One afternoon last year, I sat in a shaded courtyard outside Nakuru with a group of smallholder farmers, a robotics engineer, and a policy advisor, working through an early draft of what would become the African AI-Blockchain Ethics Charter. We were eight months into a consultation that would eventually involve thirty stakeholders — technologists, regulators, and people whose relationship to technology is measured in harvests rather than hertz. We had reached the section on autonomous agricultural machinery, and a farmer named Grace asked the engineer a question I have carried with me ever since. "When the machine decides," she said, "who is responsible for the mistake?" The engineer talked about sensor fusion, redundancy, and fail-safe modes. Grace listened patiently, nodded, and then said something that reframed the entire afternoon: "You are describing a brain. I am asking about a conscience."
I thought about Grace again when I read the transcript of Micron's most recent earnings call. Sanjay Mehrotra, the company's chief executive, had spent a portion of that call not on the data center, not on the high-bandwidth memory that has made the storage industry euphoric, but on something he called Physical AI — the idea that machines which move through the physical world will need memory and storage on a scale we have never had to provision for at the edge. He described a future in which an autonomous vehicle at Level 4 or above requires more than 200 gigabytes of memory and several terabytes of storage, an order of magnitude beyond the cars on the road today. He placed the moment of mass adoption at the end of the decade.
It was a beautiful narrative. It was also, I suspect, a careful one. And because I have spent my professional life both auditing the systems that hold value and watching narratives detach from the things they describe, I want to read it the way I would read a smart contract — tracing what it promises against what it can actually enforce. There is a moral code behind every forecast, and the work of an honest reader is to trace it.
Micron is one of the three remaining large-scale memory manufacturers, an integrated device maker that designs and fabricates DRAM and NAND flash in its own facilities. The industry it inhabits is peculiar: memory is the second-largest category in semiconductors, roughly a hundred and sixty billion dollars a year, and yet it is profoundly cyclical. In a boom, gross margins can exceed forty percent; in a bust, they can go negative. This is the context in which a chief executive speaks to investors, and it is the context in which a phrase like "the next growth market after the data center" acquires a certain weight. Words spoken to shareholders are never merely descriptions. They are instruments, and they are calibrated.
The term "Physical AI" is not Micron's invention. It belongs to Nvidia, whose chief executive uses it to describe systems that perceive, reason, and act in the physical world — robots, autonomous vehicles, and the industrial machines that will increasingly run alongside them. When a memory manufacturer borrows that vocabulary, it is doing more than adopting a phrase. It is signaling an ecosystem alignment, a bet that its fortunes will rise with the platforms that animate bodies rather than the data centers that train minds. That alignment is the subtext of the entire narrative, and it is worth holding in view.
What Micron is proposing, in essence, is a story about diversification. The data center has been the engine of memory demand for the past several years, and it will remain so. But the data center is a concentrated market. Its buyers are a handful of cloud providers, and its most lucrative segment, high-bandwidth memory, is dominated by a competitor. Automotive and robotics, by contrast, are fragmented markets, spread across tiers of suppliers and manufacturers, and they reward different virtues: reliability, longevity, certification. Micron happens to lead in exactly those virtues. The Physical AI story, then, is not a random forecast. It is a forecast told from a position of strength, and the strength is not incidental to the story — it is the reason the story is being told at all.
There is an idea in computer architecture that the industry has largely stopped discussing in public, though every engineer knows it intimately: the memory wall. Processors have grown faster than the memory that feeds them, and so the true constraint on a computation is often not how quickly a chip can think but how quickly it can be handed something to think about. For decades we hid this imbalance behind caches and prefetching and cleverness. The arrival of artificial intelligence has made it impossible to hide. A model is, at bottom, a great deal of arithmetic performed on a great deal of data, and the arithmetic is now cheap while the movement of data is not. When computation was scarce and data was abundant, we optimized for arithmetic. Now arithmetic is abundant and the movement of data is scarce, and the whole industry is being quietly reorganized around that inversion.
This is why I read Micron's two-number framing — more than 200 gigabytes of memory, several terabytes of storage — as the most technically honest sentence in the call. The memory wall, not the compute ceiling, is the real bottleneck at the edge. If the problem were merely capacity, one number would suffice. The fact that memory and storage are named separately tells you that both are growing, and that the growth is not interchangeable. Memory is where a model lives while it reasons; storage is where the world it has learned about is kept. An autonomous vehicle is not a car with a computer in it. It is a mobile data center with wheels, and it must do its inference locally, in real time, without the luxury of a round trip to the cloud.
Why local? Because latency is a safety property. A decision that must be made in a fraction of a second cannot wait for a network. And because bandwidth is finite: a fleet of vehicles streaming raw sensor data to a central facility would saturate any network we could build, and the cost of that transmission would exceed the cost of the silicon that avoids it. The edge is not a fashion. It is a physical necessity, and physical necessities are where storage demand becomes structural rather than cyclical. When a demand curve stops depending on fashion and starts depending on physics, it becomes something an investor can hold for a decade. That is the real promise buried inside the phrase Physical AI, and it is a promise about architecture, not about marketing.
The specifications themselves are less interesting than the qualifications that surround them. To put memory into a car, a manufacturer must satisfy a gauntlet that has nothing to do with capacity. AEC-Q100 for integrated circuits, AEC-Q104 for multi-chip modules, ISO 26262 for functional safety up to the ASIL-D level, and — perhaps most demanding of all — a commitment to supply the same part for fifteen years. A consumer electronics vendor may discontinue a component after eighteen months. A car maker cannot. A vehicle built today will still be on the road in 2040, and its memory must be replaceable, traceable, and available. A chip that is abandoned is a car that cannot be repaired, and a car that cannot be repaired is a liability that no manufacturer will accept.
The real barrier to entry in automotive memory is not the ability to make a fast chip; it is the willingness to make the same chip, flawlessly, for fifteen years. This is a different kind of moat than the one that protects high-bandwidth memory, which is a manufacturing and packaging race decided in quarters. The automotive moat is built from patience, documentation, and process discipline — the unglamorous virtues that markets reward only slowly and then all at once. I have seen this pattern before in my own work. When I helped structure a DAO-governed royalty system for a group of Kenyan artists, the hard part was never the smart contract. The hard part was the governance: the willingness to keep a promise after the excitement had faded. The same is true of a fifteen-year supply commitment. It is a promise, and the promise is the product.
Here I should be specific, because the narrative depends on it. The Level 2 and Level 3 vehicles on the road today carry roughly eight to sixteen gigabytes of DRAM and between 128 gigabytes and a terabyte of NAND. The Level 4 specification Micron describes — north of 200 gigabytes of memory — is not a doubling. It is a leap of an order of magnitude, consistent with what automotive analysts have been quietly modeling for years. To meet it, a vehicle will need LPDDR5X or its successor for main memory, and UFS 4.0 or automotive-grade NVMe for storage, all of it qualified to survive temperature extremes, vibration, and a decade and a half of service. The engineering is demanding, but it is not speculative. It is the kind of problem the industry knows how to solve once the volume justifies the investment.
There is a hidden implication in the 200-gigabyte figure that the call did not spell out. A single system-on-chip cannot usefully address 200 gigabytes of memory; that number implies a central computing architecture with multiple domain controllers, and it implies that bandwidth — not merely capacity — is the constraint. LPDDR5X exists precisely because moving data quickly matters as much as storing it. The vehicles of the late 2020s will be judged on how fast they can feed their own minds, and the winners in memory will be those who understand that the edge demands a three-dimensional specification: capacity, bandwidth, and reliability, held in tension. A storage provider that optimizes for only one of those dimensions will find itself selling a commodity into a market that has stopped buying commodities.
To understand why Micron tells this particular story, one must understand the geography of its strengths. In high-bandwidth memory, the stacked DRAM that feeds AI accelerators, Micron is the third of three players. SK Hynix leads with more than half the market; Samsung follows; Micron trails, by roughly half a year to a year in the current generation and, by most accounts, into the next. High-bandwidth memory is a packaging race — through-silicon vias, advanced interposers, the willingness to co-design with foundry partners — and it is a race Micron is running from behind, against competitors who began the climb earlier and have held the summit longer.
In automotive memory, the picture inverts. Micron holds more than forty percent of the automotive memory market, a position of genuine dominance. Its competitors there are not only the Korean giants but also the established automotive semiconductor firms, and none of them commands the share that Micron does. This is the geography that explains the narrative: a company that is a challenger in the data center and a champion at the edge will naturally tell a story about the edge. There is nothing cynical about this. It is simply how strategy works, and reading it correctly is the difference between being informed and being persuaded. A chief executive does not choose which future to describe at random. He chooses the one in which his company is strongest, and the honesty of the story is not diminished by the fact that it happens to flatter the teller.
The same logic applies to the customer base. Data center memory is sold to a small number of very large buyers, which gives those buyers considerable leverage and exposes the seller to their capital-spending cycles. Automotive memory is sold through a long chain of tier-one suppliers and original equipment manufacturers, which spreads both risk and pricing power. A portfolio tilted toward the edge is, all else equal, a portfolio with less concentration risk and a smoother earnings profile. When Micron talks about Physical AI, it is talking about smoothing a notoriously jagged revenue curve, and smoothing a revenue curve is, for a cyclical company, worth more than almost any single product.
There is a dimension of this story that has almost nothing to do with silicon and everything to do with borders. Advanced memory — high-bandwidth DRAM, high-performance chips — is subject to export controls that have progressively closed the Chinese market to American suppliers. Micron is an American manufacturer, which places it, unusually, on the protected side of the fence. It receives substantial support under the CHIPS Act for domestic fabrication, and its competitors face restrictions that it does not. The net effect of the past several years of policy on Micron has been, if not positive, at least not negative. In an industry where a single regulatory decision can erase a market, that asymmetry is worth a great deal.
This is where the Physical AI narrative reveals a second layer. Automotive and robotics memory is largely immune to the export-control regime, because its customers are concentrated in North America, Europe, and Japan. A growth story built on the edge is a growth story that does not depend on access to a market that is closing. That is not an accident. When a chief executive chooses which future to describe on an earnings call, he is also choosing which risks to describe, and the edge is a future with fewer of them. The narrative is, in this sense, a form of de-risking dressed as a forecast — a way of telling investors that the growth he is promising will not be interrupted by the politics he cannot control.
The countervailing force is China itself, where domestic manufacturers are advancing quickly in both DRAM and NAND. They are unlikely to threaten Micron's automotive position soon, because the certification gauntlet takes years to run and reputations in automotive are built on decades of reliability rather than on a single competitive product. But in the mature segments, where price is the only language and reliability is assumed rather than rewarded, they are already a pressure. The high-end moat holds. The low end erodes. Both are true at once, and a careful reader holds both without needing to resolve them into a single verdict.
I have spent enough of my career around capital formation to recognize a certain kind of sentence. When a company that the market treats as cyclical begins to describe itself in the language of secular growth, it is not only forecasting demand. It is managing its multiple. Memory manufacturers have historically been valued with a discount precisely because their earnings swing violently with the price of a commodity. If Physical AI can persuade investors that a durable, growing, less cyclical demand curve now sits beneath the business, then the discount can narrow, and the same earnings can support a higher price. This is not deception. It is the ordinary work of a chief executive speaking to the people who own his shares, and it is the reason earnings calls are the most carefully composed documents in finance.
But it matters for the reader, because it means the narrative should be judged on two separate axes. On the first axis — does the underlying demand exist? — the answer is almost certainly yes. Autonomous vehicles and robots will need more memory and storage, and the need will be structural rather than cyclical. On the second axis — when does that demand arrive, and how much of it will Micron capture? — the answer is genuinely uncertain, and the uncertainty is the part the narrative softens. A reader who hears only the first axis will mistake a direction for a schedule, and a direction is not a schedule.
There is a quieter technical thread here that I find more interesting than the headline numbers. The 200-gigabyte figure may not have to be met with soldered memory alone. Compute Express Link, an interconnect standard that allows memory to be pooled and shared across a system, is maturing, and it offers a path to large memory footprints that does not require every byte to be resident on the board. If automotive architectures adopt pooled memory — and there are good reasons to think they might, from cost to upgradability to the ability to add memory after a vehicle has shipped — then the demand curve looks different from the one a simple capacity model predicts. The edge may not demand more memory so much as it demands memory that can be addressed differently, and that distinction is where the next decade of architecture will be fought. I would not build a thesis on a single number. I would build it on a direction, and the direction is toward flexibility.
Here the blockchain world and the semiconductor world, which so often talk past each other, are converging on the same question. An edge device that perceives and acts is a device that must be trusted — and trust, at scale, is expensive. The reason I spent years auditing token contracts is that I wanted to know what a system actually enforces, as opposed to what it claims. The same discipline applies to a robot. If a machine makes a decision that harms someone, we will want to know why, and answering that question requires an auditable record of what the machine saw, what it believed, and what it did. That record must live somewhere. Increasingly, I suspect, it will live in the very storage Micron is describing — terabytes of it, retained for years, in a form that can be verified rather than merely asserted.
This is the human story inside the technical one, and it is the story I care about. Storage is not a commodity when it holds the evidence of a decision. The automotive moat is built on fifteen-year supply commitments, and those commitments exist because a car is a promise that outlives the transaction that created it. The blockchain community learned the same lesson the hard way, in code that could not be changed and governance that could. Preserving the human story in digital ledgers is not sentiment. It is a specification, and specifications are where values become enforceable.
Everything above is an argument that Micron's direction is correct. Now let me argue that its timing may not be, because the value of an analysis is not in agreeing with the thing it examines.
The narrative rests on two demand pillars, and both are dated in the future tense. The first is Level 4 autonomy, which has been "three to five years away" for most of my adult life. The obstacles are not primarily technical — they are regulatory, economic, and human. A city must decide that it trusts a machine to drive, an insurer must price the risk, and a public must accept the consequences of the inevitable failures. Each of those is a slow, political process, and each can delay the storage demand by years without any change in the underlying engineering. The silicon may be ready long before the society is, and it is the society that decides when the demand arrives.
The second pillar is humanoid robotics, and here the uncertainty is even larger. In 2024 and 2025, robots are a demonstration, not a product. Whether they become a mass market by 2030 depends on assumptions — about cost curves, about use cases, about whether a general-purpose humanoid is even the right form factor — that no one can responsibly forecast. The market for robot memory could be enormous, or it could be a rounding error, and the honest answer is that we do not know. I watched a version of this in 2021, when I helped ten Kenyan artists launch a collection that sold out in forty-eight hours and then watched the community that had formed around it dissolve when the speculation moved on. The technology was real. The demand was real. The durability was the thing that had been assumed rather than tested, and durability is the only part that matters.
So when a chief executive places a demand inflection at the end of the decade, the honest translation is this: this is a story that will influence the valuation for the next three to five years and the earnings only after that. A narrative that is priced today against earnings that arrive in 2030 is not a forecast; it is a wager, and the house always knows the odds better than the player. The reader who wants to profit from Physical AI must decide which of those two clocks he is actually reading. Walking away from the hype to find the soul does not mean rejecting the technology. It means refusing to confuse the two clocks, and refusing to let a distant promise substitute for a present analysis.
I will not tell you what to buy. I will tell you what I will watch, because watching is the only honest form of prediction I have learned.
I will watch whether the storage configurations in the first commercially deployed Level 4 vehicles actually approach the 200-gigabyte figure Micron describes, or whether clever architectures and pooled memory bring the number down. I will watch whether the automotive certification pipeline — the slow, patient gauntlet of fifteen-year commitments — becomes a bottleneck that protects incumbents or a cost that drives manufacturers toward cheaper alternatives. I will watch whether the humanoid robot becomes a product before it becomes a memory of a trend. And I will watch, most of all, whether the industry remembers that the machines it is building will make decisions about human lives, and that the storage which records those decisions is not merely a component but a form of stewardship.
Grace asked the engineer about a conscience. The answer, it turns out, is not a chip. It is a combination of law, design, and the willingness to be accountable for what a system does. Storage is where that accountability is kept, and the company that keeps it well will earn something more durable than a higher multiple. Building libraries where others build empires has always been the harder path, and it is the one worth walking — even at the edge of the world, even on a timeline no earnings call can promise.