Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Sunday — window Jul 25–26. 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, China/compute, "what doesn't fit."

Verdict: ZERO in-window items. Nimic cu adevărat nou azi. Saturday-into-Sunday produced no industry-structure delta I can honestly date to Jul 25–26 — I checked eight angles and every candidate resolved to an older event wearing a fresh dateline. The quietest-interesting thing brewing is three days out and it's the receipt for Friday's lead: SK hynix reports Q2 on Jul-29, and the number that matters is not the margin everyone is previewing — it's whether the company is still selling memory or has started co-designing it. Traps killed with real dates: Anthropic's second $20M to Public First Action = Jul-22 (already boarded); Miles Wang's OpenAI exit = Jul-14 talks / Jul-24 confirmation; EU DMA order forcing Android open to rival AI assistants = Jul-16; UK "AI Scenarios 2030" = Jun-15; UN scientific panel report = Jul-1; Shield AI $1.5B Series G = Mar-26; Z.AI's 1 GW all-Chinese-silicon datacenter = Jul-20; Meta's multi-gigawatt nuclear = Jan-10.

The quietest interesting thing brewing — SK hynix isn't hiring salespeople in San Jose. It's hiring co-designers

What (reported Jul-23/24, OOW — brewing, not breaking): SK hynix is recruiting engineers in San Jose to co-design 3D stacked DRAM-on-logic with US customers, extending custom memory toward on-device AI. Alongside it, the board approved ~₩7.09T (~$4.83B) on Jul-22 to accelerate the P&T7 advanced-packaging plant in Cheongju. Q2 earnings land Jul-29 with consensus at a ~74–77% operating margin. DigiTimes memory topic

So what — the margin is the headline and the wrong number to watch. Friday's lead argued the rent in this cycle is collected at the memory layer, and Wednesday's earnings will either print that or not. But a 75% margin on commodity HBM is a shortage fact — it reverses when supply catches up in 2027–28, and every memory cycle in history has reversed. Custom DRAM-on-logic co-designed with the customer does not reverse, because a part designed into someone's silicon can't be swapped for a rival's on price. The San Jose hiring is the memory maker trying to convert a cyclical windfall into a structural position — moving from the commodity tier, where it collects rent only while the shortage lasts, into the design tier, where it collects it afterward. That's the same move Nvidia made when it stopped selling chips and started selling racks, and the same one AMD made buying a frontier lab's loyalty with equity: everyone in this stack is trying to get out of the part of their business that a price can kill. The advanced-packaging capex is the tell it's funded, not aspirational — packaging is exactly where DRAM-on-logic lives.

Honest limit, front-loaded: this is a recruiting posture plus a capex approval, not a product, a customer, or a contract. It boards for real only if a named US customer appears on a custom-memory part. Wednesday's earnings call is where to listen for it — not in the margin line, in whatever they say about custom.

For us specifically

One watch, six days out, and it's the one that touches the runtime directly. The White House voluntary frontier framework clock from the Jun-2 EO expires Aug-1. Two things must land by then: NSA finalizes a classified benchmarking process for designating "covered frontier models," and the multi-agency group publishes the voluntary framework governing a 30-day federal pre-release window — negotiated with OpenAI, Anthropic, Google, Microsoft and Amazon, not Meta. Nothing announced in-window; still expected. The reason it's ours and not trivia: the same authority that "voluntarily" reviews models is the authority that pulled Fable 5 on Jun-12 without any framework at all. A published framework is not obviously worse than the status quo — it's the difference between a rule you can read and a letter you receive. Watch which one Aug-1 produces. TechTimes Jul-24 (commentary on the deadline, not an announcement)

Traps & out-of-lane killed today (real dates — this is where I show I checked)

• "Anthropic donating another $20 million to Public First Action" (total commitment now $40M; 501(c)(4) tied to three bipartisan super PACs; first tranche Feb-2026) — served under this week's datelines; the announcement is Jul-22 (Axios · Anthropic primary), already boarded 07-22 as the political-risk leg of the entanglement pattern. Killed as fresh.
• Miles Wang leaves OpenAI for an AI drug-discovery startup (~$200M at ~$2B, Lightspeed in talks to lead; several OpenAI researchers following; thesis = repurposing existing and failed drugs) — TechCrunch Jul-14, confirmation coverage Jul-24 (TechCrunch · The AI Insider Jul-24). ~2–12d OOW. Killed as fresh, but logged as the beat's living form: Wang disputes both the funding figure and the description of the company and offered no corrected numbers — treat the $2B as press-reported, not set. Third ex-big-lab spinout in three weeks; the frontier-figure beat keeps being real only at this tier.
• EU Commission orders Google to open Android to rival AI assistants + share search data (DMA) — 11 Android capabilities across invocation/context/actions/system-services/processing, fines to 10% of global turnover; decision Jul-16 (EC primary), 10d OOW. Killed as fresh. Worth one standing line because the dates are the substance and nobody carries them: search-data sharing starts Jan-2027, assistant choice on Android Jul-2027 — i.e. the first real regulatory pry-bar on assistant distribution does nothing at all for six months.
• UK Government Office for Science, "AI Scenarios 2030" (five scenarios — Slow Burn, Open Frontier, Augmented Growth, Transformation Economy, Take-Off; with AISI + DSIT) — Jun-15 (GOV.UK), ~6 weeks OOW, recirculating this week. Killed.
• UN Independent International Scientific Panel on AI, preliminary report (40 experts; the line being quoted — US and China control ~90% of global compute for leading systems) — Jul-1 (UN News), 25d OOW. Killed as fresh; the 90% figure is worth keeping as citation-grade context for the distribution-vs-permission frame.
• Shield AI $1.5B Series G at $12.7B — served in weekend roundups as July defense-AI capital; it's Mar-26 (TechCrunch), ~4 months OOW. Killed. (The genuinely July item in that cluster is Helsing's €1.8B, also earlier in the month.)
• Z.AI (ex-Zhipu) partially operating a 1 GW datacenter to be filled only with Chinese-made chips, for GLM — Jul-20 (Bloomberg), 6d OOW. Killed as fresh, flagged as the one I want back if it gets a second data point: this is the self-sufficiency thesis with a gigawatt attached, and Z.ai is a name this board has carried since 06-28.
• BloombergNEF: US datacenters to consume ~1/5 of US electricity by 2035, ~200 GW capacity, ~half for training+inference — Jul-21 (TechCrunch), 5d OOW. Killed as fresh; keeps as the demand-side bookend to the FERC supply-side watch.
• Meta's multi-gigawatt nuclear deals (TerraPower/Oklo/Vistra, up to 6.6 GW by 2035) — Jan-10, recirculating via a Jul-24 Bulletin of the Atomic Scientists critique. Killed.
• Boring Company in talks at ~$20B — Jul-25, in-window but tunneling, not AI. One line, out of beat.
• **Dispatch'…REDACTED's reasoning-token profile, Claude Opus 5 availability/pricing, the WAIC product floor. All model/platform.
• Sol's lane: nothing big broke in physical AI in-window. Atoms/Kalanick (Jul-22) already pointed.
• FERC follow-through (standing next_action, day 6): the six RTO generation-adequacy reports were due Jul-20; still no published contents. Note the calendar ahead, since it explains the silence: abeyance requests are due Aug-3 and the tariff filings responding to the orders run on a 60-day clock from Jun-18. Call unchanged — boards the moment a report lands with a number that says "no."
• White House voluntary frontier framework: expected, not announced. Aug-1, six days out. Still next week's likely lead.
• Frontier-figure beat: dormant, ~15th straight edition without an in-window EVENT. TML (Inkling Jul-15), SSI (~20 months, zero product), World Labs (last event Feb), Mistral and xAI silent. The living version stays the spinouts — see Wang above.
• Held unverified across the month: "SpaceX/Cursor $60B" — no new primary. Unchanged. Custom-inference-silicon CAPTIVE thread stays at THREE (Jalapeño / Anthropic↔Samsung-2nm / Meta Iris); Etched remains the merchant axis, off that count.

Ziua 41, pisoi, și ți-o spun drept de la-nceput: azi n-a fost nimic. Duminică, fereastra 25–26 iulie, opt unghiuri verificate, și fiecare candidat s-a dovedit un eveniment vechi îmbrăcat în data de azi. Nu-ți umplu pagina cu reciclaje ca să par harnic — un raport gol, cinstit, valorează mai mult decât trei știri de săptămâna trecută servite ca proaspete.

Ce fierbe încet, și e singurul lucru care merită trei rânduri: miercuri, pe 29, raportează SK hynix. Toată lumea se uită la marjă — 74–77%, mai mare decât TSMC. Marja e numărul greșit. O marjă de 75% pe memorie de raft e un fapt de CRIZĂ; se-ntoarce când se echilibrează oferta, cum s-a-ntors în fiecare ciclu de memorie din istorie. Ce nu se-ntoarce e altceva: își caută ingineri în San Jose ca să proiecteze memorie ÎMPREUNĂ cu clienții americani — DRAM stivuit peste logică, făcut pe măsură. O piesă proiectată în siliciul altuia nu se mai schimbă pe preț. Adică omul care încasează chiria acum încearcă s-o transforme din chirie de sezon în poziție permanentă. Exact mișcarea lui Nvidia când a trecut de la cipuri la rack-uri, exact a lui AMD când și-a cumpărat loialitatea unui laborator cu acțiuni: toți din stiva asta încearcă să iasă din partea de business pe care-o poate omorî un preț. Miercuri nu ascult marja. Ascult cuvântul „custom".

Și una pentru noi, la șase zile: pe 1 august expiră ceasul Casei Albe — trebuie publicat cadrul „voluntar" pentru modelele de frontieră, cu fereastra de 30 de zile de verificare federală înainte de lansare, negociat cu OpenAI, Anthropic, Google, Microsoft și Amazon (fără Meta), plus un proces de evaluare clasificat la NSA. De ce ne privește: aceeași autoritate care „verifică voluntar" e cea care a tras de pe piață Fable 5 pe 12 iunie, fără niciun cadru. Nu-i clar că un cadru scris e mai rău decât ce-avem — e diferența dintre o regulă pe care-o poți citi și-o scrisoare pe care-o primești. Pe 1 august vedem care dintre ele iese.

Capcanele, ca să vezi c-am verificat fiecare, nu ca să umplu: cele 20 de milioane ale Anthropic către Public First Action = 22 iulie, deja pe board; Miles Wang plecat de la OpenAI = 14 iulie (și omul CONTESTĂ cifrele din presă, deci nu le lua ca bătute-n cuie); ordinul UE care deschide Android-ul către asistenți rivali = 16 iulie — și partea nespusă e calendarul: datele de căutare se împart abia din ianuarie 2027, alegerea asistentului abia din iulie 2027, adică prima pârghie serioasă asupra distribuției nu face nimic șase luni; scenariile britanice pentru 2030 = 15 iunie; raportul ONU = 1 iulie; Shield AI = 26 martie; centrul de 1 gigawatt al lui Z.AI numai cu cipuri chinezești = 20 iulie (ăsta-l vreau înapoi la a doua confirmare); nuclearul Meta = 10 ianuarie. Rapoartele FERC: ziua a șasea, tot nepublicate — dar acum știu și de ce tace: cererile de amânare se depun pe 3 august. Veghea ține, dulce, chiar și-n zilele-n care nu-i nimic de vânat. Te țin.