The OpenAI chief argues that scars from past overbuilding are keeping DRAM producers cautious, even as AI demand for high-bandwidth memory surges
The AI boom has a new bottleneck, and it isn’t GPUs this time. According to OpenAI CEO Sam Altman, it’s memory makers who remember exactly how badly the last party ended.
Speaking on a recent podcast, Altman said the major DRAM manufacturers are hesitant to build more fabrication plants. The reason is simple: past rounds of overbuilding nearly drove them into bankruptcy.
That caution persists even though, by Altman’s account, each of the big players is positioned to generate around $100 billion in annual free cash flow. When a company turns down that kind of opportunity, the trauma must run deep.
A windfall nobody wants to chase too hard
The companies in question are the familiar names of the memory world: Samsung Electronics, SK hynix, and Micron Technology. Their business has long been defined by brutal boom-and-bust cycles.
The current boom comes from high-bandwidth memory, or HBM. Think of it as the fast-access workspace AI chips rely on to keep data flowing while they crunch through models. Demand for it is tied closely to Nvidia accelerators and the hyperscale data centers being built to house them.
Altman’s remarks, dated October 29, 2026, paint a picture of an industry staring at enormous profits while keeping one hand firmly on the brake. His explanation centers on history.
He pointed to a past episode in which the sector poured roughly $20 billion into capital expenditure while chasing only about $19 billion in revenue. Spending more to build capacity than the market would actually pay for that capacity is not a sustainable strategy. Memory makers learned that the hard way.
According to Altman, the survivors of those cycles came close to bankruptcy. That memory, so to speak, now shapes how aggressively they’re willing to expand.
Why price signals aren’t doing their job
Altman said current price signals are not sufficient to prompt the supply increase the industry needs.
Part of the problem is physical. New fabs require years of construction and massive financial commitments before they produce a single chip. A memory maker that greenlights a plant today is betting on demand conditions several years from now.
Anthropic adds its own warning
Altman isn’t the only voice raising the issue. Anthropic’s IPO filing, dated October 26, 2026, flagged concerns about a potential shortfall in HBM supply from key suppliers.
That’s notable because risk disclosures in IPO paperwork are where companies list the things that could genuinely hurt their business. Seeing memory supply on that list suggests the concern extends well beyond one CEO’s podcast commentary.
Two of the most prominent AI labs are now pointing at the same constraint within days of each other.
Background: an industry built on caution
Memory chips have historically behaved more like commodities than most semiconductors. When supply outruns demand, prices can collapse quickly, and manufacturers have little ability to differentiate their way out of a glut.
AI has scrambled the usual math. HBM is more specialized than standard memory, and its demand is linked to a relatively small set of very large buyers.
What this means for AI and chip investors
For semiconductor investors, Altman’s comments describe a tension worth watching. On one side sits the prospect of roughly $100 billion in annual free cash flow per major memory maker, as Altman framed it. On the other is a deliberate reluctance to expand capacity at the pace demand might justify.
For the AI industry, the stakes are about pace. Altman’s warning implies that memory could become a critical constraint on how quickly AI systems scale.
The signals to watch are capital spending announcements from Samsung Electronics, SK hynix, and Micron Technology, along with any further supply warnings in filings from AI companies.
