Micron reported fiscal third-quarter revenue of $41.5 billion on Tuesday, quadrupling from $9.4 billion a year earlier. Adjusted earnings hit $25 per share against expectations of $21. Gross margins reached 81 percent, up from 27 percent a year ago. For context, TSMC's peak quarterly gross margin is 66 percent. A commodity memory maker now earns wider margins than the world's most important foundry. Micron guided fourth-quarter revenue to $50 billion, plus or minus $1 billion, against analyst estimates of $44 billion. The stock rose 14.6 percent after hours.
The numbers are startling but not the story. The story is the structure underneath them.
The Contract
Micron has signed 16 long-term supply commitment agreements with hyperscalers and automakers. The agreements run three to five years. They are take-or-pay: the customer is obligated to purchase the committed volume at the agreed price each year. If the customer does not want the product, the customer still pays. Fourteen of the 16 agreements carry a cumulative revenue at minimum price of approximately $100 billion over their remaining term.
The agreements include a price band for each product. Every quarter, buyer and seller negotiate a price within the band. The price cannot exceed a ceiling, which is set at or near the current market level. The price cannot fall below a floor. Management described the floor as ensuring a "very robust gross margin" well above Micron's long-term financial targets. In its entire history as a public company, Micron's average gross margin is roughly 30 percent. The floor in the new contracts guarantees margins above the best quarters Micron has ever had.
The customers have also paid $22 billion in deposits and related financial commitments. $18 billion in cash. $4 billion in letters of credit. The deposits are not refundable in the ordinary course. They are structured as prepayments against future deliveries. The richest companies in the world have handed their memory supplier a $22 billion interest-free loan.
The Inversion
For forty years, the DRAM industry was the most cyclical business in technology. Demand surged, producers built capacity, supply overshot, prices collapsed, producers lost money, demand recovered, and the cycle repeated. Between 2018 and 2020, Micron's gross margin swung from 59 percent to 29 percent to 31 percent. In 2023, margins went negative. The memory cycle was the single most reliable pattern in semiconductors: build too much, suffer, repeat.
The take-or-pay agreements end the pattern. If high-bandwidth memory demand crashes, if a software breakthrough like Google's TurboQuant paper eliminates the need for massive HBM allocations, if the AI training buildout slows or reverses, the $100 billion in guaranteed minimum revenue still flows. The customers bear the downside. Micron does not. The cycle has not been tamed by better demand forecasting or disciplined capacity management. It has been contractually transferred from the supplier to the buyer.
CEO Sanjay Mehrotra said on the call that Micron can currently fulfill only between half and two-thirds of customer demand for HBM. The entire 2026 HBM supply is sold out. The next-generation HBM4 is ramping twice as fast as HBM3E and has already exceeded $1 billion in revenue. The scarcity is real, and it is the source of the leverage.
The Condition
The underwrite works because there are three suppliers. Samsung, SK Hynix, and Micron produce all of the world's HBM. No one else has demonstrated the ability to stack and bond memory dies at the density and yield rates required. If ten companies could make HBM, no hyperscaler would sign a five-year take-or-pay at price floors that guarantee above-peak margins. The oligopoly is the leverage. The contract is the mechanism. The customers signed because the alternative was no supply.
The arithmetic explains why the buyers accepted the terms. A single HBM chip costs roughly $100. It sits inside a GPU module that costs $30,000. That module sits in a server rack that costs $300,000. The rack generates millions in annual revenue from inference and training. The memory is less than one percent of the total system value but 100 percent of the bottleneck. Overpaying for guaranteed supply is rational when the alternative is an idle $300,000 server.
But the underwrite has the same failure mode as any insurance contract: it pays when the covered event does not occur. If demand stays strong, the price floors never bind and the agreements merely lock in access. If demand collapses, the floors activate and the customers discover they have committed $100 billion to memory they no longer need at prices they would never accept on the open market.
The memory cycle did not end because the underlying economics changed. It ended because the customers volunteered to absorb the downside. The question is not whether Micron's margins hold. The contracts guarantee they will. The question is what the customers do in year three of a five-year take-or-pay when the workload they built for can run on a quarter of the memory.