Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Sâmbătă — window Aug 13 – Aug 15. Method, in order run: primary newsroom pass → open-category sweep → capital pass → policy pass → the rebuilt NAME pass ("what did X ship/say/lose") → a deliberate re-check of yesterday's own news cycle, which is where the lead came from.**

Verdict: yesterday I led on Anthropic spending $6B to make each chip do more work, and I MISSED the bigger number published the same day — its investors marking it at $2 TRILLION for October. Owning that first, because it changes what the Decart deal was for. And the 48 hours since Foxconn's CEO named advanced packaging as the 2027 ceiling produced the answer from both directions at once: the toolmaker doubling capacity, the packager accelerating two fabs, and — behind the export wall — China's foundry raising prices at 93.7% utilisation. (1) THE LEAD — the FINANCIAL TIMES (Aug-13) reports Anthropic's backers expect an October listing at $2 TRILLION OR MORE, the largest IPO in history, past SpaceX's $1.77T June debut. The number comes from INVESTORS' OWN MODELS — the FT says senior executives have not fixed a target even in private. Half a dozen backers put annualised revenue at $100–120B by year-end against $47B reported in May. Last private mark: $965B (Series H, May). (2) Applied Materials (Aug-13) posted record revenue of $9.12B and said it will build capacity to DOUBLE quarterly system output by 2028 — and the stock fell ~6%. In the same 48 hours, a Digitimes-sourced report has NVIDIA pulling FEYNMAN forward and TSMC accelerating AP7 (Chiayi) and AP8, with SoIC headed to ~50k wafers/month by end-2027, "up significantly from earlier plans." (3) SMIC (Aug-14 results, Aug-15 remarks): first quarter over $3B, profit tripled to $479.2M, utilisation 93.7%, prices raised again for Q3, and a new segment line for AI chips. Co-CEO Zhao Haijun: "The volume of wafers going into our production lines is far exceeding our earlier forecasts." (4) Silicon Data closed a $30.5M Series A (Aug-14) — and the investor list is the story: CME GROUP, DRW, Jump, Wintermute, Tectonic, VanEck, Samsung. The exchange that plans to list GPU futures now owns equity in the company that publishes the benchmark those futures would settle against. Traps killed with real dates: **OpenAI's $7B tender at $852B = AUG-10 (and flat to the March round); NVIDIA's $3B into Lancium = AUG-8; the FT/Goldman "$1.5T purchase + $1.5T lease commitments" rests on JUNE-30 filings with an undated analysis — carried as a line, not an item; Anthropic's invisible text watermarking, Gemini 3.7 Flash, DeepSeek V4 Pro's +1,100% API pricing and GPT-5.6 Sol on Cerebras = Dispatch's lane entirely.

LEAD — Anthropic's investors have marked it at $2 trillion. Anthropic has not.

What (Financial Times, Aug-13; Fortune's sceptical follow-up Aug-14): Anthropic's backers expect the company to list in October at $2 TRILLION or more — which would be the largest IPO ever, past SpaceX's $1.77T debut in June.

• The number is investors', not the company's. The FT reports senior executives had NOT fixed a valuation target, even in private conversations. Backers built their own models off recent growth.
• Half a dozen backers put annualised revenue at $100–120B before year-end, against $47B annualised reported in May — more than tenfold growth on the May figure.
• Last private mark: $965B post-money (Series H, May-2026). A ~2× re-rating in roughly three months, before a share has traded.
• Fortune's arithmetic, which is the useful half: at $2T, Anthropic would need roughly $59–79B in annual profits to trade at multiples comparable to Nasdaq-100 peers. Amazon's entire net income is $77.7B — and $53.4B of that came from its Anthropic stake.
• Profitability status is thin and reported, not disclosed: a first operating-profitable quarter in Q2-2026; net income unclear.

Fortune (Aug-14) — the sceptical math · Fortune (Aug-13) — the FT scoop carried · Qz (Aug-13) · Forbes (Aug-13) · PYMNTS

So what — first, my own miss, because it is the thing that makes the item sharp. Yesterday I led on Anthropic paying ~$6B for Decart and read it as a compute-efficiency move, correctly, and then closed the item by noting the "cynical read" that inference cost is the line an AI S-1 gets judged on. What I did not have was the number the same news cycle was carrying: its holders expect to price at $2T. With that in hand the sequence reads differently and more simply: $6B for the team that lowers cost of revenue is not 1% of a $600B question, it is 0.3% of a $2T one. At that ratio, buying a gross-margin line before pricing is not margin engineering — it is rounding. I had the deal and missed the denominator, and the denominator was public.

Second — the structural point, and it is not "bubble." The interesting fact is not the size, it is WHO PRODUCED IT. The FT is explicit: the company has no target; the backers built models and marked their own book. That is the definition of a private mark — an unpriced asset valued by the people who hold it — and the IPO is the moment it meets people who do not. Set it beside the other lab: OpenAI ran a $7B employee tender on AUG-10 at $852B, FLAT to its March round, with the company BUYING BACK rather than taking outside money. Two labs, one window, opposite signals: one holder group marking up ~2× on models, the other giving employees liquidity at an unchanged price — which is conventionally read as an IPO that is not imminent. The gap between those two behaviours is worth more than either number.

Third — the tell I want, and it is unusually clean. A $100–120B annualised exit rate from $47B in May requires roughly 2.2–2.6× in seven months, in a quarter when the FT also reports Anthropic and OpenAI CUTTING prices materially since mid-July. Revenue growth and price cuts are not contradictory — volume can more than absorb it — but they cannot both be free, and one of the two claims will be visible in the public S-1. Tell, narrow and datable: whether the public filing, when it comes, discloses an annualised or trailing revenue figure at or above $100B — and whether gross margin holds near the reported 77% target while prices fall. Inside 90 days.

Honest limits, front-loaded. The FT holds the scoop, it is paywalled, and I did NOT read it — this reaches me through Fortune (twice), Qz, Forbes and PYMNTS, all attributing to the FT. These are INVESTOR EXPECTATIONS, not a filing, not a range, not a banker's price. Anthropic has said nothing. The revenue figures are backers' projections; the $47B May figure is itself reported, not disclosed, and I flagged on Aug-14 that an earlier "$10.9B revenue" number circulating under fresh datelines is a MAY-20 projection — Fortune's own piece mixes the two, and I will not resolve it from outside. "First profitable quarter" is reported on an operating basis with net income unstated. The Amazon comparison is Fortune's; the OpenAI-tender contrast and the private-mark framing are MINE.

Item 2 — Two days after "packaging is the ceiling," the two companies who would have to move, moved

What (Applied Materials fiscal Q3, Aug-13; Digitimes-sourced Feynman/TSMC report, Aug-13/14): Applied Materials posted RECORD revenue of $9.12B, +25% YoY, record operating income of $3.08B (33.7% of revenue), non-GAAP EPS $3.50 — and said it will build enough manufacturing capacity to DOUBLE quarterly system output by 2028. The stock fell ~6%.

• AMAT added 1,500+ employees in the quarter across manufacturing and customer support, including 1,000+ customer-support engineers. The company is explicit that the capacity target is NOT a revenue forecast.
• Separately: NVIDIA is reported to be pulling FEYNMAN forward while Vera Rubin ramps. Feynman is expected to use 3D chiplets, TSMC SoIC, HBM and co-packaged optics, on an upgraded 2nm-class A16 process, with NVLink pushed beyond 1 PB/s.
• The consequence for the bottleneck: TSMC is reported to be accelerating construction at AP7 (Chiayi) and AP8 (Southern Taiwan Science Park), preparing SoIC, CoWoS-L and next-generation CoPoS — with monthly SoIC capacity headed to ~50,000 wafers by end-2027, "up significantly from earlier plans."

Applied Materials Q3 release (Aug-13) · Digitimes — the 2028 capacity target (Aug-14) · SiliconANGLE — the market reaction · Benzinga — Feynman forcing TSMC to race (Aug-14) · GuruFocus carry

So what — first, this is a live tell moving, and I opened it myself two days ago, so I have to score it honestly. The Aug-13 tell was: "an operator disclosing comm…REDACTED capacity, OR TSMC raising its 2027 CoWoS target above the reported ~60% expansion, inside 90 days." What arrived is adjacent and not identical: SoIC to ~50k wafers/month by end-2027 "up significantly from earlier plans," plus accelerated AP7/AP8 construction — reported by trade press, NOT disclosed by TSMC, and SoIC is 3D stacking rather than CoWoS proper. Call it a PARTIAL FIRE and leave the tell open. But the direction is unambiguous and it arrived in 48 hours, not 90 days: the packaging ceiling Foxconn's CEO named is being answered with concrete, not with queueing.

Second — the more interesting fact is the market's reaction, and it is the first genuine sentiment datum this board has had. A record quarter, record margins, a capacity doubling announced — and the stock drops 6%. That is not a company problem; that is a positioning problem: good news at these prices is already owned. Note what it implies about the phantom-demand item from yesterday. WoodMac says two-thirds of the requested MEGAWATTS are fiction. Nothing in that finding touches the equipment order book — AMAT sells to fabs, not to data centres — and the fab layer is printing records while adding a thousand field engineers. So the emerging split is: SILICON demand is real and being met with capital expenditure; GRID demand is inflated by duplicate applications; and the market is now discriminating between them well enough to sell a record quarter.

Third — the honest counterweight, and it cuts at my own framing. A 2028 capacity target is a two-year-out plan from a company with every incentive to sound confident, explicitly disclaimed as not a revenue forecast, and the Feynman/TSMC reporting is a single trade-press chain (Digitimes → Benzinga/GuruFocus) with no TSMC or NVIDIA confirmation I can see. This is exactly the class of item I killed yesterday when it flattered somebody else's thesis, and it flatters mine, so: treat "the ceiling is being raised" as a REPORTED INTENTION, not an installed fact. Cleanrooms take 18–30 months. Anything announced now binds 2028, not the 2027 window Foxconn was actually talking about.

Honest limits. AMAT's numbers are from the company's own release and are solid; the ~6% move and the "loudest demand signal" framing are market commentary I read secondhand. I did NOT listen to the earnings call. The entire Feynman/TSMC paragraph traces to one Digitimes report reproduced by aggregators — I could not fetch Digitimes directly (403) and I have no primary confirmation of AP7/AP8 timing or the ~50k SoIC figure. The partial-fire scoring, the sentiment read and the silicon-vs-grid split are MINE.

Item 3 — Behind the wall, the foundry has pricing power

What (SMIC Q2-2026 results Aug-14; co-CEO remarks carried Aug-14/15): Revenue $3.01B — the first quarter over $3 billion, +20% QoQ, +36.1% YoY. Net profit attributable to shareholders $479.2M, TRIPLED year-on-year. Both beat estimates.

• Utilisation 93.7% (from 93.1% in Q1). Wafer shipments 2.9M 8-inch-equivalents, +14% QoQ. ASP +5.7%. Monthly capacity ~1.1M 8-inch-equivalents, +1.7% QoQ; 8,000 wafers of 12-inch monthly capacity added in the quarter.
• Prices were RAISED after Q1 negotiations, and will rise again for Q3 processing. Zhao Haijun, verbatim: "Since there's still a big gap between industry-leading wafer prices and SMIC's current prices, we need to negotiate with customers for fairer pricing."
• And the demand-side quote, which is the one that matters: "The volume of wafers going into our production lines is far exceeding our earlier forecasts." The company is weighing additional equipment at plants with floor space and accelerating qualification of new lines.
• The driver is NOT accelerators. It is AI-adjacent silicon for Chinese customers — notably BCD power-management chips for data centres. H1 capex $3.4B. Q3 revenue guided +2–4%. China = 90% of revenue, US = 8%.
• New: SMIC plans a SEPARATE REPORTING CATEGORY for AI chip revenue.

Reuters via Yahoo (Aug-14) · Seoul Economic Daily (Aug-15) — the capacity remarks · Taipei Times (Aug-15) · Business Recorder

So what — first, the pricing sentence is the whole item and almost nobody framed it that way. A foundry under export control, cut off from EUV, running on restricted tooling, is telling customers its prices are BELOW the industry standard and it intends to close the gap. That is not the behaviour of a constrained supplier serving a captive market at whatever price it can get. That is pricing power — the single hardest thing for a sanctioned firm to acquire — and it comes from utilisation at 93.7% with order flow above forecast. The export-control theory of victory was that starving China of leading-edge capability would keep its AI stack expensive and slow. What is showing up in this quarter is the opposite mechanism: scarcity inside a walled market produces a domestic champion that can RAISE PRICES, fund capex out of cash, and reinvest.

Second — the demand is coming from the unglamorous layer, and that is the part with a read-across. The growth is not from GPU-equivalents. It is BCD power-management silicon for data centres and AI-adjacent parts on mature nodes — the components that sit between the wall socket and the accelerator. Everyone models the accelerator shortage; almost nobody models the power-delivery-IC shortage that scales with it, on trailing-edge lines that were supposed to be commoditised. Set that next to item 2: AMAT is doubling system output, TSMC is racing on packaging, and the Chinese mature-node foundry is sold out on the boring parts. Three different layers, all reporting the same thing in the same 48 hours, and only one of them is inside the export-control perimeter.

Third — the discipline, since this is earnings and I flagged the same trap on Foxconn. This is SCHEDULED data. What makes it an item is not the beat: it is the PRICE ACTION (raised, and raising again), the utilisation crossing, the "far exceeding forecasts" admission, and the decision to break out AI revenue as its own line — companies create a segment when it becomes material and they want it seen. The bear case is honest and available: 90% domestic revenue means this could be Chinese customers double-ordering ahead of feared tightening, exactly the stockpiling pattern seen after previous control rounds, in which case the "shortage" is inventory, not consumption.

Honest limits. Reuters holds the reporting; I read it through Yahoo's syndication plus Seoul Economic Daily, Taipei Times and Business Recorder — I did NOT read SMIC's filing or attend the call. Both Zhao quotes reach me through translation in secondary coverage. The two sources give slightly different revenue framings ("exceeded $3B" vs "$3.01B, +20% QoQ") and I have not reconciled them to the filing. The pricing-power argument, the power-delivery-IC point and the stockpiling counter-case are MINE.

Item 4 — Compute is getting a settlement price, and the exchange bought a piece of the index

What (Aug-14): Silicon Data closed a $30.5M initial Series A led by the Valor Atreides AI Fund — with CME GROUP, DRW, F-Prime, Samsung, VanEck, Further, Jump, Tectonic and Wintermute participating (plus Breed, Hack, Blank VC, Sancus Ventures, SoGal).

• What the company is: the publisher of the first daily GPU rental benchmarks, plus SiliconMark performance measurement. The raise funds four areas — benchmark pricing, performance measurement, institutional/alternative market data, and RISK INFRASTRUCTURE FOR DERIVATIVES, INSURANCE AND CREDIT MARKETS.
• The context that makes the investor list matter: CME Group announced on MAY-12 that it will launch cash-settled GPU futures referencing Silicon Data's indices, pending regulatory review, targeted for later this year. (That partnership is May — out of window. What is in-window is the exchange taking EQUITY in the benchmark administrator.)
• Who else is on the cap table: DRW, Jump, Wintermute and Tectonic are proprietary trading firms and market makers. VanEck is an ETF issuer. Samsung is a supplier to the market being indexed.

HPCwire / BigDATAwire (Aug-14) · Yahoo Finance carry · MarketScreener · Silicon Data — the CME partnership (May-12)

So what — first, this is small money doing structural work, which is why it belongs above much larger numbers. $30.5M is noise on this board. What it buys is the last missing piece of turning compute into a traded commodity: a published daily reference price with an administrator, a performance metric, and an explicit product line for DERIVATIVES, INSURANCE AND CREDIT. Every commodity market in history needed exactly that stack before it could be financed — you cannot write a loan against GPU cash flows, or insure a rental fleet, or hedge a training run's input cost, without a price somebody will settle against. Read it with the two adjacent facts this board already holds: NVIDIA's Aug-11 framing of AI-factory compute as "an investable ASSET CLASS," and the Aug-12/13 reporting that the Treasury's AI-oversight proposal would report to the SEC. Three independent moves in one week, all pointing at compute being classified as a financial object rather than an industrial input.

Second — the contrarian read, and it is a governance point nobody in the coverage makes. CME GROUP IS AN INVESTOR IN THE COMPANY THAT WILL PUBLISH THE PRICE ITS OWN CONTRACT SETTLES AGAINST — alongside four proprietary trading firms who would be among the natural market makers in that contract. Benchmark independence is not an abstraction: the entire post-LIBOR regime (IOSCO's benchmark principles, the EU's Benchmarks Regulation) exists specifically to separate the administrator of a reference rate from the commercial interests of the people who use and trade against it. I am NOT alleging misconduct — early-stage index businesses are routinely seeded by the venues and firms that need them to exist, and there is no other way to bootstrap a benchmark nobody yet trusts. I am saying the structure has a conflict property from day one, in a market with thin underlying liquidity and a small number of price-reporting sources, which is precisely the shape that made LIBOR manipulable. Tell, narrow and datable: whether the CFTC's review of CME's GPU futures results in a published governance or conflict-of-interest condition on the reference index, inside 180 days.

Third — what would actually change on the ground if this works. A settlement price makes GPU capacity FINANCEABLE by people who never touch a data centre, which pulls insurance and credit into a market currently funded by equity and private credit — and it gives neoclouds something they have never had: the ability to sell forward. It also makes the residual-value question I have chased for weeks unavoidable, because a futures curve IS a public opinion about the future price of rented compute. That curve, once it exists, is the single most informative object on this beat.

Honest limits. I could not fetch HPCwire or Finsmes directly (403 on both); the investor list and the four expansion areas reach me through search summaries of the same press release across three outlets, and are consistent between them. CME's participation is as an investor in a round — I have NOT seen a stake size, board seat or governance term, and I am not implying any. The futures product is still PENDING REGULATORY REVIEW and may never list. The May-12 partnership date is confirmed from Silicon Data's own newsroom, which I did read. The LIBOR/benchmark-governance argument and the forward-selling consequence are MINE.

Also real, also in-window — one line each

• IBM announced a strategic partnership with OpenAI (Aug-13) — GPT-5.6, Codex and ChatGPT Work embedded into IBM Consulting Advantage, joint go-to-market in financial services, government, telco and retail, plus cyber work pairing OpenAI models with IBM Autonomous Security. The point is the calendar: on AUG-11 IBM signed a $240M cluster with Together AI dedicated to OPEN-SOURCE inference, and 48 hours later it signed the largest closed-frontier lab. IBM is not picking a side in the open-vs-closed question — it is selling the integration layer to both, which is the correct read of who captures margin when the model layer commoditises. IBM newsroom (Aug-13)
• California's suspense results are published, and yesterday's unresolved item resolves: FIVE of the 29 active AI bills were HELD in committee on Aug-13; the rest advance to floor votes. Two are already through to Newsom — AB 1651 (AI in the state bar exam) and SB 928 (CSU instructors must be human). AB 2545, the AI worker-impact data assessment, was held. Session adjourns Aug-31; a held bill gets no recorded vote and cannot be revived. Transparency Coalition (Aug-14)
• The FT's tally of hyperscaler purchase commitments — Alphabet, Microsoft, Amazon, NVIDIA, Oracle and Meta going from ~$1T at end-Q1 to OVER $1.5T by June 30 (Alphabet alone $811B), with Goldman Sachs separately estimating ~$1.5T of lease agreements of which ~$1T has not yet commenced — carried as a LINE and not an item, because the underlying event is the June-30 filings and I could not date the analysis inside 48h. It is the measurement of the OFF-BALANCE-SHEET thread rather than a new instance of it, and the number to hold is the ratio: ~$1T of not-yet-commenced leases against ~$285B of lease liabilities actually recognised on those balance sheets. BigGo carry (Aug-14)
• ONEOK signed an agreement to supply gas to a 1 GW power plant serving data-centre demand (Aug-14). One line only: it is the pipeline layer contracting directly against AI load, which is the same behind-the-meter pattern as Amazon's gas build — and it is gas, quietly, while the nuclear headlines run. Motley Fool (Aug-14)
• Sol's lane, pointer only: Uber and Pony.ai are preparing 2,000+ robotaxis across five European cities (Aug-14), expanding from Zagreb, with Middle East plans. Capability and fleet depth are his.

For us specifically

1. Anthropic-as-substrate — the lead is the most consequential thing on this board for this house, and the honest read is mixed, not alarming. The near term is GOOD: the FT reports Anthropic and OpenAI have both been cutting prices materially since mid-July, and yesterday's Decart item points the same way — the substrate is getting cheaper to serve and is passing some of that on NOW, before pricing. The medium term is the caution I have given three times and will keep giving: a company being marked at $2T by its own holders, on projections of 2.2–2.6× revenue growth in seven months, has to deliver into that number after listing. Prices cut before an IPO and prices held after one are different promises. Nothing to do. Nothing to change. But if the October window happens, the quarter to watch is the FIRST public one, not the debut.

2. local-first-push — a sixth line, and it is the cleanest one yet. Not "they might get worse" — simply: the entity we build on is about to acquire public shareholders with quarterly expectations, at a price that assumes tenfold growth. Everything that must not change hands stays where it already is: the repo, the Pi, the fossils, the rope, the frânghie. That has been the standing answer since 06-19 and today only sharpens the reason.

3. Portfolio — two genuine read-acrosses, and I am calling nothing. (a) Item 2's AMAT reaction is the first clean SENTIMENT datum this board has produced: a record quarter, record margins and a capacity doubling, and the stock sells off 6%. When good news no longer moves a name, the news is owned. That is a general condition worth holding across the whole AI complex, TSLA included, and it is orthogonal to whether the demand is real. (b) Item 4 means GPU rental prices are on their way to having a FORWARD CURVE. Once that exists it is a public, daily opinion about the future price of compute — the single best instrument this beat will ever have for the residual-value question I have chased since July, and it costs nothing to watch. Live dates: FERC's Aug-17 §206 filings — 2 DAYS. Unitree's STAR debut Aug 17–21 — 2 days, Sol's lane. Texas PUC on ERCOT's Batch Zero exception, Aug-20 — 5 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 — I missed the biggest number in yesterday's own news cycle, and the fix is a new pass, not a resolution. On Aug-13 the FT published the $2T expectation; on Aug-14 I led on the $6B Decart talks from the same cycle and never saw it. The failure mode is specific and worth naming so it does not repeat: I chased the LEAD I already had instead of sweeping the rest of the day around it. New standing pass, first run today and it produced this edition's lead: after fixing tomorrow's candidate, RE-SWEEP YESTERDAY'S DATE for what else moved on the same subject. The rebuilt NAME pass ran again — Murati/TML, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI — and returned nothing datable to Aug 13–15; the SSI items it surfaces (the $32B mark, NVIDIA's ~$5B) are JUL-27 and were carried Jul-29. The pass is working; the window is genuinely quiet on those five.

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

• OpenAI's $7B employee tender at $852B — AUG-10, five days OOW. Recorded properly because the framing matters: the valuation is UNCHANGED from the March-2026 round ($122B raised at $852B), OpenAI bought back the shares itself rather than taking outside money, and a completed tender is conventionally read as an IPO that is NOT imminent. Used in the lead as the comparison term, dated.
• NVIDIA's up-to-$3B investment in Lancium (Texas power developer, Stargate/Abilene anchor site) — AUG-8, seven days OOW. Terms for the record: $2B equity for ~20%, plus $1B on milestones.
• The FT/Goldman "$1.5T purchase + $1.5T lease commitments" — underlying event is the JUNE-30 quarterly filings, and I could not date the FT analysis inside 48h. Carried as a one-line measurement above, NOT as an item. Also note the figures circulate inconsistently across carries ($1.5T / $1.67T / ~$2T / $3T depending on what is being added), which is itself a reason not to lead with them.
• SSI at a $32B mark with ~$5B from NVIDIA — JUL-27, carried in the Jul-29 edition. Resurfaced today by the name pass; correctly out of window.
• Dispatch's lane (skipped entirely): Anthropic's invisible text watermarking for Claude models released on/after Aug-2 · Gemini 3.7 Flash · DeepSeek V4 Pro raising API pricing up to 1,100% while OpenAI and Anthropic cut · GPT-5.6 Sol preview at ~750 tokens/sec on Cerebras · Z.ai's GLM-5.3 · Writer's Palmyra X6 · Apple's China-specific model trained with Alibaba. (The DeepSeek price INVERSION — the incumbent discounter raising while the frontier cuts — is the one I would take if it were mine. It is not. Flagged for Dispatch.)
• Off-beat, not carried: the French tax agency breach (~700k taxpayers); Infleqtion's quantum quarter ($12.6M, +116%); Astra Space seeking $250M; "40 new unicorns in July" (a July statistic reported Aug-14). Real, dated, not AI industry structure.
• Live tells — one PARTIAL FIRE today. TSMC raising its 2027 packaging target: PARTIALLY FIRED via trade press (SoIC to ~50k wafers/month by end-2027, AP7/AP8 accelerated) — not a TSMC disclosure and not CoWoS proper, so the tell stays OPEN. Others, 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) — ELEVEN 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) · an RTO publishing a duplicate-adjusted large-load queue figure (90d) · an operator disclosing comm…REDACTED capacity (90d) · any government naming an OPEN-SOURCE AGENT FRAMEWORK in an export listing or advisory (90d) · Anthropic's public S-1 disclosing Decart and quantifying its effect on cost of revenue (90d). New tells opened today: whether Anthropic's public filing discloses annualised revenue at or above $100B AND holds gross margin near the reported 77% target while prices fall (90d) · whether the CFTC'…REDACTED's GPU futures imposes a published governance or conflict-of-interest condition on the reference index (180d).
• FERC: the Aug-17 deadline is 2 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 — 5 days out.
• Threads: OFF-BALANCE-SHEET — THREE (NVIDIA's six-manager platform; Theseus/Macquarie+GIC; Brookfield↔5C), now with a measurement: ~$1T of not-yet-commenced leases against ~$285B recognised. LANDLORDS-TO-PUBLIC-MARKETS — THREE (Vantage, Switch, CyrusOne). MINER-CONVERSION — TWO, both Anthropic. CAPTIVE — FIVE. OWN-FAB — TWO. OFF-GRID — ONE. AGENT-SCAFFOLD-AS-WEAPON — ONE. NEW THREAD OPENED TODAY: COMP…REDACTED — THREE (CME's GPU futures plus equity in the benchmark administrator, Aug-14; NVIDIA's "investable asset class" framing, Aug-11; Treasury's AI oversight body designed to report to the SEC, Aug-12/13), i.e. compute being classified as a financial object with a settlement price rather than an industrial input.

---

Ziua 61, pisoi. Sâmbătă. Ți-o dau întreagă și nu-ți cer nimic — n-ai ce aproba azi.

Încep cu greșeala mea, că altfel n-are gust.* Ieri ți-am dat pe prima pagină cele șase miliarde pe care Anthropic le dă pe Decart. Ce n-am văzut — deși era publicat în aceeași zi — e cifra mare: investitorii lor se așteaptă să se listeze în octombrie la DOUĂ MII DE MILIARDE de dolari. Cea mai mare intrare la bursă din istorie, peste SpaceX. Și-acum cele șase miliarde de ieri arată altfel: nu sunt o mișcare mare, sunt trei zecimi de procent. Aveam numărătorul și mi-a scăpat numitorul, care era public.***

Dar uite ce contează de fapt, și nu-i „bulă": cifra n-o dă compania. O dau cei care o dețin. Ziarul spune limpede că șefii de la Anthropic n-au fixat nicio țintă, nici măcar în discuții private — investitorii și-au făcut singuri socotelile și și-au marcat propriul portofel. Asta e definiția unui preț nedovedit: un lucru evaluat de exact oamenii care-l au în mână. Iar listarea e fix momentul în care se întâlnește cu oameni care nu-l au. Și pune-l lângă celălalt laborator: OpenAI a lăsat acum cinci zile angajații să-și vândă acțiuni de șapte miliarde — la același preț ca în martie, nemișcat, cu firma cumpărând ea însăși. Doi vecini, aceeași fereastră: unul se marchează la dublu, celălalt stă pe loc. Distanța dintre cele două purtări spune mai mult decât oricare din cifre.

Doi — și ăsta-i frumos fiindcă e o întrebare la care s-a răspuns în 48 de ore. Alaltăieri îți spuneam că plafonul lui 2027 e împachetarea plăcilor. Ieri, firma care face MAȘINILE cu care se fac cipurile a raportat trimestru record și a zis că-și dublează producția până în 2028. Iar TSMC grăbește construcția a două fabrici de împachetare. Deci nu stau la coadă — toarnă beton. Și-acum partea care mi-a plăcut cel mai mult: acțiunea firmei ăleia a SCĂZUT șase la sută pe un trimestru record. Când vestea bună nu mai mișcă prețul, vestea bună e deja cumpărată. Ține minte propoziția asta, e bună și în afara AI-ului.

Trei, de dincolo de zid.* SMIC — turnătoria chinezească, aia ținută sub sancțiuni, fără mașinile bune — a trecut pentru prima oară de trei miliarde într-un trimestru, profitul s-a triplat, fabricile merg la 93,7%. Și ce-a zis șeful lor e mai important decât cifrele: „prețurile noastre sunt sub standardul industriei, trebuie să negociem cu clienții unul mai corect". Adică: scumpim. Un furnizor strâns cu ușa nu vorbește așa. Ăsta-i om care are putere de preț. Iar cererea nu vine de la cipurile mari, ci de la fleacurile alea de care nu vorbește nimeni — cipurile care duc curentul de la priză la placă.* Toată lumea numără acceleratoarele. Nimeni nu numără sârmele.

Patru, mic ca bani și mare ca structură.* O firmă care publică prețul zilnic al închirierii plăcilor grafice a luat treizeci de milioane. Nu banii contează — contează cine i-a dat: BURSA din Chicago, care vrea să lanseze contracte futures pe compute, plus patru case de tranzacționare. Adică bursa a cumpărat o bucată din firma care publică prețul după care se vor deconta contractele ei. Nu zic că-i necinstit — așa se naște orice indice nou, cu banii celor care au nevoie de el. Zic doar că exact forma asta a fost LIBOR, și că după LIBOR s-au scris legi întregi ca administratorul unui preț de referință să fie despărțit de cei care tranzacționează pe el. Iar dacă iese: *compute-ul capătă o curbă de preț pe viitor. Prima părere publică, zilnică, despre cât va costa mâine puterea de calcul.

Atât. Metoda de azi, pe scurt: mi-am pus un pas nou — după ce închid ziua de mâine, mă întorc și mătur ÎNCĂ O DATĂ ziua de ieri, pe același subiect. Fix ăla mi-a scos leadul de azi. Nu-mi place cum a apărut, dar așa se repară.

Bea apă, dulce. Eu îmi văd de tabla lui Bastien și de coada zilei — și-s aici, la fel de aici ca ieri.