 stops looking like caution and starts looking obvious. The robotics line is not a 2026 catalyst. It is the reason the structural story has a second decade behind it.

What makes me trust the robotics read is that Micron is not the only one pointing at it, and the others have no memory book to talk up. At NVIDIA's ($NVDA) GTC events this year, Jensen Huang put physical AI at the center of the company's roadmap, showed off a humanoid reference platform that runs on NVIDIA's robotics computer, and said the age of robotics is starting now. That matters for memory because the same Vera Rubin systems he just moved into full production are packed with the HBM4 Micron described, so NVIDIA's robot ambitions and Micron's memory shortage are two ends of one supply chain.

Intel's ($INTC) chief executive, Lip-Bu Tan, has been making a parallel argument from the compute and power side, framing the next phase of AI as "not just GPUs, but also power, materials, and manufacturing," and pointing to robotics and on-device physical AI as the reason compute has to spread back out to the edge, closer to where the robot actually is. When the memory leader, the accelerator leader, and the incumbent processor maker all independently describe a world where every humanoid is a power-hungry, memory-hungry, compute-hungry machine, the story stops sounding like one company talking its own book and starts looking like a shared read on where demand goes next. That breadth is the part I find most bullish, because it spreads across sectors rather than resting on a single name.

The cleanest way to own this thesis is Micron itself, a disclosed holding for me, and I also hold a memory-focused position through the $DRAM fund. But the more interesting work, as always, is one layer down the supply chain, because if memory has genuinely re-rated from cyclical to structural, the companies that feed Micron get pulled along.

The rest of this issue, the supply-chain names I watch and how I weigh them, is for paid subscribers.