Altman warns Samsung of AI memory demand volatility

At an Oct. 31 meeting with about 50 presidents of Samsung affiliates, Sam Altman said “in the short term there are ups and downs” for AI demand, according to South Korea’s Financial News, which said it had obtained the session transcript and reported the remarks on Nov. 11.

The group included presidents from Samsung’s electronics, finance and biotech affiliates and met at the company’s training center in Yongin. Altman was asked whether AI, memory-chip and data-center demand could saturate or slow over the next five to 10 years.

What did Altman say about near-term demand?

His answer split the horizon. He said the market had double-counted some demand and that “the only two places actually driving the big increase in demand are OpenAI and Anthropic,” a concentration that could produce a supply shock and an unsettled period.

Financial News offered a matching observation: some AI infrastructure announcements show the same project appearing in several tallies.

A single buildout can therefore appear in more than one headline forecast without an equivalent rise in physical demand. Altman tied the issue to this year’s bubble debate over neoclouds, the newer cloud companies that rent out GPUs for AI.

The paper’s quotations are its Korean rendering of the exchange, and neither OpenAI nor Samsung has commented publicly on them. Taiwan’s DigiTimes picked up the report on Nov. 12. Altman’s remarks came with no numerical capacity, revenue or price forecast.

How large could long-term demand become?

Altman projected enormous long-term growth as models write custom software in real time and AI moves into robots, spacecraft and scientific work. He offered one usage marker: early 2020’s top token user consumed about 100,000 tokens per month, while this year’s heaviest user consumed 100 billion per month.

The report did not translate those usage figures into memory shipments or revenue.

The long term still looked enormous. Near term, far murkier.

Where are the bottlenecks?

Asked for the biggest bottleneck, Altman identified a broad set of constraints:

• Supply chains
• Financing
• AI safety and social acceptance
• Enterprise adoption

He then named chip supply chains and data-center infrastructure as the biggest limits over the next several years and said more supply-chain investment was needed. That put physical supply and infrastructure ahead of financing, safety and adoption in his multi-year ranking.

What does OpenAI’s cost model require?

On profitability pressure from OpenAI’s spending, Altman told the executives that vertically integrated infrastructure, cheaper than any other company’s, would let OpenAI sell tokens at low margins and at large scale.

The approach places the commercial burden on controlling infrastructure cost while keeping token margins low and volume large. Altman’s visit was read in Korea as a push for more DRAM supply.

Can memory makers plan around the signal?

Samsung, SK hynix and Micron have been planning capacity around a multi-year shortage, according to the report. That makes Altman’s warning commercially material: two companies were driving the reported increase in demand, while some infrastructure announcements could inflate near-term tallies. For suppliers, multi-year shortage planning and the demand caveat pull in opposite directions.

From the supplier side, Winbond’s chairman made a similar point this week, saying cloud buildouts are funded through 2028 but memory prices have hit their limit, according to an MLQ.ai report.

The tension is explicit: cloud buildouts have funding extending to 2028, while a supplier says pricing has reached its limit. The reported role of OpenAI and Anthropic makes them the critical demand drivers, even as Altman cautions that the market may be counting some of the same demand twice.

No new supply contract accompanied Altman’s visit, leaving Samsung, SK hynix and Micron to reconcile existing multi-year plans with demand that may prove less uniform. The next commercial signal will be whether cloud buildouts funded through 2028 generate enough memory orders to support added capacity without pushing prices beyond the limit Winbond’s chairman described.