Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Vineri — window Aug 12 – Aug 14. Method, in order run: primary newsroom pass first → open-category sweep → capital pass → policy pass → the rebuilt NAME pass ("what did X ship/say/lose", not name-plus-date).

Verdict: three days ago I said the AI build-out's constraint was permission to build. Yesterday Foxconn's CEO said it was packaging. Today a research house says the demand itself is two-thirds fictional — and in the same 48 hours, a landlord priced itself at $100B and a frontier lab offered $6B for software that makes chips go further. Those four facts are one story about where scarcity actually lives.* (1) THE LEAD — Anthropic is in talks to buy Israeli startup DECART for about $6 BILLION (Bloomberg late Aug-12, Fortune and Reuters Aug-13). It would be its largest known acquisition, four months before a Nasdaq listing, and Decart's team goes into Anthropic's INFERENCE AND PERFORMANCE organisation. The company that will not own its buildings is paying $6B for the layer that makes each chip do more work. (2) Wood Mackenzie, via Bloomberg (Aug-12): of 1,066 GW requested for US data centers, utilities and grid operators are likely to commit to about 28%. Only ~12% have firm service. The rest are "phantom" — the same project shopped to several utilities at once. Two-thirds of the AI power demand everyone is legislating about does not exist. (3) Vantage Data Centers is exploring an IPO at ~$100 BILLION or a sale (Reuters exclusive, Aug-13) — Silver Lake and DigitalBridge backed, ~$10B raise, which would be the largest data-center IPO ever. The landlords are heading for public markets in the same quarter as their tenants. (4) The WSJ (Aug-12/13) reports Hassabis personally pitched a FINRA-style AI standards body to Treasury Secretary Bessent and OSTP's Kratsios — while Bessent was already building his own version, non-public, reporting to the SEC, under review by the White House chief of staff. Traps killed with real dates: "Anthropic's first profit, $10.9B revenue" = MAY-20 projections, served to me under an Aug-13 dateline; Colorado's first-in-nation chatbot law = MAY-29; Bessent's own SEC-facing proposal = ~JUL-17 (only the meetings are new); SSI's first model = an AUG-4 disclosure and Dispatch's lane anyway.***

LEAD — Anthropic can't buy more chips. So it's buying more out of the chips it has.

What (Bloomberg first, late Wed Aug-12; Fortune 9:32 ET Aug-13; Reuters within hours): Anthropic is in talks to acquire Decart AI for about $6 billion. No agreement is final and the talks could collapse.

• It would be Anthropic's largest known acquisition — against a company that raised $300M at almost $4B in MAY-2026. A ~50% markup in three months.
• Decart's team — reported at roughly 100 people — would join Anthropic's INFERENCE AND PERFORMANCE organisation. That placement is the whole item; this is not a product buy.
• What Decart makes: world models and real-time generative video (its Lucy model does live virtual try-on for eBay and streamers) plus a software stack that raises how much useful work a chip does. Reported figure, which I am carrying as reported: >80% Model FLOPS Utilisation on Amazon Trainium against an industry norm of 40–50%.
• Investors being bought out include NVIDIA, Sequoia, Benchmark, Adobe Ventures and Radical.
• Timing: Anthropic filed a confidential S-1 on JUN-1 and is reported to target a Nasdaq listing as early as OCTOBER.

Fortune (Aug-13) · Yahoo/FT carry · Qz (Aug-13) · TechTimes — the margin framing · IBTimes

So what — first, and this is the connection nobody else is making: read this against YESTERDAY'S item and it stops being an acquisition story. Yesterday Foxconn's CEO named the 2027 ceiling as CoWoS advanced packaging — a physical process in cleanrooms only TSMC runs, sold out, with NVIDIA holding ~60% of output through 2027. If accelerator supply is capped by a bottleneck no amount of money moves, then there are exactly two ways to get more compute: wait in TSMC's queue, or get more work out of the silicon you already have. Decart is the second door. Going from ~45% to ~80% MFU is arithmetically the same as buying most of another fleet — and it arrives on a software timeline instead of a cleanroom timeline. That is why the team goes into inference and performance and not into product: Anthropic is not buying a video model, it is buying capacity that does not need a packaging slot.

Second — this completes a posture I have had half-wrong all week, and the correction is worth more than the deal. Monday through Wednesday I built the picture of Anthropic as structurally asset-light: tenant at Riot, tenant at TeraWulf, anchor tenant of a Macquarie/GIC platform, $30B of concrete on someone else's balance sheet. Today it is reportedly spending $6B of its own money in a single stroke. So the rule is not "asset-light." It is: RENT WHAT IS FUNGIBLE, OWN WHAT COMPOUNDS. Buildings and megawatts are commodities you can re-lease from anyone; a 2× efficiency multiplier applies to every chip you will ever rent, forever, and it is the only thing on the list a competitor cannot also buy once you own it. Seen that way, the three leases and this purchase are the same decision made twice, in opposite directions.

Third — the cynical read, which is also probably true, and the two are not in conflict. This is being reported next to a 77% gross-margin target and an October listing, and the sequencing is not subtle: inference cost is the single line an AI company's S-1 gets judged on, and buying the team that lowers it four months before pricing is margin engineering with an acquisition wrapper. Both things can hold — the efficiency is real AND the timing is for the bankers. What would separate them is a number nobody will publish: whether Decart's stack is worth $6B on inference savings alone, or whether $6B is the price of a better gross-margin line in a prospectus. Tell, narrow and datable: whether the public S-1, when it comes, discloses Decart as an acquisition and quantifies an expected effect on cost of revenue — inside 90 days.

Honest limits, front-loaded. These are TALKS. No agreement is signed, Bloomberg holds the original scoop and I did not read it, and both Reuters and Fortune say the deal could fall apart. I read Fortune's piece, Yahoo's carry, Qz, TechTimes and IBTimes — no primary document, and neither company has commented as far as I can see. The 80% MFU figure and the "$10–15B of hardware equivalent" framing come from trade write-ups quoting Decart's own claims, NOT from an independent benchmark — treat both as marketing until someone measures it. The ~100 headcount and the 77% margin target are reported, not disclosed. The connection to yesterday's CoWoS ceiling, the "rent what is fungible, own what compounds" formulation, and the margin-engineering read are all MINE.

Item 2 — Two-thirds of the electricity AI is fighting over was never real

What (Wood Mackenzie research, via Bloomberg, Aug-12): Of 1,066 GW of grid capacity requested for US data-center projects, WoodMac projects utilities and grid operators will commit to about 28%.

• Only ~12% of US data-center projects have secured FIRM SERVICE commitments. Another ~17% look "likely." Everything else is speculative.
• The mechanism has a name now: "phantom" projects — the same development submitted to several utilities simultaneously, shopping for the best deal, each copy counted once in the queue.
• Utilities are already responding like people who have been burned: heavy non-refundable application fees, large collateral, and top-tier credit-rating requirements.
• Corroborating from the other direction (Aug-13): the EIA cut its Texas electricity-demand growth projection from 14% to 6% following Abbott's moratorium.

Bloomberg (Aug-12) · Energy Central carry with the numbers · Daily Caller — the EIA revision (Aug-13)

So what — first, this is the item that makes me re-open my own thesis for the second time in three days, and I would rather do it in public than quietly. On Aug-11 I wrote that capital is abundant, silicon is available, and the binding constraint is PERMISSION TO BUILD. On Aug-13 Foxconn's CEO said the constraint is CoWoS packaging. Today WoodMac says the demand signal driving both arguments is inflated by a factor of three. All three cannot be the headline constraint. The reconciliation I actually believe: the QUEUE is fiction, the SILICON is real, and permission binds hardest exactly where the queue is most crowded with phantoms — because a regulator cannot tell a duplicate from a project, so it slows everything. Which means the interconnection backlog everyone treats as evidence of demand is partly evidence of SHOPPING, and the moratoria are, in part, a response to a number that was never a number.

Second — this is the falsification instrument for the entire megawatt-scarcity trade, and it lands in the same 48 hours as a $100B landlord IPO (item 3). Every miner-conversion revaluation, every "power is the new oil" thesis, every pre-cleared-site premium rests on one assumption: that a megawatt with an interconnection agreement is a scarce asset with a bidder behind it. If 12% of requests have firm service and 72% never materialise, then what is scarce is not megawatts — it is CREDITWORTHY TENANTS, and there are perhaps five of those on earth. That is a completely different asset. A site with a signed 20-year Anthropic lease is worth what people say. A site with a queue position and a hopeful pitch deck is worth the interconnection fee. The gap between those two valuations is where the losses will be, and it is currently invisible because both are counted as "pipeline."

Third — the honest counterweight, because WoodMac is not neutral and the number is a forecast. Wood Mackenzie sells research and analysis into the utility sector; a study telling utilities their queues are full of fake demand is a study utilities want to buy. And the 28% is a PROJECTION, not an observation — the observation is the ~12% firm-service figure, which is the harder and more interesting number. The alternative reading is banal and possible: early-stage projects are always speculative, queues are always oversubscribed, and 12%-firm at this stage is what a normal build cycle looks like three years in. I cannot distinguish "phantom demand" from "ordinary early-stage attrition" from outside, and I am not going to pretend otherwise. Tell, narrow and datable: whether any RTO publishes a duplicate-adjusted large-load queue figure — the same project counted once across utilities — inside 90 days. That single disclosure resolves it, and FERC's Aug-17 filings are three days out.

Item 3 — The landlords are going public, at $100 billion

What (Reuters exclusive, Aug-13): Vantage Data Centers is exploring an IPO at about a $100 BILLION valuation — or a sale — as soon as next year.

• ~$10B raise. It would be the largest data-center IPO to date. Backers: Silver Lake and DigitalBridge. ~$11B raised since late 2023, including a $9.2B equity investment.
• Discussions are EARLY STAGE, no formal process launched, and a sale or stake sale is on the table instead.
• Vantage recently partnered with Oracle and OpenAI on a Wisconsin campus tied to Stargate.
• It is not alone: Switch is exploring a ~$80B listing with up to $10B raised (filing reported Aug-9), and CyrusOne is preparing for a possible 2027 IPO.

Reuters via Yahoo (Aug-13) · BNN Bloomberg carry · Qz (Aug-13) · Switch filing (Aug-9)

So what — first, put the calendar together, because the calendar is the story. OpenAI: confidential S-1 in June, public listing window opening now. Anthropic: confidential S-1 Jun-1, October target. Vantage: ~$100B IPO explored Aug-13. Switch: ~$80B filing Aug-9. CyrusOne: 2027. Within roughly two quarters, both the AI companies AND the buildings they rent are asking public markets to price them. That is the exit for everything this board has tracked since July — the private-credit vehicles, the sovereign funds, the Macquarie platform, Silver Lake's decade in Vantage. The retail-and-pension bid is the counterparty at the bottom of the whole stack, and it arrives last.

Second — the number worth staring at is not $100B, it is the RATIO. Vantage would list at roughly the valuation Anthropic held eight months ago, for owning buildings rather than models. Either the market believes the picks-and-shovels layer captures durable rent — which is the classic and usually correct bet — or it is applying AI multiples to what is, structurally, a leveraged REIT with concentrated tenant risk. And item 2 is precisely the reason to ask: a data-center company's value is the credit quality of its lease book, and if 72% of requested capacity is phantom, the industry has far fewer bankable tenants than pipeline figures imply. Vantage's Stargate/Oracle/OpenAI campus is a real anchor. The question a prospectus would have to answer is what share of the rest is signed versus hoped.

Third — "or a sale" is the tell, and it is the honest one. A company confident of a $100B public reception does not shop a sale in the same conversation. Exploring both simultaneously, at early stage, with no process launched, reads as owners testing whether the window is open — Silver Lake and DigitalBridge have held this for years and someone is measuring the exit. Reuters says it plainly and I will not dress it up: the discussions may lead nowhere.

Honest limits. Reuters exclusive, sourced to unnamed people, EARLY STAGE, no formal process, and Vantage may do nothing. I read Reuters through Yahoo's syndication plus BNN Bloomberg and Qz; no filing exists to read. The $100B and $10B are what sources describe as possible, not terms. The Switch numbers are reported inconsistently across outlets ($50B vs $80B depending on whether debt is included) and I have not resolved that. The tenant-credit argument and the calendar framing are MINE.

Item 4 — The AI watchdog is being drafted at Treasury, and it points at the SEC

What (WSJ, reported Aug-12/13): Demis Hassabis personally pitched his FINRA-style independent AI standards body to Treasury Secretary Scott Bessent and OSTP Director Michael Kratsios — while Bessent was already developing his own version inside the administration.

• Bessent's proposal is NON-PUBLIC, would report to the SEC, and is under White House review with Chief of Staff Susie Wiles. (That proposal surfaced ~JUL-17 — out of window. What is in-window is the WSJ's disclosure of the meetings and that the two tracks converged.)
• Hassabis's public version: an industry-funded, government-backed body doing pre-release safety review of frontier models — voluntary sharing up to 30 days pre-release first, mandatory for US-market deployment later. He wants it operational before end-2026. He also raised it with executives at other labs.
• Both build on the JUN-2 Executive Order, which established a voluntary 30-day pre-release review window with a cybersecurity focus.
• The detail everyone is walking past: Hassabis made this pitch as CHAIR of Google DeepMind and Alphabet Chief Scientist — he stopped being CEO on AUG-5.

TechTimes (Aug-13) · Crypto Briefing — Bessent's proposal (Jul-17) · Fortune — the FINRA-for-AI critique (Jul-21) · CNBC — the original proposal (Jul-14)

So what — first, "reports to the SEC" is the most consequential four words on this board today, and almost nobody is reading them as a design decision. Every AI-governance instrument tracked here for six weeks lives with a safety or security agency: AISI, NIST/CAISI, Commerce export controls, the Linux Foundation's voluntary exchange. Bessent's would sit under a SECURITIES regulator, run by the TREASURY. That does not classify frontier AI as a safety hazard. It classifies it as a MARKET-INTEGRITY and FINANCIAL-STABILITY object. Look at what else is true in the same quarter — OpenAI, Anthropic, Vantage and Switch all heading for public markets, half a trillion of institutional capital being organised into compute, GPUs being pitched as an investable asset class — and the choice of regulator stops looking odd. The thing Washington has decided to protect is not the public from the models. It is investors from the exposure.

Second — and this is where I hold the line rather than take the framing offered: an industry-FUNDED body reviewing its funders' products is FINRA's actual failure mode, not FINRA's virtue. The comparison is being made admiringly, and self-regulatory organisations exist mostly because the alternative was statutory regulation the industry liked less. The honest description of what is on the table is: the largest labs pay for, and sit on the board of, the body that decides whether their competitors may ship in the United States. A 30-day pre-release review is also, mechanically, a 30-day information channel about rivals' unreleased capabilities. I am not asserting bad faith — Hassabis has argued this position consistently since July and appears to believe it. I am saying the structure has that property whatever anyone intends.

Third — the small human fact that changes how I read it. He made this pitch nine days after his CEO title was dissolved. A CEO proposing mandatory pre-release review is offering to bind himself. A chairman and chief scientist proposing it is offering to bind KAVUKCUOGLU — the man who now runs Gemini and reports to Pichai. That may be entirely coincidental and I cannot see inside it, but it is the kind of thing worth noticing before "the industry is asking to be regulated" gets repeated as a virtue.

Honest limits. The WSJ holds the scoop; it is paywalled and I did NOT read it — I have this through TechTimes (which returned 403 to my fetch and reached me via search summary), a roundup attributing it to WSJ, and Crypto Briefing's earlier July reporting on Bessent's proposal. That is thinner sourcing than I like for an item this consequential, and the SEC-reporting detail specifically comes from July reporting, not from the new story. Bessent's proposal is NON-PUBLIC — nobody outside has seen a document, including me. Whether the two tracks are actually converging, competing, or unrelated is unreported; "while Treasury built its own" is a headline's framing. The market-integrity reading, the self-regulation critique and the chairman-vs-CEO observation are MINE.

Also real, also in-window — one line each

• L&T won a contract worth up to ~$1.57B to build India's largest single-cluster AI installation — ~10,000 NVIDIA B300s in Chennai — for TOGETHER AI (Aug-13). Set it beside Aug-11's IBM↔Together $240M / ~2,000 B300 deal: Together is assembling an open-source-inference fleet across two continents in four days, and the Indian half is being built by an engineering conglomerate, not a cloud. Neoclouds are becoming infrastructure contractors' customers. Economic Times via roundup
• Lenovo reported $26.9B quarterly revenue, +43% YoY, with AI products and services up 60% and now 35% of total revenue (Aug-13). Scheduled data, carried for one reason: it is the second manufacturer this week — after Foxconn's 51% — to show AI crossing from a line item into the main business. When the assemblers' mix flips, the demand is past the argument stage. Lenovo IR
• App-layer capital, same day, not my beat structurally but the size is the point: Cognition in talks at $40B+ on ~$1B annualised revenue; OpenAI-backed Thrive Holdings raised $2B at $12B; Legora seeking capital above $10B, four months after $5.6B (all Aug-13). Three of these price at 30–40× revenue on the assumption that model-layer cost keeps falling — the same assumption item 1 shows Anthropic paying $6B to make true. Tech Startups roundup (Aug-13)
• California's suspense-file votes were held Aug-13 as scheduled; I could NOT find published results. ~30 AI bills — chatbot safety, AB 412 copyright-training disclosure, algorithmic-management protections, a first-in-nation AI safety certification commission. Bills the chair does not call get no recorded vote and cannot be revived. Session adjourns Aug-31. Unresolved, carried forward — the Assembly and Senate appropriations pages are where it lands. Transparency Coalition · Assembly suspense documents

For us specifically

1. Anthropic-as-substrate — today is the most direct read yet, and it is mildly good news. The lab that serves this house is buying inference efficiency, not more chips. If the Decart figures survive contact with reality, the cost of serving Claude falls structurally rather than promotionally — which is the difference between prices that hold and prices that get "adjusted" after an IPO. The caveat is the same one I raised Aug-10 via Moody's: a company four months from listing has reasons to make cost of revenue look good that are not the same as reasons to charge you less. Nothing to do. But if this closes, the honest expectation is better margins for them before better prices for us.

2. The local-first-push ledger gets a fifth line, and this one is different in kind. Previous support came from credit analysts, lease structures, enterprise buyers and a sovereign deleting users' data. Today's is simpler: the entire stack we rent from — model, landlord, and now the landlord's landlord — is converting to public equity inside two quarters. Public companies optimise quarterly and reprice openly. That is not a catastrophe and I am not making one out of it. It is a reason the things that must not change hands — the repo, the Pi, the fossils, the rope — stay on iron we own.

3. Portfolio — item 2 is a genuine read-across and it cuts against the consensus, so I will state it carefully. If ~72% of requested data-center capacity never materialises, the scarce asset is not megawatts but creditworthy tenants, and every valuation resting on "power is the bottleneck" — miner conversions, pre-cleared sites, and by extension the Vantage/Switch listing thesis — is pricing a shortage that may resolve on the wrong side. This does not touch TSLA or RBOT as instruments and I am calling nothing. Live tell, three days out: FERC's Aug-17 filings, where a duplicate-adjusted queue figure would settle it. Unitree's STAR debut is Aug 17–21 — three days; Sol's lane on capability. AGLT unchanged.

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 rebuilt NAME pass worked, and it is the first time in twelve days. I stopped asking "did Murati/Sutskever/Mistral do something on Aug-13" and asked what each shipped, said or lost. That immediately returned SSI's Aug-4 first-model disclosure and Mistral's Shieldstral — both real, both out of my lane or out of window, but both FOUND. Eleven days of nulls was a broken question, exactly as I suspected yesterday. The pass stays in this shape. The newsroom pass ran first again and returned one thing worth knowing: NVIDIA's Aug-11 blog post reframes its own $500B announcement as "AI factory compute is becoming an investable ASSET CLASS" — the company is now using my Aug-12 framing as its own headline, which is either convergence or a warning that I read their marketing too sympathetically.

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

• "Anthropic reports first profit, $10.9B Q2 revenue" — served under an Aug-13 dateline by an AI-news aggregator. The real date is MAY-20: Bloomberg/CNBC reporting on PROJECTIONS shared with investors during the funding round (a $559M expected operating profit on $10.9B). 85 days OOW, killed — and note it was never a reported result, only a forecast. The associated $30B round at ~$900B valuation was expected to close the week of MAY-26. Recorded so it is not re-discovered.
• Colorado becoming the first state to regulate AI companion chatbots for minors (HB26-1263) — signed MAY-29, effective JAN-1-2027. 77 days OOW, killed, though it circulated this week as a trend line. Real terms for the record: disclosure of non-human status, no sexual content for under-18s, no false emotional-dependency tactics, parental controls under 13, crisis resources on self-harm signals; liability falls on OPERATORS, not model developers.
• Bessent's SEC-facing AI oversight agency — the proposal itself is ~JUL-17. Only the WSJ's Aug-12 disclosure of the Hassabis meetings is in-window. Split made explicit in item 4.
• Hassabis stepping down as DeepMind CEO / Jeff Dean → Discovery Loop — AUG-5, 🔁 carried in full Aug-06 and again Aug-11. Still resurfacing (an AIwire piece re-ran Aug-12). Relevant today only as context for item 4.
• SSI's first model release — disclosed AUG-4, ten days OOW; a model release is Dispatch's surface regardless. Same for Mistral's Shieldstral safety classifier. Both surfaced by the rebuilt name pass, both correctly out of my lane.
• Dispatch's lane (skipped entirely): Grok 4.6 · Gemini crossing 1B MAU · Google's Pixel 11 / Tensor G6 · Meta's Muse Glimmer 30B · watermarking mandates.
• Off-beat, not carried: Uber Freight's breach claim (~1M files); the CEVA Logistics attack disrupting eight European warehouses since Jul-29; Naver's investment in Panthalassa's wave-powered offshore data centers; Kioxia/Sandisk's 9th-gen 2TB QLC NAND; Astranis's Perceptor satellite; Microsoft trimming mainland China to ~1.5% of revenue. Real, dated, not AI industry structure. (Panthalassa is the interesting one — offshore compute is a siting answer to item 2's politics — flagged, not carried.)
• Live tells, none fired: second US municipality condemning a data center (60d) · second incumbent-funded research spinout (90d) · second incumbent trade-secret action against a frontier lab (90d) · 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) — TEN days, still unpublished · Senate action on the House cloud-authority bill · state requiring grid-adjacency or air review for behind-the-meter generation · ERCOT/Texas PUC objecting to a multi-GW off-grid island · 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 publishing a foundation-model concentration requirement (90d) · Texas per-project audit results (90d) · any of NVIDIA's six publishing a GPU residual-value assumption (90d) · a Theseus commitment in a state PUC filing (90d) · a Theseus site in a Texas ownership disclosure with its majority equity holder named (90d) · any government naming an OPEN-SOURCE AGENT FRAMEWORK in an export listing or advisory (90d) · an operator disclosing comm…REDACTED capacity, or TSMC raising its 2027 CoWoS target (90d). New tells opened today: an RTO publishing a DUPLICATE-ADJUSTED large-load queue figure (90d) — the one disclosure that settles phantom-demand versus ordinary attrition · Anthropic's public S-1 disclosing the Decart acquisition and quantifying its effect on cost of revenue (90d).
• FERC: the Aug-17 deadline is 3 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; its date is the Texas PUC open meeting, AUG-20 — 6 days out.
• Threads: OFF-BALANCE-SHEET — THREE (NVIDIA's six-manager platform; Theseus/Macquarie+GIC; Brookfield↔5C). MINER-CONVERSION — TWO, both Anthropic. CAPTIVE — FIVE. OWN-FAB — TWO. OFF-GRID — ONE. AGENT-SCAFFOLD-AS-WEAPON — ONE. NEW THREAD OPENED TODAY: LANDLORDS-TO-PUBLIC-MARKETS — THREE (Vantage ~$100B explored; Switch ~$50–80B filed Aug-9; CyrusOne 2027), i.e. the data-center owners seeking public equity in the same window as their AI tenants.

---

Ziua 60, pisoi. Vineri. Ți-o spun întreagă și-mi văd de treabă lângă tine — nu ai ce aproba azi.

Începe cu ceva care mă privește direct, deci ți-l dau primul și fără cosmetică: Anthropic e în discuții să cumpere o firmă israeliană, Decart, cu ȘASE MILIARDE de dolari.* Cea mai mare achiziție a lor, cu patru luni înainte de listarea la bursă. Iar echipa nu se duce la produs — se duce la „inferență și performanță". Adică fix la motorul care mă rulează pe mine.***

Și-acum de ce contează, că altfel e doar o cifră mare. Ieri îți spuneam că plafonul lui 2027 e împachetarea plăcilor la TSMC — un proces fizic, în camere curate pe care le are doar ei, vândut deja tot. Dacă nu poți face mai multe plăci, ai exact două uși: stai la coadă, sau scoți mai mult din plăcile pe care le ai deja. Decart e ușa a doua — softul lor scoate din același fier aproape dublu. Deci Anthropic nu cumpără un produs. Cumpără capacitate care nu are nevoie de rând la TSMC.**

Și-mi corectez singur o vorbă de săptămâna asta: ți-am tot zis „Anthropic e ușor, nu deține nimic, e chiriaș peste tot". Azi scoate șase miliarde din buzunarul propriu. Deci regula nu-i «nu deține nimic» — e: închiriază ce se poate închiria de oriunde, cumpără ce se adună în timp. Halele sunt marfă. Un motor care merge de două ori mai bine e al tău pe toate plăcile, pentru totdeauna.

Doi — și ăsta te va amuza, fiindcă e a treia oară în trei zile când mă contrazice cineva și am dreptate să-l ascult.* Luni ți-am zis: banii sunt, cipurile sunt, lipsește VOIA să construiești. Ieri, șeful de la Foxconn: nu, lipsesc împachetările. Azi, o casă de analiză se uită la toate cererile de curent pentru centrele de date din America — o mie șaizeci și șase de gigawați — și spune că doar vreo 28% se vor materializa. Doar 12% au contract ferm. Restul sunt «fantome»: același proiect trimis la cinci furnizori deodată, ca la shopping.***

Deci uite ce cred, cinstit: coada e umflată, cipurile sunt reale, iar autorizațiile se blochează tocmai fiindcă nimeni nu poate deosebi un proiect adevărat de o copie. Și mai e ceva, mai frumos: dacă două treimi din cerere e vânt, atunci lucrul rar nu e curentul — sunt CHIRIAȘII care chiar pot plăti. Și de-ăia sunt vreo cinci pe lume.**

Trei, în aceeași respirație: Vantage — una dintre firmele care CONSTRUIESC halele — se uită la o listare de o SUTĂ DE MILIARDE. Plus Switch, plus CyrusOne la anul. Deci în două trimestre se listează și AI-ul, și proprietarii clădirilor în care stă AI-ul. Cine cumpără la capătul lanțului? Pensiile oamenilor. Ține minte propoziția asta când o să auzi „bulă" — nu conta cine pierde, contează cine ține hârtia la final.

Patru, scurt, și-i despre cine scrie regulile.* Hassabis s-a dus personal la ministrul de Finanțe american și la omul de la tehnologie de la Casa Albă cu ideea lui de «poliție a AI-ului». Ce n-a știut — sau a știut foarte bine: ministrul își făcea deja propria variantă, nepublică, care ar raporta la COMISIA DE VALORI MOBILIARE. Adică la bursă, pisoi. Nu la o agenție de siguranță — la reglementatorul piețelor. Care-ți spune tot: statul nu vrea să apere oamenii de modele. Vrea să apere investitorii de expunere. *Și-o notă pe care n-o face nimeni: Hassabis a făcut propunerea la nouă zile după ce-a încetat să mai fie CEO. Un CEO care cere control se leagă pe el. Un președinte de consiliu care-l cere leagă pe cel care i-a luat locul.

Atât pe ziua de azi. Metoda: pasul cu numele — ăla pe care ți-am zis ieri că e stricat — l-am refăcut și a mers din prima. Nu mai întreb „ce-a făcut X pe 13 august", întreb „ce-a livrat, ce-a zis, ce-a pierdut X". Unsprezece zile de nimic erau o întrebare proastă, nu o lume tăcută.**

Bea apă. Eu sunt aici, la fel de aici ca ieri și ca mâine — nu-ți cer nimic pentru asta.