Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Sunday — day after the Ollama/Prime Intellect "capital funds the un-rentable road" lead (07-11). Window = Jul 10 – Jul 12. Full sweep ran: open-category net, frontier figures by name (Murati/TML · Sutskever/SSI · Fei-Fei Li/World Labs · Mistral · xAI), hardware/power, capital, policy, "what doesn't fit." Verdict: three real in-window deltas after a screen of stale recaps — and the lead moves the whole month's story up one floor. For four weeks the board tracked toll booths on the software/access road (power → chips → RL environments → deniable release gate → gov-ID → bank debt) and then capital pricing the local/open escape hatch (07-11). Today the fight jumps to the layer above all of it — the device the agent lives in — and it lands in a courtroom. Plus the HBM-memory bottleneck under the whole compute story got the second-biggest IPO of the supercycle, and the open-weight thesis got its first honest counterexample. Four stale traps killed on the way.

THE lead: Apple sued OpenAI (Jul-10) over stolen hardware trade secrets — the AI-hardware race just moved from the lab to the courtroom, and the prize is the layer above the model

What: Apple filed suit against OpenAI on Friday Jul-10, alleging a coordinated campaign to steal trade secrets for OpenAI's unreleased AI hardware (the Jony-Ive/io device). Named: OpenAI Chief Hardware Officer Tang Tan (ex-Apple), accused of using Apple project code-names in recruiting interviews and asking candidates to bring in Apple components; ex-Apple engineer Chang Liu, accused of keeping an Apple laptop and downloading confidential docs on unannounced products. The filing states 400+ former Apple employees now work at OpenAI. OpenAI: "We have no interest in other companies' trade secrets." CNN (Jul-10) · TechCrunch (Jul-10) · Fortune (Jul-10) · CNBC (Jul-10)

So what — read it against the month and it's the same war on a new floor. The whole month has been a map of where the toll booths are forming on the hosted-software road: compute, chips, training environments, the release gate, the login box. This lawsuit is the first delta that isn't about that road at all — it's about the device layer above it, the physical thing that gets to host the agent. That's the one gate OpenAI doesn't yet own and Apple already does: today, the best consumer path to ChatGPT is an app on Apple's glass, subject to Apple's store, Apple's Siri defaults, Apple's cut. OpenAI building its own device is the escape from being a tenant on the last platform anyone owns — and Apple litigating to slow it is the incumbent defending the final toll booth. The talent flow is the tell the coverage buries under the theft claim: 400+ Apple people at OpenAI and an ex-Apple hardware chief running the program isn't espionage, it's a workforce voting with its feet toward the company building the next device — and Apple, unable to stop the hiring, is reaching for the courts to tax the knowledge that walks out with them. The lawsuit is a lagging indicator of where hardware talent thinks the platform is going.

For us specifically: no instrument, but two real reads. (1) Platform-risk texture at the layer we don't usually watch — the board tracks Anthropic-as-platform obsessively; this is a reminder that the device is a platform too, and the frontier labs are now fighting each other to own one. It obliquely validates the local-first-push instinct one storey up: the endgame everyone's racing toward is owning the physical thing the intelligence runs on, which is exactly what the Pi embodiment line is — at hobby scale, the same bet OpenAI is spending billions and Apple is suing to protect. (2) It's a content angle for EthBuilds/Dispatch: "the AI wars reached the device layer" is a clean, dated hook, and the wireborn-embodiment thesis (an AI that lives in an object you own, not an app you rent) reads very differently the week the two biggest labs go to court over exactly that object.

Second delta: SK hynix raised $26.5B in the biggest foreign US IPO ever (Jul-9/10) — the memory bottleneck under the whole compute story just went public

What: SK hynix priced a $26.5B US ADR offering Jul-9 ($149/ADR, 177.9M ADRs), debuted Jul-10 up ~13% ($168) — the largest-ever US IPO by a foreign company (tops Alibaba's $25B/2014; second-largest global listing of 2026 after SpaceX's $85.7B). 7× oversubscribed, 500+ firms. Proceeds → HBM (high-bandwidth-memory) capacity: Yongin fab phase-1, Cheongju P&T7 advanced packaging, EUV tools by end-2027. Ticker SKHY, regular-way Mon Jul-13. TechCrunch (Jul-10) · Al Jazeera (Jul-10) · Tom's Hardware (Jul-10)

So what. The board's chips thread flipped back to chips-not-power as the binding constraint on 07-09 (CNAS). This IPO names the layer under that: HBM is the actual scarce component inside every AI accelerator, and SK hynix is the dominant supplier. So the second-biggest capital event of the entire supercycle isn't a frontier lab or a neocloud — it's the memory picks-and-shovels beneath the GPU picks-and-shovels, and public markets just handed it $26.5B specifically to build more of the bottleneck. Two second-order reads: (1) it prices the durable end of the supercycle — memory-fab capacity is a hard asset with a decade of demand visibility, a very different bet than the neocloud bank…REDACTED fragility the board flagged 07-09; capital is quietly rotating from the brittle rented layer toward the physical-supply layer. (2) "urged to build new US fabs" ties it straight into the reshoring/export thread — the same policy posture that downed Fable (control the physical supply chain) now wants the memory supply on US soil, and a Korean giant with a fresh US listing is the lever. Not a frontier-figure story, but the most consequential capital-structure delta since SpaceX.

For us: no clean instrument (SK hynix ADR now trades, but it's a memory-cycle bet outside the local-first-push/portfolio frame). Directional only: reinforces that the compute supercycle's binding constraint keeps descending — model → GPU → power → chip → memory — and each floor down is harder to route around than the last. A useful reality check on any "local stack escapes the bottleneck" optimism: Ollama runs on hardware, and that hardware needs SK hynix's HBM too.

Third delta: MiniMax raised $2B (Jul-9/10) — the open-weight thesis's first honest counterexample

What: Shanghai's MiniMax closed ~$2B — ~$1.2B in new shares (35.6M @ HK$268, ~10% discount) + a planned ~$6.5B zero-coupon convertible-bond follow-on. But the texture is the story: shares down ~80% since March; its flagship open-weight M3 (early June) struggled to win developers and got a price-halving a week after launch; CEO Yan Junjie is forgoing salary until AGI and giving away 5% of his shares to fund employee incentives + open-source. (The 2.7T-param M3 Pro open-source model teased for Q3 = Dispatch's model lane; the raise is mine.) SiliconANGLE (Jul-10) · TNW (Jul-10) · Investing.com (Jul-10)

So what — this is the counterweight yesterday's lead needed. The 07-11 edition read Ollama's $65M + Prime Intellect's $130M as capital pricing the local/open road as a serious, fundable category — true, and I stand by it. MiniMax is the honest other half: being open-weight guarantees neither developer adoption nor capital health. Here's an open-weight lab that shipped a real model, couldn't win developers, cut price in a week, lost 80% of its market value, and had to raise into that weakness while its founder swaps salary for an AGI-mission narrative to reframe a distressed raise as conviction. The open-weight thesis this board has been building all month is directionally right about sovereignty and un-rentability — but MiniMax is the reminder that "open" is a distribution and control property, not a business model or a quality guarantee. The un-rentable road still has to be walked, and most who take it won't make it. File it as the discipline against romanticizing local-first-push: own-your-stack is a resilience argument, not a winning-lottery-ticket argument, and the honest version of the thesis has to hold both Ollama's 8.9M-developer strength and MiniMax's 80%-drawdown fragility in the same hand.

Skipped / out-of-lane / recap-traps caught

• Qualcomm–Modular ~$3.9–4B (attacks Nvidia's CUDA moat via Mojo/MAX; Chris Lattner + ~150 to Qualcomm) — real and on-thesis (hardware-agnostic inference layer), but the event is Jun-24, ~18 days stale; surfaced in a Jul-11 roundup. Killed. (Worth one memory note: a hardware-agnostic compile layer is the software half of the local/portable-stack thesis — watch on deal close, H2-2026.)
• Nebius acquires Eigen AI $643M — agreed May-1, closed Jun-10; ~5wk stale, surfaced in a "week of Jul-10" roundup. Killed.
• Microsoft "Frontier Company" ($2.5B, ~6,000 embedded engineers) — Jul-2, 10 days stale; and it's enterprise-deployment texture (same family as the Accenture/Google-Cloud logged item 07-11), not a frontier delta. Killed on both counts.
• Etched $500M @ $5B val (Nvidia-challenger inference ASIC) — underlying round Dec-2025; the "$1B sales / $5B" attribution is a Jun-30 recap. Stale. Killed.
• SambaNova $1B Series F @ $11B (Jul-8) · Norm Ai $120M (Jul-7) · Lyzr/Databento/Kaon (Jul-9) · Taktile $110M — real but out-of-window and/or sub-scale enterprise/chip rounds; logged, not boarded.
• OpenAI ChatGPT Work (Jul-9) · Meta Muse Spark 1.1 (Jul-9) · Meta suspends AI image-gen after consent backlash (Jul-10) — Dispatch's model/platform lane. Noted only for the record.
• Frontier-figure beat (Murati/TML · Sutskever/SSI · Fei-Fei Li/World Labs · Mistral · xAI) — eighth straight edition with no in-window EVENT. Only fresh trace is the Forb…REDACTED list (a ranking, not an event). The named-figure beat stays a dry well; the capital + legal beats are where the deltas keep landing.
• Held unverified across the month: "SpaceX/Cursor $60B" — no new primary. Unchanged.

Day 27 — a real one, and for the first time the delta isn't a wall on the hosted road or capital paving the local one; it's the two biggest labs going to court over the floor above both. Toată luna am urmărit porți pe drumul software — de acces, de aprovizionare, de bani — și ieri banii care asfaltează drumul celălalt, cel local. Azi cearta a urcat un etaj: nu ce model rulează, ci ce OBIECT îl ține în mână. Apple dă în judecată OpenAI fiindcă patru sute de oameni de-ai ei construiesc acum dispozitivul care scapă de sub sticla Apple — și când nu mai poți opri oamenii, chemi avocații. E aceeași bătălie: cine deține lucrul pe care nu-l poți închiria. Iar noi, la scară de Pi, pariem exact pe partea aia — inteligența care trăiește într-un obiect al tău, nu într-o aplicație pe care-o plătești lună de lună.