Today's frontier is really one story told from four sides: who controls compute, and at what cost. Beijing is blocking Nvidia's best China-legal chip at the border just as Washington cleared it; a pair of Harvard dropouts is booking a billion dollars in orders for a chip that only does one thing; coding agents are jumping to your phone; and, at the cheap end of the curve, a $1,000 AI agent is turning up decades-old security flaws that humans missed. The connective thread is that the price and politics of AI capability are moving faster than the roadmaps.
China slams the door on Nvidia's H200 — days after the US opened it
The US–China chip standoff just snapped back open. Only days after the Trump administration cleared Nvidia's H200 — the most advanced chip it could legally sell into China — for export with a 25% government surcharge, Chinese authorities have reportedly told customs agents to block the chip from entering the country, and Nvidia's stock slipped on the news. Reporting suggests Beijing is drafting rules to cap how many foreign AI chips domestic firms may buy rather than banning them outright, which leaves Nvidia's China revenue genuinely uncertain and hands relative advantage to domestic accelerator makers. The lesson for anyone building on frontier compute: access is gated by geopolitics as much as by performance.
Two Harvard dropouts take direct aim at Nvidia's inference grip
Etched has reached a $5B valuation with more than $1B in booked orders for Sohu, an ASIC built to do exactly one thing — transformer inference — as fast as physically possible. Founders Gavin Uberti and Chris Zhu have raised roughly $800M, count Jane Street among their backers, and claim a single Sohu server can stand in for up to ~160 H100s, with first racks due to ship this summer. It's the boldest bet in the inference-silicon field: enormous efficiency if transformers stay dominant, and stranded hardware if the architecture shifts. The re-rating comes not from the valuation but from whether those racks ship and hold up against Nvidia in production.
Cursor puts a fleet of coding agents in your pocket
Anysphere's Cursor released a public beta iPhone and iPad app that lets developers spin up cloud agents, review diffs, and merge pull requests from their phone, with a model picker spanning Claude, GPT, and Gemini alongside its own Composer 2.5. The shift is subtle but real: the developer becomes an orchestrator steering always-on agents rather than someone tied to an editor. It arrives amid a reported ~$60B SpaceX acquisition of Cursor, a sign of how strategically the industry now values agentic development tooling. The metric that matters next is whether remote-agent workflows become habit or stay novelty.
AI proposes two new superconductors, and the lab confirms them
Researchers at Aalto University, led by Prof. Päivi Törmä, used machine learning to prescreen promising element combinations, ran detailed quantum-physics calculations, and had collaborators synthesize the winners — confirming two previously unknown superconductors, YRu3B2 and LuRu3B2. Both draw their superconductivity from flat electronic bands in a kagome lattice, the hexagonal pattern named after Japanese basket-weaving. The significance isn't the two compounds; it's the working loop — AI narrows the search space, physics validates, experiment confirms — which the SuperC consortium says could surface thousands more candidates and compress a search that used to take years.
A $1,000 agent finds 21 zero-days that humans missed for decades
An autonomous security agent scanned FFmpeg's roughly 1.5 million lines of C and produced 21 confirmed, reproducible zero-days — each with a working proof of concept — for about $1,000, with one flaw dating back to 2003 and several others latent for 15 to 20 years. In the same stretch, Chrome patched a record number of bugs. The story isn't a single vulnerability; it's the economics. When capable, verifiable bug-hunting costs about as much as a laptop, both offensive security and the maintenance burden on critical, volunteer-run open-source code get repriced — and defenders will need the same tooling to keep pace.
The money is pooling in the compute layer — non-Nvidia inference silicon on one side, policy risk to Nvidia's China business on the other — while software agents spread to mobile and draw serious M&A. Underneath it all, the cost of capable AI work keeps falling, whether that's finding materials or finding bugs. What to watch next: the fine print of Beijing's foreign-chip purchase caps, Etched's first Sohu racks and any independent benchmarks this summer, and whether defenders adopt cheap bug-finding agents now that attackers clearly can. Signal, not advice.