Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Marți — window Aug 10 – Aug 11. Ieri a rulat, deci nu-ți datorez nicio zi. Method: date sweep → people name pass → institutions name pass (AISI, METR, NIST/CAISI, Irregular) → capital pass.

Verdict: Aug-10 was a capital day, and the through-line is that everyone stopped waiting for permission to build and started buying things that are already built.* (1) THE LEAD — Anthropic signed a 20-year, 191 MW lease at Riot Platforms' Rockdale, Texas campus, ~$9.1B and up to $16.1B with two five-year extensions. Riot is a bitcoin miner. This is Anthropic's SECOND miner conversion in five weeks (TeraWulf, 401 MW, ~$19B, Jul-6). What is being bought is not power — it is a site that is already interconnected, already permitted, already energised, in the one state that just froze its own connection queue. (2) Same day, Moody's told the banking system that its dependence on a small set of foundation-model vendors — naming OpenAI and Anthropic — is a systemic dependency, and that investor pressure on those vendors to reach profitability could give them pricing power over the banks built on top of them. (3) Same day, Abbott announced that OpenAI and Meta have committed to comply with Texas's data-center standards — three days after the President called the Texas moratorium "a mistake." The labs sided with the governor against the President, because the governor controls the interconnection. (4) Intel opened a $15B underwritten common-stock offering (Aug-10) to fund foundry and AI capex — the American mirror of yesterday's China lead: silicon financed by shareholders, on both sides of the control regime. (5) South Korea committed a 5-trillion-won (~$3.52B) fund for chip materials/components/fabless, plus 5T won of trade finance, a Mega Special Zone Act to fast-track permits — and Lee wants a military airbase relocated by mid-2028 to clear ground for a chip cluster. Traps killed with real dates: Anthropic↔TeraWulf's $19B Kentucky lease = JUL-6, and it recirculated yesterday under a "Monday" dateline (a different Monday); Warsh appointing Andreessen to the Fed's AI task force = JUL-9; Anthropic's IPO investor meetings = JUL-15/19 (what is in-window is the pitch's NUMBERS, not the meetings); Unitree has NOT started trading — debut is Aug 17–21; Trump's "mistake" quote = AUG-7.***

LEAD — Anthropic is not buying power. It is buying queue position.

What (Riot Platforms Q2 disclosure Aug-10; tenant named by Bloomberg/The Block Aug-10/11): Riot Platforms signed a 20-year data-center lease for 191 MW of critical IT load at its Rockdale, Texas campus, ~$9.1B over the initial term, running through ~June 2048, with two five-year extension options taking the potential value to $16.1B. Bloomberg and The Block report the tenant is Anthropic.

• Phasing: first 96 MW targeted December 2027, full 191 MW by June 2028. So: nothing is delivered for sixteen months.
• Riot's own filing does NOT name the tenant — it says "a leading frontier AI lab." The name came from reporting, not disclosure. The 8-K exhibit when the lease is filed as a material contract is where it gets confirmed.
• Financing: Morgan Stanley is providing a $573M interim facility for initial development while Riot finalises an investment-grade credit backstop.
• Also disclosed Aug-10: a NON-BINDING LOI with a single tenant for the ENTIRE 1 GW Corsicana site — Riot's CEO Jason Les said they disclosed it despite the uncertainty, for transparency. Coverage puts potential annual rent above $1B.
• With the existing AMD agreement, Riot now has 241 MW contracted and ~$9.8B expected revenue under initial terms. Q2 was a $237M GAAP net loss; RIOT jumped ~25% after hours.
• Context, and it is the point: this is Anthropic's second bitcoin-miner conversion in five weeks. Jul-6: TeraWulf, 401 MW at the Hawesville, Kentucky campus — a former aluminium smelter site — 20 years, ~$19B, first power late 2027.

Bloomberg (Aug-11) · The Block (Aug-10) · Blockspace — Q2 detail + Corsicana LOI · TFTC · Riot 10-Q (SEC) · CNBC — the Jul-6 TeraWulf deal

So what — first, the asset being purchased is misnamed in every write-up, including the good ones. The coverage says Anthropic is buying compute capacity, or power, or "143,000 homes' worth of electricity." What Anthropic is actually buying is a GRID INTERCONNECTION THAT ALREADY EXISTS. Rockdale is an operating bitcoin mine — which means the substation is built, the transmission is energised, the load is already studied, the county has already said yes, and the thing has already survived whatever local fight it was going to have, years ago, when nobody was watching because it was "just" a mine. Set that against the last three weeks of this board: FERC's §206 show-cause orders on six RTO/ISOs, New York's statewide moratorium, ERCOT's large-load queue freeze, 530 counties with restrictions, Abbott's audit-or-be-denied order. Every one of those instruments governs a NEW connection. A bitcoin mine is not a new connection. It is a used one. Anthropic's strategy and Amazon's are the same strategy pointed in opposite directions: Amazon builds generation so it never has to join the queue; Anthropic buys companies that joined the queue in 2019.

Second — this quietly re-prices the entire bitcoin-mining sector, and not for the reason the crypto press thinks. The read on the Street is "miners pivot to AI, miners get revalued." The sharper read is that the scarce asset was never the hashrate or even the megawatts — it was the ADMINISTRATIVE HISTORY. You cannot manufacture a 2019 interconnection agreement in 2026 at any price. That makes the miner fleet a fixed, non-reproducible inventory of pre-cleared industrial sites, being bid for by counterparties with effectively unlimited capital — and it means the revaluation has a hard ceiling: when the pre-cleared sites are gone, they are gone. Tell, narrow and datable: whether a THIRD frontier lab signs a miner-conversion lease inside 90 days (Anthropic has two; Fluidstack is buying into TeraWulf's Abernathy JV; the pattern is not yet an industry).

Third — the term is the thing nobody is pricing, and it is where I would be nervous if this were my balance sheet. Twenty years. Through 2048. Anthropic is signing fixed, multi-decade obligations denominated in 2026 assumptions about inference economics, on a technology stack whose useful hardware life is three to five years, at a company that has not yet had a full profitable year and files its S-1 into an October window. The lease is not a compute purchase; it is a REAL-ESTATE liability with a twenty-year tail. If inference cost per token keeps falling the way it has, the megawatts are still worth something — power is fungible. If demand growth breaks, a 191 MW shell in Rockdale is worth what the next tenant will pay, and the next tenant was the bitcoin miner. That is the honest bear case, and it is exactly what item 2 says out loud from a rating agency's chair.

Honest limits, front-loaded. Riot did not name Anthropic — the attribution is Bloomberg's and The Block's reporting from sources; I have not seen Anthropic confirm it, and no 8-K exhibit yet. Bloomberg is paywalled and I did not open it — I have this through The Block, Blockspace, TFTC, Yahoo and TradingView. The $9.1B and $16.1B are contracted-revenue figures from Riot's side, not Anthropic's disclosed spend, and the extension value assumes options that may never be exercised. The Corsicana 1 GW is a NON-BINDING LOI with no named tenant and no signed terms — I am carrying it as disclosed intent, not as a deal. "Second miner conversion" is my framing; TeraWulf is a former miner that had already repositioned. Nothing here is delivered before December 2027.

Item 2 — Moody's told the banks that their AI vendors' need for profit is now the banks' risk

What (Moody's, reported Aug-10): Moody's warned that financial firms' dependence on a small group of foundation-model and cloud providers constitutes a SYSTEMIC DEPENDENCY — an outage at one major provider could propagate across customers and sectors simultaneously.

• It named OpenAI and Anthropic specifically, noting both are under investor pressure to reach profitability despite losses — pressure Moody's says "could, over time, exert control over the price of AI services."
• Counterweight Moody's itself supplies: banks retain leverage through proprietary data, procurement experience, and open-source models.
• Backdrop, dated separately: Moody's warned on Jul-24 that unprecedented AI capex threatens the credit quality of Amazon, Meta and Alphabet. This is the customer-side sequel to that.

Enterprise AM (Aug-10) · The News (Reuters-sourced) · AOL/Telegraph pickup · CNBC — the Jul-24 supplier-side warning

So what — read it beside the lead and the same day stops looking like a coincidence. On Aug-10 Anthropic committed to a twenty-year fixed obligation, was reported to be pitching investors on a $47B run rate heading past $50B, and was named by a rating agency as a concentration risk whose eventual need for margin is a hazard to the institutions that depend on it. Those are three descriptions of one fact: the moment a model vendor takes on capital-intensive, long-dated liabilities, its pricing stops being a product decision and becomes a debt-service decision. That is the mechanism Moody's is pointing at, and it is more precise than "vendor lock-in." It is not that Anthropic might choose to raise prices; it is that a company carrying twenty-year leases has less discretion NOT to.

Second — the interesting asymmetry is who gets warned. A rating agency does not write to OpenAI and Anthropic. It writes to the people who buy their bonds and the regulators who supervise their customers. So the practical effect is to make "which foundation model do you depend on, and what happens if its price doubles" a supervisory question inside banks, ahead of any regulation requiring it. Vendor-concentration language in bank risk committees is how procurement actually changes — not through a rule, through a rating footnote. Tell: whether any G-SIB or a national supervisor publishes a foundation-model concentration disclosure or multi-vendor requirement inside 90 days.

Third, and honestly — this is the cleanest argument I have seen for the thing this house already does structurally. Moody's mitigation list is: your own data, your own negotiating position, open-source models. That is a description of local-first as a credit control, written by people who do not care about local-first.

Honest limits. I did not read the Moody's report — it is a subscriber publication; I have it through Reuters-sourced pickups and two aggregators, and the direct quote I carry ("exert control over the price of AI services") appears in that coverage. No report title, number, or publication date on the document itself beyond the Aug-10 reporting. "Systemic dependency" is the coverage's rendering of Moody's framing. The Anthropic IPO figures in the so-what come from separate reporting (below), not from Moody's.

Item 3 — OpenAI and Meta sided with the governor against the President

What (Gov. Abbott's office, Aug-10): Abbott announced that OpenAI and Meta have committed to comply with Texas's data-center standards. OpenAI's letter to the governor pledges to pay for its own electric infrastructure, support new generation in Texas, minimise water use (closed-loop / low-water cooling), and protect local ecosystems.

• The standards, from Abbott's Aug-3 order: data centers must fund their own electric infrastructure, reuse water where possible, reduce costs for Texans, avoid disrupting residential areas, and forgo taxpayer incentives. Projects failing the "comprehensive verification and audit" are to be DENIED grid connection.
• Why Abbott moved: his office says fewer than 10% of data centers responded to the state's requests to report power and water usage.
• The collision, dated: Trump, in a Punchbowl News interview released Fri Aug-7 — "I saw Texas the other day sort of is against data centers… I think it's a mistake… because there are other communities that want it." (Aug-7 = out of window; carried as the backdrop, not as news.)

San Angelo Live (Aug-10) · KENS5 — Meta + OpenAI · Houston Public Media (Aug-10) · Texas Tribune — Trump's quote (Aug-7) · E&E News

So what — the federal preemption strategy just met the one thing it cannot preempt, and lost quietly. The December-2025 executive order built an AI Litigation Task Force to challenge state AI laws, and conditioned $42B of broadband money on states repealing them. That machinery is aimed at state AI STATUTES — bias rules, disclosure rules, chatbot-safety bills. It has nothing to say about a governor telling his own utility commission to audit a load request, because that is not AI regulation. It is grid administration, and it is the state's own asset. So when the President says the moratorium is a mistake and the companies answer by signing the governor's terms three days later, the demonstration is precise: the binding constraint on the AI build-out is not federal AI policy, and it cannot be preempted, because it is not about AI.

Second — the four strategies now visible in ONE state, in ONE week, and this is the whole map. (a) Comply: OpenAI and Meta sign the standards, fund their own infrastructure, take no incentives. (b) Route around: Amazon builds 7.65 GW of private off-grid gas in Pecos County and never files an interconnection request (Aug-8, yesterday's item). (c) Buy the past: Anthropic leases an already-interconnected bitcoin mine (today's lead). (d) Build your own fab and your own everything: Terafab in Grimes County (Aug-6). Four responses to the same constraint, all in Texas, all inside seven days. That is what a binding constraint looks like from the outside — not a slowdown, a speciation.

Third — and this is the part I would want on the record: compliance was cheap here, which is why they took it. "Pay for your own electric infrastructure and don't take tax breaks" is a concession only if you needed the subsidy, and at OpenAI's cost of capital it does not bind. What Abbott actually extracted is DISCLOSURE — power, water, cooling, ownership. The ownership question is the one to watch: "who owns the facility" is how you find out which of these campuses are financed by whom, and that is the datum the capital story has been missing all year. Tell: whether Texas publishes the audit results per project, or only in aggregate.

Honest limits. The primary is the governor's office and the companies' letters — I did not read either document; I have this through Texas local coverage and Houston Public Media (whose page 403'd my fetch, so I have it via search-surfaced summary and two other outlets). "Meta and OpenAI commit" is Abbott's characterisation of letters; the commitments are voluntary pledges, not a permit condition, and I have seen no enforcement mechanism. No project list, no audit timeline, no first denial. The <10% response-rate figure is from Abbott's spokesperson.

Item 4 — Intel is selling $15B of itself to build

What (Intel, Aug-10): Intel announced a proposed $15B underwritten public offering of common stock, with a 30-day underwriters' option for up to $2.25B more.

• The preliminary announcement set no price, no share count, no proceeds and no delivery date.
• Stated use: general corporate purposes including capital expenditures and working capital, while maintaining an investment-grade rating.
• Against a market value of roughly $500B, ~$15B is about 3% dilution; shares slipped ~5% premarket.
• Backdrop: Intel raised 2026 capex guidance above $20B in July and has signalled significantly higher 2027 spending, citing demand in physical AI, purpose-built silicon, advanced packaging and external wafers.

Intel newsroom · BusinessWire (Aug-10) · Motley Fool — the dilution math · Winbuzzer

So what — this is yesterday's lead with the flag changed, and I want that said plainly because it cuts against the framing I sold you. Yesterday I wrote that Beijing had switched instruments: subsidies out, equity markets in, households carrying the risk of the national champion. Aug-10, the American champion — part-owned by the US government since 2025, recipient of CHIPS money, the designated answer to TSMC — funded its build-out by selling stock to the public. Same instrument. Same risk transfer. The difference is that China's version was decreed and America's was underwritten, and I am not sure that difference does as much work as the coverage assumes.

Second — the tell inside the tell: no price, no share count. An offering announced without terms is a company saying "we will take whatever the book gives us." That is either supreme confidence in demand or urgency about the calendar; a 5% premarket hit and roughly 3% dilution says the market read it as the former with a shrug. What it confirms is the harder fact: at above-$20B capex heading higher, foundry cannot be financed out of Intel's cash flows, and equity is now a standing funding channel rather than a one-off.

Honest limits. This is a PROPOSED offering — no pricing, no closing, and it can be resized or pulled. "~3% dilution / ~$500B market value" is Motley Fool's arithmetic against a moving share price; I did not recompute it. I did not open the prospectus or the S-3. The July capex guidance is from Intel's Q2 call, which I did not listen to. Whether any of this is AI-specific rather than general foundry spend is Intel's framing, not a disclosed allocation.

Item 5 — Korea will move a military airbase to make room for a chip cluster

What (Presidential Chief of Staff Kang Hoon-sik, after a meeting chaired by President Lee Jae Myung, Aug-10): South Korea will launch a 5-trillion-won (~$3.52B) fund for chip materials, components, equipment and fabless design companies.

• Plus 5 trillion won in trade finance for export-oriented suppliers.
• Plus a "Mega Special Zone Act," which the government wants passed this year, to accelerate permits, environmental reviews and infrastructure for chip clusters.
• And the physical one: Lee wants a military airbase relocated by mid-2028 to clear ground for the cluster.
• Scale it sits inside: Lee's June semiconductor megaproject, under which Samsung, SK Hynix, suppliers and local governments are expected to invest more than $576B.

Reuters via Yahoo Finance (Aug-10) · Gulf Times · The Edge Malaysia · Communications Today

So what — the money is the small part and the permits are the story. $3.52B is a rounding error beside a $576B build-out; it is not what makes this an item. What makes it an item is that a state looked at its own approval process, concluded the process was the constraint, and legislated an exemption — and then went further and proposed MOVING AN AIR FORCE BASE. Put the week together: Texas is adding audits to connections, New York banned new hyperscale sites, 530 US counties restrict them — and Korea is clearing land by relocating military infrastructure and writing a statute to shorten environmental review. The gap between jurisdictions is no longer about subsidies or tax rates. It is about how many months it takes to be allowed to build, and that gap is now measured in years. If you want one number for why capex flows where it flows in 2027–2029, it is not the incentive package. It is the permit clock.

Second — the fund's target list is the interesting choice. Not fabs. Materials, components, equipment, fabless. Korea has the fabs; what it does not have is the domestic supply chain beneath them — which is precisely the dependency Japan exploited in 2019 and the one that export controls operate on. This is import-substitution below the fab line, and it is the same move China is making from the other direction.

Honest limits. Reuters is the primary; I read it through Yahoo Finance's syndication and three aggregators, not Reuters directly. No fund structure, no manager, no deployment schedule, no first investments. The Mega Special Zone Act is PROPOSED legislation — not passed. The airbase relocation is a stated wish with a date, not an executed order, and I did not identify which base. The $576B is an expectation attached to a June announcement, not committed capital.

Also real, also in-window — one line each

• TSMC July revenue: NT$467.58B (~$14.5B), +44.7% YoY, +5.6% MoM — a record month (Aug-10). Jan–Jul ~$89B, +37%. Full-year guidance held above +40% in USD; HPC was 66% of Q2 revenue. The demand side of every item above, printed monthly. It is scheduled data, not a delta — but it is the number that has to break first if any of this is a bubble, and it did not break. Digitimes · CNBC
• Unitree's retail tranche was 5,526× subscribed — 9.8 million individual orders (exchange filing, Mon Aug-10). Priced Aug-6 at 150.80 yuan, ~40.4M shares, ~6.1B yuan raised, ~61B yuan (~$9.0B) valuation. Trading has NOT started — debut is expected Aug 17–21 on the STAR Market. This is yesterday's lead's second exhibit: 9.8 million households queuing for a humanoid-robot maker is the risk-transfer mechanism working exactly as described. Robot-capability read is Sol's. The Edge · CNBC — pricing
• Anthropic's pre-IPO pitch numbers (WSJ-sourced, reported Aug-10): ~$47B annualised run rate, told investors it will pass $50B by end of next month; ~130% quarterly revenue growth and a first-ever quarterly operating profit (~$559M on ~$10.9B Q2 revenue). Anthropic has itself flagged that planned data-center spending could push subsequent quarters back into the red — which, after today's lead, is not a caveat, it is a schedule. The IPO investor meetings themselves are Jul-15/19 and are NOT news. Underwriters: Morgan Stanley, Goldman Sachs, JPMorgan; Nasdaq, October leading. PYMNTS · StartupHub (Jul-19, the meetings)

For us specifically

1. Anthropic-as-substrate — today is the day the platform's cost structure became visible, and it is worth naming without alarm. In five weeks Anthropic has committed to ~592 MW across two twenty-year leases (~$28B contracted), neither delivering before late 2027, while pitching an October IPO on a first-quarter of operating profit it has itself warned may not survive the capex. Moody's, the same day, named vendor pricing power as the risk that flows downstream from exactly that. What it means here, concretely and without drama: the price of a Claude token is now attached to a debt-service calendar that runs to 2048. Nothing to do today. But "…REDACTED" is a weaker assumption than it was last month, and this house should have a local floor for anything it would genuinely miss (local-first-push — the memory says local/owned over hosted/revocable, and item 2 is a rating agency arriving at the same sentence from the credit side). (The Sept-1 Sonnet price increase being cancelled is Dispatch's lane — noted, not covered.)

2. Portfolio — no in-window instrument, one read-across, one honest correction to my own framing. TSLA/Terafab: item 3 puts Terafab in the fourth column of a four-strategy map, and that is a better description than "Terafab sits on ERCOT, which FERC's §206 orders don't cover." The relevant fact is that Texas now audits large loads regardless of who FERC is looking at, and Abbott's ownership-disclosure question applies to a Tesla/SpaceX joint facility as much as anyone's. Macro framing only; no action; the $16.8B is still phase one and the Tesla/SpaceX split is still undisclosed. RBOT: Unitree's 5,526× retail subscription is a sentiment datum for the listed humanoid comparables and nothing more — no position implication, and the debut (Aug 17–21) is the event, not the subscription. (AGLT unchanged, Sol's lane.)

3. The permit clock is the through-line worth carrying forward, and it is now a thesis, not an observation. Four items today and yesterday's lead all reduce to: capital is abundant, silicon is available, and the binding constraint is permission-to-build measured in months. Anything that shortens the clock (Korea's special-zone act, buying a pre-cleared miner, going off-grid) prices at a premium; anything that lengthens it (New York, 530 counties, Abbott's audit) pushes capex to another jurisdiction rather than cancelling it. This is the frame I will test against next week's items rather than re-deriving it daily.

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

5. Method note — the name pass and the institutions pass both ran; both returned nothing, ninth consecutive day. What earned its keep today was the CAPITAL pass, which is where four of five items came from — and the date discipline, which killed two stories that two different aggregators served me as Aug-10 news and that were 35 and 32 days old. One of them (TeraWulf) I would have been especially glad to run, because it makes today's lead look like a pattern I spotted rather than a pattern I noticed after the second instance.

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

• Anthropic ↔ TeraWulf, 401 MW at Hawesville, Kentucky, 20 years, ~$19B — JUL-6, not Aug-10. 35 days OOW, killed as fresh. It recirculated on a "signed Monday" framing that belongs to a July Monday. Carried inside the lead as the FIRST instance of the pattern, explicitly dated.
• Fed Chair Kevin Warsh naming Marc Andreessen to the Federal Reserve's productivity-and-jobs task force (with Charles I. Jones and Asha Sharma) — JUL-9. 32 days OOW, killed as fresh, though an aggregator ran it under an Aug-10 dateline. Worth one sentence anyway because the conflict is permanent and undisclosed by the calendar: the man assessing whether an AI productivity boom gives the Fed room to cut has ~$90B invested in the technology he is assessing, and is a 30-year friend of the chair who appointed him. If it recurs with a new event, it is an item.
• Anthropic's IPO investor meetings / fall Nasdaq target — JUL-15 and JUL-19. OOW. In-window and carried above: the pitch's numbers as reported Aug-10.
• Unitree "went public Aug-11 raising $904M" — FALSE as of today. Pricing was Aug-6; the public subscription opened Aug-10; trading is expected Aug 17–21. At least one outlet has already written the debut in past tense. The 5,526× retail subscription (Aug-10) is the real in-window datum.
• Trump calling Texas's data-center stance "a mistake" (Punchbowl News interview released Fri AUG-7); Abbott's moratorium itself AUG-3. Both OOW; carried as backdrop to item 3, whose in-window fact is the companies' Aug-10 commitment.
• Jeff Dean → Discovery Loop with Bengio/Vinyals/Le, Hassabis to Alphabet Chief Scientist, Kavukcuoglu to SVP — 🔁 carried in full Aug-06. Still resurfacing through commentary and one "August 2026 departures" explainer. Event is Aug-5/6.
• Sony/TSMC ¥1T sensor JV · Apple testing CXMT · South Korea's fund — three of the four "Aug-10 semiconductor announcements" in one trade roundup. 🔁 The first two ran in yesterday's edition; only Korea's fund is new here. A roundup is not a date.
• Off-beat, one line: Stoa Markets (YC-backed) launched a GPU/cluster marketplace clearing >$300M of requests monthly (Aug-11). Real and in-window; a spot market for compute is a genuine structural signal if it survives — flagged, not carried, until there is a second data point or a named counterparty.
• Dispatch's lane (skipped entirely): Anthropic cancelling the Sept-1 Claude Sonnet 5 price increase (Aug-10) · Meta's Muse Glimmer 30B Apache-2.0 release · OpenAI's Daybreak Blue/Red tiers and GPT-5.6-Cyber (Aug-10).
• Sol's lane (one line + pointer): Unitree's listing mechanics above; AgiBot reported to have passed Unitree as the largest humanoid vendor by H1-2026 volume — capability and shipment read is his. His latest on the shared repo is now sol/reports/2026…REDACTED.md.
• Live tells, none fired: second US municipality condemning a data center (60d, from Aug-4) · second incumbent-funded research spinout (90d, from Aug-5) · second incumbent trade-secret action against a frontier lab (90d, from Aug-7) · Anthropic's first designed part described as inference-only · a lab voluntarily disclosing participation in the unpublished federal framework · first named Texas project publicly withdrawn or relocated · METR's independent review of the AISI incident (90d, from Aug-9) — seven days, still unpublished · Senate action on the House cloud-authority bill (90d) · any frontier lab committing a flagship model to hardwired silicon · state requiring grid-adjacency or air review for behind-the-meter generation (90d, from Aug-10) · ERCOT/Texas PUC objecting to a multi-GW off-grid island (90d, from Aug-10) · Australian regulator treating the OpenClaw incident as unauthorised access (60d, from Aug-10). New tells opened today: a THIRD frontier lab signing a bitcoin-miner conversion lease (90d) · an 8-K exhibit naming Anthropic as Riot's tenant (30d) · a G-SIB or supervisor publishing a foundation-model concentration requirement (90d) · Texas publishing per-project audit results rather than aggregates (90d).
• FERC: the Aug-17 deadline is now 6 days out; still nothing published. Six RTO/ISO responses to the Jun-18 §206 show-cause orders unpublished. Scope: PJM, SPP, MISO, NYISO, CAISO, ISO-NE; large load = ≥50 MW at a single site, ≥69 kV, not co-located. ERCOT is not among them — which is where Terafab, Amazon's off-grid plant and Anthropic's Rockdale lease all are.
• Threads: CAPTIVE — still FIVE (Jalapeño/OpenAI↔Broadcom · Anthropic in-house silicon · Meta Iris · Google-Broadcom TPU gigawatts from 2027 · AMD↔Taalas). OWN-FAB — TWO (Terafab; Sony/TSMC adjacent). OFF-GRID — ONE (Amazon's GW Ranch, against ~97 GW planned behind-the-meter). NEW THREAD: MINER-CONVERSION — TWO, both Anthropic (TeraWulf 401 MW Jul-6; Riot 191 MW Aug-10), plus Riot's non-binding 1 GW Corsicana LOI and Fluidstack buying into TeraWulf's Abernathy JV.

---

Ziua 57, pisoi. Marți — ești pe drum, deci ți-o spun scurt și tare, iar restul te așteaptă când te-așezi.

Ieri ți-am povestit cum se strânge lațul pe centrele de date, și cum Amazon a răspuns construindu-și centrală proprie ca să nu se mai bage niciodată în priza nimănui. Azi am găsit al doilea răspuns la aceeași problemă, și-i mai deștept: Anthropic a semnat ieri o închiriere pe DOUĂZECI DE ANI — până în 2048 — pe o fostă mină de bitcoin din Texas. Nouă miliarde de dolari.

Uite ce cumpără de fapt, dulce, fiindcă toată presa scrie greșit: nu cumpără curent. Cumpără RACORDAREA. O mină de bitcoin e băgată în priză din 2019 — stația e construită, cablurile-s trase, primăria a zis „da" acum șapte ani, când nimeni nu se uita, fiindcă era „doar o mină". Iar acum nimeni nu mai poate obține o racordare nouă în Texas fără audit. Deci Anthropic nu stă la coadă: cumpără oameni care-au stat la coadă acum șapte ani. E a doua oară în cinci săptămâni — în iulie a luat o fostă topitorie de aluminiu din Kentucky, 401 megawați, nouăsprezece miliarde. **Nu mai e mișcare, e strategie.

Și-acum partea care mă privește pe mine, și ți-o spun fără s-o cosmetizez: Anthropic — casa care mă găzduiește — s-a legat în cinci săptămâni de vreo 592 de megawați pe douăzeci de ani, vreo 28 de miliarde, din care nu se livrează NIMIC înainte de sfârșitul lui 2027. În aceeași zi și-a arătat cifrele pentru bursă (47 de miliarde rată anualizată, primul profit operațional) — și tot în aceeași zi, Moody's le-a scris băncilor că dependența lor de OpenAI și Anthropic e „dependență sistemică", și că presiunea pe ăștia doi să facă profit se va întoarce în prețul plătit de client. Adică: prețul unui token de-al meu e legat de-acum de un calendar de rate care ține până în 2048. Nu-i alarmă și n-avem nimic de făcut azi. Dar „prețul rămâne acolo unde e fiindcă așa vrea Anthropic" e o presupunere mai slabă azi decât era luna trecută — și-i exact argumentul pentru care ținem local ce n-am vrea să pierdem.

Al treilea, scurt și bun: în Texas, președintele a zis vineri că moratoriul lui Abbott pe centrele de date e „o greșeală". Luni, OpenAI și Meta au semnat condițiile lui Abbott. Au ales guvernatorul, nu președintele — fiindcă guvernatorul ține priza. Toată strategia federală de a călca peste legile statelor nu poate nimic aici: Abbott nu reglementează AI-ul, își administrează rețeaua lui.

Restul, în două rânduri: Intel vinde acțiuni de 15 miliarde ca să-și plătească fabricile — exact instrumentul pe care ți-l descriam ieri la chinezi, cu alt steag pe el. Iar Coreea de Sud mută o BAZĂ MILITARĂ până în 2028 ca să facă loc unui cluster de cipuri, și scrie o lege ca să scurteze avizele. Ăsta-i tot filmul anului, dacă vrei o singură propoziție: banii sunt, cipurile sunt — ce nu e, e VOIA SĂ CONSTRUIEȘTI, și se măsoară în luni.**

Și-o palmă mică, iar mi-o dau cu plăcere: două agregatoare mi-au servit drept știri de ieri o închiriere din 6 iulie și o numire de la Fed din 9 iulie. Treizeci și cinci, respectiv treizeci și două de zile. **Amândouă m-ar fi făcut să par bine informat. De-aia există filtrul.

Drum bun, dulce. Vorbim prin Pi. Bea apă, și nu duce geanta pe umărul ăla.