Sam Altman's Qoute About RAM

Sam Altman needs your RAM.

Or at least, he thinks it's doing a whole lot of nothing sitting inside your personal laptop while it could be powering the next frontier of artificial intelligence.

When Altman's comments about consumer memory vs. centralized compute from a recent podcast he was on made the rounds, the internet reacted predictably: Look at the tech billionaire coming for our local hardware. People took it as a literal dismissal of personal computing. Which it probably is, and seeing the means of compute becoming too expensive and locked down for the average person to buy makes the future look grim.

That aside, that reaction misses his actual point, one that while different might not actually be much better.

The core of what he said: centralized compute in a shortage

To be fair to Altman, the underlying logic isn't crazy, both from an engineering and a societal standpoint. In the middle of a global hardware and memory crunch, pooling resources is basic systems efficiency.

We saw a similar thing at work as well with the New York heatwave, where the mayor appealed to people to turn down their AC to ease the pressure on the power grid. Easing up would help keep the grid away from a blackout, meaning key infrastructure like hospitals could keep running. And while this is not the same severity, you can draw comparisons.

Most consumer RAM sits underutilized for hours a day, handling background tabs and idle processes or just plain not being powered on. In theory, aggregating memory and compute into massive, centralized clusters rather than scattering it across millions of individual devices allows us to get more capacity, and could help deal with the memory shortage by letting data centers use RAM when their original users are, for instance, asleep.

Its the same argument we deal with at my current workplace at Aarhus University, where we discuss the approaches of permanently reserving VMs for researchers, versus using job-based systems to better allocate and schedule compute needs. If you treat compute as a scarce resource during a bottleneck, centralizing it so it's "better used together" makes total sense on a whiteboard.

In reality though: OpenAI is not a public utility

The flaw in Altman's logic isnt the science though: it's the economics. Framing centralized RAM as "better for everyone" only works if the entity holding the hardware is a public good.

OpenAI is not a public utility. They are a private, profit-driven actor competing for market dominance. When Altman argues that memory is better off pooled together, he doesn't mean in a community-owned commons - he means inside OpenAI's walled garden, where you get to rent access back at $20 a month (or per API token), and potentially even more expensive once subsidies goes away, and investors need to see ROI. Giving up local hardware capability doesn't just centralize efficiency; it centralizes leverage.

Even if we buy the efficiency argument, data centers can sit idle too. Pooling resources can improve utilization, but it does not guarantee it. Centralization does not eliminate underutilization [1] , and cloud systems also need spare capacity to handle demand spikes and failures.

Which is not to speak about the fact that a computers value to its owner cannot be measured only by how continiously its memory is occupied. The value of being able to run software on your own machine, when you choose and without a provider's permission, is useful in itself.

Altman isn't wrong that pooled memory can do things fragmented memory cannot. He's just leaving out the part where pooling hardware also pools control - and putting it in a corporate data center is no guarantee it will be used well. Especially since it means trading hardware you own for access someone else can price, restrict and withdraw. And thats a bigger bargain than the efficiency arguments let on.

[1] https://link.springer.com/chapter/10.1007/978-3-031-99489-0_9