MU · Forward model · Cloud Memory

Assumptions & reasoning

• Modelled on capacity rather than growth because that is Micron's actual constraint: HBM is sold out ahead of production, so what ships is what the fabs and the packaging lines can make. The 10-Q says it plainly — demand is growing 'at a rate greater than our ability and the industry's ability to increase supply'.
• The petabyte figure is NOT disclosed. It is $13,769M divided by an assumed $6.04 per gigabyte, roughly a 2x premium to commodity DRAM in the same quarter. Micron gives bit-growth commentary in prepared remarks but no volumes in the filings, so treat the level as a calibration and argue with the price instead.
• Price rises 9% a quarter and glides to -4%, rather than falling from the first quarter as this model used to assume. Industry reporting through August 2026 has 2027 DRAM and HBM capacity at Micron, Samsung and SK Hynix already sold out under multi-year contracts, and Meritz put supplier fulfilment at only 75-80% of demand in the second half of 2026, possibly falling into the 60s during 2027. New capacity is scheduled but late: SK Hynix's first Yongin cleanroom opens in February 2027, its second fab not until 2029, and Samsung's first Yongin fab starts up in 2029.
• The upside is capped as well as floored, and by the same disclosure. Micron's largest agreements carry a ceiling price approximating the CALENDAR Q2 2026 market price, which is roughly where price already is. So the contracted book cannot ride a further spike — only the uncontracted minority and mix can. That is why the drift decays fast rather than compounding at 9% for years.
• Bit supply still nearly triples over the horizon while price ends below where it starts, so revenue grows about 80% and Cloud Memory goes from a third of Micron to about 43% of it. That mix shift is the strongest claim in this model and it survives every scenario.
• Utilisation glides DOWN from 95% to 88%, which reads oddly for a sold-out product. It is the cycle: capacity brought on for a peak runs slack when the peak passes, and the last three memory downturns were all made worse by fabs that could not idle cheaply.
• The reversion is timed to that build schedule rather than to a view about prices. Additional wafer starts from the second half of 2027 through 2028 are what turn the cycle, and the same forecasters who call 2027 the tightest year on record put conventional DRAM into excess capacity from about 2029 — which is where this model has revenue rolling over. HBM itself is expected to stay tighter for longer, which is why Cloud Memory keeps the gentlest price decline of the four lines.