Sam Altman: Put Scarce RAM in Data Centers — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Cheeky Pint interviewed Sam Altman about local inference, and at the 29-minute mark he said putting scarce DRAM in home computers for occasional use was “quite terrible.”
- Altman quantified the utilization argument: personal computers use their DRAM only 5% of the time, while he wants “maximum utilization in the cloud” in a supply-constrained environment.
- OpenAI has some open-source models, but its primary business remains closed-source models running in data centers, which the article presents as the business model Altman is defending.
- Data-center RAM is treated as more worthwhile when used 24/7, an argument resting on data-center buildouts not slowing; home RAM instead serves spreadsheets, browsing, gaming and occasional local AI.
- GPT-6 Luna anchors the commercial stake: the article says local AI could let people pay less for subscriptions to models such as GPT-6 Luna if their home hardware is powerful enough.
- A reader cited a report saying AI data centers were buying scarce GPUs and memory while sitting idle over 70% of the time, and suggested the purchases also denied competitors access.
Why it matters: OpenAI’s closed-source, data-center business has a direct stake in Altman’s allocation argument: the article says capable home hardware could reduce subscriptions to models such as GPT-6 Luna. Concentrating scarce DRAM where OpenAI runs models supports that subscription model, while his 5%-versus-24/7 utilization comparison supplies the efficiency rationale.
Ask SkimNews
