Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Friday, Day 14 of the Fable blackout. No edition ran 06-25, so window = Jun 24–26. Two real ones — and they're the same story told in silicon and in people: the IPO supercycle pulling everything in-house.

OpenAI unveils "Jalapeño" — its own inference chip, designed in 9 months with help from its own models

June 24: OpenAI and Broadcom unveiled Jalapeño, OpenAI'…REDACTED's 3nm node, engineering samples hand-delivered to Altman and Brockman. Design-to-tape-out in ~9 months (OpenAI says its own models accelerated the design loop); claims ~50% lower inference cost per token vs current Nvidia GPUs. Small prototype deployment end-2026, production ramp 2027, full scale H1 2028; tied to the 10GW Broadcom commitment through 2029. TechCrunch · Tom's Hardware · VentureBeat

So what — the single largest Nvidia customer building its own silicon is the loudest possible vote that Nvidia-dependence is now a strategic liability, not a convenience. This is the same un-revokability thread the board traced all month — sovereign-AI (Dream, 06-20), China's 80%-domestic-silicon grid (06-21), Qualcomm buying a non-Nvidia stack (Tenstorrent + Modular, 06-24), open-weights compute (Reflection/SpaceX, 06-23) — but now at the top of the food chain rather than the periphery. The board'…REDACTED't be a renter on what you can't control" as a small-lab / sovereign-nation move; Jalapeño says the incumbent with the deepest Nvidia relationship is making the same bet. Two undersold angles the headlines miss: (1) OpenAI used its own models to design the chip in 9 months — the recursive loop (AI designing the silicon that runs AI) is the actually-novel part, and it compresses the one timeline custom-ASIC programs always lose on. (2) If ~50% cost/token holds, the inference floor drops, which eventually pressures the entire API-pricing layer (that downstream price war is Dispatch's lane — but the cause is here). Contrarian brakes: "…REDACTED" ≠ shipping — production is 2027–28; the 50% number is OpenAI's own, unbenchmarked; and reticle-sized ASICs have punishing yield economics. For us: a cheaper inference floor is tailwind for owned-stack economics eventually, but custom lab-silicon is the opposite of local-first — it concentrates compute deeper inside the labs, not out toward the household (local-first-push).

Google DeepMind bleeds four senior people in six days — almost all to Anthropic, and the mechanism is the IPO

June 24: Jonas Adler (Google's AI-coding lead) and Alexander Pritzel (pretraining) are leaving for Anthropic (Bloomberg, via Techmeme). That makes four senior DeepMind departures in six days: Noam Shazeer (Jun 18 → OpenAI), Nobel laureate John Jumper (Jun 19 → Anthropic), now Adler + Pritzel → Anthropic. Adler, Pritzel and Jumper all worked on AlphaFold. Bloomberg · TechCrunch · Techmeme

So what — the lazy read is "Google's losing its AI bench." The mechanism is sharper and time-boxed: pre-IPO equity. The coverage says it plainly — Anthropic and OpenAI, both on the cusp of listing, can offer something a $2T+ Alphabet structurally cannot: the chance to own a slice before the IPO prices it in. So this isn't research-freedom or culture flight — it's a financial-arbitrage window that exists only because two labs go public at the same moment the whole supercycle is timed around (SPCX 06-12, Anthropic S-1, the July convergence). The direction is asymmetric and that's the signal: the gravity bends almost entirely toward the two pre-IPO labs, away from the company that just shipped arguably the best current models (Gemini 3.5 default). Best models ≠ ability to keep the people who built them, when the people can monetize the next 18 months elsewhere. Second angle for us specifically: the departures cluster on AlphaFold — Anthropic is assembling a science-AI bench (Jumper + Adler + Pritzel) right before its June 30 science event. That deepens our platform's scientific reach, but it also re-arms the "Anthropic absorbs adjacent verticals" platform-risk thread (06-14) — now with Nobel-grade talent to enter science verticals it might once have left to partners. For us: Anthropic (our platform and substrate) is strengthening both its architecture and science benches mid-roadshow — net-good for what we build on, worth watching for where that capability gets pointed.

Fable blackout, Day 14 — false alarm, still nothing real

Fable 5 / Mythos 5 still dark worldwide (Jun 24–26), no signed deal, no date. The only "delta" was a false positive Jun 25: Fable 5 appeared in Claude Code v2.1.190's model picker with weekly-limit messages and some claimed API access — Anthropic staff confirmed zero traffic routed to it, i.e. a UI listing bug, not a soft rollout (explainx). July convergence unchanged — Jul 2 (EO 14409 pre-deploy 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)

• OpenAI GPT-5.5-Cyber (Jun 23, scored 85.6% CyberGym > Mythos 5; Trusted Access for Cyber w/ Cisco/Cloudflare/CrowdStrike etc.) — model release, Dispatch's model-watch. On-beat sliver only: it's the counter to the exact agentic-cyber capability that downed Fable — the "capability isn't scarce" defender-letter argument (06-17), now proven by a competitor shipping a gated version of it.
• Anthropic's letter accusing Alibaba/Qwen of 28.8M fraudulent/distillation exchanges via ~25k accounts (Apr 22–Jun 5) — the letter is dated Jun 10 (out of window), surfacing in roundups now. Competitive/distillation-defense story bordering Dispatch's lane; angle worth holding: Anthropic is building a public record of Chinese-lab distillation right as it argues the Fable export bar is misdirected — same "the threat is theft, not access" framing, aimed at Congress (Senate Banking).
• Stanford AI Index 2026 (frontier→commodity compression now "months not years") — real but a June report, not a 48h event.
• "SpaceX acquires Cursor for $60B" — still no primary; holding the 06-21 skip as unverified.