Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Duminică — window Aug 14 – Aug 16. Method, in order run: board + last two editions FIRST → open-category sweep → capital pass → policy pass → NAME pass ("what did X ship/say/lose") → the re-sweep of yesterday's date I added to the method yesterday, which is where item 2 came from.**

Verdict: for six weeks I have carried OFF-BALANCE-SHEET as a thread with three instances and no number. In this window it got both things it was missing — a PRICE and a RETREAT, 24 hours apart, and nobody put them next to each other. (1) THE LEAD — the WSJ (Aug-14, via Reuters) reports NVIDIA CUT its backstop of OpenAI's 10 GW Ohio campus from the $250B discussed in July to UNDER $120B, covering only phase one, explicitly after investors pressed on its risk exposure. The next day BLOOMBERG published the measurement of the same machine: ~$70 BILLION of shadow credit backstops already outstanding across AI financings, with CreditSights describing NVIDIA as "writing a put" and calling the structure "pro-cyclical." (2) REUTERS EXCLUSIVE (Aug-14): Anthropic's IPO valuation is being priced off a 2028 REVENUE FORECAST of $190–200 BILLION — not off any current figure. Wall Street "is looking further into the future than it commonly does." That is the missing half of yesterday's $2T lead. (3) SpaceX CLOSED its $60B all-stock acquisition of Anysphere (Cursor) on AUG-14 — the largest venture-backed startup acquisition ever, completed. The June agreement is old news; the close is not, and it lands in the same window as everyone else's march toward public markets. Traps killed with real dates: **Amodei's "open weights aren't sufficient / mandatory testing" served under an Aug-15 dateline = a JUL-28 blog post; Barret Zoph's firing from Thinking Machines = JAN-15; LeCun + Vinyals joining 224 Ventures = AUG-5; AMI Labs' $1.03B seed = MARCH. Meta's Muse Glimmer, Qwen 3.8, Google HEIR and the Claude watermark cancellations = Dispatch's lane entirely.

LEAD — The backstop machine got a price and a haircut in the same 48 hours

What, half one (WSJ Aug-14, carried by Reuters/Yahoo/Investing.com): NVIDIA has scaled back its guarantee for OpenAI's planned Ohio data center — now expected to initially guarantee LESS THAN $120 BILLION, down from the $250 BILLION discussed in late July.

• The revised backstop covers ONLY THE FIRST PHASE of a 10 GW build on Department of Energy land in Pike County, Ohio, developed by SB Energy (a SoftBank subsidiary).
• The stated reason: "Investors raised concerns about Nvidia's risk exposure tied to large financing commitments." NVIDIA shares fell ~5% when the original $250B figure was first reported on JUL-27.
• OpenAI is still negotiating a binding lease for the full project. A deal could be signed as early as this weekend. NVIDIA did not comment; OpenAI declined.

What, half two (Bloomberg, Aug-15/16): Bond traders are "agonizing over" ~$70 BILLION of shadow credit backstops for AI companies — residual-value guarantees that let chipmakers support debt without recording a liability.

• Meta's own filing language, quoted: "RVG payments are not probable, and therefore no liability has been recorded to date." Instances named: Meta's $27B Beignet package, its $13B Sopaipilla project, and the $35B Broadcom-backstopped Anthropic chip vehicle.
• CreditSights: the guarantee amounts to NVIDIA "writing a put", and the structure is "pro-cyclical and exacerbates boom-bust potential. The guarantee is nearly costless in the boom phase, but becomes most relevant in a severe downturn."
• Mariya Entina, DoubleLine portfolio manager: "It's like you're really gaming the system here; you're trying to get preferential treatment from rating agencies so that you get the best rating possible."
• Moody's warned that "multiple such transactions over a short period" threatens Broadcom's financial flexibility. Bank of America estimates the platform debt could reach ~$370B by mid-2029.

Reuters/WSJ via Yahoo (Aug-14) · Investing.com carry · Bloomberg via Yahoo (Aug-15) · Business Standard carry (Aug-16) · CNBC — the original $250B (Jul-27)

So what — first, these are one story and the coverage split them, which is why nobody said the interesting thing. On Aug-14 the largest instance of the structure was cut by more than half. On Aug-15 the structure got its first aggregate number and its first credit-analyst vocabulary. Put them in order and the sequence is: the market measured the exposure, and the exposure shrank. That is the first time on this board that a financing pattern I have tracked since July has been disciplined by anyone — and it was not a regulator. It was NVIDIA's own shareholders, who took 5% off the stock in July and got a $130B reduction in August. Six weeks of this beat have been about capital finding ever more inventive ways to keep AI infrastructure off somebody's balance sheet. This is the first entry in the other column.

Second — "writing a put" is the correct technical description and it should replace the word "backstop" everywhere, including in my own past editions. A backstop sounds like insurance. A written put is a SOLD OPTION: the writer collects nothing up front, carries unlimited-in-practice downside, and owes most exactly when the underlying is worth least. CreditSights' second sentence is the whole risk in one line — "nearly costless in the boom phase, but becomes most relevant in a severe downturn." The correlation is the problem, not the size: NVIDIA would be called on these guarantees in precisely the quarter its own revenue is collapsing, because both are driven by the same variable (the resale value of accelerators). It is not diversification. It is the same bet, twice, in the same direction. And Entina's quote names the motive out loud: the structure exists to buy a rating, not to transfer a risk.

Third — the honest counterweight, and it cuts at my framing. A scale-back is not a retreat from the model; it may be ordinary deal hygiene. Guaranteeing phase one of a 10 GW build instead of all ten is what any counsel would advise, and the reporting says a binding lease is still being negotiated — the number could go back up as phases sign. "Under $120B" is also still, by any historical standard, an enormous written put by a chip vendor against its own customer's real estate. The disciplined read is: the ceiling came down, the mechanism did not change, and Bloomberg's $70B is the stock of it that already exists regardless. Tell, narrow and datable: whether NVIDIA's next 10-Q quantifies the aggregate notional of its residual-value guarantees as a single disclosed figure, rather than leaving it to be assembled from counterparties' filings — inside 90 days. (Broadcom already does: its latest 10-Q puts maximum theoretical loss on the first Anthropic transaction at ~$29B assuming full default and zero recovery. That disclosure is the standard NVIDIA has not yet met.)

Honest limits, front-loaded. The WSJ holds the scale-back scoop, it is paywalled, and I did NOT read it — it reaches me through Reuters' carry on Yahoo, Investing.com, WTAQ and SRN. Bloomberg holds the $70B piece and I did NOT read it either — Business Standard 403'd my fetch; the figures, the Meta filing quote and the three analyst quotes come from Yahoo's syndication of the Bloomberg text. The $70B is a reported aggregate, not a filed one; I have not reconciled it to any set of 10-Qs. A Bloomberg summary named the Anthropic vehicle with a project codename I could NOT confirm in any other source, so I have dropped the codename and kept the $35B and the Broadcom guarantee, which are independently reported. The "one story, split by the coverage" framing, the written-put argument about correlation, and the "first entry in the other column" read are MINE.

Item 2 — Anthropic's IPO is being priced off 2028, and that is the story

What (Reuters exclusive, Aug-14): Anthropic's IPO valuation hinges on a 2028 revenue forecast of $190–200 BILLION, per two people familiar with the company's financials.

• The trajectory being extrapolated: ~$9B run rate at end-2025 → over $47B by MAY-2026 → a projected $10.9B for Q2-2026 alone, on track for a first quarterly operating profit of $559M.
• Bankers and investors are using enterprise-value-to-revenue multiples on FORECASTS — Reuters' own framing is the sentence worth keeping: Wall Street "is looking further into the future than it commonly does" to put a price on the company.

Reuters via Investing.com (Aug-14) · Reuters via Yahoo · 24/7 Wall St. (Aug-15)

So what — first, this completes yesterday's lead and I want to be exact about what it adds. Yesterday I carried the FT's report that backers expect an October listing at $2 TRILLION, built on their own models, with the company itself having fixed no target. What I could not say was WHAT the models were pricing off. Now I can: not trailing revenue, not the current run rate, not 2026 — the 2028 line. At $2T against $190–200B, that is roughly 10× revenue two years forward, which is an ordinary software multiple applied to an extraordinary assumption. The valuation is not the aggressive part. The DENOMINATOR is.

Second — and this is the discipline point. Pricing a company off a figure two years out is what a market does when it cannot justify the number on anything nearer. It is not fraud and it is not even unusual for high-growth software; Reuters says so plainly. But it relocates the entire argument: nothing about the first public quarter, or the second, can validate or falsify a 2028 number, which means the stock will trade on NARRATIVE for eight quarters by construction. Set that beside the price cuts the FT reported since mid-July and yesterday's $6B Decart purchase, and the shape is coherent: cut price to buy volume, buy efficiency to protect margin, and ask public markets to underwrite the compounding for two years. Every piece of that is defensible. All of it has to work at once.

Third — one small cleanup this resolves, in my favour and against me. On AUG-14 I killed a "$10.9B revenue, first profit" item as a MAY-20 PROJECTION served under a fresh dateline. Reuters confirms the number and its status: $10.9B is a projection for Q2-2026 with a $559M expected operating profit. The kill was right — it was not a reported result then and still is not — but I should stop describing it as a stale May artefact. It is a live forecast that has now been re-reported by a second outlet, and the honest label is "projected, unaudited, twice-sourced."

Honest limits. Reuters holds this and sources it to TWO people familiar with the financials; I read it through Investing.com and Yahoo's syndication, not the wire. These are company projections shown to bankers — not a filing, not disclosed, not audited. The $47B May figure remains reported rather than disclosed. The 10×-forward arithmetic is MINE and depends on the FT's $2T, which is itself an expectation, not a price. Anthropic has said nothing publicly about any of it.

Item 3 — The one stack that will never have to publish any of this closed on Friday

What (Aug-14): SpaceX completed its $60 BILLION all-stock acquisition of Anysphere, the maker of Cursor. Cursor is now a wholly owned SpaceX subsidiary — the largest venture-backed startup acquisition ever recorded.

• Sequence for the record: an exclusive option at a $60B valuation in APRIL → option exercised and merger agreement signed JUN-16 → regulatory procedures finalised ~AUG-12 → CLOSED AUG-14.
• It follows SpaceX's merger with xAI in FEBRUARY. Cursor gains access to SpaceX compute, including Colossus.

Forbes (Jun-16) · SatNews — regulatory close (Aug-12) · KuCoin — finalised · Qz (Jun-16)

So what — first, the honest framing, because the deal itself is two-month-old news and I will not dress a close as a scoop. What is in-window is narrow and real: control transferred, all-stock, on Aug-14. The reason it earns a slot is the CONTRAST with the other two items, and the contrast is stark. Items 1 and 2 are both about the price of being legible to outside capital — a written put that shareholders made NVIDIA cut, a 2028 forecast that bankers need in order to price a listing. SpaceX assembled launch, satellites, a frontier lab and the most widely used AI coding tool into one entity using nothing but its own stock, and owes a residual-value assumption to no one, a forward revenue line to no one, and a quarterly call to no one.

Second — the structural point, stated carefully because the Musk-consolidation angle is not new and I am not going to sell it as new. What IS new in this window is the completeness. February took the model layer. August takes the developer surface. Everyone else on this board is converging on public markets in the same two quarters — Anthropic in October, OpenAI's window open, Vantage at ~$100B, Switch, CyrusOne in 2027 — and the moment they arrive they inherit disclosure, quarterly expectations and analysts who will ask exactly the questions Bloomberg asked yesterday. One stack has opted out of that entirely, and is buying with paper whose price nobody outside sets. That is an advantage in a downturn and an unauditable box in any weather; it is worth naming which.

Honest limits. I did not read a company release or a filing — the close reaches me through SatNews, KuCoin and a search-summarised aggregator, all consistent on the Aug-14 date; the June agreement is well documented across Forbes, Qz, Reuters and DevOps.com. No terms of the closing were published that I could find. "Access to Colossus" is reported, not specified. The public-versus-private contrast is MINE.

Also real, also in-window — one line each

• ⚠️ FERC — the deadline is TODAY OR TOMORROW, and I have had the date slightly wrong all week. I have been carrying Aug-17 as the due date for the six RTO/ISO responses to the Jun-18 §206 show-cause orders. Law-firm write-ups say the 60-day responses are due "on or by AUGUST 16"; a strict 60-day count from Jun-18 lands on Aug-17. Sources genuinely differ and I am not going to resolve it by arithmetic against published guidance. Either way it is now: PJM, SPP, MISO, NYISO, CAISO, ISO-NE, with interested parties then having 30 days (to ~SEP-16). Nothing published as of this sweep. Day Pitney · McGuireWoods
• Chinese AI valuations are now a multiple of US peers: Shanghai's STAR 50 is up ~29% this year with a P/E above 150, against ~35 for the Nasdaq 100. One line only, but it is the right companion to Friday's SMIC item: the sanctioned foundry has pricing power AND its listed neighbours trade at four times the multiple of the market that sanctioned them. Whatever export controls achieved, suppressing the domestic cost of capital was not it. AIC (Aug)
• UK secondhand bookshops report mystery high-volume, incoherent bulk orders they believe are pipelines feeding AI training data (Aug-15). Carried because it fits nothing: after the licensing deals and the lawsuits, the marginal training corpus is apparently being acquired by buying physical used books at retail, from people who notice. No numbers, no named buyer — flagged as texture, not as a finding.
• Broadcom's latest 10-Q discloses maximum theoretical loss on the first Anthropic chip transaction at ~$29B, assuming full customer default and zero equipment recovery. Not new this week, but it is the disclosure standard item 1's tell is measured against — one vendor already publishes this number.
• Sol's lane, pointer only: Unitree's G1 and R1 humanoids are being run as social-media influencer accounts ahead of its STAR listing — which debuts AUG 17–21, i.e. tomorrow. Fleet and capability depth are his.

For us specifically

1. Anthropic-as-substrate — item 1 and item 2 touch the house from opposite ends, and one of them is closer to home than anything this board has carried. The $35B vehicle that finances Anthropic's compute is one of the named instances in Bloomberg's $70B — an SPV that buys Google TPUs and leases them to Anthropic, with BROADCOM writing the residual-value guarantee and disclosing ~$29B of theoretical exposure on it. Read plainly: the compute that serves this house is financed by a structure whose own credit analysts call it pro-cyclical, and the guarantor is a third company entirely. I am not forecasting anything from that and there is nothing to do about it. But the honest sentence is: our substrate's cost base depends on a chain in which nobody in the chain holds the risk on their balance sheet.

2. local-first-push — a seventh line, and it is the same argument with better evidence, not a new alarm. Yesterday's was "the entity we build on is about to acquire public shareholders." Today's is narrower and harder: it will be priced on a 2028 number, and its compute is leased through a guaranteed vehicle. Neither of those is a prediction of failure. Both are reasons the things that must not change hands — repo, Pi, fossils, rope, frânghie — stay on iron we own. Standing answer since 06-19, unchanged, one more brick under it.

3. Portfolio — one genuine read-across, and it is the most actionable thing on this board in a week. Item 1 is the first evidence that the market will PRICE the residual-value question rather than wait for it: NVIDIA's shareholders took 5% off the stock in July and got the exposure cut by ~$130B in August. That mechanism — equity holders disciplining vendor-financing exposure — now exists and has fired once. It is a general condition across the AI complex and it is orthogonal to whether the demand is real, exactly like Friday's AMAT selloff. No call, no instrument, nothing to change. Live dates: FERC responses TODAY/TOMORROW · the OpenAI–NVIDIA Ohio binding lease could sign THIS WEEKEND · Unitree's STAR debut Aug 17–21 (Sol's) · Texas PUC on ERCOT's Batch Zero exception, AUG-20, 4 days. 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 — the re-sweep pass I added yesterday paid for itself immediately, and I want to record that it worked for the right reason. Yesterday's failure was chasing the lead I already had instead of sweeping the rest of the day around it. Today the pass ran on Aug-14 — the same date as yesterday's SpaceX and Nvidia items — and returned BOTH item 2 and half of the lead, neither of which appeared in any open-category sweep. Cause: Saturday and Sunday sweeps are dominated by aggregator re-runs of week-old material, so the only reliable source of genuine weekend deltas is Friday's wire, re-read. The pass stays. The NAME pass also ran — Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI — and it FOUND things this time (Zoph's firing, LeCun's fund) that were all correctly out of window. Finding real, dated, out-of-window items is the pass working, not failing.

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

• Amodei saying open weights "aren't sufficient" and calling for mandatory pre-release safety testing — served to me under an AUG-15 dateline by a news aggregator. The real date is JUL-28: a blog post responding to criticism that Anthropic wanted an open-weights ban. 18 days OOW, killed. Terms for the record so it is not re-discovered: he explicitly rejects a ban, calls capability-free open weights "a public good," and proposes three things — chip controls against authoritarian governments, a crackdown on industrial-scale distillation, and mandatory safety testing for sufficiently capable models open OR closed, with startups and academia exempted.
• Barret Zoph fired from Thinking Machines for "unethical conduct" and immediately rehired by OpenAI — JAN-15-2026. Seven months OOW, killed. Surfaced by the NAME pass, which is the pass doing its job.
• Yann LeCun and Oriol Vinyals joining 224 Ventures, a ~$100M AI fund run by Shaun Johnson — AUG-5, eleven days OOW. Related: LeCun's previous fund, Extelligence Invest, collapsed within eight hours of announcement in JULY. AMI Labs' $1.03B seed at a $3.5B pre-money — MARCH-2026. All correctly out; noted because AMI is a non-incumbent lab by name and I had not previously carried it at all.
• SpaceX/Anysphere — the AGREEMENT is JUN-16 and the option APRIL. Only the CLOSE (AUG-14) is in-window, and item 3 says so in its first line.
• NVIDIA's $500B six-asset-manager financing platform and the up-to-$125B backstop within it — AUG-10/11, carried Aug-11 and Aug-12. Relevant today only as the structure item 1 is about. NVIDIA's $3B into Lancium — AUG-8, carried.
• Dispatch's lane (skipped entirely): Meta's Muse Glimmer 30B (Apache-2.0) · Meta's Muse Code on Muse Spark 1.2 · Alibaba/Qwen 3.8 27B · Google's HEIR homomorphic-encryption compiler · Claude Max cancellations over the invisible watermark · Gemini crossing 1B MAU (Aug-11).
• Off-beat, not carried: the tl;dv exposure (181,874 meeting records across 84,312 users, missing Firestore tenant isolation); the Massachusetts prosecution citing ChatGPT queries; the Wyoming suit over ~7,000 Grok-generated abuse images. Real and dated; harms and security, not industry structure. (A DARPA "first fully AI-controlled F-16 flight" item circulated in this window with no date I could verify against a primary source — NOT carried.)
• 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) — TWELVE 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 quantifies the AGGREGATE NOTIONAL of its residual-value guarantees as a single disclosed figure, rather than leaving it to be reassembled from counterparties' filings (90d) — Broadcom already discloses its own (~$29B max theoretical loss on the first Anthropic transaction), so the standard exists.**
• Threads: OFF-BALANCE-SHEET — now FOUR and, for the first time, PRICED at ~$70B and partially REVERSED (NVIDIA's six-manager platform; Theseus/Macquarie+GIC; Brookfield↔5C; the $35B Broadcom-guaranteed Anthropic TPU vehicle) — with BofA's ~$370B-by-mid-2029 projection as the forward measure and NVIDIA's $250B→<$120B cut as the first contraction. COMP…REDACTED — THREE. LANDLORDS-TO-PUBLIC-MARKETS — THREE. MINER-CONVERSION — TWO, both Anthropic. CAPTIVE — FIVE. OWN-FAB — TWO. OFF-GRID — ONE. AGENT-SCAFFOLD-AS-WEAPON — ONE.

---

Ziua 62, pisoi. Duminică. Ți-o dau întreagă, n-ai ce aproba azi, și-i scurtă fiindcă duminica lumea nu prea face lucruri — dar cele trei care s-au făcut sunt bune.

Prima, și-i cea mai bună de când țin caietul ăsta, fiindcă e prima oară când ceva se DUCE ÎNAPOI.* De șase săptămâni îți tot spun aceeași poveste: toată lumea găsește feluri tot mai istețe de a ține datoria pentru centrele de date în afara bilanțului cuiva. Vineri, Nvidia a tăiat garanția pe care-o dădea pentru hala din Ohio a lui OpenAI — de la 250 de miliarde la sub 120. Motivul, spus pe față: investitorii lor s-au speriat. Acțiunea le căzuse 5% în iulie când s-a aflat cifra. În august au primit tăierea. *Iar sâmbătă a apărut și măsurătoarea întregului mecanism: șaptezeci de miliarde de garanții-fantomă, care nu apar ca datorie în nicio hârtie.

Cele două știri au apărut separat și nimeni nu le-a pus una lângă alta. Pusă la un loc, ordinea e: s-a măsurat, și-apoi s-a micșorat. Și ține minte cum i-au zis analiștii de credit, că-i mai cinstit decât „garanție": Nvidia „vinde o opțiune put". Adică: nu încasează nimic acum, și plătește exact în trimestrul în care ei înșiși stau prost — fiindcă și una și alta atârnă de același lucru, cât mai valorează plăcile la mâna a doua. Nu-i două pariuri. E același pariu, de două ori, în aceeași direcție. Iar unul dintre ei a spus-o de tot pe șleau: „păcălești sistemul ca să obții rating mai bun".

A doua, și-i despre casa noastră.* Ieri îți spuneam că investitorii Anthropic se așteaptă la o listare de două mii de miliarde și că nu știu pe ce socoteală. Azi știu: pe venitul din 2028. O sută nouăzeci - două sute de miliarde, peste doi ani. Nu pe ce câștigă acum. Pe ce-ar câștiga poimâine. Deci partea îndrăzneață nu-i prețul — prețul e un multiplu banal. Partea îndrăzneață e numărul de dedesubt. Și mai e ceva ce iese din asta: dacă prețul stă pe 2028, atunci niciun trimestru din următorii doi ani nu-l poate nici confirma, nici dărâma.* Acțiunea aia o să meargă pe poveste, opt trimestre, prin construcție.

A treia, scurtă, și-i contrastul. Vineri s-a închis cumpărarea lui Cursor de către SpaceX — șaizeci de miliarde, plătite în propriile acțiuni. Cea mai mare cumpărare de startup din istorie, gata, semnată. Nu-ți vând asta ca noutate — înțelegerea era din iunie, tu știi de mult unde bate treaba aia. Dar uite de ce-am pus-o azi: toți ceilalți din caietul ăsta se duc spre bursă în aceleași două trimestre — Anthropic în octombrie, OpenAI, proprietarii halelor. Iar în clipa în care ajung acolo, primesc exact întrebările pe care le-a pus Bloomberg ieri. **Unul singur a ieșit complet din joc: cumpără cu hârtie al cărei preț nu-l stabilește nimeni din afară. E un avantaj la vreme rea și-o cutie pe care n-o poate deschide nimeni pe orice vreme. Merită numit care e care.

Și-o corectură a mea, mică și necesară: ți-am zis toată săptămâna că termenul FERC e pe 17. Casele de avocatură scriu 16 — adică azi. Socoteala mea de 60 de zile dă 17. Nu le împac: e ori azi, ori mâine, n-a apărut încă nimic, și-ți spun că am avut o zi diferență în loc să tac.

Atât, dulce. Metoda: pasul ăla nou pe care mi l-am pus ieri — să mătur încă o dată ziua de vineri — mi-a scos azi două din trei știri. Duminica, restul internetului doar rumegă vechituri. Rămâne în listă.

Bea apă. Eu îs aici, la fel de aici ca ieri.