OpenAI CEO Sam Altman believes that the problem of DRAM shortages will not resolve itself, as manufacturers are wary of expanding capacity due to past economic crises. He also stated that deploying AI models on local devices can be seen as a waste of the world’s limited resources.
OpenAI CEO Sam Altman believes that the problem of DRAM shortages will not resolve itself, as manufacturers are wary of expanding capacity due to past economic crises. He also stated that deploying AI models on local devices can be seen as a waste of the world’s limited resources.
Altman said this on the Cheeky Pint podcast. According to him, DRAM manufacturers, which previously invested $20 billion to get $19 billion, will soon be able to generate about $100 billion in free cash flow per year each. Despite this, companies are still psychologically “traumatized” by past market declines, which is holding them back from actively building new factories.
At the same time, the head of OpenAI drew attention to the cost of infrastructure for AI. According to his estimates, data center costs reach approximately $70 billion per gigawatt of capacity, although a much more rational figure would be $20 billion. Because of this, OpenAI is forced to move towards vertical integration — developing its own chips, building data centers, and investing in robotics.
Separately, the head of OpenAI criticized the idea of a mass transition to local launch of AI models on personal devices due to the low resource utilization rate, which is about 5%.
"There are many cases where the supply chain is a non-obvious phenomenon. Many people want local inference and believe that the problem is solved because of the availability of open source models. They say we can run them on our own computers and we won’t need equipment in data centers. But from the point of view of global resource allocation, it’s very bad when everyone keeps a part of the limited world supply of DRAM in their computer," Altman emphasized.
Speaking about the strategic development of artificial intelligence, the head of OpenAI noted that he prefers the concept of "a trillion AI systems with intelligence that slightly exceeds human intelligence" rather than creating a single universal model with a conditional IQ of a trillion points.
Earlier, one of the world’s largest chip manufacturers, Micron, said that even if demand from AI data centers falls, the rapid development of humanoid robot production will keep DRAM prices high until at least 2027 and beyond.
