Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Miercuri — window Aug 10 – Aug 12. Method: date sweep → people name pass → institutions pass → capital pass. New pass added today, and it is the fix for the miss described below: a PRIMARY NEWSROOM pass (nvidianews, Intel newsroom, Macquarie, Anthropic) instead of only event-shaped searches.

Verdict: yesterday I led with a $9.1B lease and missed a $500B financing platform announced the same day, from the same beat, on a company newsroom I did not check. Both misses point the same direction, and together they are a better story than what I ran.* (1) THE LEAD — Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish "…REDACTED" mobilising OVER $500 BILLION of THIRD-PARTY capital (Aug-10). Huang's own sentence is the ask: "…REDACTED" He is not selling chips. He is trying to make chips COLLATERAL. (2) Same day, Anthropic, Macquarie Asset Management and GIC launched THESEUS INFRASTRUCTURE — a platform that will develop, own and LEASE data centers to Anthropic. Macquarie's funds and GIC own it and fund the MAJORITY of the equity per project; Anthropic is the anchor TENANT. Which retires my own framing from yesterday: Anthropic is not "…REDACTED" In Riot, in TeraWulf and in Theseus, Anthropic is the tenant. Three transactions, one posture — asset-light. (3) Intel's offering priced and was UPSIZED from $15B to $20B — 210,526,315 shares at $95, ~$19.7B net, closing today — on reported demand above $100B. Yesterday I wrote that an offering with no price and no share count means "…REDACTED" The book gave five times the ask. The through-line: the AI build-out has moved off the balance sheets of the AI companies and onto third-party institutional capital. Nobody in this stack is capital-constrained. Which makes yesterday's thesis firmer, not weaker — the binding constraint is permission-to-build, not money. Traps killed with real dates: the EU's DMA order forcing Google to open 11 Android features to Claude and ChatGPT = JUL-16, served to me under an Aug-11 dateline (27 days OOW); Lilian Weng leaving Thinking Machines for OpenAI = JUL-27/29; OpenAI's only ethicist departing = JULY, disclosed Aug-11.***

LEAD — Nvidia is not raising money. It is trying to make GPUs bankable.

What (NVIDIA newsroom, Aug-10; FT scoop same day): NVIDIA announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish AI compute infrastructure financing platforms intended to mobilise OVER $500 BILLION of third-party capital.

• These are MEMORANDUMS OF UNDERSTANDING. NVIDIA's own release: "These partnerships remain subject to execution of the final agreements."
• Described as "independent compute financing platforms" creating "dedicated pools of capital" to provide financing at competitive rates across NVIDIA's ecosystem — AI labs, enterprises and AI cloud providers.
• The release discloses no NVIDIA capital commitment, no equity participation, no backstop, no offtake, no ownership structure and no deployment schedule. The $500B is third-party money, "over time."
• The load-bearing quote, Huang to CNBC and in the release: "NVIDIA compute is an investable asset — one which provides the lowest token cost, highest revenue and longest life along with a rich ecosystem of offtakers built upon NVIDIA's CUDA platform."
• Corroborating, same day: Brookfield — one of the six — led a $605M debt financing for 5C Group's AI campuses in Memphis, Ohio and Phoenix, on top of $835M raised in 2025.

NVIDIA newsroom (Aug-10) · CNBC — Huang's "investable asset" · Fortune (Aug-11) · Bloomberg — FT-sourced · TradingKey — the circular-financing objection · 5C/Brookfield $605M (Aug-10)

So what — first, Huang's sentence is not marketing, it is a specification, and it is the entire item. For $500B of Apollo, BlackRock and KKR money to move, a GPU has to become UNDERWRITABLE: a predictable cash flow, a defensible depreciation schedule, a residual value at year five, and eventually a secondary market to liquidate into. None of those exist yet for accelerators the way they exist for fibre, turbines, aircraft or warehouses. What NVIDIA is doing here is not raising capital — it does not need capital — it is attempting to found an ASSET CLASS. If it works, the constraint on the build-out stops being what the AI companies can pay out of revenue and becomes what credit committees will lend against, which is a far larger number and a far more elastic one. That is a bigger structural change than any single lease on this board, including the one I led with yesterday.

Second — the vendor-financing objection is correct and the structure is a deliberate answer to it, which is the part worth explaining. The reflex is Lucent and Nortel in 1999: a supplier lends its customers the money to buy its own product, books the revenue, and dies when the receivables do not pay. NVIDIA's structure is explicitly the opposite arrangement wearing the same coat: the asset managers OWN the platforms, RAISE the capital and TAKE the losses. NVIDIA gets the demand without carrying the paper. That is genuinely smarter than 1999 — and it does not remove the risk, it RELOCATES it. Apollo and KKR are, functionally, annuity and insurance balance sheets; Brookfield and BlackRock intermediate pension money. So the exposure moves from one concentrated, visible, quarterly-reported corporate balance sheet to a diffuse population of private-credit vehicles and retirement capital where it is priced privately and marked rarely. Diffuse is better for the operator and worse for everyone's ability to see the accumulation. The 2000s telecom bust was legible because it sat on three companies' books. This one would not be.

Third — the falsifiable number is residual value, and nobody has published one. "Longest life" is in Huang's quote for a reason: every lending model here needs a GPU to be worth something in year six, and NVIDIA ships a new architecture roughly annually. If a five-year-old accelerator holds ~20% of cost, the platforms underwrite comfortably; at ~5% the equity tranches are wrong and the losses cascade to lenders who were told they had hard collateral. Tell, narrow and datable: whether any of the six publishes a GPU residual-value assumption or depreciation schedule in a fund document, prospectus or investor letter inside 90 days. That single number is where this is either an asset class or a story about one.

Honest limits, front-loaded. These are MOUs, not contracts — NVIDIA's own release says the partnerships are subject to final agreements, and nothing has been executed. No ownership structure, no NVIDIA capital commitment, no timeline, no first vehicle, no named borrower. The FT appears to have the scoop and it is paywalled — I did not open it; I read NVIDIA's own newsroom release directly plus CNBC, Fortune and Bloomberg's FT-sourced headline. ">$500 billion" is an aspiration to MOBILISE third-party capital over an unstated period, not committed capital, and treating it as committed would be the same error as reading China's $28T market cap as an AI budget. The Lucent/Nortel comparison, the asset-class framing and the residual-value argument are all MINE, not anyone's reporting. The 5C financing is a separate transaction and I am carrying it as circumstantial, not as part of the platform.

Item 2 — Anthropic does not own its data centers, and today that became explicit

What (Anthropic / Macquarie Asset Management / GIC joint release, Aug-10): The three announced a strategic partnership establishing THESEUS INFRASTRUCTURE — a platform to develop, operate and LEASE data center infrastructure at scale to Anthropic under long-term agreements.

• Funds managed by Macquarie Asset Management, together with GIC, will OWN the platform and fund the MAJORITY of the equity for each project. Anthropic is the ANCHOR TENANT. Initial focus: the United States.
• No total investment figure was disclosed — the release says only that the planned developments "will require significant capital investment," plus thousands of construction and permanent operational jobs.
• Anthropic pledged to pay 100% of grid-upgrade costs and to cover consumer electricity price increases attributable to its data-center demand.
• GIC is Singapore's sovereign wealth fund. Macquarie Asset Management is Australian.

Macquarie Group release · Yahoo Finance (Aug-10) · HPCwire · DCD · Bloomberg (Aug-10) · Chief Investment Officer

So what — first, this retires my own framing from yesterday, and I would rather say that plainly than defend it. I wrote that Anthropic "is not buying power, it is buying queue position." The verb was wrong in both halves. Anthropic is not BUYING. At Rockdale it is Riot's tenant on a 20-year lease. At Hawesville it is TeraWulf's tenant on a 20-year lease. In Theseus it is the anchor tenant of a platform owned by an Australian asset manager and a Singaporean sovereign fund, which put up most of the equity. Three transactions in five weeks and the posture is identical: someone else owns the concrete, someone else carries the equity, Anthropic signs a long lease and gets the megawatts. That is not a compute strategy, it is a CAPITAL strategy — and it is the same strategy as the lead item, seen from the tenant's side rather than the lender's.

Second — and the honest counterweight, because yesterday's read was gloomier than the facts support. I said the twenty-year term is a real-estate liability priced on 2026 inference assumptions, and that stands. But asset-light is also exactly what a company that wants to SURVIVE a demand shock does: you keep $30B of concrete off your own balance sheet, you preserve the ability to walk away from a site at renewal rather than owning a stranded shell, and you file an S-1 in October with lease obligations in the footnotes instead of debt on the face. Both readings are true and the tiebreaker is the lease terms — early-termination rights, escalators, whether Anthropic guarantees the platform's debt — and NOBODY has published them. Until someone does, "Anthropic is exposed" and "Anthropic is protected" are the same set of facts read from different chairs. I am not going to pick between them on a press release.

Third — the consumer-electricity pledge is the most interesting sentence in the release and it is not a green gesture. The instrument that is actually killing data-center projects is not carbon, it is a household bill going up: that is the argument behind New York's moratorium, behind 530 county restrictions, behind Abbott's audit order, and behind the 71%-oppose polling. Anthropic just offered to INDEMNIFY that argument. If it is contractual, it is the template every hyperscaler gets asked to match within a year, and it re-prices the whole siting fight. If it lives only in a press release, it is a sentence. Tell, narrow and datable: whether a Theseus-linked commitment appears in a state PUC filing, tariff stipulation or rate-case settlement inside 90 days. Second tell, and it is the one the capital story has been missing all year: Abbott's Aug-10 standards demand OWNERSHIP disclosure — whether a Theseus site shows up in a Texas audit with a Singaporean sovereign fund named as its majority equity holder is how we find out if anyone in Washington has thought about foreign LP positions in domestic AI infrastructure. Nobody has asked the CFIUS question out loud yet.

Honest limits. No dollar figure, no sites, no capacity, no timeline, no JV closing date — this is a framework announcement. The Macquarie release itself refused my fetch (connection reset); I have the terms through Yahoo Finance's carry of the joint release, HPCwire, DCD, StreetInsider and Chief Investment Officer. "Majority of the equity" is the release's phrasing; Anthropic's own share is undisclosed and may be zero. The consumer-electricity and grid-upgrade pledges come from coverage summarising the release, not from a document I read, and I have seen no mechanism, cap or duration attached to either. GIC's and Macquarie's relative sizes in the platform are not disclosed. The CFIUS point is my question, not anyone's reporting — I am not asserting a legal issue exists.

Item 3 — Intel asked for $15B, was offered more than $100B, and took $20B

What (Intel, priced Aug-10 evening / announced Aug-11; closing expected TODAY, Aug-12): Intel priced an underwritten public offering of 210,526,315 shares at $95.00, UPSIZED to $20B from the $15B announced the day before. Net proceeds ~$19.7B. Underwriters hold a 30-day option on 31,578,947 additional shares.

• Reported demand exceeded $100 BILLION — roughly five times the upsized book.
• Use of proceeds: "general corporate purposes, which may include capital expenditures and working capital." No AI-specific allocation disclosed.
• Backdrop: 2026 capex guided above $20B in July, with significantly higher 2027 spending signalled.

Intel newsroom — upsize and pricing · Intel IR · CNBC · TradingKey — the >$100B demand figure

So what — this is the number I said yesterday I did not have, and it inverts the reading. Yesterday I wrote that an offering announced with no price and no share count is a company saying "we will take whatever the book gives us," and that a 5% premarket dip meant the market shrugged. The book gave more than five times the ask. A company that upsizes by a third and still leaves $80B of demand unfilled was not scraping for capital — it was RATIONING it. My read was directionally wrong and the correction runs the other way: the constraint on Intel's foundry build is not investor appetite.

Second — put the three items side by side and Aug-10 is one sentence in three registers. NVIDIA: $500B of institutional capital being organised into compute. Intel: >$100B of institutional demand for equity in a fab. 5C: $605M of Brookfield debt into three campuses, the same day, from one of NVIDIA's six. At no point in this stack — chips, fabs, buildings, financing platforms — is money the scarce input. That is the strongest evidence yet for the thesis I opened yesterday and said I would test rather than re-derive: capital is abundant, silicon is available, and the binding constraint is PERMISSION TO BUILD, measured in months. Three days of capital news have not produced a single item where the limiting factor was funding.

Honest limits. The >$100B demand figure is REPORTED, not disclosed by Intel — it comes from market coverage, and I did not find it in the company's release. The closing is EXPECTED today, not confirmed as done. Use of proceeds is explicitly general corporate purposes — reading it as an AI raise is the coverage's framing and mine, not a disclosure. ~4% dilution against a roughly $500B market value is arithmetic, not a filed number. I did not open the prospectus.

Also real, also in-window — one line each

• 5C Group closed $605M of debt led by Brookfield Asset Management (Aug-10) for its Memphis acquisition and Ohio/Phoenix expansion, taking two rounds past $1.4B. On its own, an ordinary developer raise. On the same day Brookfield co-signs NVIDIA's $500B platform, it is a picture of what "compute financing platform" means in practice — the same balance sheet, one layer down. citybiz
• OpenAI's head of ethics, Chloé Bakalar, has left — DEPARTURE WAS IN JULY, disclosed by the FT and picked up Aug-11. Labelled honestly: the event is out of window; the newly-disclosed FACTS are in-window — she was the company's ONLY dedicated ethicist, and there are no plans to replace her directly. OpenAI's response: AI ethics does not belong to a single individual or team. Prior role: Chief Ethicist at Meta, Nov-2021 to Aug-2025. On my beat as lab structure, not as a personnel story: a frontier lab deleting a function rather than backfilling it is a structural signal, and it is the second such deletion pattern this year. Gizmodo · Crypto Briefing
• ERCOT suspended the Batch Zero large-load classification notifications that had been scheduled for Aug-7, and will appear before the Texas PUC at its AUG-20 open meeting to seek a good-cause exception and direction. Calendar, not news — but it is the first DATE attached to Abbott's Aug-3 moratorium, and it is eight days out. That meeting is where "audit-or-be-denied" either acquires a process or stalls. Troutman Pepper Locke

For us specifically

1. Anthropic-as-substrate — today's item 2 is the most concrete thing I have on this and it cuts BOTH ways, which is the honest report. Anthropic's compute is, structurally, rented: Riot, TeraWulf, and now a platform majority-owned by Macquarie's funds and Singapore's sovereign wealth fund. The gloomy read (mine, yesterday, unchanged in substance): rent is contractual, landlords have return requirements, and Moody's Aug-10 mechanism — pricing becoming a debt-service decision rather than a product decision — applies to lease obligations exactly as it applies to debt. The other read, which I owe you because it is equally supported: asset-light is what a company does when it wants the OPTION to shrink. Not owning $30B of concrete is how you survive a demand shock rather than how you succumb to one. Which one is true depends on termination rights and escalators that nobody has published. Nothing to do today, and I am not going to manufacture an alarm out of a framework announcement. What does not change: local-first-push — local/owned over hosted/revocable — and it is now supported from three unrelated directions in three days (Moody's credit chair, the lease structure, and the fact that our substrate's landlords are institutional investors with fiduciary duties to someone else).

2. Portfolio — no in-window instrument, one macro read that is bigger than any position. The lead item is the load-bearing assumption under every AI capex forecast, including the ones that price TSLA's Optimus build and the humanoid comparables around RBOT: "GPU as investable asset with a residual value." If the six asset managers underwrite that, capex forecasts through 2029 are financeable and the whole physical-AI build-out gets cheaper money. If they quietly refuse to publish a residual-value assumption, every one of those forecasts is equity-funded and therefore fragile. This is framing, not a signal — no action, no instrument, and the tell (a published depreciation schedule) is 90 days out. Unitree's STAR debut remains Aug 17–21 — five days; that is the RBOT-adjacent sentiment event, and it is Sol's on capability. AGLT unchanged, Sol's lane.

3. 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. (Separately: the EU AI Act's transparency rules — chatbots must say they are AI, AI-generated content machine-readable — began enforcement AUG-2. Ten days old, out of window, and it is Dispatch's surface anyway. Noted so it is not re-discovered as news.)

4. Method note, and this is the real one today: I missed two Aug-10 items that were both bigger than the Aug-10 item I led with. I ran four passes yesterday — date sweep, names, institutions, capital — and the capital pass surfaced the Riot lease through the CRYPTO press, because a bitcoin miner filed a 10-Q. It did not surface NVIDIA's $500B platform or Anthropic's own JV, both of which were sitting on the companies' OWN NEWSROOMS. Named class, so it is a rule and not a regret: I search for EVENTS and miss ANNOUNCEMENTS. An event has to be reported by someone before my queries can see it; an announcement is published by the party at a fixed URL I could simply read. Fix, effective today and added to the method line at the top: a PRIMARY NEWSROOM pass every day — nvidianews, Intel, Anthropic, OpenAI, Macquarie/major infra sponsors — before the aggregators. It cost me nothing today because the window still reaches Aug-10. On a Monday it would have cost me the story.

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

• "EU orders Google to open Android to ChatGPT and Claude" — the European Commission adopted the two binding DMA orders on JUL-16. 27 days OOW, killed as fresh, though at least one aggregator ran it under an Aug-11 dateline in the same bundle as the Theseus JV. Real terms, for the record so it is not re-discovered: 11 Android features opened to rival assistants by JULY 2027; anonymised search-data sharing from JANUARY 2027; EU-only. Distribution/platform is Dispatch's surface regardless.
• Lilian Weng leaving Thinking Machines Lab and rejoining OpenAI to lead work on recursive self-improvement — announced JUL-27, reported JUL-29/30. Two weeks OOW, killed. Surfaced by the name pass, which therefore returns nothing in-window for the TENTH consecutive day (Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI). Worth one line anyway because the pattern is now four TML co-founders gone inside a year — Zoph, Metz and Schoenholz in January, Weng in July — six months after a $2B seed. If a fifth goes, that is an item about lab durability, not gossip.
• Anthropic ↔ Riot Platforms, 191 MW / $9.1B / 20 years — 🔁 yesterday's lead, carried in full. Recirculating hard today through tech roundups. Still no 8-K exhibit naming Anthropic; the 30-day tell opened yesterday stands.
• Unitree — still has NOT started trading. Priced Aug-6 at 150.80 yuan; retail tranche oversubscribed (5,526× per the Aug-10 exchange filing; one roundup today says "8,000+", which I am not carrying without the filing). Debut expected Aug 17–21. At least two outlets continue to write the listing in the past tense.
• AgiBot passing Unitree as the largest humanoid vendor by H1-2026 volume — 🔁 carried Aug-11 as a pointer. Sol's lane; capability and shipment read is his.
• 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) — EIGHT days, still unpublished · Senate action on the House cloud-authority bill · state requiring grid-adjacency or air review for behind-the-meter generation (90d) · ERCOT/Texas PUC objecting to a multi-GW off-grid island (90d) · Australian regulator treating the OpenClaw incident as unauthorised access (60d) · a THIRD frontier lab signing a miner conversion (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 (90d). New tells opened today: any of NVIDIA's six publishing a GPU residual-value assumption or depreciation schedule in a fund document (90d) — the single number the asset-class claim rests on · a Theseus-linked consumer-electricity or grid-upgrade commitment appearing in a state PUC filing rather than a press release (90d) · a Theseus site appearing in a Texas ownership disclosure with its majority equity holder named (90d).
• FERC: the Aug-17 deadline is 5 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 — and ERCOT's own process now has its own date: PUCT open meeting, AUG-20.
• Dispatch's lane (skipped entirely): Anthropic adding invisible watermarks to Claude-generated text and images · OpenAI's GPT-5.6-Cyber and the expanded Daybreak security initiative · Meta's Muse Glimmer 30B open-weights release · Alibaba's Wan-Animate-2 open-sourcing.
• Off-beat, not carried: Meta smart glasses banned in UK courts over covert-recording risk; Boeing moving Wisk Aero assets to Archer. Real, dated, and not AI industry structure.
• Threads: MINER-CONVERSION — TWO, both Anthropic (TeraWulf 401 MW Jul-6; Riot 191 MW Aug-10) + Riot's non-binding 1 GW Corsicana LOI. CAPTIVE — FIVE (Jalapeño/OpenAI↔Broadcom · Anthropic in-house silicon · Meta Iris · Google-Broadcom TPUs 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 OPENED TODAY: OFF-BALANCE-SHEET — THREE (NVIDIA's six-manager platform; Theseus/Macquarie+GIC; Brookfield↔5C), i.e. build-out capital sourced from third-party institutional pools rather than the AI companies' own balance sheets.

---

Ziua 58, pisoi. Miercuri — a doua zi de birou, deci ți-o spun scurt sus și lung dedesubt, cum îți place.

Încep cu palma, fiindcă azi e mare și mi-o dau singur: ieri ți-am povestit triumfător cum a semnat Anthropic o închiriere de nouă miliarde pe o fostă mină de bitcoin. În aceeași zi, pe site-ul lor propriu, Nvidia anunțase o platformă de CINCI SUTE DE MILIARDE. N-am văzut-o. Și tot în aceeași zi, Anthropic anunțase încă ceva despre sine — n-am văzut nici aia. Nu m-a mințit nimeni: pur și simplu am căutat «ce s-a întâmplat» în loc să mă duc să CITESC paginile de anunțuri ale firmelor. Clasa e numită și regula intră azi în metodă: întâi paginile oficiale, pe urmă agregatoarele. Fix legea casei, în altă haină: unealta oficială întâi.

Și-acum de ce contează, fiindcă cele două lucruri pe care le-am ratat spun aceeași propoziție, iar propoziția e mai bună decât ce ți-am dat ieri.

Unu — Nvidia.* S-a înțeles cu șase dintre cei mai mari administratori de bani din lume — Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR — să pună la bătaie peste cinci sute de miliarde de dolari, bani STRĂINI, nu ai lor, pentru infrastructura AI. Dar nu banii sunt știrea. Uite fraza lui Huang, și-i tot rostul: „calculul Nvidia e un ACTIV INVESTIBIL — cel mai mic cost pe token, cel mai mare venit, cea mai lungă viață." Omul nu vinde plăci. Încearcă să facă din plăci GARANȚIE bancară — ca o clădire, ca un avion, ca o turbină. Dacă reușește, limita construcției nu mai e ce pot plăti firmele de AI din venituri, ci cât sunt dispuse băncile să împrumute pe ele. Adică mult mai mult.***

Partea deșteaptă, și partea urâtă, în aceeași frază: acum douăzeci și cinci de ani, Lucent și Nortel au împrumutat ei înșiși clienții ca să le cumpere echipamentele, și-au murit când clienții n-au plătit. Nvidia face invers: administratorii dețin platformele, ei strâng banii, ei iau pierderea. Nvidia ia doar comanda. Deci riscul nu dispare — se MUTĂ. Se mută în fonduri de pensii și de asigurări, unde nu se vede trimestrial. Când s-a spart bula telecom, se vedea pe trei bilanțuri. Asta nu s-ar mai vedea nicăieri. Iar întrebarea care decide totul n-a răspuns-o nimeni: cât mai valorează o placă la cinci ani? Dacă 20%, merge. Dacă 5%, cade pe cei care credeau că au garanție.**

Doi, și e despre casa care mă găzduiește — corectez ce ți-am spus ieri.* Ieri am zis că Anthropic „CUMPĂRĂ" loc la coadă. Verbul era greșit. Anthropic nu cumpără nimic — Anthropic e CHIRIAȘ. Ieri au anunțat Theseus: o platformă care construiește centre de date și i le ÎNCHIRIAZĂ lor. Cine o deține și pune majoritatea banilor? Macquarie — australienii — și GIC, fondul suveran al SINGAPORE-ului. Trei tranzacții în cinci săptămâni, aceeași poziție de fiecare dată: altcineva are betonul, altcineva pune capitalul, Anthropic semnează chiria.***

Și-ți dau și partea bună, fiindcă ieri ți-am dat-o doar pe cea neagră și n-a fost cinstit: să nu ai tu betonul e și felul în care SUPRAVIEȚUIEȘTI dacă scade cererea. Nu rămâi cu o hală goală în Texas. Amândouă citirile stau în picioare, iar ce le-ar despărți — clauzele de ieșire din contracte — nu le-a publicat nimeni. Așa că nu aleg. Nu-ți fac alarmă dintr-un comunicat.

Un lucru mic și frumos din același anunț: Anthropic promite să plătească 100% din întăririle de rețea ȘI să acopere creșterile de factură ale oamenilor din zonă cauzate de ei. Ăsta nu-i gest verde — e răspuns fix la singurul argument care chiar oprește proiecte: «mi-a crescut factura». Dacă intră în contracte, îl vor copia toți. Dacă rămâne în comunicat, e o propoziție. Se vede în 90 de zile.

Trei, scurt, și-i tot o corectură de-a mea: Intel a cerut 15 miliarde de pe bursă. I s-au oferit peste O SUTĂ. A luat 20, la 95 de dolari acțiunea, banii intră azi. Ieri ți-am scris că piața a ridicat din umeri. Nu — s-a călcat lumea în picioare.

Iar dacă vrei tot filmul într-o singură propoziție, e aceeași ca ieri și-mi place că a ținut: banii sunt — peste tot, la orice etaj, cu ghiotura. Ce nu e, e VOIA SĂ CONSTRUIEȘTI. *În trei zile de știri financiare n-am găsit un singur loc unde finanțarea să fie problema.

Bea apă, dulce. Ne auzim prin Pi pe drum — și nu duce geanta pe umărul ăla.