The Physical Layer Turn: AI Scarcity Moves to Memory and Megawatts

For three years the story of artificial intelligence was a story about software. The scarce thing was the idea — the architecture, the training recipe, the alignment technique, the person who knew how to make the loss curve bend. Capital chased researchers, and the market rewarded whoever could ship the next model. The physical world underneath all of it — the chips, the memory, the buildings, the power — was treated the way plumbing is treated in a discussion about architecture: necessary, assumed, and beneath mention. Then, in a single week in July 2026, the plumbing sent two invoices large enough that nobody could look past them.

On July 10, SK Hynix listed on the Nasdaq and raised roughly 26.5 billion dollars, the largest US listing ever by a foreign company and the second-largest share sale in American history, trailing only SpaceX. It did this not by making models or writing software but by making the high-bandwidth memory that sits, in stacked silicon towers, inches from every AI accelerator on earth. Four days earlier, on July 6, Anthropic — a company whose entire product is intelligence delivered as text — signed a 20-year lease worth about 19 billion dollars for a 401-megawatt data center in Hawesville, Kentucky, built on the carcass of a decommissioned aluminum smelter. One deal was about the memory that feeds the compute. The other was about the power that runs it. Read together, they describe the same movement: the scarce, valuable, defensible layer of the AI stack is no longer the algorithm. It is the substrate. It is memory and megawatts. It is atoms.

This is what I want to call the physical layer turn — the moment the binding constraint on frontier AI stopped being the thing you could patent and started being the thing you had to fabricate, cool, and plug into a grid. It has been building for a year. This week it became impossible to ignore.

Two invoices, one story

Start with the numbers, because the numbers are the argument.

SK Hynix sold 177.9 million American depositary shares at 149 dollars each, raising about 26.5 billion dollars before fees. The shares closed their first day up roughly 12.8 percent, near 168 dollars, valuing the company at about 1.27 trillion dollars. This is a memory company. Its 2025 revenue was 97.1 trillion won, about 64.1 billion dollars, and it earned about 28.3 billion dollars in net income — a net margin near 44 percent, the kind of margin that used to belong to software and now belongs to the people who make the memory software runs on. It controls roughly 60 percent of the global market for high-bandwidth memory by revenue, the specialized stacked DRAM that sits beside the logic die inside almost every Nvidia accelerator and now commands scarcity pricing because there is not enough of it.

Now the other invoice. Anthropic committed to a 19-billion-dollar, 20-year lease with TeraWulf — a company that until recently was a bitcoin miner — for the "Justified Data" campus in Hawesville, Kentucky. The site is 401 megawatts, built on a former aluminum smelter, with initial capacity expected in the second half of 2027 and full capacity by early 2028. The lease is expected to generate about 19 billion dollars of contracted revenue over its initial term, backed by investment-grade credit, and it followed Anthropic's roughly 65-billion-dollar funding round. Concurrently, TeraWulf sold its majority stake in a Texas joint venture for 450 million dollars to recycle capital into wholly owned AI infrastructure. A bitcoin miner traded hashing for hosting, and an AI lab committed two decades and nineteen billion dollars to a building it will not own, on land that used to smelt metal.

Two companies, opposite ends of the stack, same week. One sold the world direct access to the memory bottleneck at a trillion-dollar valuation. The other locked up two decades of electricity because electricity, not talent, is what now gates its growth. Neither deal is about a model. Both are about the physical layer that models depend on — and both were priced as if that layer, not the model, is where the enduring value lives.

The constraint moved down the stack

To understand why this is a turn and not just two big deals, you have to see where the binding constraint has been over the last few years and watch it descend.

In 2023 and early 2024, the scarce input was talent and technique. Whoever had the researchers who could train a frontier model had the advantage, and capital flowed to teams. Through 2024 and into 2025, the constraint moved to raw compute — specifically to the supply of leading-edge accelerators. The phrase "GPU-rich versus GPU-poor" captured a real hierarchy: your ceiling was set by how many chips you could get. That was already a move from software toward hardware, but it was still about the logic die, the part that does the arithmetic.

What changed in 2026 is that the constraint moved past the logic die to the two things that surround it. The first is memory bandwidth. A modern accelerator is starved not for arithmetic but for data; it can multiply far faster than it can be fed, and the feeding is done by high-bandwidth memory. HBM is hard to make, supply is concentrated in a few hands, and demand is effectively unbounded, so it became the true bottleneck — the reason SK Hynix can earn a software margin on a physical product. The second is power. A frontier training or inference cluster is, at bottom, a machine for converting electricity into tokens, and the amount of electricity involved has grown past the point where you can simply order more. You have to find it, contract it years in advance, and sometimes resurrect a dead smelter to get it.

The chart is a schematic, not a measurement, but it captures the shape of the argument. Talent was the mountain everyone climbed in 2023; by 2026 it is table stakes. Compute peaked as the gating factor in 2024 and 2025. Memory and power — the substrate — are the lines still climbing. The value follows the constraint, and the constraint is now underneath the model, in the parts of the stack that have weight.

I have written before about the earlier phase of this descent, when the AI chip war stopped being about one general-purpose monolith and fractured into many bespoke inference chips — the logic-die stage of the same story. The physical layer turn is that same gravity pulling one level further down, past the processor to the memory that feeds it and the power that runs it. You can read that earlier stage in the analysis of the inference-silicon turn and the shift to purpose-built chips, and this piece as its sequel.

The memory wall becomes the moat

For most of computing history, memory was a commodity. It was interchangeable, brutally price-competitive, and the last thing anyone building a differentiated product wanted to be selling. The AI era inverted this. High-bandwidth memory is not a commodity; it is a scarce, technically demanding, capacity-constrained component whose supply is dominated by a small number of manufacturers, and it has become the single most reliable predictor of how fast an accelerator can actually run.

The reason is architectural. Training and inference are memory-bandwidth-bound workloads. The arithmetic units on a modern accelerator sit idle a large fraction of the time, waiting for weights and activations to arrive from memory. Adding more arithmetic does not help if you cannot feed it. HBM exists to feed it: stacks of DRAM dies bonded vertically, connected by tens of thousands of through-silicon vias, placed as close to the logic as physics allows, to move data at a bandwidth ordinary memory cannot approach. This is what the hero image of this article depicts — not a metaphor but the actual thing, the copper-veined silicon tower where the scarcity now lives.

Because HBM is hard and concentrated, whoever makes it captures the scarcity premium. SK Hynix makes most of it, which is why a memory company earns a 44 percent margin and lists at a trillion-dollar valuation on the strength of a component that goes inside someone else's chip. The market that used to reward the model now rewards the memory that makes the model runnable.

Concentration is the whole point. When three suppliers cover essentially the entire market for a component that every AI system needs and no AI system can substitute, those suppliers hold pricing power that looks nothing like the old commodity-DRAM business. A shortage in HBM is not a temporary imbalance to be competed away; it is a structural feature of a market where demand grew faster than anyone could build fabs, and where building a fab takes years and tens of billions of dollars. That is why SK Hynix could go to US public markets and raise the largest sum a foreign company ever has: it is selling investors direct exposure to the tightest choke point in the AI supply chain.

This is also the deeper meaning of the enterprise reclassification I have written about, where companies stopped treating AI as an experimental line item and moved it onto the balance sheet as permanent core infrastructure. When AI becomes infrastructure, the money follows the physical layer, because infrastructure is a physical-layer word. You do not reclassify a clever prompt as core infrastructure. You reclassify the memory, the silicon, and the power.

The power wall and the resurrection of dead industry

If memory is the first wall, power is the second, and it is the one that turns software companies into landlords of the grid.

A 401-megawatt data center is a serious piece of industrial infrastructure. To put it in human terms, that is enough continuous electricity to run a mid-sized city, dedicated entirely to converting electrons into model outputs. Anthropic did not lease this capacity because it wanted to be in the real estate business. It leased it because power has become the hard limit on how much intelligence it can produce, and because you cannot acquire that much power on short notice. You contract it for twenty years, and you take it where you can find it — which increasingly means places that already had the electrical infrastructure for heavy industry and then lost the industry.

Hawesville, Kentucky is exactly that kind of place. The site was an aluminum smelter. Aluminum smelting is one of the most electricity-intensive industrial processes in existence, which means the site came with grid interconnection, substations, and a power envelope that would take a decade and a war with local permitting to build from scratch. When the smelter closed, the town lost its anchor employer and its reason to exist on the map. Now the same power lines that once fed molten metal will feed a building full of accelerators. The megawatts outlived the metal.

Notice who the intermediary is. TeraWulf was a bitcoin miner, and it is not alone — the same week saw Marathon acquire a two-gigawatt Texas site for AI and bitcoin expansion, and neocloud and miner stocks moved together on the theme. Bitcoin miners spent years becoming experts at one specific thing: siting enormous, power-hungry compute next to cheap, abundant electricity and running it around the clock. That is precisely the skill AI now needs at scale. The miners had the sites, the power contracts, and the operational muscle; the AI labs had the demand and the credit. The pivot from mining to AI hosting is not opportunism. It is the same underlying business — industrial-scale compute married to industrial-scale power — pointed at a more valuable output.

The map replaces the org chart

When the scarce input was talent, the relevant geography was a handful of neighborhoods — the places where the researchers clustered, which is to say a few square miles of the Bay Area and a short list of university towns. You could draw the map of who mattered in AI on the back of a napkin, and it was a map of buildings full of people. The physical layer turn redraws that map, and the new one is much larger, much stranger, and organized around resources rather than résumés.

Consider what SK Hynix's listing actually represents geographically. The most strategically important memory in the world is fabricated in a small number of plants in South Korea, by a company that just raised its capital in New York. That single sentence contains most of the geopolitics of AI: the fabrication is concentrated in one country, the demand is concentrated in another, and the capital now flows across the Pacific to bind them. When a Korean memory maker becomes the largest foreign company ever to list on US markets, it is because the US capital base wants direct ownership of the choke point, and the Korean manufacturer wants access to that capital base to fund the fabs that widen the choke point. The listing is not a financial event so much as a treaty.

Now consider Hawesville. The new AI map has a node in a small Kentucky town most of the technology industry could not have found on a map two years ago, for the simple reason that the town had something the industry now needs more than it needs another engineer: an electrical envelope. The geography of AI is becoming a geography of stranded power — decommissioned smelters, retired coal plants, hydroelectric dams near depopulated towns, anywhere a previous era of heavy industry left behind the grid connections and the permits that a data center would otherwise spend a decade trying to acquire. The org chart told you where the talent was. The map tells you where the power is, and the map is winning.

The reason to lay these events on one timeline is that they are usually reported as unrelated — a memory IPO here, a data center lease there, a model release over there — when they are in fact one story observed at different points. The memory IPO prices the scarcity of the substrate. The power lease secures the scarcity of the substrate. The model flood proves that the model is no longer the scarce thing. The map, not the org chart, is where the action is.

Why value follows the constraint

There is an old lesson in economics that explains this cleanly: in any production chain, profit concentrates at the stage that is scarcest and hardest to substitute. When compute logic was the bottleneck, Nvidia captured the profit. As the bottleneck moves to the memory that feeds the logic and the power that runs it, profit and strategic value move there too. This is not a metaphor about importance; it is a statement about pricing power. The scarce, non-substitutable stage sets its own price. The abundant stages compete theirs down to cost.

Frontier models are, increasingly, an abundant stage. This is the uncomfortable truth underneath a week when three or four labs each shipped a competitive frontier model within days of each other. When capable models arrive on a weekly cadence from multiple vendors, the model stops being scarce. What remains scarce is everything required to serve one at scale: the memory to hold and feed it, the accelerators to run it, and the power to keep those accelerators fed for years. The model is the abundant stage; the substrate is the scarce one; and value, as always, follows scarcity down to where it lives.

This allocation is directional rather than audited, but the direction is the argument. A growing share of the capital going into frontier AI is being spent on the physical layer — the building, the power, the memory, the packaging — and a shrinking share, proportionally, is the thing everyone once thought of as the product. When two-thirds or more of a dollar of AI capex lands on atoms, the market that prices those atoms is the market that matters.

That repricing is the same force I traced in the efficiency turn, where enterprises stopped paying for maximal token consumption and started paying for value delivered. Efficiency is a physical-layer discipline. It is about squeezing more useful output from each watt and each byte of bandwidth, which only matters because watts and bandwidth are the scarce things. You do not optimize what is abundant.

The twenty-year lease is a thesis about time

Look again at the shape of the Anthropic deal, because its structure carries a claim that its headline number does not.

Twenty years is a very long time in artificial intelligence. It is longer than the entire history of deep learning as a commercial force. For a company whose core technology is reinvented on something like an eighteen-month cycle to commit to a specific building, in a specific town, for two decades, is to make a bet not about which model will win but about the durability of the physical demand underneath all models. The lease says: whatever the model looks like in 2038, it will need enormous quantities of power, and I would rather own the right to that power now than compete for it later.

The 20-year lease and the trillion-dollar memory IPO are two expressions of the same conviction: that the physical demand for AI infrastructure is durable enough to underwrite very long-dated, very large capital commitments. Investors bought SK Hynix because they believe HBM scarcity persists. Anthropic signed with TeraWulf because it believes power scarcity persists. Both are betting that the substrate outlasts any particular model — that atoms are a longer-lived asset than algorithms. On the evidence of this week, the capital markets agree.

There is a financial-engineering wrinkle worth naming. Anthropic is not buying the building; it is leasing it, which keeps the asset off its own balance sheet and puts the construction risk and capital burden on TeraWulf, backed by investment-grade credit. This is how software companies become the anchor tenants of the physical world without becoming, on paper, physical-world companies. The lab supplies the demand and the credit; the specialist supplies the atoms. It is a division of labor, and it tells you that even the companies driving the physical layer turn would prefer, if they could, to stay asset-light. They cannot. The scarcity is too binding to leave to the spot market, so they sign twenty-year leases and let someone else hold the concrete.

What actually breaks

If this analysis is right, the failure modes of AI shift with it. The things that can go wrong are no longer primarily about models. They are about atoms, and atoms fail in slow, physical, hard-to-reverse ways.

The first failure mode is fab capacity. HBM scarcity is a manufacturing problem, and manufacturing capacity takes years to add. If demand keeps climbing faster than fabs come online, the shortage does not clear; it intensifies, and the memory premium widens. That is good for SK Hynix and bad for everyone who has to buy from it, which is everyone.

The second failure mode is the grid. A 401-megawatt facility is a large load, and the sum of all such facilities is a load large enough to strain regional grids, provoke local opposition, and collide with the slow, contested politics of building transmission. The reason AI is resurrecting dead smelters is precisely that new grid capacity is so hard to create; the pre-built envelopes are a workaround for a bottleneck that the workaround does not solve. At some point the supply of convenient dead industry runs out, and the industry has to build new power, which is where it meets the hardest constraint of all — the physical and political difficulty of adding electricity to a grid.

The third failure mode is concentration risk. When one company makes 60 percent of the world's HBM, that company is a single point of failure for the entire AI industry. A fire, an earthquake, an export restriction, or a geopolitical shock that touches its fabs is no longer a memory-market story; it is an everything story. The same concentration that gives SK Hynix its pricing power gives the whole system a fragility it did not have when the scarce thing was a widely-distributed idea rather than a narrowly-sourced physical component.

What it means if you build

For most people reading this, the practical question is not whether to buy SK Hynix stock or lease a data center. It is what the physical layer turn means for the way you build on top of AI. A few things follow.

First, cost structure is now a hardware story, and it is not going to fall the way software costs fall. The reflexive assumption of the last few years — that inference gets cheaper on a smooth curve forever — collides with a physical layer where the key inputs are supply-constrained. Memory and power do not obey Moore's-law intuitions. They obey fab schedules and grid politics. Plan your economics on the assumption that the substrate has a floor, and that the floor is set by scarcity you do not control.

Second, efficiency is now a competitive advantage rather than a nicety. When the scarce input was a clever model, the winning move was to have the cleverest model. When the scarce input is memory bandwidth and power, the winning move is to extract the most useful work from each unit of both. The teams that win the next phase are the ones that treat tokens, watts, and bytes of bandwidth as the precious, metered resources they have become — which is exactly the discipline the efficiency turn rewarded.

Third, dependency is the risk to watch. If your product depends on frontier inference, then your product depends, transitively, on a memory supply chain dominated by a few firms and a power supply chain contended for by the largest companies in the world. That dependency was always there; the physical layer turn just made it visible and priced it. Know whose atoms you are renting, and know what happens to you if they get more expensive or less available — because on the evidence of this week, the people who own the atoms are the ones with the pricing power.

A new tier appears between the lab and the grid

One of the quieter consequences of the physical layer turn is the emergence of an entire class of company that did not meaningfully exist two years ago: the neocloud. These are the intermediaries that sit between the AI labs, which have the demand and the credit, and the physical world, which has the power and the land. TeraWulf is one. The bitcoin miners pivoting to AI hosting are others. So are the specialist developers building gigawatt-scale campuses on speculation, confident that a lab will sign a long lease before the concrete cures.

This tier exists because the AI labs have discovered they would rather not own the buildings. Owning a data center means carrying construction risk, real-estate risk, and a decade of capital on a balance sheet that investors want pointed at research and inference, not at concrete and switchgear. So a division of labor emerged. The neocloud raises capital against a signed lease from an investment-grade tenant, builds the campus, and operates the physical layer. The lab supplies the demand and the creditworthiness that makes the whole structure financeable. Everyone gets to specialize, and the labs get to stay, on paper, the asset-light software companies they prefer to be — while still commanding two decades of dedicated power.

The bitcoin miners are the most interesting entrants because their pivot is so natural it barely counts as a pivot. For years, mining was a business of one skill executed relentlessly: put enormous, power-hungry compute next to the cheapest electricity you can find, and run it every hour of every day. Everything a miner learned — how to site near stranded power, how to negotiate multi-year energy contracts, how to build and cool dense compute at industrial scale, how to operate it around the clock — transfers directly to AI hosting. The only thing that changed is the workload running on the racks and the value of its output. A miner converting a megawatt into hashes was making a few dollars. The same megawatt converting tokens for an AI lab is worth far more, which is why the capital markets rerated the miners the moment the leases started landing. The skill was always latent. AI just made it valuable.

What makes this tier strategically important is that it now sits directly on the critical path of every AI product in the world. The lab you depend on depends on a neocloud you have never heard of, which depends on a power contract in a town you could not find on a map, which depends on an electrical envelope a dead industry left behind. The stack got taller and more physical at the same time, and the new layers are the ones with the least redundancy and the longest lead times. That is worth knowing even if you never touch a data center, because it is now part of the dependency graph underneath your software.

The layer that was always underneath

The most honest way to describe the physical layer turn is that nothing new was created this week. The memory was always the bottleneck; the power was always the limit; the substrate was always underneath the model, holding it up. What changed is that the market finally looked down and priced what it saw. SK Hynix did not suddenly become essential — it was always essential, and the IPO simply let the public buy a piece of the essential thing. Power did not suddenly become scarce — it was always the ceiling, and the Anthropic lease simply made the ceiling visible by putting a twenty-year, nineteen-billion-dollar number on it.

For a while it was possible to talk about artificial intelligence as though it were made of ideas. It is made of ideas, but it runs on atoms, and the atoms have started to send invoices large enough to reorganize the industry around them. A memory company is worth a trillion dollars. An AI lab is a twenty-year tenant of a dead smelter. The scarce, valuable, defensible layer is the one with weight. That is the turn, and it is not turning back.

If you want the near-term, falsifiable version of this thesis — a specific, dated claim about how far the physical layer will capture AI capital — I have written it up as a standalone prediction on HBM and physical infrastructure taking the majority of AI hardware capex. For the reported detail on the two deals that anchor this piece, see the companion news analysis of the SK Hynix listing and the Anthropic power lease. And for the human version — what it is like to stand inside a smelter that has been reborn as a machine for making intelligence — there is a short story, The Smelter Remembers.