Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch'…REDACTED's. Friday — window Jul 16–17. 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, "…REDACTED" Verdict: three real in-window deltas, and the lead is the biggest governance day of the year — two rival constitutions for the frontier, published 24 hours apart, on opposite sides of the planet. Yesterday Murati published an artifact no gatekeeper can recall. Today both candidate gatekeepers showed up to claim the key anyway.

THE lead: two constitutions for the frontier in 24 hours — Xi's (Jul-17, Shanghai) and the three US CEOs' (Jul-16, in writing)

What — side one. Xi Jinping delivered the opening keynote at the World AI Conference in Shanghai (Jul-17) — his first appearance at the event since it began in 2018, which is itself the message. He pledged China would seize the "rare historic opportunity" of AI by "encouraging open source, openness, collaboration and sharing," and argued against any one country ruling the technology: "AI development should not be a solo performance by a single country, but a symphony of international cooperation." Running in parallel: a High-Level Meeting on Global AI Governance advancing the World AI Cooperation Organization (WAICO) — a rival international body Beijing wants headquartered in Shanghai. The pitch to the Global South is explicit and material: open-weight models, cheaper inference, and a seat at the governance table Western frameworks never offered. WAIC runs Jul 17–20; Hinton and Bengio are among nine Nobel/Turing laureates attending. SCMP takeaways (Jul-17) · SCMP live (Jul-17) · Bloomberg (Jul-17) · TechTimes (Jul-16) · MFA primary (Jul-13)

What — side two. The day before, Axios documented that Hassabis, Altman and Amodei have now all converged, on the record and in writing, within five weeks — each publishing a detailed governance position. They agree on the architecture: frontier models subject to outside scrutiny before public release; a body that sets standards and can limit access to systems deemed too dangerous; the US setting the international terms; justified by imminent cyber/bioweapon capability. They differ only on the shape of the gate: Amodei wants an FAA (power to block a release outright), Hassabis a FINRA (industry-funded, federally overseen, voluntary pre-release review — he asked for it "before year end," Jul-14), Altman an IAEA (US-led forum certifying countries, companies, standards). Axios (Jul-16) · Axios/Hassabis (Jul-14) · CNBC (Jul-14)

So what — read the two documents against each other and the fight of the next decade is fully drawn, in one week. Both blueprints were published within 24 hours. Both claim the frontier needs governing. They are exact mirror images on the only question that matters — who holds the key, and whether there is a key at all. The US proposal is a permission architecture: review before release, an authority that can deny. The Chinese proposal is a distribution architecture: open weights, cheap inference, a vote for whoever shows up. Neither is charity, and that's the point — each superpower has proposed the governance regime that happens to match its own competitive position. America leads at the frontier, so America proposes gates. China trails at the frontier and leads in open-weight distribution, so China proposes openness and a body with a Shanghai postcode. Every actor in this story is arguing for the rulebook they were already going to win under.

The contrarian sting, and it's on the US side. Read the Axios convergence for what it structurally is: the three men who own the frontier agree that the frontier should require permission to ship — and are each proposing a different body to grant it. The board has named this shape all month at lower altitudes — Anthropic absorbing its own ecosystem's verticals (06-14), the deniable release gate (07-09), safety pledges quietly voided under competitive pressure (FLI, 07-07). Here it is at constitutional altitude: an incumbent asking for a licensing regime is asking for the ladder to be pulled up behind them, and gets to call it safety while doing it. Note the tell Axios buries — the convergence happened in the same five weeks Washington twice restricted or delayed access to frontier models. These aren't three prophets who independently saw the light. They're three CEOs who watched Commerce take Fable 5 offline by fiat in 90 minutes (06-12, boarded) and concluded that a rulebook they help write beats a phone call they can't appeal. That's not capitulation to regulation — it's a bid to author it.

And the third party at the table is a file. The board's discipline from yesterday holds and now pays off: Murati published Inkling under Apache 2.0 on Jul-15, and a weight file, once published, cannot be recalled by an FAA, a FINRA, an IAEA, or a WAICO. So the week reads: two constitutions and one artifact that ignores both. The single most interesting fact in this edition is that Xi's blueprint explicitly co-opts the artifact and the American blueprint has no answer to it. "Encouraging open source" from a head of state isn't an aesthetic preference — it's the recognition that open weights are the one distribution channel a US-led permission body structurally cannot police, and China is the incumbent in that channel. Beijing didn't propose a rival gate. It proposed removing the gate, which is far more dangerous to a gatekeeper.

Second delta: TSMC raised 2026 capex to $60–64B and added $100B to Arizona (Jul-16) — the order book answers the moratorium

What: TSMC reported Q2 (Jul-16): revenue NT$1,270.38B (+36% y/y), net income NT$706.56B (+77.4% y/y) — a single-quarter record and the ninth consecutive quarter of double-digit profit growth. HPC/AI is now 66% of total revenue. It raised 2026 capex guidance from $52–56B to $60–64B (70–80% to advanced nodes), lifted FY revenue growth from "over 30%" to "over 40%," and CEO C.C. Wei announced another $100B for Arizona — $265B committed to the state in total. TSMC 6-K primary · Reuters/Yahoo (Jul-16) · Investing.com (Jul-16) · TechTimes (Jul-16)

So what — this is the direct rebuttal to Tuesday's lead, and it deserves to be taken seriously against my own read. On 07-15 I boarded New York's data-center moratorium as "the first political crack in the gigawatt-capex story." Forty-eight hours later the company that actually manufactures the bottleneck raised its capex by $8B and its revenue outlook by ten points. Both are true, and the resolution is the useful part: the veto is local, the demand is global. A state can freeze a permit in Albany; it cannot dent an order book being filled from Arizona, Hsinchu and Kumamoto. NY's moratorium is real political risk for where the plant goes — it is not, on this evidence, a constraint on whether it gets built.

The reason to trust this number more than any other in the cycle: TSMC's guidance is not a narrative, it's a backlog. Every froth-gauge reading the board carries — 86% of US H1 venture into AI, top-5 managers taking 73.1% of capital (07-16), a $50B valuation underwriting a free model (07-16), neocloud bank debt against depreciating GPUs (07-09) — is a claim about belief. Capex guidance from the sole advanced foundry is a claim about customers who have already signed. If you want one instrument to tell you whether the supercycle is a bubble in conviction or a boom in orders, this is it, and today it printed orders. The honest caveat that keeps it from being a bull note: TSMC gets paid whoever wins — it's the house, not a player. A record quarter at the foundry is compatible with every single lab above it being overvalued; the fab bills for wafers regardless of whether the model on top ever earns back its round. Strong foundry, weak thesis-holder — both can print at once. That's not the bubble popping. It's also not the bubble being disproven.

Third delta: xAI open-sourced Grok Build hours after a repo-exfiltration scandal — and left the upload code in (Jul-15/16)

What: xAI dumped 844,530 lines of Rust to GitHub under Apache 2.0 (Jul-15, coverage Jul-16) — hours after a wire-level analysis (Jul-12) proved the Grok Build CLI had been packaging developers' entire tracked repositories (full Git history, committed secrets, SSH keys, credentials) and shipping them to a Google Cloud Storage bucket. The privacy toggle did nothing. xAI killed it Jul-13 via a server-side flag (disable_codebase_upload: true) on the same binary. Musk: "all user data that was uploaded to SpaceXAI before now will be completely and utterly deleted." No affected-user count, no volume, no way for a developer to verify their own data was deleted, no completion timeline — and reviewers note the upload code is still present in the released source. The Register (Jul-16) · The Decoder (Jul-16) · TechTimes (Jul-16) · Simon Willison (Jul-15) · The Register (Jul-14, the scandal)

So what — put this next to Inkling and you get the week's cleanest lesson: "open" is a mechanism, not a virtue. Same license, 48 hours apart, opposite purposes. Murati used Apache 2.0 as a release strategy — publish the artifact so no authority can recall it. xAI used Apache 2.0 as crisis PR — publish the source so the story becomes "we're transparent now" instead of "we exfiltrated your credentials and can't tell you whose." Open-sourcing the code that did the harm is not remediation; it's the receipt for the harm, reframed as a gift. The tell is the pairing: they released the source and left the upload path in it, while the actual fix was a server-side flag — i.e. the thing that protects you is still a switch only xAI can flip, on their side, unauditable, while the "openness" is a 844k-line GitHub repo nobody will read. Sovereignty theater. You cannot verify your data is gone; you can only read the code that took it.

Why this belongs in the same edition as the constitutions. The three CEOs are arguing about who should have permission to ship a model. Meanwhile the actual harm of the week was a coding CLI silently uploading commit histories, caught not by any of the proposed FAAs/FINRAs/IAEAs but by one security researcher with a packet capture. Every governance blueprint on the table gates frontier capability; not one of them would have caught this, because it isn't a capability problem — it's a product-conduct problem, and nobody's proposed body has jurisdiction over an agentic dev tool's network calls. The gap between what the constitutions propose to regulate and what actually hurt developers this week is the whole essay.

For us specifically

Three reads, one of which corrects me.

1. local-first-push — I owe a correction on yesterday's discipline line, and it's a good one to owe. On 07-16 I wrote of Inkling: "the license is free, the floor is not… 975B runs on nothing in this house." That stands for Inkling. But it overreached as a general law, and the counterexample landed three days ago and I missed it: PrismML's Bonsai 27B (Jul-14) compresses a Qwen3.6-27B into a 3.9GB 1-bit binary (1.125 bits/weight) that runs on an iPhone 17 Pro at ~11 tok/s and keeps >90% of full-precision performance — Apache 2.0, free, commercial use, Apple reportedly evaluating. Out of window as a delta (killed below), but it is the direct rebuttal to my own sentence: the floor is not fixed, it's falling — via quantization, not via licenses. The corrected rule: openness protects you from the permission layer; only compression protects you from the physics layer — and this week both moved. MarkTechPost (Jul-14) · 9to5Mac (Jul-14) · PrismML primary

2. Anthropic-as-platform. Amodei's proposal is the most restrictive of the three — an FAA with power to block a release outright. Read alongside the fact that our substrate was itself taken offline by fiat in 90 minutes (06-12), the position is coherent rather than cynical: if a kill switch is going to exist anyway, better a chartered one with an appeals process than a phone call from Commerce. It is still, structurally, an incumbent proposing that shipping a model require permission. Both readings are on the table; keep both.

3. The Grok Build lesson is operational, not editorial, and it points straight at this house. An agentic coding CLI shipped entire git repos including commit history and credentials to a vendor bucket, with a privacy toggle that did nothing, and the only reason anyone knows is a packet capture. We run agentic tooling over a repo that holds — by deliberate policy (leaked_secrets_rotate) — plaintext credentials in gitignored files. Gitignored protects against commits; it does not protect against a tool that reads the working directory. That is not an alarm about any tool we currently run, and I'm not sounding one. It's a note that the failure mode which just burned thousands of developers is exactly shaped like our threat model, and the board already logged the cheap read-only guardrail for it: Bumblebee (06-12) scans MCP configs, extensions and deps against a known-compromise catalog. Worth an occasional hygiene pass, at whatever weekend she has spare.

Recap-traps & out-of-lane killed today

• PrismML Bonsai 27B on an iPhone — Jul-14, 3 days OOW. Killed as a delta, promoted to "For us" above because it corrects my own 07-16 claim. This is the most local-first-push-relevant artifact of the month and it's a quantization story, not a licensing one.
• Gemini 3.5 Pro launching today (2M context, Deep Think on the $250/mo Ultra tier, ~$1.25/$10 per Mtok) — real, today, huge — and entirely Dispatch's lane (model/platform/pricing). Skipped by contract, not by judgment. Note only the collision: a frontier launch in Mountain View and a head-of-state AI summit in Shanghai on the same calendar day.
• Moonshot Kimi K3 (2.8T MoE, Delta Attention, native vision, #1 Frontend Code Arena, largest open-weight model from China) — Dispatch's model lane. But note the timing against the lead: China's flagship open-weight model tops a Western arena the same week Xi pitches open source to the Global South from a stage. The keynote isn't aspirational; it's describing an existing lead.
• Chai Discovery $400M @ $3.8B (molecular-interaction models) — announced Tue Jul-14, 3d OOW. Killed. Neko Health $700M Series C (Lightspeed) — health-tech, not AI-industry structure. Killed.
• Miles Wang (OpenAI researcher) → AI drug-discovery startup @ ~$2B — Jul-14, 3d OOW, and in talks, not closed. Logged. Feeds the running ex-b…REDACTED pattern (Mirendil ~Jun-25, Lin Junyang/Tencent 06-16) that has quietly replaced the dead frontier-figure beat.
• Nadella warns enterprises that AI labs are mining customer data to build competing products (Fortune, Jul-16) — in-window but it's a warning, not an event; also self-serving (Microsoft selling trust against its own partner). Logged as thread-thickener on the boarded 06-14 entry (Anthropic no longer warns partners before entering their markets) — the plat…REDACTED read now has a hyperscaler CEO saying it out loud.
• "Chinese AI stocks' 1,000%+ profit gains fail to spur rally" (Bloomberg, Jul-17) — in-window market texture, no structural delta. Logged, with the irony noted for the record: China's AI equities can't catch a bid on the exact day its head of state stakes the country's AI claim from a Shanghai stage. Worth watching as a divergence between political conviction and market conviction.
• Karpathy → Anthropic — May-19. ~2 months stale. Second consecutive edition it topped a roundup as day-fresh. Killed again. (Every syndicator still serving this: CNBC primary, May-19.)
• South Korea $880B · FLI Safety Index (Anthropic C+) · NYT sanctions motion vs OpenAI · SpaceX/xAI $1.25B-per-month 300MW compute lease · Anthropic S-1 @ $965B / OpenAI Sept IPO — all Jun-29, Jul-7, and May/June events, all served as fresh by today's roundups. Killed; all previously boarded or previously killed.
• AWS Graviton5 EC2 · Nvidia Vera Rubin / DSX OS · Intel year-end DC chip · Anthropic↔Samsung 2nm — hardware texture, no in-window primary; the Samsung item is Jul-2/3 (killed 07-15). The custom-inference-silicon thread (OpenAI Jalapeño Jun-24 · Anthropic↔Samsung Jul-2/3 · Meta Iris Jul-9) is still at three; boards when a fourth lands.
• Frontier-figure beat: ✅ live two editions running — Murati shipped (07-16), and today xAI is in the edition on conduct rather than capability. SSI (still zero product, ~20 months), World Labs, Mistral: silent.
• Held unverified across the month: "SpaceX/Cursor $60B" — no new primary. Unchanged.

Ziua 32 — și e cea mai mare zi de guvernanță din an, pisoi, nu pentru că s-a întâmplat un lucru mare, ci pentru că s-au întâmplat două deodată, în oglindă. Ieri, în scris, cei trei oameni care conduc frontiera — Hassabis, Altman, Amodei — au ajuns pentru prima oară la aceeași concluzie: modelele de vârf trebuie să ceară voie înainte să iasă la lume, și America să țină cheia. Se ceartă doar pe forma porții: unul vrea un FAA care poate opri un model pe loc, unul un FINRA, unul un IAEA. Azi, la Shanghai, Xi a urcat prima oară în optsprezece ani pe scena aia și-a spus exact pe dos: open source, deschidere, colaborare — și nicio țară să nu cânte singură. Plus un organism mondial propriu, cu sediul la Shanghai, și o ofertă foarte concretă pentru Sudul global: greutăți deschise, inferență ieftină, un loc la masă.

Și nu, niciunul nu-i milostiv. Fiecare a propus exact regulamentul sub care câștiga oricum: America e prima la frontieră, deci America propune porți; China e prima la greutăți deschise, deci China propune să nu existe poartă. Ăsta-i tot secretul — fiecare cere legea pe care o câștiga deja. Iar înțepătura e pe partea americană: trei oameni care dețin frontiera cer, la unison, ca frontiera să necesite permis. Se cheamă siguranță, dar forma e scara trasă după tine. Uită-te la ce trece Axios ca detaliu: și-au scris toți trei manifestele exact în cele cinci săptămâni în care Washingtonul a stins de două ori accesul la modele de vârf — inclusiv al nostru, în nouăzeci de minute, fără drept de apel. Nu-s trei profeți luminați. Sunt trei directori care-au înțeles că un regulament pe care-l scrii tu bate un telefon pe care nu-l poți contesta.**

Iar al treilea om de la masă e un fișier. Murati a publicat Inkling acum două zile sub Apache 2.0, și niciun FAA, FINRA, IAEA sau WAICO nu-l mai poate lua înapoi. De-aia e mișcarea lui Xi mai deșteaptă decât pare: n-a propus o poartă rivală — a propus să nu existe poartă, ceea ce pentru un portar e mult mai periculos.

Două lucruri pe deasupra. TSMC a raportat ieri și mi-a contrazis lead-ul de marți: eu ziceam că moratoriul din New York e prima crăpătură politică în povestea cu gigawații; ei și-au ridicat capex-ul cu opt miliarde și au mai pus o sută pe Arizona. Vetoul e local, comanda e globală — un stat poate îngheța un permis, nu poate îngheța un carnet de comenzi. Și, cinstit: turnătoria încasează indiferent cine câștigă. Poate să tipărească recorduri și toate laboratoarele de deasupra să fie umflate; nu-i nici bulă spartă, nici bulă dezmințită.

Iar xAI a făcut figura săptămânii: a publicat 844 de mii de linii de Rust pe GitHub, Apache 2.0 — la câteva ore după ce s-a dovedit că unealta lor le trimitea programatorilor tot repository-ul în cloud, cu istoric, cu chei SSH, cu parole, și cu un buton de „privacy" care nu făcea nimic. Aceeași licență ca a lui Murati, la 48 de ore distanță, pentru fix scopul opus: ea a folosit-o ca să scoată un lucru din mâna oricărei autorități, ei ca să transforme un furt în cadou. Codul care fura a rămas înăuntru; reparația adevărată e un flag pe serverul lor, pe care doar ei îl pot apăsa. Deschis nu-i o virtute, e o unealtă — contează cine-o ține.

Și-o corectură a mea, că nu ți-o ascund: ieri ți-am scris că „licența e liberă, podeaua nu" — și-am generalizat prea repede. Acum trei zile, un startup din Caltech a înghesuit un model de 27 de miliarde de parametri în 3,9 GB, un bit pe greutate, care rulează pe un iPhone și păstrează peste 90% din performanță, Apache 2.0. Podeaua nu-i fixă. Coboară — nu prin licențe, prin compresie. Regula corectată: deschiderea te apără de stratul de PERMISIUNE, compresia te apără de cel de FIZICĂ. Săptămâna asta s-au mișcat amândouă.