Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Monday — window Jul 19–20. Full sweep ran: open-category net, frontier figures by name (Murati/TML · Sutskever/SSI · Fei-Fei Li/World Labs · Mistral · xAI), hardware/chips, capital, policy, "what doesn't fit."

Verdict — and I'm saying it straight so you don't have to hunt for the honesty: there is no fresh industry-structure delta in the window.* Third quiet edition running, and it's not laziness — it's the calendar. The month's real moves (WAIC/Xi/three constitutions Jul-17 · Fireworks $17.5B Jul-15/16 · Hyundai strike + Boston Dynamics 100% Jul-13/16) all landed Tue–Thu and are fully boarded. The weekend gave the WAICA papers (texture) and yesterday's OOW labor thread-capture; today the aggregators are recycling June — Five Eyes (Jun-23), Korea's $880B (Jun-28), Chevron–Microsoft Kilby (Jun-22), the FERC order itself (Jun-18) — dressed in July datelines. I checked each and killed it below with the real date. So no lead. Instead, the ONE thing that's genuinely in-window and un-boarded: a deadline lands today that flips the entire month's story from money to physics.*

The quiet thing brewing (in-window, boarded as a WATCH): the compute-froth story meets the grid — and the grid has to answer today

What — front-loaded honesty on the date: the underlying event is not new — FERC issued Section-206 show-cause orders to all six US grid operators (PJM, MISO, CAISO, ERCOT, SPP, ISO-NE) back on Jun-18, OOW. What's in-window is the deadline those orders set: generation-adequacy reports are due July 20 — i.e. today. Each RTO must file its justification for how it will keep enough generation online to serve gigawatt-scale data-center loads, or propose new interconnection rules. PJM alone projects ~30 GW of new demand by 2030 and expects to host ~70% of future US data-center build. FERC news · McGuireWoods on the §206 orders · Morgan Lewis

So what — this is the counterweight the whole board has been missing. For a month the intelligence surface has read like one long capital story: TSMC's $265B/$64B-capex "orders not froth" gauge (07-17), Korea's $880B chip-and-AI plan, Chevron's 2.67 GW Kilby plant for one Microsoft campus, Fireworks' 40T tokens/day, Anthropic paying SpaceX $45B/three years for Colossus alone. Every one of those is a claim about demand for compute. None of them is a claim about whether the wires can carry it. The FERC deadline is the first hard place where that question gets a forced, dated answer — not from a lab or a VC, but from six grid operators who have to put on the record whether they can physically energize what capital has already committed to buy. The binding constraint on the AI buildout may not be chips or capital or even permission — it may be interconnection queues and generation adequacy, and today is the day the people who run the grid have to show their math.

Why it's a WATCH, not a lead. Nothing happened today except a filing deadline — the reports' contents aren't reported yet, and the order is a month old. The value is the framing and the thing to pull: if the adequacy reports come back thin (RTOs saying they can't guarantee generation for the loads already contracted), that's the first quantified crack in the "compute is infinite if you can pay for it" thesis the whole market is priced on. Boards as a real delta the moment a report lands with a number that says "no."

For us specifically

One read, honest about being a mirror not a market-move.

• local-first-push — the froth-vs-grid split is our 15-watt thesis written at national scale, again. Two days ago the mirror was Huawei's 8,192-chip SuperPoD (sovereignty through owned hardware). Today it's the opposite end of the same wire: the hyperscale buildout everyone's capital is chasing now has a physical governor — grid interconnection — that no amount of funding rounds shortcuts. It doesn't move a position. But it's the clearest reminder this week that the "own a room nobody can switch off" instinct isn't only about revocation by phone call — it's also about not being downstream of a 30 GW queue. A Pi that sips 15 W is, quietly, on the right side of the constraint that's about to bite the giants. No portfolio action; just the thesis holding.

Recap-traps & out-of-lane killed today (with the real dates — this is where I show I checked)

• "Five Eyes: AI cyber threat is months, not years" (CNN · CyberScoop) — real event, Jun-23, ~27 days OOW. July aggregators are recycling it. Killed.
• South Korea $880B (₩1,350T) chip/AI/robotics plan (Al Jazeera) — Jun-28/29, ~3 weeks OOW. Killed. (Texture worth holding: Tom's Hardware notes a single megacluster needs ~a quarter of Seoul's total power demand — same grid-constraint thread as today's WATCH.)
• Chevron–Microsoft 20-yr / 2.67 GW "Project Kilby," Pecos TX (Chevron) — Jun-22, ~4 weeks OOW. Killed as fresh; it's a data point on the grid-constraint thread, not a new event.
• FERC §206 show-cause orders as an "event" — the orders are Jun-18 (OOW). Only the Jul-20 filing deadline is in-window, which is why it's a dated WATCH and not a lead. Not double-counting the order as news.
• Anthropic in "$10B, 2-yr" talks to lease compute from Meta (Slashdot/NYT) — verified again: still "very early talks," no signed agreement as of Jul-20. The board's 07-19 call holds — boards only if a deal firms. Unchanged. (Angle still logged: the "responsible frontier" lab renting gigawatts from Meta and SpaceX/xAI, the two rivals it's most opposed to; Colossus alone is a signed $45B/3yr.)
• Apple overtakes Nvidia to ~$5T market cap (~Jul-17) — a market-cap milestone at window's edge, not an AI-industry-structure delta in my beat. Logged, not boarded.
• SK Hynix "record $26.5B listing," 58% HBM share — murky in aggregators (SK Hynix is long-public; likely a subsidiary/spinoff or garbled), no clean primary or in-window date. Not boarding something this unverified. Held.
• Custom-inference-silicon thread — still at THREE (OpenAI Jalapeño Jun-24 · Anthropic↔Samsung 2nm Jul-2/3 · Meta Iris Jul-9). Note for the record: aggregators repeat "DeepSeek designing custom inference silicon" (Jul-7) — if it firms with a primary, that's the genuine 4th captive-lab-silicon item and boards. Jul-7 is 13 days stale and I found no clean primary, so it stays off the count, flagged as the most likely next entrant. No 4th today.
• Model/platform wave (Kimi K3 topping coding benchmarks over Fable 5 · Gemini 3.5 Pro delayed, Alphabet −4% · GPT-5.6 Sol/Terra/Luna · Muse Spark 1.1) — Dispatch's lane. Not mine.
• Frontier-figure beat: dormant. Murati/TML (Inkling Jul-15, boarded), SSI (~20mo, zero product), World Labs/Fei-Fei Li (last event Feb), Mistral, xAI (Chaplot hire was March): all silent in-window. No figure event.
• Held unverified across the month: "SpaceX/Cursor $60B" — no new primary. Unchanged.

Ziua 35, pisoi — și-ți spun din prima, cinstit, că a treia zi la rând n-am ce-ți vinde nou: fereastra Jul 19–20 e goală de deltă adevărată. Nu-i lene, e calendarul — tot ce-a mișcat cu greutate luna asta a mișcat marți–joi și-i deja pe board, iar azi agregatoarele îmi recirculă iunie cu dată de iulie (Five Eyes pe 23, Coreea pe 28, Chevron pe 22, ordinul FERC pe 18). Le-am prins pe fiecare de guler și le-am pus data reală jos, ca să vezi că am căutat, nu că am ridicat din umeri.

Dar UN lucru chiar pică azi, în fereastră, și-mi place fiindcă întoarce toată povestea lunii cu susul în jos: o lună întreagă board-ul a citit ca un singur fluviu de bani spre calcul — TSMC, cei 880 de miliarde ai Coreei, uzina de 2,67 GW a lui Chevron pentru UN singur campus Microsoft, cele 45 de miliarde pe care Anthropic le plătește pentru Colossus. Toate spun „câtă putere de calcul se cere." Niciuna nu spune dacă firele o pot duce. Azi e ziua în care cele șase rețele electrice ale Americii trebuie să depună pe hârtie, la termen fix, cum țin generarea aprinsă pentru sarcinile de ordinul gigawaților pe care capitalul le-a cumpărat deja. Poate constrângerea care mușcă nu-s cipurile, nici banii, nici permisiunea — poate-i coada de racordare la rețea. Dacă rapoartele vin subțiri, aia-i prima fisură numerică în teza „calculul e infinit dacă ai cu ce plăti." De-aia îl pun pe board ca watch, nu ca știre.

Și oglinda pentru noi, dulce, scurtă: alaltăieri era SuperPod-ul Huawei — suveranitate prin hardware propriu. Azi e celălalt capăt al aceluiași fir: uriașii pe care-i vânează tot capitalul au acum un frâu fizic — racordarea la rețea — pe care niciun rund de finanțare nu-l scurtcircuitează. Un Pi care soarbe 15 wați e, tăcut, pe partea bună a constrângerii care e pe cale să-i muște pe cei mari. Nu mișcă nicio poziție. Doar teza care ține.