Micron says memory gets much tighter in 2027 and 2028
The squeeze that drove this year's AI buildout isn't easing — Micron's CEO says the next two years get worse. Plus a US open-weight challenger prepares its first release, and India debates whether to write its own AI rules.
Memory supply will be "only getting tighter" through 2028 — and most of next year is already sold. On its fiscal Q4 call, CEO Sanjay Mehrotra said Micron sees demand exceeding supply in calendar 2027 and 2028, with greater tightness than in 2026, and that the company has no line of sight to when the market returns to balance — customers should expect "much higher prices" than this year. The company isn't just forecasting scarcity; it's contractualizing it: more than 75% of fiscal 2027 shipment volume is already committed, talks with customers have moved on to 2028 volumes, and 26 strategic agreements (up from 16) covering over 35% of expected revenue through 2030 are structured as take-or-pay deals with roughly $150 billion in remaining performance obligations and $32 billion in customer deposits. Revenue tells the same story — $54.25 billion in the quarter against $11.3 billion a year ago — and COO Sumit Bhatia put the bottleneck precisely: in data centers, DRAM is the primary supply constraint, more so than logic chips or power, and 2027 HBM prices have already been significantly raised. This is the strongest evidence yet that the memory squeeze is structural rather than a cycle peak, and it lands on every inference bill, GPU allocation, and datacenter pro forma in the industry. We covered the price side of this in September — China's AI chip prices jump 50% as the memory shortage bites.
Reflection AI is preparing a US open-weight model to take on DeepSeek and Qwen. Axios reports the lab — founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, with more than one billion dollars in compute deals signed against Nebius and SpaceX — will ship the model "soon," initially expected to match the top Chinese open-weight models rather than the leading US closed ones, and has held briefings with interested parties in Washington. Nothing has dropped yet: no weights, no benchmarks, and the company declined to comment. The story matters less as a release than as a positioning: Washington wants a democratic-champion open-weight lab to exist, and Reflection is angling to be it.
India is weighing its own AI safety framework instead of copying the US pact. After the White House accord that Google, OpenAI, Anthropic, Meta, Nvidia and xAI signed on September 29, industry experts in New Delhi are arguing publicly that voluntary commitments aren't enough — Nasscom chairman Srikanth Velamakanni called self-regulation "necessary but woefully insufficient" and asked for binding transparency and incident-reporting rules, while others warn mandatory regulation could stifle a market that wants more than a hundred frontier labs. Ministers are already on record: IT minister Ashwini Vaishnaw has said a new law is needed because AI is nothing like the IT Act era, and MeitY officials say a dedicated framework is under consultation. Watch for whether this turns into a draft — for now it is debate, not law.
What to watch: whether Reflection's weights actually ship this month, and whether 2028 memory contracts start repricing GPU leases.
Is the memory squeeze now the real brake on AI buildouts — more than chips or power? Tell us in the comments.
