Every AI model you’ve ever used, from the large language models answering your questions to the systems generating images and writing code, needs one thing that rarely gets talked about: memory. Specifically, a type called High Bandwidth Memory, or HBM. It’s the component inside AI accelerator chips, like those made by NVIDIA Corporation (NASDAQ: NVDA), that determines how fast data can flow between a processor and its memory. Think of it less like the storage drive on your laptop and more like the working space on your desk. The bigger and faster that space is, the more complex the work you can do at once. And right now, there is nowhere near enough of it.
Only three companies on earth make HBM at scale: South Korean giants Samsung Electronics (KRX: 005930.KS) and SK Hynix (KRX: 000660.KS), and one American company. That company is Micron Technology (NASDAQ: MU), headquartered in Boise, Idaho, and it just reported the most profitable quarter in its 48-year history.
The numbers are hard to put in context without stopping to read them twice. Revenue for fiscal Q3 2026, the quarter ending May 28, 2026, came in at $41.46 billion, up from $9.30 billion in the same period a year earlier. That is not a seasonal bump or a one-time contract. Non-GAAP diluted earnings per share hit $25.11, well above the analyst consensus of $20.28, and gross margins rose to 84.6% from 37.7% a year ago. For context, most manufacturing businesses consider 30% gross margins healthy. Micron is now keeping 85 cents of every revenue dollar before operating costs.
The driver behind all of it is the AI infrastructure buildout. Micron’s CEO Sanjay Mehrotra disclosed that the company can currently fulfill only between half and two-thirds of customer demand for HBM, and the company’s entire 2026 HBM supply is sold out under multi-year contracts. It has collected $22 billion in customer cash deposits, essentially prepayments from hyperscalers desperate to lock in supply. Those deposits come from the largest cloud and data center operators in the world, companies that have collectively committed over $725 billion in AI infrastructure spending for 2026.
The HBM market itself is still dominated by SK Hynix, which commands roughly 50 to 62% of global HBM market share. Samsung holds the next largest slice. Micron overtook Samsung on some allocations in 2025, representing one of the bigger competitive surprises of this memory cycle. Its next-generation product, HBM4, is already in volume shipment and is ramping at twice the speed of the previous HBM3E generation, with over $1 billion in HBM4 revenue already shipped. That matters because the fastest ramp wins supply contracts, and supply contracts right now translate almost directly into profit.
There is a second story running alongside the financial one, and it has more to do with geography than quarterly results. Micron is the only U.S.-based manufacturer of advanced memory chips. Samsung and SK Hynix build their chips in South Korea, and the world’s most advanced chip foundry, TSMC, is in Taiwan. Micron has announced plans to invest $200 billion in U.S. semiconductor manufacturing and research and development, spanning facilities in Boise, Idaho, Clay, New York, and Manassas, Virginia, with support from the CHIPS and Science Act. The federal government has specifically flagged Micron’s domestic status as a national security consideration, given that the U.S. currently produces only around 10% of the chips it consumes.
That combination of policy tailwinds and structurally tight supply is clearly reflected in investor sentiment. Micron shares surged more than 19% following the earnings release. Analysts quickly responded, with firms such as JPMorgan issuing post-earnings notes that included a price target of $1,540. The company’s forward guidance was equally assertive, projecting fiscal Q4 revenue of $50 billion, plus or minus $1 billion, compared to just over $11 billion in the same quarter a year earlier.
The company expects the total addressable market for HBM to grow at a compound annual rate of roughly 40% through 2028, rising from approximately $35 billion in 2025 to around $100 billion. Supply constraints, the company says, will persist into at least 2028, driven by the physical complexity of building new fabs and the manufacturing demands of HBM itself, which requires significantly more cleanroom space per chip than conventional memory. What that means for the broader AI buildout is straightforward: the companies spending hundreds of billions on data centers are all competing for a resource that one American company, for now, cannot make fast enough to meet the demand.
