Micron just had the biggest quarter in its history. The number that matters is the one about wafers.

Start where the memory actually comes from. In a fabrication plant in Boise, or Hiroshima, or Taichung, a polished disc of silicon three hundred millimetres across enters a line of machines that will spend roughly three months turning it into something that can hold a thought. The disc is called a wafer. The plant can start only so many of them a week — a number fixed years ago, when the building was financed and the tools were ordered, and almost impossible to raise on short notice. Everything Micron told Wall Street on Wednesday evening, every record in a quarter that was nothing but records, rests on that fixed number of wafers and the brutal arithmetic of what the company chose to print on them.

The numbers are worth stating plainly, because they are the largest in the company's history. Micron reported $41.46 billion in revenue for its fiscal third quarter, up 346% from the same quarter a year earlier, when the figure was $9.3 billion. Net income was $28.24 billion. The adjusted gross margin — the share of each dollar of sales left after the cost of making the product — was 84.9%, a number memory companies historically dream about in the good years and never sustain. DRAM, the working memory in every server and phone, brought in $31.3 billion of the total; NAND flash storage accounted for $9.9 billion. Then the company guided to roughly $50 billion in revenue for the current quarter, against the $43 billion analysts had pencilled in. The stock rose as much as 9% in after-hours trading. Sanjay Mehrotra, Micron's chief executive, called it proof of "the strategic value of memory in the AI era."

The chokepoint is the wafer, not the chip

Here is the part the margin number hides. A memory fab does not have a knob marked "more." It has a fixed throughput of wafer starts, and every product it makes competes for the same silicon. The product driving these results — high-bandwidth memory, or HBM, the stacked DRAM that sits beside an AI accelerator and feeds it data — is extraordinarily hungry for wafers. By the industry's own rule of thumb, a bit of HBM consumes more than three times the wafer area of a bit of ordinary DRAM, because the dies are larger, stacked a dozen high, and yield worse for being so. That ratio is the whole story. Every wafer Micron commits to HBM for a data centre is a wafer it cannot commit to the commodity DRAM that goes into a laptop or a car. The memory did not get more expensive because demand rose in the abstract. It got more expensive because a finite number of wafers is being routed, deliberately, toward the customer who pays the most.

This is why the same boom shows up as a triumph on Micron's income statement and as a grievance on a PC maker's bill of materials. Earlier this month a colleague on this desk traced how Apple began describing memory price increases as "unavoidable"; the unavoidability is real, and it lives here, in the wafer-allocation decision, not in any single negotiation. When three companies — Micron, Samsung and SK hynix — control essentially all the world's DRAM, and all three tilt their lines toward HBM at once, the commodity market they leave behind tightens by construction. Goldman Sachs has put the 2026 supply-demand gap for DRAM at close to 5%, the most severe in fifteen years. A 5% shortage does not sound like much until you remember that memory pricing is violently sensitive at the margin: a few points of scarcity can double a contract price, which is precisely the mechanism printing Micron's 85% gross margin.

Twenty-two billion dollars, paid in advance, to hold a place in line

The most revealing figure in the release was not the revenue. It was a smaller one, in the commentary: under what Micron called strategic customer agreements, the company expects to receive about $22 billion in cash deposits and related financial commitments. Read that slowly. Micron's customers — the hyperscalers and accelerator makers who need HBM — are handing the supplier billions of dollars up front, before the chips exist, to guarantee they get them. In any other year this would be remarkable. In the logic of a chokepoint it is the most natural thing in the world. When the binding constraint is wafer capacity that cannot be expanded for years, money stops being the way you buy a product and becomes the way you buy a position in the queue.

When the constraint is wafer capacity that cannot be expanded for years, money stops buying the product and starts buying a place in line.

There is a precedent for this, and it is worth naming because it tells you where the memory business is heading. For a decade the leading-edge logic foundries — TSMC above all — have run on exactly this model: capacity is scarce, lead times are long, and the customers who want guaranteed supply prepay for it, sometimes years out. Memory was supposed to be different. It was the commodity corner of the industry, boom-and-bust, sold on the spot market, where no one prepaid for anything because there was always another fab spinning up to flood the market and crush the price. The $22 billion is the sound of that assumption breaking. HBM has turned a slice of the memory business into something that behaves like a foundry — allocated, contracted, prepaid — and the companies that make it are discovering they can charge accordingly.

The obvious objection is that capacity problems solve themselves: if memory is this profitable, build more fabs. Micron is trying. But a leading-edge memory fab costs on the order of fifteen to twenty billion dollars and takes years to equip and qualify, and the most constrained step — the advanced packaging where the stacks are assembled — has its own queue for its own tools, made by its own short list of suppliers somewhere far upstream. Micron spent $7.1 billion on capital projects in the quarter and signalled more to come, but a dollar of capital expenditure today is a wafer two or three years from now. In the meantime the company is selling everything it can make, which is why it told investors it intends to return essentially all of its excess cash to shareholders rather than chase the shortage with a crash build it might later regret. That sentence is the discipline the industry learned the hard way, and it is also a quiet admission that the bottleneck will not clear soon.

The moat is a packaging line, not a design

The product at the centre of all this is HBM4, the next generation, and Micron was specific about its progress: high-volume shipments to a lead customer's platform — Nvidia's, by every indication — with qualification samples out to others, and a twelve-high stack, twelve DRAM dies bonded into one package, ramping, the company said, twice as fast as the previous generation did. More than a billion dollars of HBM4 has already shipped. The detail that matters to anyone trying to understand the moat is the one about the stack. Bonding twelve thinned dies into a single part without warping them, drawing heat out of the middle of the sandwich, and keeping yield high enough to sell at a profit is a manufacturing problem of the first order — the kind that took the industry years to solve at eight high and is harder still at twelve. The barrier to entry in HBM is not the design. It is the packaging line, and there are perhaps three companies on earth that can run one at scale.

How much rides on how little

Which returns us to the wafer, and to the fragility a record quarter is built on. The entire AI build-out — the trillion dollars of data-centre capital, the accelerators, the bond sales and the IPOs — now passes through a memory supply chain held by three firms, fed by a fixed number of wafer starts, served by a packaging step only those firms can perform, and increasingly rationed by who paid in advance. None of this was designed as a bottleneck. It accreted the way chokepoints always do, one rational decision at a time: build fewer, larger fabs because they are more efficient; consolidate because memory's boom-and-bust punished everyone who didn't; specialise the most advanced capacity because that is where the margin is. Each step made sense. The sum is a world in which the speed of artificial intelligence is set, more than anyone planning it intended, by how many discs of silicon three companies can start each week, and what they decide to print on them.

Micron's guidance says it expects to start a great many, and to keep its lines tilted toward the customers paying the most for as long as they keep paying. That is a rational answer to the demand in front of it. It is also a bet that the demand is real and durable rather than a buildout running ahead of revenue — a bet the whole memory industry is now making with capacity that takes years to install and cannot be unbuilt if the call is wrong. For now the wafers go to the data centre, the deposits keep arriving, and the laptop buyer pays more for less. The chokepoint is doing exactly what chokepoints do. The only open question is who is standing on the wrong side of it when the cycle turns.