The earnings call on June 24 had the kind of numbers that used to belong to software companies, not chipmakers. Micron Technology reported revenue of $41.5 billion for its fiscal third quarter, up more than 300 percent from a year earlier, with a gross margin of about 85 percent, a level the memory industry has never seen. The company guided to roughly $50 billion of revenue in the current quarter. But the number that dominated the call was the one behind the results: Micron’s contracted backlog has passed $100 billion.
The backlog is the accumulation of 16 strategic customer agreements, 14 of them signed, under which customers have committed to buy at least $100 billion of memory at floor prices over the term of the contracts. The agreements carry cash deposits of around $18 billion, and most include take-or-pay provisions, meaning the hyperscale customers must pay for capacity even if they do not take delivery. For a business that has spent its history riding the most violent cycles in technology, the structure is a transformation: revenue that is contracted, priced and largely non-cancelable years into the future.
The driver is artificial intelligence, and specifically the memory that AI consumes. High-bandwidth memory, the vertically stacked chips that feed AI accelerators, has become the industry’s bottleneck, and Micron has shipped more than $1 billion of its latest generation, HBM4, which it says is ramping twice as fast as its predecessor. The company says it can meet only about half to two-thirds of customer demand for HBM, and its orders stretch into 2028. The entire 2026 HBM allocation is already sold at fixed prices.
The financial results have turned Micron into one of the market’s most dramatic stories. The stock has risen more than 600 percent over the past year, and the company was reclassified by FTSE Russell from a value stock to a pure growth name, effective June 29, a change that pushed it into the indexes where growth investors concentrate their buying. Barron’s, in a widely discussed column last month, called Micron the new Nvidia, arguing that the memory maker would outpace the chip giant as the AI trade’s center of gravity shifts.
The Barron’s thesis rests on a structural argument about how AI consumes resources. For the past three years, the constraint on AI has been compute, the chips that train and run models, and Nvidia has been the beneficiary. The argument for Micron is that the constraint is shifting to memory. AI inference, the phase where models answer questions and do work, is growing faster than training, and inference consumes memory at a different rate, one that analysts say will keep HBM and DRAM tight for years.
The numbers support the thesis in the current quarter. Micron’s data-center business generated revenue of more than $25 billion in the quarter, up fivefold from a year earlier, and the company’s cloud customers have signed contracts that lock in prices through the ramp. The take-or-pay structure means the revenue is not just forecast but owed. For a cyclical industry where forecasts have always been guesswork, the contracts are a new kind of certainty.
The risks are the mirror image of the opportunity. Memory remains a capital-intensive business, and Micron is spending about $10 billion a quarter on new capacity, building fabs in the United States, Taiwan, Singapore and Japan. If AI demand slows, or if model architectures become less memory-hungry, the contracted floor protects revenue but not margins, and the industry’s history suggests the excess capacity will eventually be built, priced and punished. Insider selling has also picked up to its fastest pace in a decade, a signal that some executives are taking profits near the top.
The valuation debate has become the market’s central argument about memory. Bears say the stock has priced in years of flawless execution, and that a memory cycle that has never lasted this long will eventually turn. Bulls say the structure of the business has changed, that contracted revenue and take-or-pay floors have turned Micron from a cyclical into something closer to a utility, and that the market has not fully accepted the change. Both sides agree the next few quarters of data will settle it.
The shift from training to inference is the argument’s engine. Training runs are enormous but finite, and they have dominated the industry’s attention for three years. Inference is smaller per task but continuous and growing, and it is spreading from data centers into phones, cars and industrial systems. Each new inference deployment consumes memory at a steady rate, and the analysts who see Micron as a structural winner argue that this consumption, not the flash of a training run, will define the next decade of demand.
The concentration of the backlog is worth noting. The 14 signed agreements are with a small group of hyperscale customers, and the take-or-pay floors that protect Micron’s revenue also concentrate its exposure: if one large customer’s build-out slows, the revenue is protected by contract, but the next round of orders will be negotiated in a different environment. The company’s ability to sign the next $100 billion, not the first, will determine whether the re-rating holds.
For the industry, Micron’s order book is the clearest evidence yet that memory has become the AI economy’s chokepoint. The narrative that began as a novel argument, that memory, not compute, would be the binding constraint, has become the consensus view of the analysts and customers who matter. The $100 billion backlog is the number they point to, and it is the number that will determine whether Micron is a cycle stock that got lucky or a franchise that changed its own rules.
