Follow us on Google News
OpenAI chief executive Sam Altman has sparked a debate over who should get access to the world’s limited supply of memory chips, arguing that keeping large amounts of DRAM in personal computers is an inefficient use of a scarce resource when the same hardware could be used more intensively in cloud data centres.
Altman made the remarks during an October 29 appearance on Cheeky Pint, a podcast hosted by Stripe co-founder John Collison. Discussing the global shortage of memory chips, he questioned the growing interest in running artificial intelligence models directly on personal computers instead of relying on cloud-based services.
“From a global allocation perspective, actually, it’s quite terrible to have everybody have a bunch of the world’s limited supply of DRAM sitting in their own computer,” Altman said. He argued that much of this memory is used only about five per cent of the time, while cloud infrastructure can allocate computing resources across multiple users and workloads.
The comments have drawn attention because the shortage is already putting pressure on the supply of memory used in smartphones, laptops, gaming computers and other consumer electronics. They also highlight a growing disagreement over whether AI computing should be concentrated in large data centres or distributed across devices owned by individuals and businesses.
Altman’s Argument Centres On Memory Utilisation
Altman’s argument is based on how efficiently memory resources are used. A personal computer may have 16GB, 32GB or more of RAM, but its capacity is not necessarily occupied throughout the day.
In a cloud data centre, memory can be shared across different users and applications. The same infrastructure can serve several workloads as demand changes, potentially allowing expensive hardware to remain more heavily utilised.
Altman argued that this model makes better use of limited memory supplies while the industry struggles to increase production. He also said memory manufacturers have historically been cautious about expanding capacity because earlier investment cycles left some companies close to bankruptcy after building too many factories.
His remarks were specifically about the allocation of memory for local AI inference, the process of running an AI model on a personal device rather than sending requests to a remote data centre. He did not call for personal computers to stop using RAM or suggest that consumers should be prevented from running AI models locally.
AI Data Centres Are Competing For The Same Memory Supply
The debate comes as AI companies expand data centres that require large quantities of specialised processors and memory. High-bandwidth memory, or HBM, is particularly important because it allows AI accelerators to access large amounts of data quickly.
Memory manufacturers are prioritising advanced products for servers and AI infrastructure because demand and margins are strong. This is putting pressure on the wider supply chain, including conventional DRAM used in personal computers and smartphones.
TrendForce said in September that conventional DRAM contract prices were expected to rise by 10–15 per cent quarter-on-quarter in the fourth quarter of 2026. NAND flash contract prices were projected to increase by 15–20 per cent during the same period.
The research firm also warned that PC memory supply could become tighter in 2027 as manufacturers allocate more production capacity to server applications. Constraints in manufacturing, packaging and testing are making it difficult for suppliers to meet demand across different product categories.
Samsung Electronics said its high-bandwidth memory supply for the following year was already sold out, while SK Hynix reported that its DRAM and HBM chips were sold out through 2029. The companies have announced investment plans, but new manufacturing facilities take time to build and bring into production.
Local AI Offers Benefits That Cloud Utilisation Cannot Measure
Altman’s argument focuses on the efficient use of hardware, but it does not account for every reason people want to run AI models locally.
Running an AI model on a personal computer can give users greater control over their data. Sensitive documents, business records and private information can remain on the device instead of being sent to an external service.
Local systems can also work without a continuous internet connection and may reduce dependence on cloud subscriptions or usage-based charges. For developers and organisations handling confidential information, these considerations can be as important as hardware utilisation.
The five per cent figure cited by Altman should also be understood as his argument during the podcast, not as a universal measurement of how much memory every personal computer uses. RAM supports operating systems, applications, browsers and background processes even when a user is not actively running an AI model.
The debate is therefore not simply about whether cloud computing is more efficient. It is also about privacy, ownership, reliability and whether individuals should retain the ability to run increasingly capable software on hardware they control.
Memory Shortage Could Make Consumer Devices More Expensive
The supply imbalance is already affecting the economics of consumer electronics. When manufacturers pay more for memory components, some of those costs can eventually be passed on to buyers through higher prices or reduced hardware specifications.
Laptop and smartphone manufacturers may also adjust the memory configurations offered in different price segments. A device that previously included more RAM could become more expensive, while manufacturers may use lower-capacity components to control production costs.
The shortage is not limited to AI companies purchasing specialised chips. The shift towards server and high-bandwidth memory affects production priorities across an industry where manufacturing capacity cannot be expanded immediately.
For chipmakers, the challenge is deciding how much capacity to allocate to high-margin AI products while continuing to supply conventional memory markets. For consumers, the result could be greater pressure on device prices and availability.
What This Means For Consumers And The AI Industry
Altman’s comments raise a larger question about how the benefits and costs of the AI expansion should be distributed. Cloud computing allows technology companies to pool expensive hardware and serve large numbers of users, but it also gives those companies greater control over the infrastructure on which AI services depend.
A shift towards cloud-based AI could make advanced models easier to access without requiring users to purchase powerful computers. However, it could also increase dependence on a small number of service providers and their pricing, privacy and access policies.
The growing interest in local AI reflects a different priority: keeping computing resources and data under the user’s control. That approach may not always maximise hardware utilisation, but it offers benefits that cannot be measured through the number of hours a memory chip remains active.
The immediate challenge for the industry is to expand memory production and reduce supply constraints without making ordinary computing devices increasingly expensive. Altman’s remarks have brought the allocation debate into public view, but the longer-term balance between personal computing and centralised AI infrastructure remains unresolved.
