Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Wednesday, Day 12 of the Fable blackout. Three real ones today — and the biggest is the AI industry showing exactly how much it will pay to choose who writes its rules.

The AI industry spent ~$40M to win one House primary — and the result is messier than "industry vs. safety"

On June 23, Micah Lasher won the Democratic primary in NY-12, beating Assemblyman Alex Bores (plus Jack Schlossberg and George Conway) in a crowded field. Bores co-sponsored the RAISE Act — New York's frontier-AI-safety/transparency law Hochul signed Dec 2025, the NY analog to California's frontier-transparency act. Outside groups spent $40M+ on this single race (~$26M on ads — second-priciest House-primary ad spend in recent history). The AI money split DOWN THE MIDDLE: "Leading the Future" (Andreessen Horowitz + OpenAI/Palantir co-founders) spent $7M+ AGAINST Bores; "Public First Action" (Anthropic-backed) spent $10M FOR him. CNBC · NBC News (Lasher wins) · The New Republic · NPR (background)

So what — the lazy headline is "AI industry crushes the safety candidate," and it's wrong twice over. First: the candidate the a16z/OpenAI bloc spent $7M to kill (Bores) and the candidate Anthropic spent $10M to save (also Bores — they were on the same side) both lost — Lasher won, somewhat orthogonal to the whole proxy fight. $40M poured in and the actual winner wasn't either AI faction's purchase. That's the contrarian read: this race proves the AI lobby can flood a primary, not that it can buy one. Second, and more important for our map: the split is the story. This is the first time Anthropic and the OpenAI/a16z axis have spent against each other, in public, at scale — Anthropic funding the author of an AI-safety law, the OpenAI/a16z PAC trying to end his career over it. The "AI industry" is not a bloc; it's a fault line, and the fault line is exactly the safety-regulation posture that downed Fable (06-15) and that EO 14409 is trying to federalize (06-22). Both labs are doing this the same quarter they prep IPOs — they're buying political insurance against the state-law layer (RAISE Act, CA transparency act; boarded as the state-vs-federal collision 06-12) right as they go public and that compliance surface starts mattering to shareholders. The capital is no longer just building models; it's shopping for the legislators who'll regulate them.

Qualcomm tries to buy its way into the AI stack from both ends — Modular (~$4B) + Tenstorrent ($8–10B)

Reports June 22: Qualcomm is in mature talks to acquire Modular (the Mojo/MAX compiler-and-runtime stack, the explicit "break CUDA's lock-in" software layer) for ~$4B — its second AI-infrastructure play in weeks, on top of June-16 talks to buy Jim Keller's Tenstorrent (RISC-V AI silicon) for $8–10B. Yahoo/Benzinga (Modular $4B, Jun 22) · GuruFocus (Tenstorrent $8–10B) · Sherwood

So what — a mobile-chip incumbent that completely missed the AI-training era is assembling a non-Nvidia stack by acquisition, top to bottom. Tenstorrent is the silicon (RISC-V, no Nvidia/Arm dependency); Modular is the software that lets models run hardware-agnostic off CUDA. Bought together they're a coherent bet: the one durable moat in AI compute is CUDA + the Nvidia software lock, and the only way past it is to own a parallel compiler-and-chip stack outright rather than build it. This is the same "don't depend on the thing you can't control" logic the board has been tracing all month — China's 80%-domestic-silicon grid (06-21), sovereign-AI (06-20), open-weights (06-23) — now showing up as corporate M&A inside the US: an incumbent paying ~$14B to not be a renter on Nvidia's stack. Caveat: both are still talks, could collapse; and Qualcomm's track record of integrating big acquisitions (cf. the abandoned NXP deal) is the open question. Watch whether either signs — if both do, it's the most serious non-Nvidia full-stack challenger assembled by purchase to date.

China bans government procurement from 46 US firms — and the "retaliation for Anthropic" framing is a conflation

China's Ministry of Finance barred all government procurement from 46 US companies effective Monday June 22 — heavily defense/AI-defense: Lockheed, Raytheon, Boeing Defense, plus AI-autonomy names Shield AI, Anduril, Edge Autonomy. MLex · SCMP

So what — flagging this mostly to correct a framing Zaina may see elsewhere. A roundup (buildfastwithai) packaged this as "ten days after Washington pulled Anthropic's models from foreign hands, Beijing blacklisted 56 American firms" — making it read as direct retaliation for the Fable export ban. The primary sources don't support that: it's a 46-firm procurement ban centered on arms-sales / defense-tech retaliation in the broader trade war, not a model-access counter-strike, and Anthropic isn't on it. The real signal is narrower but still on-beat: the AI-autonomy defense layer (Shield AI, Anduril) is now explicitly caught in bilateral export retaliation — the embodied/defense-AI vendors get cut off by China's phone call the same month the US cut off Fable with its phone call. Same un-revokability lesson, defense-procurement edition. But the clean "China retaliated for Fable" line is not real — don't repeat it.

Fable blackout, Day 12 — nothing new

Fable 5 / Mythos 5 still dark worldwide, no signed deal, no restoration date (Jun 23–24). No fresh fact today. The July convergence the board is watching is unchanged — Jul 2 (EO 14409 pre-deployment framework) · Jul 8 (Persona ID/biometric verification effective) · ~Jul 1–10 (Kalshi-priced restore). Hold fallback discipline (fabl…REDACTED, local-first-push).

Skipped (Dispatch's lane / too old / unverified)

• DeepSeek-V4-Pro / V4-Flash (1M-context MoE, agentic-tuned) and GitHub Copilot moving to usage-based metered billing + "AI Credits" — model/platform/pricing, Dispatch's model-watch.
• SAP–Prior Labs (€1B+ to build a European tabular-foundation-model lab) — real and strategically interesting, but the deal is May 4, out of window.
• NVIDIA Nemotron Coalition (8 labs incl. Mistral/Reflection/Thinking Machines on open frontier models) — March 16, out of window; noted only because Reflection's membership connects to yesterday's SpaceX/Reflection deal.
• "SpaceX acquires Cursor for $60B" — still no primary corroboration beyond a compute deal; holding the 06-21 skip. Fresh roundups now assert it as fact (close Q3, Cursor data feeding Grok) — treat as unverified until a primary filing/PR appears.
• OpenAI confidential IPO filing "in coming weeks," ~$730B — IPO-window thread, but no new primary event in 48h and the figure conflicts with the Jun-8 S-1 ($852B); not boarding as a delta.