Post

Here is why AI compute and consumer hardware are about to get significantly more expensive, despite what most people think. Common belief says hardware and AI token prices will decline as computing efficiency improves and memory supply increases. But that overlooks the structural reality of the supply chain. At a recent SpaceX earnings call, @elonmusk pointed out that memory supply grows by roughly 20% a year, while demand is growing by 200%. Many expected cheap memory supply from China's CXMT to ease the pressure, but Apple recently received a quote from CXMT that was actually higher than Samsung's. Having talked directly with semiconductor fab engineers, this is entirely expected. Demand is so far ahead of supply that CXMT has no incentive to price aggressively. This gap widened massively after the agent boom earlier this year. Agents consume significantly more tokens by default, causing global inference usage to skyrocket. Better performance brought in a wave of new users, as shown by Codex rapidly crossing 10M users. Meanwhile, compute efficiency gains have been very small. AI labs chose to increase model sizes to raise intelligence, consuming at least double the compute and accelerating demand further. On the supply side, hardware bottlenecks take years to fix. SK Hynix's Yongin cluster only starts its first fab in 2027, with full operations planned for 2033. Even with massive construction starting today, it takes about 5 years for actual supply to land on the market. This is an unsolvable short-term bottleneck. AI inference costs will inevitably keep rising, and consumer hardware will follow. This is why I keep telling everyone to buy the personal hardware now.#AIMemoryBullTest #FedHawksVsWeakJobs #SpaceXUnlockRebound