WORLDTECHOpenAI's Altman calls home PC RAM 'quite terrible' versus data centers
OpenAI CEO Sam Altman has argued that it is "quite terrible" for people to keep RAM in their own computers when cloud data centers could use it, according to Tom's Hardware. The comments, made in a podcast interview, bear on a memory supply that is already strained.
What Altman said in the interview
Speaking in an interview with the YouTube channel Cheeky Pint, Altman described local inference, and by extension DRAM sitting in home computers, as "quite terrible". He tied the view to how the world's limited memory supply is allocated, arguing that spreading it across personal machines that sit in use only a small share of the time is wasteful.
Altman said that when supply is tightly constrained, "you want maximum utilization in the cloud", and added that many people's intuition runs the opposite way.
Why the argument favours OpenAI's model
Tom's Hardware notes that OpenAI does have some open-source models, but that its primary business lies in closed-source models that run in data centers. Read that way, the comments defend the company's business model. The outlet also says the argument rests on the idea that data center buildouts will not slow down.
The reasoning further treats RAM working in AI data centers around the clock as more worthwhile than RAM used for spreadsheets, web browsing, gaming and occasional local AI.
Concerns about owning hardware
According to Tom's Hardware, the remarks have fuelled enthusiast worries that people could end up renting an agent and a computer in the cloud rather than owning their hardware and software. The outlet says local AI could one day let people pay less for subscriptions if they can run models at home, but it does not claim that laptops and gaming rigs threaten OpenAI today.
The report places the comments against the wider RAM shortage, noting that memory now costs considerably more than it did a year or so earlier.
Other topics in the interview
The conversation also covered this week's PHP vulnerabilities, sandboxing of models in training, and OpenAI's research in mathematics.
Open questions
- Will memory supply stay tight enough for data-center priority to shape consumer hardware?
- Will OpenAI change its stance on local inference in response to enthusiast concern?