HBM4 Margins and the Collapsing Cost per Million Tokens

The HBM4 rent lasts one cycle

The HBM4 memory price premium disappears by 2028. Consensus prices it as a rent that holds until 2030. The difference is worth two years of margin, and the capex curve says who is right.

In the quarter ended 3 September 2026 Micron turned 69.5% of revenue into net income: $54.23 billion in sales, $43.75 billion in operating income, $37.7 billion in net income, as The Next Platform^[1] reported on 2 October 2026.

A memory maker with a software publisher's margin.

This is a regime change, and regime changes carry an expiry date. The line that sets it sits below the profit: $10.77 billion of capex against $73.45 billion of cash.

Consensus reads the margin, the answer sits in the capex

Consensus has the wrong frame. It reads the 69.5% as proof of a physical barrier: stacking DRAM and bonding it to a logic die remains a craft for a handful of plants worldwide.

The cash says something else. With $73.45 billion on hand, Micron has the money for almost four new fabs, and commits $10.77 billion of it. The brake sits upstream of physics: it sits in the choice of how much supply to bring to market.

The Next Platform puts it plainly: memory makers have learned from the OPEC countries with oil, and keep scarcity tight to avoid a bust after the boom.

That kind of discipline holds as long as the club stays at three and the premium stays enormous. An enormous premium calls in capacity: it is the script of every memory cycle, and DRAM history has repeated it for thirty years.

Three data points with a value, a date and a measure

One data point makes news. Three make a trajectory.

The new one comes from the quarter ended 3 September 2026: net margin at 69.5%, revenue at 4.8x the year before, net income at 12x. Micron's own accounts measure it, and TradingView^[2] logs the same numbers as record profits driven by AI demand.

The second comes from this desk's accounting class, the one that tracks advanced packaging: planned CoWoS capacity goes from 13,000 to 120,000 wafers per month. Nine times the capacity, and the premium on that bottleneck compressed while the capacity arrived.

The third measures the opposite end of the chain: inference cost per million tokens for a GPT-4-class model has dropped roughly a thousandfold in three and a half years, with the price of intelligence down about 40x a year in the same archive.

The curve says one clear thing. Downstream the price collapses, upstream the rent climbs: two opposite directions on the same supply chain last one cycle.

Cliff event: capacity lands all at once in 2028

Memory capacity arrives in steps, and that explains the price jump better than any demand model.

The Next Platform uses the pregnancy image for fabs: the work stays serial, from foundation to yield, and splitting it across nine teams shortens little. Equipment orders signed in 2026 become wafers in 2028, all together.

Cliff event: HBM capacity started in 2026 enters production as a block in 2028 → the premium per stack falls to double digits → makers' net margin drops back below 55%.

Three conditions make that date operational:

• lithography and test equipment orders signed by 2027
• qualification of a second HBM4 supplier at a large accelerator buyer
• a buyer splitting volume between two suppliers instead of rewarding one

The 69.5% margin is the best incentive ever offered to Samsung and SK Hynix to add lines. A premium that size pays for the chase two years in advance.

Three categories that change shape by 2029

Samsung and SK Hynix go from scarcity managers to share contenders. The first to break ranks wins the customer, and the club of three loses the pricing power that holds the premium up today.

Nvidia and AMD change the bill of materials of their accelerators. The Next Platform notes that GenAI rests on stacked DRAM from these three suppliers: when memory cost per board falls, the margin returns to whoever sells the system, or it ends up in the list price.

Inference vendors live the third transformation. A capacity contract signed today on 2026 memory cost becomes a competitive weight in 2028, when the rival buys the same stacks at half the price.

The position, and the data point that breaks it

This desk's position: the 69.5% measures a supply choice, and a choice changes within a quarter.

The counterargument deserves respect. HBM4 brings a logic base die, built on a foundry process and customized per customer: closer to a logic product than to catalogue DRAM. Qualification at a large buyer takes months and ties supply across product generations.

That reasoning holds for the single supplier already inside. The premium pays for qualification: with margins this size, opening a second and a third channel becomes the first line item in every large buyer's procurement plan.

What would change my mind: flat capex alongside margins above 65% for four straight quarters. That combination describes a structural rent, and my thesis falls.

What to do now, role by role

For a technology leader the work is architectural. A stack designed around scarce memory (huge batches, aggressive quantization, everything in the cloud) loses its point when the HBM stack costs half as much. Inference close to the data comes back into play.

For a fund the uncomfortable bet sits downstream, in vertical applications with proprietary data: they pay for the token and collect the collapse in its price.

For a strategy leader the test is blunt: a three-year plan that budgets expensive memory through 2030 assumes a world that ends sooner.

For technology buyers the rule stays one: short duration, renegotiable price, adjustment clause. Locking three years of supply today at peak price hands the seller the good part of the curve.

Forecast, horizon, kill signal

Forecast on the technology: high confidence, because capacity follows the premium. Forecast on market timing: medium confidence, because the date depends on the decisions of three boards.

Micron reports a quarterly net margin below 55% in a quarter whose accounts land by 31 December 2027. Confidence: 65 out of 100. Horizon: 454 days. Kill signal: four consecutive quarters with net margin at or above 65%, through the quarter closing in autumn 2027.

Market indicator: MU down within the horizon.

The interim evidence comes from the investment line, quarter by quarter. TheStreet^[3] frames Micron's memory boom as facing a new test tied to artificial intelligence: the test reads in the capex, and the answer arrives before 2030.