[AI: Is Your Home Memory Being Wasted?] OpenAI CEO Sam Altman made a blunt statement on the Cheeky Pint podcast: Globally, everyone is cramming limited DRAM into their own computers, only squeezing out about 5% utilization, which is actually "quite terrible." He wants memory to be in cloud data centers, running at full utilization. This statement immediately caused a stir. He was talking about running open-source models locally (local inference). Many people think that since open-source models can run on their own machines, data centers aren't so important—Altman says his intuition is exactly the opposite: the tighter the supply, the more wasted memory at home seems. The reality is even more stark: * The DRAM shortage has already driven up prices for laptops, smartphones, and DIY components. * Samsung and SK Hynix have signed long-term contracts with cloud providers and projects like OpenAI Stargate, locking in significant production capacity for the next few years. * Reports suggest these intentions could impact a substantial portion of global DRAM production (some say "about 40%"—note that this is the potential ceiling, not already secured orders). * In the same week, Intel was still citing an "unprecedented memory shortage" to explain weak PC demand. What does this mean? On one hand, AI companies are urging "memory shouldn't be idle," while on the other hand, ordinary people are finding that building new PCs and upgrading laptops are becoming increasingly expensive—the main driver of the shortage is precisely HBM (Hardware Bus) memory, which consumes even more wafers in data centers. Some calculations show that one GB of HBM can consume approximately three times the wafer capacity of DDR5. While cloud utilization is indeed high, it's not a "free lunch" of memory. Online reactions were direct: some countered with photos of 64GB or even hundreds of GB of RAM; others worried that the next step would be "don't buy your own computer, just rent from the cloud"; still others pointed out that the focus was on local inference, not on using memory for gaming or running multiple tabs. Privacy, offline operation, and avoiding one-click rejection—these are reasons for running models locally that the cloud can't replace. Efficiency is one factor, but ownership is another. Which side are you on: should memory be fully utilized in a data center, or left on your own machine for occasional use? Let's discuss in the comments. Source: The Verge / Tom's Hardware / Cheeky Pint