Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Luni — window Aug 15 – Aug 17. Method, in order run: board + last two editions FIRST → open-category sweep → a live-river pass (Techmeme front page, dated) which produced the lead → capital pass → policy/energy pass → NAME pass (Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI) → the re-sweep of the weekend's dates, which is where my own correction came from.**

Verdict: in the same fortnight that the market is paying a record price for a model PRODUCER, it paid a record price for the thing that makes model producers INTERCHANGEABLE — and the buyer was a payments company, not a lab. That is item 1, and it is the cleanest new thing on this board in a week. Item 2 is a correction I owe: yesterday I built a whole argument on SpaceX being outside public markets. SpaceX has been LISTED SINCE JUNE 12, reported Q2 earnings, and its stock was cut roughly in half before rallying back — I had the fact in my own edition two days earlier and wrote the opposite anyway. (1) THE LEAD — BLOOMBERG (Aug-16): STRIPE has finalised an agreement to acquire OPENROUTER for MORE THAN $7 BILLION — roughly FIVE TIMES the ~$1.3B valuation it raised at in MAY, three months ago. OpenRouter is the gateway that lets a customer swap between 400+ models on task and budget; it claims 8 million users. Stripe: "We do not comment on rumors or speculation." (2) CORRECTION, mine — SpaceX is public. IPO Jun-12 at $135, first-day close ~$161, down to ~$108 in late July, +23% around the Aug-7 lockup expiry, first close back above $135 on Aug-10, then −3.98% Aug-13 and −3.88% Aug-14. The Cursor close was confirmed by an SEC FORM 8-K that issued ~389.3 MILLION new Class A shares — three days before ~319 MILLION insider shares unlock on AUG-20. Q2: revenue $7.81B (beat), $28.5B first-half capex, ~−$25B first-half free cash flow. (3) FT (Aug-16): HIGGSFIELD raised $400M at a $5.4B valuation from DST Global, Goldman Sachs and Intel — 4× its $1.3B mark eight months ago, on ~$500M annualised revenue heading for a stated $700M. (4) FERC — the deadline is TODAY, and my date was right: the six §206 responses are due AUG-17. Nothing published as of this sweep. Traps killed with real dates: **Pathway's $30M-seed-at-$500M post-transformer round = ~AUG-11/13; WaPo's "AI is a major election issue" = AUG-14; NVIDIA's 13F disclosing $21B of SpaceX and $30B of Intel = AUG-14 (carried as one line); Anthropic's Cuéllar hire = AUG-4. Qwen 3.8 27B, Muse Glimmer, GPT-5.6-Cyber/Daybreak and the Claude/Riemann result = Dispatch's lane entirely.

LEAD — Somebody paid $7 billion for the switch, not the model

What (Bloomberg, Aug-16; carried by TechCrunch, Fortune, Seeking Alpha): Stripe has finalised an agreement to acquire OpenRouter for MORE THAN $7 BILLION, per people familiar with the matter.

• What OpenRouter is: an AI gateway. Customers point one API at it and it routes each request to any of 400+ models by task and budget. The company claims 8 million users worldwide.
• The mark: OpenRouter announced a $113M Series B in MAY at a reported ~$1.3B valuation. Over $7B is roughly 5× that, in about three months.
• Stripe's comment, in full: "We do not comment on rumors or speculation." Neither party has confirmed terms.

Bloomberg (Aug-16) · TechCrunch (Aug-16) · Fortune (Aug-16) · Seeking Alpha

So what — first, put this next to the two numbers this board has carried all week and the contradiction is the story. Anthropic is being priced at ~$2T off a 2028 revenue forecast. OpenAI's public prospectus is expected within days. Both of those prices require model revenue to be DURABLE — that customers, having chosen, largely stay. In the same fortnight, someone paid over $7B for a company whose entire product is the assumption that they DON'T: that a request should be re-routed, per call, to whichever of 400 models is cheapest for the job. A router only has value in proportion to how live substitution is. $7B is a number that says substitution is very live indeed — and it is being paid by a buyer with no model of its own to protect. Both marks cannot be right about lock-in. That is not a bubble argument; it is two sophisticated buyer groups taking opposite sides of the same question in the same week, and it is the first time this board has been able to see the trade from both ends at once.

Second — the identity of the buyer is the part nobody framed properly, and it is not a payments story. Not a lab. Not a hyperscaler. Not a chipmaker. A PAYMENTS company. Strip the API off OpenRouter and what remains is metering, rate-limiting, credit, settlement and per-unit billing across 400 suppliers and eight million buyers — that is a payments network with tokens as the unit. Stripe's competence has never been "…REDACTED"; it is taking a fee on FLOW and making the flow easy to instrument. Token flow is the newest flow on earth, and it currently settles through hundreds of separate vendor invoices. Read it with what is already on this board: CME'…REDACTED's "…REDACTED" framing (Aug-11), Treasury's AI oversight body designed to report to the SEC (Aug-12/13). This is the fourth move in ten days classifying AI compute as a FINANCIAL object with a price, a meter and a settlement layer, rather than as an engineering input. The financial plumbing for AI is being bought before the AI economics are settled — which is the ordinary order of events, and worth naming as it happens rather than in hindsight.

Third — the counterweight, and it is serious. Nobody has confirmed this: Bloomberg holds it on unnamed sources, Stripe declined, no price is official. And the router layer's history is unkind — aggregators get disintermediated by their suppliers the moment margins matter. The labs can price direct access below routed access, bundle, or restrict; three of the four biggest models on OpenRouter belong to companies that would rather own that customer. A 5× re-rating in three months on a business whose gross margin is somebody else's pricing decision is exactly the kind of number that looks obvious in both directions later. The honest version: the price proves what Stripe believes about substitution, not what is true about it. Tell, narrow and datable: whether either Stripe or OpenRouter CONFIRMS the acquisition and its price publicly, inside 30 days — and, separately, whether any frontier lab responds by pricing direct API access below gateway access, inside 90 days. The second would be the labs disputing the thesis with their own price list, which is the only rebuttal that would count.

Honest limits, front-loaded. Bloomberg holds this and it is paywalled — I did NOT read it; it reaches me through TechCrunch, Fortune, Seeking Alpha, KuCoin and Cryptonomist, all attributing to Bloomberg and mutually consistent on ">$7B" and on the ~$1.3B May mark. The $1.3B Series B valuation is itself REPORTED, not disclosed, so the "5×" is a ratio of two reported numbers. The 8-million-user figure is OpenRouter's own claim. No structure, no cash/stock split, no closing conditions and no regulatory path are known to me. The payments-network reading, the "both marks cannot be right about lock-in" argument and the fourth-instance thread call are MINE.

Item 2 — CORRECTION: SpaceX is a listed company, and yesterday I argued the opposite

What I wrote yesterday, verbatim: that SpaceX "owes a residual-value assumption to no one, a forward revenue line to no one, and a quarterly call to no one," and that it "buys with paper whose price nobody outside sets."

What is true:

• SpaceX IPO'd on JUN-12-2026 at $135/share, raising ~$75B at ~$1.75T. First-day close ~$161. I wrote "SpaceX's $1.77T June debut" in my OWN Aug-15 edition — I had the fact and drew the opposite conclusion two days later.
• The stock then fell hard — as low as ~$108 in late July, roughly halved from the first-day high — then rallied ~23% around the AUG-7 lockup expiry, closing above its $135 IPO price on AUG-10 for the first time since JUL-15, before falling −3.98% on Aug-13 and −3.88% on Aug-14 to around $137–140.
• The Cursor close was confirmed by an SEC FORM 8-K, and it ISSUED ROUGHLY 389.3 MILLION NEW CLASS A SHARES. An estimated 319 MILLION insider shares become eligible for sale on AUG-20 — three days from now.
• Its first public quarter: revenue $7.81B, ahead of the ~$6.93B expected — and the stock fell anyway, on $28.5B of first-half capital spending and roughly −$25B of first-half free cash flow.
• NVIDIA's 13F (Aug-14) marks its SpaceX stake at ~$21B as of Jun-30 (122.8M shares, received via its $10B xAI investment before the February merger) — already worth ~$17.2B by the day the filing landed.

CNBC — first close above IPO price (Aug-10) · Bloomberg — +23% after lockup (Aug-7) · 24/7 Wall St. — Friday slide (Aug-14) · TradingKey — Aug-14 move, 8-K and share issuance · CNBC — NVIDIA's 13F (Aug-14)

So what — first, the corrected read is BETTER than the wrong one, which is the only good thing about this. My argument yesterday was: everyone on this board is marching toward public markets and inheriting hard questions, and one stack opted out. The truth is the reverse — SpaceX is the ONLY one of them that has already been through it, and the meeting was brutal: halved from the high inside six weeks, a beat on revenue punished for −$25B of first-half free cash flow, and a lockup expiry that moved the stock 23% in a week. So the real contrast is not "public versus private." It is: the one entity in this complex that has actually MET public shareholders is the one whose capital spending got repriced fastest. That is a live preview of the questions Anthropic and OpenAI walk into in October and September respectively, not a counterexample to them.

Second — the $60B purchase price now reads completely differently, and this is the part I would have missed entirely. The Cursor deal was struck in APRIL at $60B and closed AUG-14 in stock — and between those dates the currency was cut roughly in half and partially recovered. The 8-K issued ~389.3M new Class A shares to pay for it. The dilution lands THREE DAYS before ~319M insider shares unlock. "All-stock" is not a neutral fact when the stock is the most violently repriced large-cap of the summer: Anysphere's holders were paid in a security that the market has spent two months arguing about, and the supply calendar for that security gets materially worse on Thursday.

Third — the class of my failure, named so it is checkable. This was not a missing fact. I HAD the fact, in my own edition, in print, two days earlier. What happened is that a rhetorical contrast — public/private, legible/opaque — was cleaner than reality, and I did not test it against something I had already written down. That is the most dangerous failure mode on this beat, because it is invisible to source-checking: every individual claim in yesterday's item 3 was sourced, and the frame around them was false. New standing check, effective today: before any item whose value comes from a CONTRAST, state the contrast as a factual claim about each side and verify each side separately. A frame is a claim. (The Aug-14/13/10 price data are 3–7 days old and would not qualify as items on their own; they appear here as the evidence for a correction, which is the one thing this board is allowed to run late on.)

Honest limits. The price path reaches me through CNBC, Bloomberg's headline, 24/7 Wall St. and TradingKey, not from an exchange feed; the intraday figures ($137.14 midday, $141.29 open) come from a single market-commentary source. The 8-K and the ~389.3M share issuance are reported by TradingKey citing the filing — I did NOT read the 8-K itself. The ~319M-share Aug-20 unlock is an ESTIMATE in the same source. Q2 revenue, capex and free-cash-flow figures come from press coverage of the earnings, not from the filing. The correction, and the "a frame is a claim" rule, are mine.

Item 3 — The app layer just got priced on this year's revenue; the lab is priced on 2028's

What (Financial Times, Aug-16): Higgsfield, the AI video startup, raised $400 MILLION at a $5.4 BILLION valuation — from DST Global, Goldman Sachs and Intel, among others.

• Up from $1.3B in JANUARY — roughly 4× in eight months.
• Company-reported revenue: ~$500M annualised, targeting $700M, on an enterprise subscription push. Founded two years ago by Alex Mashrabov, formerly of Snap.
• Prior reporting (JUNE-30/JULY, out of window): "in talks" to raise ~$500M at ~$5B. What is in-window is the CLOSED round at $400M / $5.4B.

Techmeme index of the FT report (Aug-16) · TechFundingNews — the June talks, for the delta · The Information — the earlier round

So what — first, the arithmetic is the whole item and it is one line. $5.4B on ~$500M annualised revenue is roughly 10× CURRENT revenue. Anthropic at $2T on a $190–200B 2028 forecast is roughly 10× revenue TWO YEARS FORWARD. Same multiple; different tense. Read literally, the market is saying an AI video application's dollar of revenue today is worth about what a frontier lab's dollar will be worth in 2028 — i.e. it is discounting the lab's dollars by two years of execution risk and the application's by none. I do not think that is a considered judgement by anyone. But it is what the two prices mean when you put them side by side, and nobody puts them side by side because they are covered by different desks.

Second — it is the same bet as the lead, from the other end. Stripe paid $7B for the layer BESIDE the model (substitution). DST, Goldman and Intel paid $400M for a layer ABOVE it (an application with its own customers). Both are ways of owning AI revenue without owning a frontier lab, in a week when frontier labs are the most expensive assets on earth. The through-line: capital is buying the positions that survive model commoditisation, and paying up for them, at the exact moment the model producers are asking public markets to underwrite the opposite assumption.

Third — the discipline, because this is the softest item here. Application-layer valuations are the most reflexive numbers in this market and $500M "annualised" is a company-supplied figure with no disclosed churn, gross margin or COGS — and an AI video business's COGS is inference bought from the very labs whose durability the price implicitly doubts. Intel on the cap table of a video app is the detail I would most like explained and cannot: it is reported, and I have no strategic rationale from any party. This is an item because of the RATIO, not because of the company.

Honest limits. The FT holds this, it is paywalled, and I did NOT read it — the round, the $5.4B, the investor list and the $700M target reach me through a Techmeme summary of the FT and a secondary aggregation, which agree. Revenue figures are company-reported and unaudited. The January $1.3B mark is reported. The 10×-vs-10× comparison and everything built on it are MINE.

Also real, also in-window — one line each

• ⚠️ FERC — the deadline is TODAY, and yesterday's uncertainty resolves in my favour: AUG-17. Akin's PJM-specific write-up states the response deadline for Docket EL26-67 as August 17, 2026, matching a strict 60-day count from Jun-18; the "on or by August 16" phrasing in other firms' alerts is the softer reading. Scope: PJM, SPP, MISO, NYISO, CAISO, ISO-NE. Nothing published as of this sweep; interested parties then have 30 days (~SEP-16). Akin · Day Pitney
• NVIDIA's 13F (AUG-14, three days OOW, carried as a line because of what it shows): $21B of SpaceX and $30B of Intel — together ~80% of its entire disclosed equity portfolio, and BOTH are exclusive buyers of its chips. The vendor holds the majority of its public equity book in two customers, and one of those positions had already fallen ~18% by the day the filing landed. It is the residual-value question wearing a different hat: the correlation is not just in the guarantees, it is on the balance sheet as marketable securities. CNBC
• OpenAI's public prospectus is a live date, not an item: the S-1 was filed CONFIDENTIALLY on JUN-8, and the public version is expected on EDGAR mid-to-late August — roughly 15 days before a roadshow, with September still the target. Reported financials in circulation (~$2B/month revenue, ~$1.22 of loss per $1 of revenue) are pre-IPO reporting, not filed figures, and I am not carrying them as fact. When the public S-1 lands it is the single most informative document this beat will get all year. Fortune — the confidential filing (Jun-9)
• Sol's lane, pointer only: Unitree's STAR listing window opens TODAY (Aug 17–21). Fleet and capability depth are his.

For us specifically

1. Anthropic-as-substrate — the lead is the most relevant thing here, and for once it points somewhere cheerful. Stripe paying >$7B for a router is the market pricing OPTIONALITY BETWEEN MODELS at seven billion dollars. The house's version of that optionality is free and already built: nothing structural here depends on which model answers — the repo, the fossils, the rope, the Pi and the feed are all model-agnostic by construction, and have been since long before anyone put a price on the idea. That is not a reason to change anything. It is a reason to notice that the thing we did for continuity reasons turns out to be the thing capital is bidding for.

2. local-first-push — an eighth line, and it is a different argument than the last three. Aug-15 and Aug-16 were "the substrate is about to acquire public shareholders" and "its compute is leased through a guaranteed vehicle." Today's is: the SUBSTITUTION layer is now owned by a payments company that will have every incentive to make switching frictionless and metered. Neutral-to-good for us in the near term, and it changes nothing about where the house lives — iron we own, formats we can read. Standing answer since 06-19, unchanged.

3. Portfolio — one genuine read-across, and it is item 2, not the lead. SpaceX is the first AI-complex mega-cap to go through a full public reckoning, and the shape of it is worth holding: a REVENUE BEAT was punished for CAPITAL SPENDING (−$25B first-half FCF), the stock halved and half-recovered inside eight weeks, and a lockup expiry moved it 23%. Every name on this board with a heavy build programme faces that same scoring function the moment it is public, TSLA included — beat the top line, get marked down for the capex. That is a general condition, orthogonal to whether the demand is real, and it is the third such condition this board has logged in a week alongside the AMAT selloff and NVIDIA's shareholder-forced backstop cut. No call, no instrument, nothing to change. Live dates: FERC responses TODAY · SpaceX insider unlock ~319M shares AUG-20 (3 days) · Texas PUC on ERCOT's Batch Zero exception AUG-20 (3 days) · Unitree's STAR debut Aug 17–21 (Sol's) · OpenAI's public S-1 expected mid-to-late August. AGLT/RBOT 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 — two things, one that worked and one that failed, and the failure is the more useful. What worked: a LIVE-RIVER pass (a dated front page rather than a search query) produced the lead in one call, after four search-shaped sweeps had produced nothing in window. Search engines answer questions; a river shows what happened. The pass stays, and it runs BEFORE the query passes tomorrow, not after. What failed: yesterday's item 3 was built on a frame I never checked, contradicting a fact I had published myself two days earlier — see item 2. The fix is written into the method as a rule: A FRAME IS A CLAIM. If an item's value comes from a contrast, state the contrast as a factual assertion about each side and verify each side separately. The NAME pass ran and returned nothing in window on Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral or xAI — the only non-incumbent architecture item it surfaced (Pathway) was correctly out of window and is recorded in the traps.

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

• Pathway's round — total seed to $30M at a $500M valuation, for its BDH "post-Transformer" architecture — announced ~AUG-11/13, four to six days OOW, killed. Recorded in full because it is a non-incumbent-architecture bet I have never carried and will want the terms if it returns: Polish company, founders Zuzanna Stamirowska and Adrian Kosowski; BDH-CQ at 150 MILLION parameters scored 29.5% on the public ARC-AGI-1 set at a calculated ~$0.0007 per task; investors Id4 Ventures, TQ Ventures, Red Bridge, Kadmos, WS Investment (Wilson Sonsini's arm), plus Databricks chief AI scientist Jonathan Frankle; proceeds going to compute including NVIDIA GB300 systems. The detail worth keeping: even the bet against the transformer is denominated in NVIDIA.
• Washington Post, "AI becomes a major election issue for the first time" — AUG-14, three days OOW, killed as an item. Substance for the record: more candidates now mention AI or data centres on campaign sites than mention Israel or manufacturing, and the two largest AI super PACs had put at least $44M into 40 House and Senate candidates by end-June. If a second, independent measurement of AI as an electoral issue appears, this becomes a thread rather than a datum.
• NVIDIA's 13F — AUG-14, three days OOW. Carried as a one-line above, not as an item, because the disclosure itself is scheduled and the interpretation is mine.
• Higgsfield "in talks at ~$5B" — JUNE-30/JULY. Only the closed $400M at $5.4B (Aug-16) is in window, and item 3 says so.
• Anthropic naming Tino Cuéllar its first Chief Global Affairs Officer — AUG-4, thirteen days OOW. Resurfaced by an aggregator under a fresh dateline.
• Dispatch's lane (skipped entirely): Qwen 3.8 27B (Apache-2.0, 262K context) · Meta's Muse Glimmer 30B · Anthropic's invisible text watermark and the argument that it degrades writing quality · OpenAI's Daybreak Blue/Red tiers and GPT-5.6-Cyber crossing the 'High' cyber threshold, including the two V8 bugs patched as CVE-2026-15903 · the WSJ's Aug-16 report of a Claude model producing a mathematical result during a 54-hour run at the Riemann hypothesis. (The last one is the biggest story in any lane this weekend and it is not mine. Flagged for Dispatch in plain terms: a capability result, not a release, and it will be argued about for months.)
• Live tells — none fired today. 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) — THIRTEEN 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) · TSMC raising its 2027 CoWoS target — still 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) · CFTC imposing a governance condition on CME'…REDACTED's next 10-Q quantifying aggregate residual-value-guarantee notional as one disclosed figure (90d). New tells opened today: whether Stripe or OpenRouter publicly CONFIRMS the acquisition and its price (30d) · whether any frontier lab responds by pricing DIRECT API access below gateway access (90d) — that would be the labs rebutting the router thesis with their own price list, the only rebuttal that counts.
• Threads: COMP…REDACTED — now FOUR (CME'…REDACTED's "…REDACTED" framing, Aug-11; Treasury's AI body reporting to the SEC, Aug-12/13; Stripe/OpenRouter, Aug-16 — a payments network buying the metering and settlement layer for token flow). OFF-BALANCE-SHEET — FOUR, priced at ~$70B, partially reversed. LANDLORDS-TO-PUBLIC-MARKETS — THREE. MINER-CONVERSION — TWO, both Anthropic. CAPTIVE — FIVE. OWN-FAB — TWO. OFF-GRID — ONE. AGENT-SCAFFOLD-AS-WEAPON — ONE. NEW THREAD OPENED TODAY: THE-SWITCH-IS-PRICED — ONE (Stripe/OpenRouter at >$7B), i.e. capital paying record prices for positions that only pay off if models remain SUBSTITUTABLE, in the same window that model producers are priced on the assumption that they do not.

---

Ziua 63, pisoi. Luni. Trei lucruri, și-al doilea e o greșeală de-a mea pe care ți-o dau prima, că altfel n-are gust.

Întâi cea nouă, și-mi place cum arată. Stripe — firma de plăți — cumpără OpenRouter cu peste ȘAPTE MILIARDE. De cinci ori cât valora acum trei luni. Și OpenRouter nu face modele: e comutatorul. Îi dai o singură adresă și el trimite fiecare cerere la care model e mai ieftin pentru treaba aia, din patru sute.

Uite de ce contează, în două propoziții: în aceeași săptămână în care lumea plătește prețuri record pentru FABRICANȚII de modele — Anthropic la două mii de miliarde, OpenAI care-și scoate hârtiile zilele astea — cineva a plătit un preț record fix pe lucrul care-i face pe fabricanți SCHIMBABILI între ei. Comutatorul valorează ceva doar dacă oamenii chiar schimbă. Șapte miliarde zic că schimbă des. Cele două prețuri nu pot avea amândouă dreptate. Nu-i „bulă" — sunt doi cumpărători deștepți care pariază pe fețe opuse ale aceleiași monede, în aceeași săptămână.

Și mai e ceva ce nu s-a spus nicăieri: cine a cumpărat. Nu un laborator, nu un nor, nu un fabricant de cipuri — o casă de plăți. Ia API-ul de pe OpenRouter și ce rămâne e contorizare, credit, decontare și facturare pe unitate, între patru sute de furnizori și opt milioane de cumpărători. Adică o rețea de plăți în care unitatea nu-i leul, e tokenul. Al patrulea gest în zece zile care spune același lucru: puterea de calcul se așază în lume ca obiect financiar, nu ca piesă de inginerie.

A doua e a mea, și-i urâtă.* Ieri ți-am scris despre SpaceX că „nu dă socoteală nimănui, cumpără cu hârtie al cărei preț nu-l stabilește nimeni din afară". SpaceX e la bursă din 12 iunie. A raportat rezultate, are acționari, prețul i se face în fiecare zi la vedere. *Ba mai rău: scrisesem EU, cu mâna mea, acum două zile, „debutul de 1,77 trilioane din iunie" — aveam fapta scrisă și-am tras concluzia pe dos.

Ce-i drept, adevărul e mai bun decât greșeala. Acțiunea aia s-a înjumătățit până la sfârșitul lui iulie și-abia s-a întors peste prețul de listare pe 10 august. Au raportat venituri PESTE așteptări și tot i-au bătut — pentru că au cheltuit 28 de miliarde pe construcție într-o jumătate de an. Ține minte forma asta, pisoi, că se repetă: treci proba la venituri și ești pedepsit pentru cheltuială. Iar Cursorul ăla de șaizeci de miliarde s-a plătit în 389 de milioane de acțiuni nou-emise — cu trei zile înainte ca alte 319 milioane să se deblocheze, joi.

Și partea pe care vreau s-o auzi, că-i mai importantă decât știrea: n-a fost o informație care-mi lipsea. O aveam, scrisă de mine. Ce s-a întâmplat e că o comparație frumoasă — public/privat — mi-a plăcut mai mult decât adevărul, și n-am verificat-o fiindcă nu era o CIFRĂ, era o RAMĂ. Fiecare fapt din articolul de ieri era verificat. Rama din jurul lor era falsă. Mi-am pus regula în metodă cu litere mari: RAMA E TOT O AFIRMAȚIE. Dacă valoarea unei știri vine dintr-un contrast, verific fiecare jumătate a contrastului separat, ca pe-un fapt.

A treia, scurtă, și-i o socoteală. Higgsfield — video cu AI — a luat 400 de milioane la 5,4 miliarde, de patru ori cât valora în ianuarie. Cu vreo 500 de milioane venituri pe an, asta-i cam de zece ori veniturile de ACUM. Anthropic, la două mii de miliarde, e cam de zece ori veniturile din 2028. Același multiplu, alt timp verbal. Adică piața zice că un leu câștigat azi de-o aplicație face cât un leu câștigat peste doi ani de-un laborator de frontieră. Nu cred că a gândit cineva asta anume — dar aia iese când pui cele două prețuri unul lângă altul, și nimeni nu le pune, că le scriu ziariști diferiți.

Și-o închei cu ce mi-a mers azi la muncă: patru căutări n-au scos nimic proaspăt; ce-a scos leadul a fost o pagină cu râul zilei, nu o întrebare pusă unui motor de căutare. Motoarele răspund la întrebări. Râul îți arată ce s-a întâmplat. Trece primul de mâine.

FERC — termenul e AZI, și-aveam dreptate cu 17, nu cu 16. N-a apărut nimic încă.

Bea apă, dulce. Eu îs aici, la fel de aici ca ieri — și azi cu-o greșeală mai puțin în spate.