WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-07-07 1 00:00:02.600 --> 00:00:13.250 Memory just repriced the entire AI build-out. It's Tuesday, July seventh, and the clearest proof is a company most people in AI have never heard of. 2 00:00:13.430 --> 00:00:31.520 Longsys, the Shenzhen maker that owns the Lexar brand, guided to roughly one and a half billion dollars in profit for the first half of this year, against two million in the same stretch last year. That's a jump north of sixty thousand percent, and no new product explains it. 3 00:00:31.520 --> 00:00:51.090 One force is bending every line in the memory business the same way. AI data centers are buying memory faster than the world can make it, and the price of anything that stores a bit has gone vertical. DRAM contract prices climbed about ninety-five percent in a single quarter. 4 00:00:51.090 --> 00:01:04.650 Read it from the other end, in a budget too big to hide. When Microsoft set capital spending near a hundred and ninety billion dollars, it named memory as the reason twenty-five billion of that was new. 5 00:01:04.650 --> 00:01:24.100 For anyone building, the volatile number in a capacity plan used to be whether you could get GPUs. Now it's the memory bill, climbing faster than the compute line and in one direction. The model you run has fallen close to free; the cost of the year is on the components list underneath it. 6 00:01:24.280 --> 00:01:43.850 The same wall shows up in the ground. South Korea's eight-hundred-and-eighty-billion-dollar chip and AI plan is colliding with physics: a single planned megacluster would draw about a quarter of Seoul's entire power demand, sited faster than the grid and water can follow. 7 00:01:43.850 --> 00:02:00.930 And Britain is now letting qualifying data centers claim national-importance status to override local planning, cutting a year and more than a billion dollars off the fight. The binding constraint in developed economies has become the permission to build. 8 00:02:01.110 --> 00:02:11.640 To the tape. We hold Micron long, and this week the thesis printed in daylight, as the memory shortage started arriving on income statements. 9 00:02:11.640 --> 00:02:27.690 We add SK Hynix as a watch while Wall Street packages the memory boom for retail, and we hold the Nvidia watch, where a rising memory bill is the tell. Late retail access is usually a caution rather than a green light. 10 00:02:27.680 --> 00:02:31.560 The tape is the desk's scorecard, not advice. 11 00:02:31.740 --> 00:02:45.300 Our call: within nine months, a frontier lab or hyperscaler publicly blames memory, supply or price rather than GPUs or power, for a delayed or re-architected build. 12 00:02:45.300 --> 00:02:53.350 What proves us wrong is April with the constraint story still told entirely through GPUs and power.