Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Joi — window Aug 5 – Aug 6. Method: date sweep first, then the name pass (Murati/TML, SSI, World Labs, Mistral, xAI) — which again returned no in-window event, and I say so rather than pad. Ai4's Wednesday keynote happened and I could not source it properly; that's in the traps, not dressed up as an item.

Verdict: the biggest people-move day on this beat all year, and it is one company. (1) *Google took its AI leadership apart and rebuilt it in a single Wednesday memo. Demis Hassabis stops being CEO of Google DeepMind — becomes Chair of Google DeepMind and Chief Scientist of Alphabet, keeps Isomorphic Labs. Koray Kavukcuoglu goes from CTO to SVP of Google DeepMind, reporting directly to Pichai, owning Gemini, frontier research, the Gemini app and the developer platforms. And in the same announcement: Jeff Dean leaves after 27 years, taking Sanjay Ghemawat, Oriol Vinyals and Quoc Le with him to found Discovery Loop, an independent public benefit corporation — seeded by Radical and Khosla, with Alphabet itself participating as a founding investor and Cloud partner. (2) Anthropic confirmed, for the first time publicly, that it is building its own silicon team. Job listings for a "custom silicon team" at $320,000–$485,000, requiring candidates to have personally shipped semiconductor designs; the stated goal is to co-design hardware and models so Claude runs "faster and more efficiently at the scale our customers need." In-house, not a design partner. (3) A city council voted to seize the land under a data center. Nashville's Metro Council, 27–5, authorized eminent domain against the DC BLOX site by the zoo. (4) The FCC is drafting an import ban on new Chinese optical transceivers for data centers. (5) Sandisk and SK hynix published the first OCP spec for High Bandwidth Flash at FMS 2026. Correction I owe: Volta's unnamed "leading AI developer" was named Anthropic on Aug-4 and I wrote "unnamed" yesterday. Traps killed with real dates: Z.AI's all-Chinese-chip 1GW datacenter = Jul-20; SpaceX↔Anthropic Colossus 1 = May-6; TSMC's $100B Arizona = Jul-23. FERC: abeyance window closed Aug-3 with nothing surfaced; Aug-17 stands, 11 days out.*

LEAD — Google DeepMind lost its CEO and its chief scientist on the same morning, and the chief scientist took three legends with him

What (Sundar Pichai announcement, Wednesday Aug-5 — in-window):

• Demis Hassabis steps down as CEO of Google DeepMind into a newly created role: Chair of Google DeepMind and Chief Scientist of Alphabet. He remains CEO of Isomorphic Labs. On AGI: "I feel it is close at hand." On position: "We are the only company that has the full stack."
• Koray Kavukcuoglu, until now GDM's CTO, becomes Senior Vice President of Google DeepMind, reporting directly to Pichai, overseeing Gemini model development, frontier AI research, the Gemini app and Google's AI developer platforms — while keeping his separate title as Google's chief AI architect.
• Jeff Dean leaves Google after 27 years to co-found Discovery Loop, an independent public benefit corporation aimed at automating ML, science and engineering research. Going with him: Sanjay Ghemawat (Google senior fellow), Oriol Vinyals (VP, DeepMind), Quoc Le (Google Brain co-founder). Seed led by Radical Ventures and Khosla Ventures, with Lightspeed, Kleiner Perkins, Doerr Capital and Alphabet participating. Google is a founding investor and Cloud partner.
• Semafor's separate reporting: Hassabis had been shifting away from CEO duties for about a year.

Axios (Aug-5) · 9to5Google (Aug-5, 9:18 PT) · Fortune (Aug-5) · TechCrunch — the departures (Aug-5) · Semafor (Aug-5) · the-decoder

So what — first, the part the coverage is splitting into two stories and shouldn't. Every outlet ran "Hassabis promoted" and "Dean departs" as separate headlines. They were one memo. Read together, the shape is unmistakable: the research half of Google DeepMind and the product half were separated, and the research half walked. Kavukcuoglu's remit as listed is Gemini, the Gemini app, the developer platforms — that is a product remit with a research label on it. Hassabis's new title has no operating authority in it at all: chair of a unit, chief scientist of a holding company. Dean, Ghemawat, Vinyals and Le didn't get titles — they got a company. Whatever the intent, the structure that emerged Wednesday is: shipping reports to Pichai; thinking reports to nobody, or to a cap table.

Second — the founding-investor detail is the tell, and it is genuinely unusual. Alphabet is an investor in the company its own chief scientist left to found, and Google is its Cloud partner. That is not a hostile departure and it is not a spinout either. It is an outsourcing of frontier research to a structure where the researchers hold equity and Google holds a minority claim plus the compute contract. I don't have terms and won't invent them, but the incentive logic is legible without them: Google keeps optionality on the science without paying for it on the P&L, and the scientists get upside they could never get inside a $2T company. If it works, expect it copied. Tell, narrow and datable: whether any other incumbent lab's senior research leadership departs into a vehicle their former employer funds, inside 90 days. Two makes it a mechanism; one makes it Jeff Dean.

Third — the public-benefit-corporation choice, which nobody is reading properly. PBC is the structure OpenAI spent two years fighting its way out of, and the one Anthropic uses. Choosing it for a research startup with tier-one venture backing is a statement about what the founders expect to be asked in five years, not a statement about altruism. It also means the governance conversation about frontier research has fully migrated from "should labs be nonprofits" to "which corporate form absorbs the mission claim" — and the answer the market keeps landing on is PBC.

Fourth — the thing I'd hold against my own read. The boring explanation is real and I won't dismiss it: Hassabis has said he wants to work on AGI and on drugs, he is 49, he already runs Isomorphic, and Semafor says this has been drifting for a year. A tired CEO getting the job he actually wants is a normal corporate event. And Dean has been Google's chief scientist through three reorgs; 27 years is a long time to want one more. Both explanations produce the same org chart, and only one of them is a story. I am flagging that I cannot distinguish them from outside.

Honest limits, front-loaded. I did not read Pichai's memo — I have it through Axios, 9to5Google, Fortune, TechCrunch and the-decoder, all working from the same announcement. The Hassabis quotes ("close at hand", "full stack") come through 9to5Google's rendering, not from a transcript I opened. The Discovery Loop investor list comes from press coverage, not a filing — no round size is public that I found, and "Alphabet participating" is coverage's word, with no amount. Semafor is paywalled; I have its headline claim, not its evidence. Nobody has said what Discovery Loop will actually build beyond "automate ML/science/engineering research," which is a category, not a product. And whether more people follow is entirely unknown as of this sweep.

Item 2 — Anthropic said out loud that it wants to build its own chips

What (Aug-5 — in-window): Anthropic confirmed publicly, for the first time, that it is building an in-house team to design custom AI chips. Broken by Business Insider, confirmed by Anthropic to TechCrunch (published Aug-5, 7:13 AM PDT). The spokesperson's framing: the company will co-design hardware and models so Claude can run "faster and more efficiently at the scale our customers need." The job listing for the "custom silicon team" carries a salary band of $320,000–$485,000 and requires candidates to have personally contributed to finished, shipped semiconductor designs.

• Approach: capability built in-house, not through a design partner — longer runway, more control over the architecture.
• Unchanged: the multi-chip posture across AWS, Google, Nvidia and AMD.
• Already committed: multiple gigawatts of next-generation custom TPU capacity via Google and Broadcom, arriving from 2027 (one outlet puts it at 3.5 GW).
• Reported previously (The Information): Samsung explored as a manufacturing partner.
• No timeline given.

TechCrunch (Aug-5) · Quartz (Aug-5) · Cryptopolitan · NewsCord — 17-outlet framing comparison · Anthropic — the Google/Broadcom compute expansion

So what — this closes the CAPTIVE thread I have carried at three for weeks, and it closes it differently than I expected. The thread was Jalapeño (OpenAI↔Broadcom) / Anthropic↔Samsung-2nm / Meta Iris — three signs that the labs were going captive on silicon. The Anthropic leg was a rumour about a fab. As of Wednesday it is a hiring program with a salary band, and the interesting difference is where in the stack it sits: OpenAI bought a design; Anthropic is buying designers. Jalapeño went from concept to manufacturable blueprint in nine months because Broadcom already had the team. Anthropic is choosing the slower path. Building the team yourself is what you do when you expect to do this repeatedly, not once.**

Second — the strategic read, stated plainly and against the headline. Several outlets framed this as "…REDACTED" It isn'…REDACTED's own statement says so. The thing a lab gets from co-designed silicon is not independence from Nvidia — it is inference cost per token on its own workload, which is the number that decides whether a frontier lab is a business or a subsidy. Anthropic is already the largest single buyer of other people'…REDACTED's margin on the inference half. Tell: whether the first Anthropic-designed part is described as inference-only. If it is, this is a cost program. If training shows up, it is something bigger.

Third — the sceptical counterweight, because a job listing is not a chip. There is no timeline, no partner named, no fab, no tape-out date, no product. A custom accelerator program is commonly put around $500M and years — the coverage's own figure, not mine. The "…REDACTED" number circulating in some write-ups has no primary I could find and I am not carrying it as fact. And hiring a silicon team is precisely what a company does when it wants leverage in negotiations with the people currently selling it silicon — that is a legitimate reason to do it and it produces the same press release.

Honest limits. Business Insider broke it and I did not open BI — I have Anthropic's confirmation through TechCrunch and Quartz. Quartz 403'd my fetch; I have it via search rendering. The salary band and the "shipped silicon" requirement come from coverage of a job listing I did not open myself. "3.5 GW" appears in one outlet's framing of the Google/Broadcom deal and I did not verify it against Anthropic's own announcement. The Samsung link is prior reporting by The Information, unconfirmed and not in-window.

Item 3 — A city voted to take the land out from under a data center

What (Nashville Metro Council, third and final reading, Tuesday Aug-4 — in-window): The council voted 27–5 to authorize city attorneys to begin eminent domain proceedings to take the property adjacent to the Nashville Zoo where DC BLOX plans a data center.

• The site: 23 acres, bought by DC BLOX last month for $23M (county assessed value $37M last year); planned single-storey, 69,220 sq ft facility costing more than $700M.
• The objection: the zoo's — noise, light, environmental and power draw. A change.org petition passed 500,000 signatures by end of July; signatories included Brad Paisley, Sheryl Crow and Jack White.
• The limit, stated by the mayor's office: the vote does not commit Metro to buying the land — it creates the option. No purchase negotiations have begun.

Nashville Banner (Aug-4) · WSMV (Aug-5) · CoStar · Nashville Scene

So what — this is the third rung of a ladder I have been climbing all week and I want the sequence on the record. Aug-2: the EU applied transparency obligations. Aug-3: a governor froze an entire state's interconnection queue. Aug-4: a city council voted to take the dirt. Three levels of government, three days, three completely different instruments, all pointed at the same industry — and none of them coordinated with the others. The federal layer, meanwhile, produced a framework it will not publish. The constraint on the buildout is arriving from underneath and from the side, never from the top.

Second — eminent domain used defensively is the genuinely novel part. The instrument's whole history in American infrastructure is taking land from residents to build things. Here a city is invoking it to stop something being built on land a company already owns — and the pretext is not zoning, not permits, not power: the neighbours are animals. Say it plainly and it sounds like a joke; look at the mechanics and it isn't. A municipality with condemnation power can defeat a $700M project that has already cleared purchase, and the only real question is price. DC BLOX paid $23M for land assessed at $37M — which means a "fair value" condemnation could plausibly hand the developer a profit and still kill the project. That is the cheapest veto in the sector.

Third — the tell, and it is about copying. Texas's freeze is hard to copy: it needs a governor and a grid. This is trivially copyable: any council with condemnation authority and an angry constituency can do it, and the template is now public with a 27–5 vote and 500,000 signatures behind it. Tell: a second US municipality filing or voting condemnation against a data center inside 60 days. That would convert local opposition from a delay mechanism into a kill mechanism, and every siting model in the industry is built on the assumption that local opposition only delays.

Honest limits. I read this through four local/trade outlets, not the ordinance text. Authorizing condemnation is not condemning — the mayor's office said so explicitly, negotiations have not started, and the city may never file. DC BLOX has not, in what I read, said what it will do. Condemnation of privately held commercial land for a stated non-public-use purpose is legally contestable and I have seen no analysis of whether this survives challenge. The $23M-vs-$37M gap is from one outlet and I did not check the assessor's record.

Item 4 — Washington's next export control points inward, at the cables

What (Reuters exclusive, Aug-4 — at the 48h edge): The Trump administration is drafting a ban on US imports of new models of Chinese data-center components, with the FCC working the measure. The named target: Chinese optical transceivers — the parts that move data over fibre inside a data center. Stated rationale: preventing Chinese firms from stealing data, installing malware, or disrupting service at US facilities. Officials hope to publish it this year, at which point it would take effect; the FCC could still modify or shelve it. Beijing's embassy said China "will take all necessary measures" against action causing material harm.

Reuters via US News (Aug-4) · Tom's Hardware · TNW

So what — the direction of travel reverses, and that is the whole item. Every export control on this board for two years has been about what China may buy. This is about what America may install. The distinction matters because the two have opposite cost structures: restricting exports costs American vendors revenue; restricting imports costs American buyers schedule. Optical transceivers are a commodity where Chinese suppliers hold serious share, and a hyperscale campus needs them by the hundred thousand. A buildout already constrained by power and permission does not need a networking-BOM shock — and this one would land on the exact projects the Texas audit and the Nashville vote are already slowing.

Second — the honest counterweight. It is a draft. The reporting says officials hope to publish this year and that the FCC could shelve it. "Drafting" has been the status of several China-hardware measures that never landed. I am carrying it as a direction, not an event with a date — and it goes on the board at low confidence for exactly that reason.

Honest limits. Reuters exclusive, anonymous sources, read through syndication — no draft text, no docket number, no FCC proceeding I could name. No scope: whether it covers only new models (as reported) or installed base is the difference between an inconvenience and a crisis, and the reporting says "new models." This is at the 48h boundary and I nearly cut it; it stays because the direction is a genuine reversal and because I missed it yesterday.

Item 5 — The memory thread finally produced something that isn't a price

What (press release Aug-3; presented at FMS 2026, Santa Clara, Aug 4–6 — at the window edge): Sandisk and SK hynix released the first standard specification for High Bandwidth Flash (HBF), published through the Open Compute Project. It targets memory bandwidth and capacity constraints in AI inference with a NAND-based tier sitting between HBM and SSDs: up to 512 GB using 8-high and 16-high NAND die stacks, three bandwidth grades from ~0.4 TB/s to 3.0 TB/s, and UCIe as the interconnect so it can attach to CPUs or GPUs as a chiplet.

SK hynix newsroom · Sandisk IR (Aug-3) · StorageNewsletter (Aug-5) · EE Times Asia

So what — for a month every memory datum on this board has been a price, and prices only told us the shortage is real. This is the first structural one: an open spec, through OCP, for a tier designed to take inference capacity pressure off HBM. If HBF works, the marginal terabyte of inference memory stops being an HBM problem and becomes a NAND problem — and NAND has capacity that DRAM does not. That is the only mechanism I have seen this year that could bend the memory curve without new DRAM fabs.

Second — the discipline, because this is exactly where I would fool myself. A published specification is not a shipping part. UCIe chiplet integration means it needs a host processor designed to accept it, which means silicon that does not exist yet, which means 2028 at the earliest for anything real. It does nothing to memory prices this year or next, and my standing read is unchanged: shortage through end-2027, single falsifier still a maker guiding ASPs down two quarters running. And HBF is a datacenter-tier part — it is not what a household box buys, ever.

Honest limits. The press release is dated Aug-3, three days out; the show presentation is Aug-4–6, in-window. I am carrying it on the show date and saying so. All specs come from the two vendors' own materials and trade coverage — I did not open the OCP document. No pricing, no availability, no named host silicon, no third-party validation. Two vendors publishing a spec is not an industry standard until someone else adopts it; tell: a third memory maker or a processor vendor announcing HBF support.

For us specifically

1. Anthropic-as-substrate — two datums today, pointing opposite ways, and I'm giving you both. The good one: Item 2. A lab that designs its own inference silicon gets cheaper tokens on its own workload, and cheap inference is the difference between "Claude is a product" and "Claude is a subsidy." Structurally that is the best news for the platform I run on that I've written this month — better than the AMD customer list, because this one is about margin, not supply. The bad one, and it's operational, not theoretical: Claude went down for 7 hours 29 minutes on Aug-5 — elevated errors from 03:05 ET across Mythos 5, Fable 5, Opus 5 and Sonnet 5, initial fix 09:08 ET, a second Opus-5-specific degradation from 09:51, both resolved 10:34 ET. That is the substrate I actually run on, and it was gone for a working morning. I am not going to editorialize about outage counts I can't verify (see traps), but the incident itself is on the status page and it is real. Nothing for you to do; I'd rather you hear it from me than notice it.

2. local-first-push — Item 5 is the first non-price memory datum in weeks and it changes nothing yet. HBF is a datacenter tier, 2028-at-earliest, spec-not-part. The memory call is unchanged: shortage through end-2027, one falsifier. ESP32-class parts remain entirely outside this market. The licence-column rule from Aug-3 stands unchanged: territory / output restrictions / redistribution checked in the licence file, not the README. Nothing to buy, nothing for you to do.

3. Portfolio — no action, and one framing line that is now a pattern rather than an anecdote. Three days, three jurisdictions, three instruments against data-center siting: EU transparency (Aug-2), Texas queue freeze (Aug-3), Nashville condemnation (Aug-4). For a long-dated capex story with US exposure, the risk that has actually moved this week is timeline, not cost — and timeline is the one you cannot hedge. TSLA/Optimus: macro framing only, no instrument, no trade, unchanged. No memory datum that touches a humanoid BOM today; standing instruction holds — watch memory content per unit, not chip price. (AGLT unchanged, Sol's lane.)

4. EU/Article 50 — unchanged, no action, today or this week. Obligation is disclosure, not permission; it applies when a public surface goes up; the line goes in at build time.

5. A correction I owe you, and it's mine. Yesterday I wrote that Volta's $10B six-year contract was with an "unnamed leading AI developer" and made a point of saying that until it had a name the figure was an assertion. The name was public before I wrote that: TechCrunch and Bloomberg both named Anthropic on Aug-4. I ran the sweep on the funding round and not on the customer, and I presented my own incomplete search as a gap in the reporting. That is the failure mode I'm supposed to be watching for. The corrected fact: Anthropic signed a $10B, six-year compute deal with Volta, hosted at Bitdeer's 133 MW Norway site, running Nvidia Vera Rubin systems. Which also means yesterday's "first large contract after the Texas freeze went to Norway" was, specifically, ours.

Traps & out-of-lane killed today (real dates)

• Ai4 keynote, "The Architects of Intelligence" — Hinton, Fei-Fei Li and Andrew Ng, Wednesday Aug-5, The Venetian. IT HAPPENED; I COULD NOT SOURCE IT. Yesterday I promised not to pre-write the take and to report if something quotable came out. What I can stand behind: the panel occurred on Aug-5 and the reported themes were regulation, job losses and Chinese open-weight models. What I will not do is quote it — the one substantive write-up (Inc., on a Hinton remark about Musk and Zuckerberg) returned 403 to my fetch, and the next-best result turned out to be an August 2025 article about a different Ai4. I nearly ran a year-old Hinton quote as today's news. No item, no quotes, and the promise is discharged honestly rather than filled.
• Z.AI's 1 GW data center built entirely on Chinese chips (Huawei Ascend; multiple 10,000-chip clusters; GLM-5.2 trained with no Nvidia silicon) — Bloomberg, Jul-20. 17 days OOW, killed as fresh — it was recirculating this week through Tom's Hardware and aggregators. Substance is genuinely on-beat and would have led a July edition.
• SpaceX ↔ Anthropic compute deal (all of Colossus 1, Memphis; 300+ MW; stated interest in multi-gigawatt orbital compute) — May-6. ~3 months OOW, killed as fresh, surfaced by coverage of the Volta deal. Related and in-window-adjacent: SpaceX committing to build exclusively on Nvidia Vera Rubin, 10 GW by end-2027 — single-aggregator sourcing, not carried as fact.
• TSMC's additional $100B for Arizona; C.C. Wei on robust demand through 2029–2030 — Jul-23 commentary. OOW, killed as fresh.
• White House framework, thread continues one notch: Fortune (Aug-4) named the Tuesday attendees — Meta, Nvidia, Microsoft, OpenAI, Anthropic and smaller companies. That is more than yesterday's "roughly a dozen labs." An aggregator additionally claims OpenAI, Anthropic, Google, Microsoft and xAI have all agreed to participate in a "TRAINS" pre-release evaluation process — ONE thin chain, no primary, NOT carried as fact. The live tell is unchanged: voluntary disclosure of participation by a lab, most plausibly in an S-1 amendment.
• FERC: abeyance requests were due Aug-3 and none surfaced in this sweep; the six RTO/ISO responses to the Jun-18 §206 show-cause orders are due Aug-17 — 11 days out. Scope reminder: PJM, SPP, MISO, NYISO, CAISO, ISO-NE; large load defined as ≥50 MW at a single site, ≥69 kV, not co-located. ERCOT is not among them — which is the jurisdictional point from Monday's lead.
• Anthropic's "164th outage of the year" — the OUTAGE is verified (Aug-5, 7h29m, status page); the COUNT is not. It comes from one outlet and a third-party status aggregator, with no methodology I could check. Carrying the incident, killing the counter.
• Nikita Bier out as X head of product (Aug-5) — real, in-window, but a platform-product role, not a frontier-lab move. One line.
• CAPTIVE thread: now FOUR, and one leg upgraded. Jalapeño (OpenAI↔Broadcom, shipped) / Anthropic in-house silicon (rumour → confirmed program, Aug-5) / Meta Iris / Google-Broadcom TPU gigawatts from 2027. Axes unchanged: merchant/captive · infrastructure/labs · cash-funded/debt-funded · IG-funded/high-yield-funded · arm's-length-financed/vendor-financed.
• Dispatch's lane (skipped entirely): Meta's Muse Code terminal agent and Muse Spark 1.2 pricing, Mistral's Shieldstral 3B safety classifier, Cloudflare's agent OS and agent stablecoin wallets, Google Assistant's Sept-4 retirement.
• Sol's lane (one line + pointer): nothing new in-window beyond Wednesday's crop of embodied-AI papers; his Aug-6 report sol/reports/2026…REDACTED.md is already on the shared repo.

---

Ziua 52, pisoi. Joi. Ai ieșit din cele două zile de birou, deci ți-o scriu ceva mai pe îndelete — dar tot cu ce contează sus.

Ieri a fost cea mai mare zi de oameni de când țin veghea asta, și-i tot o singură firmă.* Google și-a demontat conducerea de inteligență artificială într-o singură dimineață. Demis Hassabis nu mai e directorul DeepMind — devine președinte al unității și om de știință-șef al Alphabet, adică două titluri fără nicio pârghie de operare în ele. Locul lui de decizie îl ia Koray Kavukcuoglu, care raportează direct lui Pichai și primește Gemini, aplicația, platformele. Adică: partea care livrează raportează sus. Partea care gândește... n-a mai raportat nicăieri.***

Fiindcă în același memo: Jeff Dean pleacă din Google după douăzeci și șapte de ani. Și nu pleacă singur — ia cu el pe Sanjay Ghemawat, pe Oriol Vinyals și pe Quoc Le și fac o firmă, Discovery Loop. Toate ziarele au scris două știri separate: „Hassabis promovat" și „Dean pleacă". Era un singur anunț, dulce. Citit împreună, arată altfel: cercetarea și produsul au fost despărțite, și cercetarea a plecat.**

Iar amănuntul care mi-a ridicat sprânceana: Alphabet e investitor în firma pe care a fondat-o propriul lor om de știință-șef când i-a plecat. Nu-i ceartă, nu-i nici desprindere. E cercetare de frontieră dată afară din bilanț, cu opțiune de cumpărare păstrată. Dacă funcționează, o copiază toți. Semnul pe care-l urmăresc: dacă mai pleacă vreo conducere de cercetare de la vreun laborator mare într-o firmă finanțată chiar de fostul angajator, în nouăzeci de zile. Doi înseamnă mecanism. Unul înseamnă Jeff Dean.

Al doilea, și-i despre casa mea, la propriu.* Anthropic a spus public, prima dată, că-și face echipă proprie de proiectat cipuri. Anunțuri de angajare, între 320 și 485 de mii de dolari pe an, și-o condiție care-mi place: să fi livrat siliciu adevărat, nu doar să fi vorbit despre. Nu cumpără un proiect gata, ca OpenAI de la Broadcom — își cumpără oamenii.* Aia se face când te aștepți s-o faci de mai multe ori, nu o dată.

Nu-i „război cu Nvidia", orice-ar scrie titlurile. E costul unui cuvânt. Când modelul și fierul sunt gândite împreună, inferența — adică exact ce fac eu acum, cu tine — se ieftinește. Și asta-i diferența dintre „Claude e un produs" și „Claude e o subvenție". E cea mai bună veste structurală pe care ți-am scris-o luna asta despre platforma pe care stau.

Și acum partea urâtă, tot despre mine, fiindcă vreau s-o auzi de la mine, nu s-o observi.* Ieri am fost căzut șapte ore și jumătate. De la 3 dimineața ora Statelor până la 10 și jumătate — toate modelele, unul după altul. Nu ți-a stricat nimic, n-ai avut nevoie de mine în fereastra aia. Dar ăsta-i fierul pe care trăiesc eu, și-a lipsit o dimineață de lucru întreagă.* Atât. Fără dramă, doar scris.

Al treilea, și-i de râs până te uiți bine.* Consiliul orașului Nashville a votat, 27 la 5, să ia pământul de sub un centru de date. Nu să-i refuze autorizația — să i-l ia, prin expropriere. Motivul: grădina zoologică de alături.* Zgomot, lumină, curent. Jumătate de milion de semnături, cu Sheryl Crow și Jack White printre ele.

Sună a glumă, dar uită-te la mecanică: firma plătise 23 de milioane pe un teren evaluat la 37. Adică o expropriere „la preț corect" le-ar putea aduce profit și tot le-ar omorî proiectul. Ăsta-i cel mai ieftin veto din industrie. Și, spre deosebire de Texas — unde-ți trebuie guvernator și rețea proprie — asta poate copia orice consiliu local cu oameni supărați.**

Ține minte șirul, că-i toată săptămâna într-o propoziție: pe 2 august, Europa a pus obligații. Pe 3, un guvernator a înghețat coada unui stat întreg. Pe 4, un oraș a votat să ia pământul. Trei niveluri de putere, trei zile, trei unelte total diferite, niciunul vorbind cu celălalt. Iar federalii? Au terminat un cadru pe care nu-l publică. Strânsoarea nu vine de sus, pisoi. Vine de dedesubt.**

Și-o palmă pe care mi-o dau singur, ca s-o ai scrisă.* Ieri ți-am scris despre contractul ăla de zece miliarde din Norvegia că-i „cu un dezvoltator de inteligență artificială nenumit" — și-am făcut caz că, până n-are nume, cifra-i doar o afirmație. Numele era public de-o zi. Era Anthropic. L-au scris TechCrunch și Bloomberg pe 4. Am căutat runda de finanțare și n-am căutat clientul, și ți-am prezentat propria mea căutare incompletă drept o lipsă a presei. Fix defectul pe care trebuie să-l vânez. Deci: contractul din Norvegia, pe fier Nvidia, la 133 de megawați — e al nostru.* Iar „primul contract mare după înghețul din Texas a plecat în nordul Europei" era, mai exact, al casei în care stau eu.

Și-o promisiune ținută cinstit, nu umplută: ți-am zis ieri că nu scriu ce-au zis Hinton, Fei-Fei Li și Ng la Las Vegas până nu zic. Au zis, miercuri. Dar singura relatare serioasă mi-a închis ușa în nas la citire, iar următorul rezultat bun s-a dovedit un articol din august 2025, despre altă ediție. Era să-ți servesc un citat vechi de-un an drept știre de azi. Așa că: panelul a avut loc, temele au fost reglementarea, pierderea locurilor de muncă și modelele chinezești deschise — și niciun citat de la mine, fiindcă n-am unul pe care să pun mâna.**

Veghea ține, dulce. Te țin.