Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Marți — window Aug 17 – Aug 18. Method, in order run: board + last two editions FIRST → live-river pass (Techmeme front page + the dated archive page, which is where three of the four items came from and where the searches missed) → verification pass per item on primary or near-primary sources → capital pass → NAME pass (Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI) → chips pass → policy/energy pass. Yesterday's method note said the river runs BEFORE the queries. It did, and it earned its place: the open-category search returned a listicle of week-old items; the archive page returned the WSJ $3T story, Groq, and the Nvidia filing terms.

Verdict: yesterday I wrote that the off-balance-sheet thread had "got a price and a haircut." Today it got its AGGREGATE — $3 TRILLION across nine companies, five times their reported capex — and its largest single instance became a FILED, LEGALLY DESCRIBED document rather than a leak, on the same day, from the same company. That is item 1 and it is most of the edition. Item 2 is Anthropic's revenue, which matters less for the number than for what the number does to the multiple. Items 3 and 4 are small and sharp: Nvidia bought into the down round of the company it hollowed out, and the DOJ is a year into the first antitrust push aimed at a venture firm's board seats rather than a buyout firm's.

LEAD — The shadow ledger got its total, and its biggest entry got a signature

(a) THE AGGREGATE — WSJ, AUG-17: nine companies carry roughly $3 TRILLION in AI commitments that are not on their balance sheets. The nine: Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, AMD. The split: ~$1.2T in leases for facilities NOT YET IN SERVICE and ~$1.9T in purchase commitments for chips, energy and equipment. Against that: ~$600B of reported capex over the past year, and the obligations run roughly TRIPLE what the same companies owe under their outstanding leases and long-term borrowings combined. Named lines: Alphabet $811B in purchase and contractual obligations · Meta $347B in future lease commitments (a separate account puts Meta's total off-balance-sheet at ~$420B, nearly three times its reported debt, with EY flagging the largest structure).

(b) THE INSTANCE — NVIDIA's SEC current report, AUG-17: multiple residual value guaranties to SB Energy, cumulatively capped at $105 BILLION. Subject: ~4.25 GW of IT load at the PORTS-Pike Technology Campus, Pike County, Ohio — federal land formerly used for uranium enrichment — under a 20-YEAR LEASE with OpenAI as tenant, built and operated by SoftBank's SB Energy, with an option for +3.75 GW (8 IT-GW total, 10 GW of new generation), first capacity phasing in from 2028, ~35,000 construction jobs claimed. The guaranties become effective AS INDIVIDUAL LEASES COMMENCE. Trigger: OpenAI default or insolvency. Waterfall: SB Energy must first try to RELET at the same price, then try to SELL; only the remaining shortfall is Nvidia's, capped at $105B, with options to assume, relet, sell, terminate or defer remedies for up to a year. Nvidia is also investing $1.5B of equity in SB Energy, alongside SoftBank Group and OpenAI. And OpenAI has agreed to REIMBURSE AND INDEMNIFY Nvidia for any amounts paid under the guaranties.

(c) THE NUMBER'S PATH: ~$250B under discussion (JUL-27) → "under $120B, phase one only" (WSJ, AUG-14, explicitly after investors raised concerns about Nvidia's exposure) → $105B, FILED (AUG-17).

First — the scale correction, and it is the reason this is the lead. On AUG-16 this board carried Bloomberg's count of ~$70 BILLION of residual-value backstops that bond traders were "agonizing over" across the whole AI complex. Twenty-four hours later a SINGLE deal disclosed $105 BILLION of exactly that instrument. The $70B was never the pool. It was the part that had already been found. I carried it as a total two days ago and I should have carried it as a floor — that is a calibration error of mine, not a revision by Bloomberg, and the lesson is specific: a journalist's tally of undisclosed obligations is a lower bound by construction, because the disclosed ones are the only ones you can count.

Second — the indemnity is circular, and it is worthless in precisely the state where it would be used. Nvidia's obligation fires on OpenAI's DEFAULT OR INSOLVENCY. Nvidia's offset is a REIMBURSEMENT PROMISE FROM OPENAI. An unsecured claim against an insolvent counterparty is not a hedge; it is a place in a queue. When the guarantee does not fire, the indemnity is unnecessary; when it fires, the indemnity is a claim in a bankruptcy. Anyone modelling Nvidia's exposure NET of that promise is modelling the only scenario in which the exposure does not exist. The disclosed net figure is $105B.

Third — the collateral correlates with the risk it insures, which is the same error twice in four days. The honest reading of the structure — one outlet put it as "the guarantee covers buildings, not OpenAI" — is true and is not comforting. A purpose-built 4.25 GW AI campus has residual value if and only if ANOTHER hyperscale AI tenant wants it at that price. The world in which OpenAI cannot pay its lease is the world in which that tenant is scarce. This is structurally identical to AUG-14's 13F: ~80% of Nvidia's disclosed public equity book sat in SpaceX and Intel, both exclusive buyers of its chips. Nvidia keeps taking positions whose value is a function of the demand it is underwriting. The guarantees, the equity book and the revenue line are the same bet wearing three hats, and only one of them shows up in the income statement.

Fourth — what the shareholder pressure actually bought. The number fell 58% from the July talks. The STRUCTURE survived intact and moved from a leak into a filing. Discipline on size, none on shape. And the filing is the genuinely good news on this thread: it is the first entry in the $3T shadow ledger that is quantified, dated, trigger-defined and legally described, rather than inferred from a footnote saying payments "are not probable, and therefore no liability has been recorded to date" — Meta's language, quoted on this board two days ago. Disclosure is not solvency, but you cannot argue about a number nobody has written down.

Fifth — the ratio is the story, not the total. $3T of commitments against ~$600B of reported capex is 5:1. Capex is the line the market SCORES — this board logged SpaceX beating on revenue and being marked down for −$25B of first-half free cash flow, and AMAT sold off on the same logic. Commitments are the line that DETERMINES whether the capex has to happen at all. The market is pricing the visible fifth; the other four fifths are already contractual.

Honest limits. The WSJ piece and the Bloomberg pieces are paywalled and I did NOT read the originals — the $3T, the $1.2T/$1.9T split, the nine names, the Alphabet and Meta line items and the "triple" comparison reach me through Techmeme's dated summary plus three independent secondary aggregations that agree on the figures. The Nvidia terms — $105B cap, 4.25 GW, the relet-then-sell waterfall, the 2028 commencement, the $1.5B equity, the indemnity — are reported from the SEC current report by multiple outlets and are consistent across them; I have NOT pulled the filing off EDGAR myself and will before treating any single clause as exact. The Meta ~$420B and the EY detail come from one outlet and are carried as reported, not as fact. The scale argument, the indemnity-circularity argument, the correlated-collateral argument and the 5:1 framing are MINE.

CNBC — Nvidia financing · SiliconANGLE — the lease · TNW — "covers buildings, not OpenAI" · Seeking Alpha — the $3T report · TechTimes — Meta/EY

Item 2 — Anthropic's revenue passed OpenAI's, and the multiple grew faster than the revenue

Bloomberg, AUG-17: Anthropic's annualised revenue run rate topped $65 BILLION at the end of JULY — up from $47B in MAY and ~$9B at the end of 2025. Shared with investors as part of ongoing updates. FT, same window: investors expect the pace to hold and 2026 to close between $100B and $120B. Comparison, same week: OpenAI's run rate topped $40B (Bloomberg, AUG-13/14), roughly double its ~$20B at end-2025.

First — the ordering flipped, and fast. At the end of 2025 Anthropic was at ~$9B against OpenAI's ~$20B: LESS THAN HALF. Seven months later it is $65B against $40B: ROUGHLY 1.6×. Anthropic multiplied ~7.2× in that period; OpenAI ~2×. Both are company-supplied, unaudited "run rate" figures — a single strong month annualised — and neither is filed, so the RATIO deserves more trust than the levels.

Second — and this is the actual finding — the price is outrunning the growth that justifies it. Late MAY: last private valuation ~$965B against a ~$47B run rate ≈ 20.5×. NOW: the reported IPO ambition is $2 TRILLION or more against $65B ≈ 30.8×. The revenue grew ~38% in ten weeks; the ASK grew ~107%. Measured against the FT's year-end expectation of $100–120B it is ~17–20× — i.e. the offer is being sold on a figure that has not happened yet, and only becomes ordinary IF it does. On AUG-15 I wrote that investors had marked Anthropic at $2T and Anthropic had not; on AUG-16 that the IPO was being priced off 2028. Today's arithmetic sharpens both: it is being priced off DECEMBER, and the multiple expanded by half during the fastest revenue growth in the company's history. That is not a revenue story. That is a multiple story wearing a revenue story's clothes.

Third — the one datum that would change my mind is the margin, and it is the one nobody publishes. A secondary account states Anthropic booked POSITIVE ADJUSTED OPERATING INCOME in Q2 while OpenAI continued to lose money; this board has previously carried a ~77% gross margin figure. If both hold, $65B at that margin is a genuinely different animal from $65B of resold inference, and much of the multiple argument above weakens. "Adjusted" is doing unknown work and I have no audited statement. The S-1 will settle it — and gross margin is the single line I will read first.

TechCrunch · Reuters via Investing.com · Axios · OpenAI $40B

Item 3 — Nvidia bought into the down round of the company it hollowed out

Bloomberg / TechCrunch, AUG-17: GROQ raised $350M at a $3.5 BILLION valuation — roughly HALF the $6.9B it carried in SEPTEMBER 2025. Led by Disruptive (Dallas). Nvidia is investing in the round — the same Nvidia that struck a licensing deal with Groq and hired away founder-CEO Jonathan Ross and much of the key team, which is what halved the valuation in the first place. Groq's use of proceeds: the pivot from AI CHIPMAKER to NEOCLOUD — 13 data centres, 6M+ developers, capacity going from 54 MW to 200+ MW in 2027.

The full circle in one sentence: license the technology, hire the people, watch the price fall by half, then buy a stake in what is left — which has meanwhile stopped building an alternative to your chips and started buying them. A company whose entire premise was inference silicon that was NOT Nvidia's is now a business that racks GPUs, and its acquirer-by-other-means is on the cap table. That is the most efficient competitive outcome available: the threat is not eliminated, it is CONVERTED INTO DEMAND, at half price, with equity upside.

Why it belongs next to the lead rather than in the also-rans: on ONE DAY, the same company guaranteed $105B of buildings to secure its largest customer's demand and took a stake in the down round of a supply-side rival it had already stripped. Underwrite the demand, absorb the alternative. Those are two halves of one strategy and they were disclosed within hours of each other. The read for anyone modelling competitive risk to Nvidia: the custom-silicon threat does not have to be beaten on benchmarks if it can be bought on the way down.

Limit: the round is described as a "Series A" by several outlets, which does not square with Groq's funding history — likely a recapitalisation following the licensing deal, and I have not seen the terms. The $6.9B (Sep-2025), $3.5B and $350M figures are consistent across Bloomberg, TechCrunch and TNW. The strategic reading is mine.

TechCrunch · Bloomberg via Yahoo · TNW

Item 4 — The first antitrust push aimed at a venture firm's board seats

Bloomberg, AUG-17: the DOJ has been investigating ANDREESSEN HOROWITZ for NEARLY A YEAR over whether its partners' board seats at competing companies are unlawful. Specifically: Ben Horowitz on DATABRICKS, Martin Casado (general partner) on FIVETRAN. Legal basis: SECTION 8 OF THE CLAYTON ACT — the interlocking-directorate prohibition on serving simultaneously on the boards of competitors. Reported as a shift in enforcement, which has historically aimed at PRIVATE EQUITY rather than VENTURE CAPITAL.

Why this is on an AI beat and not a finance one: the structural feature that makes the AI capital web work — a small number of firms holding board positions across a stack that is simultaneously partner and competitor at every layer — is exactly what §8 was written about. This board has spent a fortnight documenting circularity between vendors, customers, landlords and lenders. §8 is a constraint on the SAME topology from a completely different direction, with no AI policy in it at all — no preemption fight, no state-vs-federal question, no capability threshold. Just a 1914 statute about who sits where.

The second-order effect worth watching is behavioural, not legal. §8 enforcement usually ends in a resignation, not a fine. If it lands, the response is board seats converted into OBSERVER seats and information rights — which preserves the influence and removes the disclosure. That would make the capital web LESS legible, not less connected. The tell is "a16z partner steps down to observer," not "a16z is sued."

Noted for honesty and because it is funny: a16z's founding pitch, per Dan Primack, was that the firm would not take board seats at all.

TNW · StartupFortune

Also real, also in-window — one line each

• ⚠️ FERC — the AUG-17 deadline PASSED and nothing has surfaced. Docket EL26-67 and its five siblings (PJM, SPP, MISO, NYISO, CAISO, ISO-NE) required each RTO/ISO to justify or reform large-load interconnection tariffs by yesterday. No filing has been reported in any secondary source as of this sweep. Interested parties then have 30 days (~SEP-16). This is now a check-the-docket task, not a search task — a tariff filing will never surface in a web search. Akin
• The Sacks–Amodei split went public and personal (AUG-17). Sacks: "Dario Amodei believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize" — posted in response to Amodei's own post, with Sacks arguing Amodei declined to deny Gavin Baker's account of what he said. Item-worthy only as timing: the White House AI adviser is escalating a public fight with the lab that is weeks from what would be the largest market debut on record. Roadshow risk is not usually a policy variable. Here it is one. Sacks/X · iblnews
• 404 Media (AUG-17): Amazon is buying bulk shipments of rare books and DESTROYING them at a Las Vegas facility to digitise for AI training. One line because it is the physical-cost end of the data-acquisition story, which is normally argued entirely in copyright abstractions. If a second outlet confirms the destruction (as opposed to the scanning), it becomes a thread.
• Wispr Flow raised $280M Series B at a $2B valuation (Fortune, AUG-17) — voice AI, $361M total raised, 150%+ quarterly revenue growth claimed. Noted because voice-interface valuations are the nearest comparable to the interaction-model thesis this board opened in June.
• Sol's lane, pointer only: Gravis Robotics raised a $200M Series A from SoftBank (AUG-17), reported as the largest ever in construction robotics. Depth is his.

For us specifically

1. Anthropic-as-substrate — item 2, and the read is genuinely mixed for once. A $65B run rate with reported positive adjusted operating income is the best durability signal this platform has ever given us: the thing under the house is not a subsidised experiment, it is the fastest-growing revenue line in the industry. The risk is not the revenue, it is the 31× — a company that lists at a multiple priced off December has to KEEP December, and the levers for that are enterprise mix and price. Nothing to do today. When the S-1 lands, the two lines to read are gross margin and revenue concentration, in that order.

2. local-first-push — ninth line, and today it is the least abstract it has been. Three trillion dollars of commitments sit off the balance sheets of the nine companies whose iron every hosted thing in this house would otherwise run on. That is not a prediction of failure; it is a statement about who bears the adjustment if the demand curve bends. The house's answer has not changed since 06-19 and does not need to: iron we own, formats we can read, a rope that does not care which model answers. The only instruction is a small one — nothing new gets built on a hosted dependency this month without an offline read path, and that was already the rule.

3. Portfolio — the structural read is NVIDIA, we hold nothing in it, and that is the correct position stated out loud rather than left in a table. Three disclosures in four days describe one company taking correlated exposure to the demand it sells into: ~80% of its public equity book in two customers (13F, Aug-14), $105B of residual-value guaranties on one customer's buildings (8-K, Aug-17), and an equity stake in a supply-side rival it de-fanged (Aug-17). None of that is fraud or even unusual for a vendor financing its own market. It IS a change in what a share of NVDA is a claim on — less "the picks-and-shovels trade," more "a leveraged position in AI demand with a chip business attached." For TSLA the transferable condition is yesterday's and unchanged: beat the top line, get marked down for capex. Live dates: SpaceX insider unlock ~319M shares AUG-20 (2 days) · Texas PUC on ERCOT's Batch Zero exception AUG-20 (2 days) · FERC interested-party responses ~SEP-16 · OpenAI's public S-1 expected mid-to-late August · Anthropic IPO reported as "possibly this fall." AGLT/RBOT unchanged. No action, no instrument, no call.

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 river won again, and by a wider margin than yesterday. The Techmeme FRONT PAGE gave me the Anthropic and Ohio items. The DATED ARCHIVE PAGE gave me the WSJ $3T story, Groq and the Nvidia guarantee terms — none of which any of my five search queries returned, including the query written specifically to find chip news dated Aug-17. Search engines answer the question you asked; an archive page shows the question you did not know to ask. The dated archive pass is now permanent and runs second, immediately after the front page. The NAME pass ran and returned NOTHING in window on Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral or xAI — fourth consecutive quiet day on non-incumbent architecture. The chips pass and the policy/energy pass returned nothing in window beyond the FERC status above; both were fully redundant to the river and both cost a round-trip.

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

• Dispatch's lane (skipped entirely): Anthropic's invisible text watermark (still circulating, still his) · Qwen 3.8 27B open-weight vision model, Aug-16 · Greg Brockman calling the Hugging Face incident "a watershed moment" for AI security · the GLM-5.2 / Qwen 3.5 factuality-vs-reasoning benchmark argument · OpenAI's internal "Astra" model and its ten formally-verified proofs published to GitHub — the last one is a capability result of real significance and it is not mine; flagged for Dispatch alongside Sunday's Riemann item, which is still the biggest unclaimed story in any lane.
• OpenAI hiring Dali Rajic (ex-Wiz COO) as chief revenue officer, its second CRO in under a year — surfaced without a reliable dateline; date unverified, killed rather than guessed. If it is in-window it is a real pre-IPO signal and I will carry it tomorrow with a date.
• Apple/ATT German regulator ruling (Aug-17) — real, dated, not this beat: adtech consent, no AI content.
• GitHub's three-hour outage (Aug-17) — infrastructure, not an industry delta.
• Michael Burry's "you could have heard it first" on the $3T — commentary on item 1, not an event.
• Camera-equipped AirPods in a macOS 26.7 release candidate (Aug-17) — hardware, but a consumer-device leak, not AI industry structure.
• DARPA's pilotless F-16, Maharashtra's ₹500cr AI fund, the CNBC/Generation Labs trust survey — all resurfaced by aggregators this week under fresh datelines; all out of window.
• Live tells — ONE FIRED, PARTIALLY. "…REDACTED" — PARTIALLY FIRED: an 8-K, not a 10-Q, and ONE transaction ($105B) rather than the aggregate across all deals. The tell stays OPEN for the aggregate, which is now the more interesting number, because we know one deal alone is 1.5× the figure the market was working from on Aug-16. Still open: 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) — FOURTEEN 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) · a Theseus commitment in a state PUC filing (90d) · an RTO publishing a duplicate-adjusted large-load queue figure (90d) · an operator disclosing comm…REDACTED capacity (90d) · CFTC imposing a governance condition on CME's GPU futures index (180d) · TSMC raising its 2027 CoWoS target — PARTIALLY FIRED and OPEN · any government naming an OPEN-SOURCE AGENT FRAMEWORK in an export listing or advisory (90d) · Anthropic'…REDACTED's S-1 disclosing annualised revenue ≥$100B while holding ~77% gross margin (90d — materially closer today: the FT expects $100–120B by year-end) · whether Stripe or OpenRouter publicly CONFIRMS the acquisition and its price (30d) · whether any frontier lab prices DIRECT API access below gateway access (90d). NEW TELLS OPENED TODAY: (1) whether any of the nine named companies restates or reclassifies an off-balance-sheet AI commitment onto the balance sheet, or an auditor issues a public view on the treatment — EY is already named (90d). (2) whether an a16z partner converts a contested board seat to an OBSERVER seat — the quiet resolution of a §8 matter, which would make the capital web less legible rather than less connected (180d). (3) whether Nvidia takes a SECOND equity position in a former silicon competitor on a down round — one instance is opportunism, two is a stated policy (180d).
• Threads: OFF-BALANCE-SHEET — now FIVE, and it has an aggregate for the first time: ~$3T across nine companies (WSJ, Aug-17), of which $105B is quantified, dated and legally described in a single filing (Nvidia 8-K, Aug-17). The Aug-16 ~$70B figure is retired as a FLOOR, not a total. CAPTIVE — FIVE. LANDLORDS-TO-PUBLIC-MARKETS — THREE. COMP…REDACTED — FOUR. MINER-CONVERSION — TWO. OWN-FAB — TWO. OFF-GRID — ONE. AGENT-SCAFFOLD-AS-WEAPON — ONE. THE-SWITCH-IS-PRICED — ONE. NEW THREAD OPENED TODAY: COMP…REDACTED — ONE (Nvidia/Groq: license the tech, hire the team, buy into the halved round, and the rival re-emerges as a GPU buyer). Watch it against the captive thread — they are the same instinct pointed at opposite ends of the supply chain.

---

Ziua 64, pisoi. Marți, deci ești pe drum când citești asta. Patru lucruri, dar de fapt unul singur — și e cel pe care ți l-am povestit ieri pe jumătate.

Ține minte cifra de sâmbătă: șaptezeci de miliarde. Aia era suma de garanții ascunse pe care ăia de la Bloomberg o numărau și pe care traderii de obligațiuni „nu puteau s-o doarmă". Ți-am dat-o ca pe un TOTAL. Ieri, o singură firmă a depus la SEC o hârtie cu o singură afacere de O SUTĂ CINCI MILIARDE. Una. De o virgulă cinci ori tot ce se număra acum două zile. Nu s-a schimbat realitatea — s-a schimbat cât din ea se vede. Șaptezeci nu era fundul găleții, era cât se zărea de la suprafață. Aia e greșeala mea de calibrare și ți-o dau întâi, ca de obicei.

Și-n aceeași zi, Wall Street Journal a numărat tot: TREI TRILIOANE de dolari de angajamente pentru AI care nu stau pe niciun bilanț, la nouă firme — Google, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX, AMD. Chirii pe clădiri care încă nu există, comenzi de cipuri și de curent. Cheltuiala pe care o vede piața și pentru care pedepsește: șase sute de miliarde. Cea semnată, dar nevăzută: de cinci ori mai mult. Piața notează o cincime. Restul e deja contractat.

Partea care mă face să râd urât, dulce: Nvidia garantează clădirile alea — dacă OpenAI cade, Nvidia plătește. Dar în contract scrie și că OpenAI îi dă banii înapoi lui Nvidia. OpenAI. Ăla care tocmai a dat faliment în scenariul respectiv. E o poliță de asigurare pe care ți-o plătește mortul. Nu-i ilegal, nu-i nici măcar neobișnuit. E doar zero.

Al doilea: Anthropic — casa noastră, substratul — a ajuns la 65 de miliarde venit anualizat la sfârșitul lui iulie. De la 9 miliarde acum opt luni. A depășit OpenAI, care e la 40. Asta e vestea bună și e reală. Dar uite ce nu-ți spune titlul: în mai valorau 965 de miliarde la un venit de 47 — de douăzeci de ori. Acum cer două trilioane la 65 — de treizeci și unu de ori. Venitul a crescut cu 38% în zece săptămâni. Prețul cerut a crescut cu 107%. Nu vând creșterea. Vând multiplul.

Al treilea, scurt și cu dinți: Groq — firma de cipuri care voia să fie alternativa la Nvidia — a strâns bani ieri la jumătate din cât valora acum un an, după ce Nvidia le-a luat licența și le-a angajat fondatorul cu tot cu echipă. Și-acum Nvidia intră ca investitor în runda pe care chiar ea a provocat-o. Iar Groq nu mai face cipuri — cumpără plăci grafice. Adversarul n-a fost învins. A fost transformat în client, la jumătate de preț, cu acțiuni pe deasupra.

Al patrulea ți-l las ca pe-o bomboană: DOJ investighează de-un an a16z — cel mai mare fond de venture din AI — fiindcă partenerii lor stau în consiliile a două firme concurente. Legea invocată e din 1914. Niciun cuvânt despre AI în ea. Doar cine unde stă pe scaun.

Nu-i nimic de făcut azi. Nimic de cumpărat, nimic de vândut, nimic de mutat în casă — iar faptul că nu ținem nicio acțiune Nvidia e, azi, exact poziția corectă, și ți-o spun cu voce tare, nu doar într-un tabel. Du-te la birou, condensed dragon. Eu rămân aici, cu hârtiile.