A talent war goes nuclear, a star founder finally ships, and the compute trade cracks. Today's brief is about where power in AI is actually concentrating — and it's moving faster than the money. The frontier labs are no longer competing mainly on who can spend the most on chips; they're competing on who can hire the handful of people who know how to build the next model, and on who owns the silicon those models run on. Around that fight, one of the most-watched startups in the field finally put a product in market, and the public "rent-out-GPUs" trade took its first real punch.
Meta raids Apple for its AI chief
The defining story of the moment is a hiring one. Meta's Superintelligence Lab has poached Apple's foundation-models chief Ruoming Pang, along with several senior researchers from Google's Gemini team, reportedly on nine-figure compensation packages. Apple's AI group is said to be in disarray, and Google — still recovering from months of departures — is again on the back foot. The takeaway is blunt: at the frontier, advantage compounds through a small number of researchers, not through headcount or capex. Meta is now spending its balance sheet on people as aggressively as on GPUs, and rivals that can't match the packages are watching their best talent walk. Expect Apple's answer to say a lot about the next phase — whether it tries to rebuild in-house or reaches for an acquisition or an outside model partner.
Mira Murati's Thinking Machines finally ships
After a year as the most-watched and most-secretive startup in AI, Mira Murati's Thinking Machines Lab shipped its first product: "Tinker," an API that lets developers fine-tune frontier open-weight models. The launch lands with a reported $12 billion valuation and no prior revenue — a vivid reminder that a star founder and a clear thesis can still command a mega-valuation on promise alone. The thesis itself is the interesting part: Murati is betting that the next layer of value sits in post-training and customization of open models, not only in the base frontier models. If she's right, the fine-tuning-and-inference tier — Thinking Machines, Together, Fireworks and their peers — becomes the picks-and-shovels business of the open-model economy.
The neocloud trade takes its first real punch
The market delivered the day's other big signal. Shares of the "neocloud" GPU-rental companies cratered — CoreWeave fell around 14% and Nebius around 17% in a single session, with IREN also hit — after Meta signalled a "Meta Compute" plan to resell its own excess GPU capacity. In one move, one of the sector's largest customers became a potential competitor, and investors finally priced the risk that has always sat under the neocloud model: a business built on renting chips to a handful of anchor tenants is fragile when those tenants start reselling compute themselves. Layer in Amazon's expanding Trainium silicon and reports of GPU-rental rates falling 30% or more, and the pressure on pure-rental margins is real. The counter-signal is that private money hasn't blinked: Together AI just raised $800 million at an $8.3 billion valuation, with Aramco and Nvidia in the round.
China ships an open frontier coder — on its own chips
The most consequential model release didn't come from a Western lab. Meituan, the Chinese food-delivery giant, open-sourced LongCat-2.0, a permissively (MIT) licensed 1.6-trillion-parameter agentic coding model with a million-token context that reportedly edges out GPT-5.5 on SWE-bench Pro and has been topping usage on OpenRouter. The headline isn't just the benchmark — it's the claim that the model was trained end-to-end on roughly 50,000 domestic Chinese chips using Huawei's software stack. If that holds, it's evidence that export controls are leaking, that near-frontier open models can be trained outside the Nvidia ecosystem, and that the best open coder in any given week can now come from an unexpected place.
AI reads the genome's dark matter
The deeptech story to sit with is DeepMind's AlphaGenome, a model that predicts how mutations in non-coding DNA — the roughly 98% of the genome that codes for no protein — affect gene regulation, reading long sequences at single-base resolution. That "dark matter" of the genome has been the hardest part of linking DNA to disease, and a foundation model that can predict its regulatory effects is the kind of fast surrogate for an otherwise intractable computation that keeps redrawing what's possible in biology.
The money view, and what to watch
The through-line across all five stories is a rotation: capital is moving away from simply renting compute, toward owning silicon, and above all toward owning talent. The neocloud selloff, the hyperscalers' in-house chips, and Meta's raids on Apple and Google are three faces of the same shift — the scarce resources at the frontier are now people and silicon, not just spend. Watch three things next: whether Gemini 3.5 Pro finally reaches general availability in July after the delay that cost Alphabet dearly; whether the neocloud selloff spreads as more hyperscalers move to resell compute; and how Apple responds to losing its AI chief — a reorganization, an acquisition, or a bet on someone else's model. Signal, not advice.