A 32GB DDR5 kit now costs $375, minimum. Not a server part — the memory you'd drop into a gaming PC or a home lab. The price didn't climb because manufacturing got harder. It climbed because the same DRAM lines feed AI datacenters, and datacenter demand wins every allocation fight. The buildout finally has a number an ordinary builder can feel at the checkout.
That number is the story the funding rounds keep hiding. The AI boom is a capital event, and capital events externalize. The cost of standing up the next training cluster doesn't stay on the lab's balance sheet — it lands on whoever was downstream of the same supply, the same market, the same scarce input. This week the bill arrived for people who never bought a GPU.
The AI boom is a capital event, and capital events externalize.
Watch how the money moves. DeepSeek is raising $7 billion in its first outside round. Stratechery now describes Google as, functionally, a capital-allocation company — funding compute, startups, and its own moonshots more than it ships products. Reuters puts US tech at a record share of total market value and, in the same breath, names the risk: concentration. Each headline reads like ambition. Stacked, they read like one pool of capital and physical supply being pulled toward a single use, with everyone else repricing around the gap. The memory counter is just where that gap shows up without a Bloomberg terminal.
The optimist's answer is that this is temporary friction. Fabs expand, supply catches up, and the productivity dividend dwarfs a year of expensive RAM. Fair, and partly true — memory cycles do mean-revert. But concentration risk doesn't expand its way out, and "temporary" is carrying a lot of weight while the capex curve keeps steepening. You don't get to call a cost temporary while you're still adding to it.
The buildout will probably pay off. The honest version just admits who's fronting the money in the meantime — and it isn't only the people who'll collect the returns.