Criticism of local AI amid pressure on memory supply
Appearing as John Collison's guest on Cheeky Pint on 29 October, Sam Altman discussed the shortage of computing resources and component prices. The conversation then turned to an idea: running open models on personal computers to reduce dependence on remote infrastructure.
OpenAI's chief calls this allocation of DRAM “quite terrible” if everyone ties up a substantial amount in a machine used only “5% of the time”. He favours using it as intensively as possible in the cloud, where the same infrastructure serves many customers. The official transcript clearly places this remark within the debate about local inference.
He is not asking individuals to stop buying RAM and is not announcing any measures to restrict access to it. He is arguing for a resource allocation strategy that favours pooling in data centres.
Full interview: Cheeky Pint, with Sam Altman and John Collison. The discussion of memory and local AI takes place around the 29-minute mark.
The 5% figure is not a measurement of our PCs
No study or quantified comparison with data centres is presented to support that rate in the passage. It is an estimate put forward by Altman.
The case for pooling resources deserves discussion, but a machine's utilisation rate does not capture its full value. A PC is also used for gaming, work, browsing and creative tasks. Memory capacity available to handle an occasional workload can be valuable even when it is not constantly in use. Conversely, a shared server must keep resources available for demand spikes and service continuity.
Manufacturers confirm pressure from AI servers
The industry context, meanwhile, is documented. In its third-quarter results published on 29 October, Samsung explains that it is prioritising memory products for AI in an environment of constrained capacity. The group links strong server demand to price increases across all applications and plans to further expand sales of HBM, server DRAM and SSDs.
On 28 October, SK hynix said it had secured long-term agreements with 14 major customers covering more than half its production capacity for 2027 and 2028. The manufacturer expects demand to exceed supply until 2030, given the technical challenges and the time needed to build new factories. This is its forecast, not a confirmed date for the end of the shortage.
These announcements show why the issue directly concerns readers building or upgrading a PC. They do not, however, allow a precise share of the retail price increase for a DDR5 kit to be attributed to any single company.
OpenAI itself is part of that demand
As early as 1 October 2025, OpenAI announced partnerships with Samsung and SK as part of Stargate. The announcement describes a capacity ramp-up targeting 900,000 DRAM wafer starts per month to support the needs of AI models.
This wording describes an industrial target. It does not mean OpenAI is already receiving that volume of chips every month, nor that it has bought that many consumer memory modules. The same announcement also outlined agreements to explore new data centre capacity in Korea, with several details still to be finalised.
Altman's position is therefore consistent with his company's investments: developing centralised infrastructure capable of serving large numbers of users. That economic interest should remain in view when assessing his argument about the best use of available memory.
The cloud does not replace every benefit of local AI
Local inference also meets needs for control over data, independence and availability without a permanent connection once the model is installed. For a concrete example, the Ollama documentation states that prompts and data from locally run models are not sent to Ollama; its cloud features can also be disabled.
These benefits come with trade-offs: the available hardware limits which models can be used, users handle installation themselves and bear the cost of the machine. The cloud makes larger resources easier to access and share. The choice therefore depends on usage frequency, model size, budget and privacy requirements.
One example (the ASUS Ascent GX10 we covered) illustrates precisely this alternative: dedicating resources to AI that runs at home or within a business. The discussion is therefore not just about the number of gigabytes in use; it is also about who owns the hardware and retains control over how it is used.
Key takeaways
Pooling memory can improve its utilisation at scale. That is not enough to dismiss personal computers or local AI, whose benefits include control over data and hardware.