OpenAI's CEO argues that running AI on personal devices leaves DRAM sitting idle while data centers could put it to work
Sam Altman has a problem with the AI model running on your laptop. It’s not that it doesn’t work. It’s that your computer barely uses the memory it’s hoarding.
The OpenAI CEO argued that local inference, meaning AI that runs directly on consumer devices instead of in the cloud, leaves scarce DRAM sitting idle on individual machines. With memory chips in short supply, he framed that idle capacity as something the industry can’t afford.
The case against your personal AI box
Altman made the comments during a podcast appearance, with research tying the remarks to October 29, 2026. His core claim was about utilization. According to Altman, individual devices put only about 5% of their resources to work.
Altman’s argument applies that logic to memory. DRAM is the short-term working memory computers use to run software, and AI models are famously hungry for it. When that memory lives inside a personal device, it sits unused whenever the owner isn’t actively prompting a chatbot.
Cloud data centers, by contrast, are built to squeeze as much work as possible out of every chip. Altman contended that centralized infrastructure makes far better use of limited memory supply. That matters more as demand for AI applications keeps climbing.
OpenAI is putting its money where its memory is
The company has struck supply arrangements with major memory producers, including Samsung and SK Hynix. Those agreements could see OpenAI procure up to 900,000 high-bandwidth DRAM wafers per month. That would account for roughly 40% of global output.
The goal, per the research, is to strengthen OpenAI’s data center capacity and push past what engineers call the “memory wall,” the bottleneck where processors outrun the memory feeding them data.
OpenAI has also gone the custom silicon route. In August 2026, the company announced “Jalapeño,” a custom chip built for AI inference. OpenAI claims it delivers 1.5 to 1.9 times better performance per watt compared with reference systems.
OpenAI is reportedly pursuing contracts for approximately 3.2 GW of power to expand its data center footprint.
A windfall for memory makers
Altman also sketched out what all this demand could mean for the companies making the chips. He estimated that major memory producers could potentially generate around $100 billion in annual free cash flow each, driven by improved economic conditions in the sector.
Altman’s estimate remains a projection, not a reported result, and should be read that way.
The local AI camp isn’t going quietly
Not everyone in Big Tech shares Altman’s view. Apple, among others, has championed local AI that runs on the device itself.
The appeal of on-device AI usually centers on privacy. If your data never leaves your phone, it can’t be exposed on someone else’s servers.
It’s worth noting where each side’s incentives point. OpenAI runs a cloud service and is reserving enormous amounts of memory for its data centers. A company that sells hardware has every reason to make that hardware feel indispensable. Both arguments can be sincere and self-serving at the same time.
What this means for the AI hardware race
For investors, the clearest signal is in semiconductors, particularly memory. OpenAI’s wafer commitments, if fully realized, would tie a major AI buyer to producers like Samsung and SK Hynix at a scale that could reshape how the memory market is priced and allocated.
That concentration cuts both ways. Memory makers gain a deep-pocketed anchor customer. But other buyers, from PC manufacturers to rival AI labs, may find themselves competing for whatever supply remains after a single company claims a large slice.
Altman’s local-inference critique also functions as a strategic argument. If the industry accepts that memory is too precious to leave idle in personal devices, the logic favors routing AI workloads through large cloud providers. OpenAI happens to be one of them.
