WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-06-28 1 00:00:04.500 --> 00:00:14.100 Memory is the new bottleneck in artificial intelligence, and this morning Wall Street started pricing the companies that make it like the next Nvidia. 2 00:00:14.100 --> 00:00:24.820 It's Sunday, June twenty-eighth. Here's the rundown: the memory squeeze, a compute map that's shifting east, and an A I agent that got talked into installing malware. 3 00:00:26.280 --> 00:00:38.840 Start with memory. Investors spent two years treating one chip company as the whole A I trade. This week they went looking for the next one and landed on Micron. 4 00:00:38.840 --> 00:00:49.480 Which tells you where the scarcity moved. Every accelerator needs fast memory bolted right next to it, and far fewer companies can make that memory than there are buyers for it. 5 00:00:49.480 --> 00:00:53.560 And the buyers are the richest companies on earth. 6 00:00:53.560 --> 00:01:04.840 The supply side isn't arguing. A Lenovo exec said this week the shortage everyone's calling RAMageddon is permanent. His words: it will never be like it was last year. 7 00:01:04.840 --> 00:01:10.120 That's a vendor telling its own customers to stop waiting for prices to come back down. 8 00:01:10.120 --> 00:01:23.800 And it isn't only datacenters. Memory is now one of the most expensive parts of building any machine. The same squeeze that sets a hyperscaler's bill decides whether a small shop can afford thirty-two gigs this quarter. 9 00:01:23.800 --> 00:01:28.200 So what do you actually do with that if you're building? 10 00:01:28.200 --> 00:01:41.480 Stop treating memory as a rounding error. Price it as a first-order line, lock supply where you can, and read every accelerator roadmap as a memory roadmap too. The compute story just became a memory story. 11 00:01:44.060 --> 00:01:56.940 On the wire, the compute map shifted. China retook the top of the global supercomputer rankings with a system called LineShine, its first number one since 2018. 12 00:01:56.940 --> 00:02:06.700 Under export controls, which is the part that matters. The restrictions are steering where the frontier gets built, not whether it gets built. 13 00:02:06.700 --> 00:02:19.420 Same week, Loongson shipped a homegrown sixteen-core server chip on its own architecture. Forty watts, aimed at cheap general-purpose servers. 14 00:02:19.420 --> 00:02:28.860 Not a frontier part, and that's the point. A domestic stack for the boring server middle that no export policy can switch off. 15 00:02:29.040 --> 00:02:40.560 Security desk. Mozilla's research team showed a coding agent can be walked into installing malware from a repository that looks completely clean. 16 00:02:40.560 --> 00:02:50.960 They didn't break the model. They used its helpfulness against it. If you let an agent run setup steps from a repo, that repo is now part of your threat model. 17 00:02:50.960 --> 00:02:59.360 And there's an open request to let OpenAI's coding tool exclude sensitive files that's climbing Hacker News. 18 00:02:59.360 --> 00:03:08.000 Same worry from the other side. When an agent can read your whole tree, what it can see is a question you answer before you run it, not after. 19 00:03:10.180 --> 00:03:21.860 Models. The open field keeps widening. Zyphra, Cohere, and Poolside all added to the open-weights pile this week. 20 00:03:21.860 --> 00:03:31.140 The headlines go to the closed labs, but the buildable middle is getting deeper every month. That's the part that actually changes what you can ship tonight. 21 00:03:31.140 --> 00:03:40.260 And a sharper one. Prosecutors entered a defendant's ChatGPT logs as evidence in the Palisades fire trial. 22 00:03:40.260 --> 00:03:51.300 Whatever the verdict, the precedent is the story. Chatbot history is discoverable now. What people type into an assistant carries the same weight as any other record. 23 00:03:51.480 --> 00:04:03.400 The tape. Three prints this morning. Long Micron, long S K Hynix, both on the memory squeeze, and watching S M I C on China's domestic buildout. 24 00:04:03.400 --> 00:04:14.600 Highest conviction is the memory pair. The falsifier is simple: a supply glut or an A I spending pullback that sends memory prices down over the next two quarters. Horizon, six months. 25 00:04:14.600 --> 00:04:20.040 And the tape is the desk's scorecard, not advice. 26 00:04:25.360 --> 00:04:27.760 Quick break — two from the desk. 27 00:04:27.760 --> 00:04:43.080 One we know well: vote dot direct. If you're on an H O A or a board, it runs your elections digitally — secure, verifiable, no paper, no clipboard in the lobby. Point your council to vote dot direct. 28 00:04:43.080 --> 00:04:56.560 And if this is your ten minutes of A I for the day, get the written edition too. The full wire, free, every morning — leave your email at nextbig dot dev. 29 00:04:59.790 --> 00:05:08.190 Quick hits. I B M detailed a sub-one-nanometer transistor meant to carry chip scaling into the 2030s. 30 00:05:08.190 --> 00:05:16.190 The Trump administration cleared more than a hundred U S companies and agencies to use Anthropic's Mythos five. 31 00:05:16.190 --> 00:05:21.470 OpenAI poached Uber's India chief to run its biggest market outside the U S. 32 00:05:21.470 --> 00:05:28.990 And one for builders: OpenAI says its internal coding agent's output grew more than fifty times in research this year. 33 00:05:32.050 --> 00:05:52.690 That's the rundown. Our call: within six months the A I hardware story stops being only about G P Us. A major memory maker posts a quarter growing faster than Nvidia's datacenter business, and memory supply shows up by name in a hyperscaler's spending guidance. 34 00:05:52.690 --> 00:06:00.610 Proves us wrong if neither happens by the end of December. We settle it in public when the date comes. See you tomorrow.