Micron says memory supply stays tight through 2028 as HBM eats DRAM capacity

Memory is not going to get easier to buy. The people selling it say so.

Micron CEO Sanjay Mehrotra told investors on 30 September that demand for the company's memory will exceed available supply for at least the next couple of years, and that 75 percent of Micron's memory output for 2027 is already spoken for. Most of the company's current sales discussions concern 2028, he said, according to a transcript of the call cited by Ars Technica. "Overall, supply-demand environment is only getting tighter," Mehrotra said. Micron plans to open new clean rooms for memory manufacturing in 2028, but warns that output ramps gradually even after first wafer starts.

Micron no longer sells consumer RAM. Its Crucial brand stopped selling memory to consumers after a December 2025 announcement, with EVP Sumit Sadana pointing to AI demand for data center memory and Micron's interest in serving "larger, strategic customers in faster-growing segments." That decision matters because the same wafer capacity that would have gone into a consumer DIMM is now being allocated to high-bandwidth memory for AI accelerators and server DRAM.

HBM takes a bigger share of every wafer

Kim Taewoo, an executive vice president at Samsung, said this week that HBM will account for almost 30 percent of DRAM manufacturers' wafer capacity in 2027, up from 20 percent this year, Reuters reported. HBM is more expensive to make and yields less usable product per wafer than conventional DRAM, so every percentage point shifted toward HBM removes more than a percentage point of supply from everything else. Micron says demand for HBM is now surpassing demand for its DRAM, and that the 2027 HBM volume is largely sold out at prices "much higher than 2026 prices."

Mehrotra also flagged a structural problem that will not be fixed by building more fabs: future node transitions deliver less productivity gain per wafer than previous ones, and the trade ratio between HBM generations, moving from 3E toward a greater mix of 4 and 4E, creates headwinds for supply growth. In other words, the usual escape route from a shortage, denser process technology, is narrowing.

The Register reported on 1 October that Micron's CEO used the same call to celebrate "much higher" prices, a detail that reframes the shortage as a deliberate allocation decision rather than an accident. India Today reported on 1 October that smartphones may get more expensive as the supply worsens, and Ars Technica notes that prebuilt PCs have already been affected, with OEMs offering lower RAM configurations at higher prices. Streaming sticks and gaming consoles have also seen price increases.

"Even after they are built, even after first wafer output, production ramps up only gradually in the clean rooms," Mehrotra said. "That's just the nature of what it takes to bring up production."

Beyond the data center

SemiEngineering reported on 1 October that automakers are racing to add AI capabilities to vehicles, a market it describes as a $1 trillion-plus opportunity for physical AI over the next decade, and that moving raw camera and radar data puts new demands on memory, interfaces and power efficiency. Paul Karazuba, vice president of product marketing for Silicon IP at Rambus, told the publication that an AI-defined vehicle uses AI to continuously interpret data and adapt behavior in real time, which requires more memory bandwidth than a software-defined vehicle.

Europe's space sector is running into the same wall from a different direction. At the Pretzl Connect 2026 press event in Budapest, Kate Underhill, future space transportation propulsion architect at the European Space Agency, told EE Times that space systems typically operate "at least 10 years behind consumer electronics" and that ESA once tried to order 20 laser diodes from a German supplier that required a minimum order of 10,000 units. Underhill said that if a satellite contains any U.S. component, the developer must comply with U.S. International Traffic in Arms Regulations, which is why European manufacturers are trying to build ITAR-free satellites.

Meanwhile, the supply chain rules governing what can be installed are tightening. SemiEngineering reported on 1 October that President Trump signed an executive order on 26 August 2026 declaring a national emergency to restrict high-risk foreign-produced equipment in the U.S. electric grid, covering hardware, software and firmware at 69 kilovolts and above. The order follows the supply chain downward, so a U.S. or European manufacturer sourcing critical sub-components from a covered jurisdiction may still find its finished product restricted. Implementing rules from the Department of Energy are due 24 December 2026.

Not everyone is waiting for the rules. U.S. utilities are already pausing procurement with foreign-linked supply chains and pressing vendors for provenance answers during active bids, according to the same report.

Security and privacy layers

OpenSSL released fixes on 29 September for a high-severity DTLS flaw tracked as CVE-2026-84782 that can leak heap memory to the other side of a connection or crash the program, The Hacker News reported on 30 September. The bug occurs when a DTLS resend timer fires while a larger handshake message is only partly sent, causing the resent message to carry leftover bytes as unencrypted handshake data. CISA scored it 8.2 out of 10 on 29 September, rating availability impact as high and confidentiality impact as low, and listing exploitation as "none" at that time. Fixed versions are 4.0.3, 3.6.5, 3.5.9 and 3.4.8; the older 3.0, 1.1.1 and 1.0.2 branches are patched only for premium support customers, and OpenSSL 3.0 stopped receiving public security fixes on 7 September.

On the AI infrastructure side, Google DeepMind published a technical update on 23 September describing how it will bring persistent, server-side memory to its Private AI Compute platform. The design keeps cryptographic keys on the user's devices while data is decrypted only inside an isolated cloud enclave, then re-encrypted immediately. DeepMind says the previous architecture was strictly stateless and wiped context when a task ended.

Two arXiv papers submitted on 28 September show the same constraint showing up in research. One, from Krishnakumar Balasubramanian and Zhaoyang Shi, models associative memory on Riemannian manifolds and shows that curvature is a design variable rather than a property of the data. The other, from Peng Xu and co-authors at the University of Illinois, presents a GPU clustering method that stores only a single factor array in high-bandwidth memory instead of three large buffers, an explicit response to memory-bound workloads.

The common thread is that memory is no longer a component you buy at the end of a design. It is the constraint that determines what gets built, what gets shipped, and what gets left out.