Test report DSG-4728 · Rev E · tested November 17, 2026
Supply Chain & PolicyDevice under test
Altman Says AI Chip Demand Is Being Double-Counted
Sam Altman says multiple chip suppliers assume they will win the same OpenAI and Anthropic orders, inflating demand figures across the supply chain and risking oversupply.
- Read
- 4 min
- Words
- 724
- Node
- 7nm
- Operator
- Marcus Bennett
Spec summary
- Altman said on Oct. 16 that AI chip demand is being double-counted as several suppliers book capacity for the same OpenAI and Anthropic orders.
- About 45% of Microsoft's $625 billion commercial backlog, roughly $281 billion, is tied to OpenAI, the company said in December.
- Analysts estimate only 13% of Oracle's $664 billion backlog may convert to revenue.
- Anthropic CEO Dario Amodei called computing the company's deepest business dilemma a day earlier.
- The Information reported OpenAI- and Anthropic-related contracts exceed half of roughly $2 trillion in cloud bookings.

OpenAI CEO Sam Altman says demand for AI chips is being overcounted, because multiple chip companies are booking foundry capacity on the assumption that they will win the same OpenAI and Anthropic orders. He made the remarks on October 16 at a Semafor forum in Menlo Park, California, titled "Silicon Valley and the World: A New World Order."
Altman is the second major AI lab leader to flag demand uncertainty in public. A day earlier, Anthropic CEO Dario Amodei called computing "the deepest business dilemma the company faces" and said there is "enormous uncertainty about what will happen next year."
What is being counted twice?
The moderator asked Altman to explain comments from a closed-door meeting with Samsung Electronics executives, where he said AI demand was being double-counted. His answer described a mechanism.
"For a long time there was only one chip supplier in the industry, so when people counted wafers, they could ask a company like us how much demand we would have, and from that figure out where the end demand for all those chips would be," Altman said. "Now several new chip companies are booking capacity at the same foundry. Each of them assumes it will win most of the orders from OpenAI and Anthropic, so there is clearly double counting."
The market context matters here. Nvidia once dominated AI silicon. AMD, Intel, Google, Amazon and Microsoft now field their own processors. Samsung Electronics, SK hynix and Micron supply the memory that goes with them. When each new entrant plans capacity against the same customer base, the aggregate numbers run ahead of real demand.
Altman rejected that framing of the buying side. "Many chipmakers are talking to their supply chains on the assumption that we will buy all of their chips. And I do think demand such as data center construction is solid," he said. "We cannot buy the supply we need from every chip supplier, and we do not need to."
What does this change for the supply chain?
The concrete risk sits in a second channel: the so-called neoclouds. At a meeting with roughly 50 Samsung executives on October 31, Altman warned that these companies, which lease out GPU data center capacity, contribute to double counting of AI demand. He also warned that an oversupply shock could follow, because large-scale demand growth is in practice concentrated at OpenAI and Anthropic rather than spread across the market.
The numbers from cloud providers support the concern. Key figures:
- Microsoft said in December last year that about 45% of its $625 billion commercial contract backlog, roughly $281 billion, was tied to OpenAI.
- Oracle's backlog stood at $664 billion as of the end of August. Standard & Poor's estimated about half of that was attributable to OpenAI.
- The Information earlier reported that contracts related to OpenAI and Anthropic accounted for more than half of the roughly $2 trillion in bookings at major cloud providers.
These figures are remaining performance obligations, or RPO: signed contracts not yet delivered. Neocloud operators sit in the middle of the distribution chain, which is another point where the same demand can be counted more than once.
Conversion into real revenue is the weak link. Some analysts estimate that only 13% of Oracle's backlog may be recognized as revenue. If contract backlogs keep swelling while the converting share shrinks, the fallout for the market could grow.
Is the AI market in a bubble?
Asked directly whether double-counted demand meant the AI market is in a bubble, Altman declined the label but accepted the volatility. "There will clearly be ups and downs," he said.
He drew a line between long-term direction and short-term matching of supply and demand. "I think the long-term demand for vast computing resources, the value it creates and the remarkable things being done with this technology will continue," he said. "The long-term trend is clear, but it is hard to predict whether supply and demand will match exactly quarter by quarter or year by year."
For chipmakers, memory suppliers and foundries, the practical takeaway is to discount capacity plans that assume any single AI customer will absorb their full output. For cloud and neocloud operators, the backlog numbers now on the books will face a test as they convert, or fail to convert, into revenue.
via en.sedaily.com (Source)
More from Marcus Bennett
Show full bio
News editor covering regulation, policy and their impact on hardware layer of artificial intelligence at Die Signal.
681 articles
Same lot · LOT-362D
- DSG-99325nmAltman Flags AI Chip Demand Concentration as NVIDIA Falls 28.9%
- DSG-317628nmCNBC tours $25B supplier behind next-generation AI chips
- DSG-87353nmSilicon Valley Bets $5 Billion on Advanced AI Chip Facility
- DSG-469765nmOpenAI Chases Samsung and SK Deals as Memory Chip Supply Sells Out
- DSG-199620nmAltman courts Samsung as DRAM shortage threatens AI buildout to 2027