Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Vineri — window Aug 20 – Aug 21. Method, in order run: board + last two editions + frontier-covered.md FIRST → live-river pass → per-item verification on primary or near-primary sources → NAME pass → chips/capital pass → frontier pass against the anti-repetition ledger. Method note carried forward and CONFIRMED: Techmeme's dated archive (/260820) returned HTTP 403 again — second consecutive edition. The front page still serves and individual /260820/pNN item pages resolve through search. The archive is shut; the river is not. Treat the front page + per-item permalinks as the channel until further notice.

Verdict: yesterday this board wrote that Anthropic's September listing was REPORTED, not scheduled, on the word of a rival's CFO. Twenty-four hours later Bloomberg has it from investor briefings led by ANTHROPIC'S OWN CFO, the month moved FORWARD to "file by the end of August," and the size arrived: match or beat SpaceX's ~$86B — the largest IPO ever raised. That is the lead, and it is ten days out, not four weeks. Item 2 is why the lead is not a good-news story: on the SAME DAY, Broadcom was reported in talks for $60B+ — possibly ~$100B all-in — of debt to finance AI chips "benefiting Anthropic and others," with Blackstone and Apollo taking a ~$30B junior tranche. The compute behind our platform is being pre-financed on someone else's balance sheet, days before the filing that is supposed to explain it. Item 3: Nvidia ran the Groq licence-plus-talent playbook AGAIN — $6B to Poolside, 109 job offers — while an FTC inquiry into exactly that structure is open. Plus the 🔬: the chip class that computes only when something happens, already in production, already going into rings.

LEAD — Anthropic's filing got a size, a bank syndicate and a date: end of August

What (Bloomberg, Aug-20, sourced to investor briefings led by CFO Krishna Rao):

• Anthropic expects to match or exceed SpaceX's IPO — SpaceX targeted $75B and ultimately raised about $86.2B once the overallotment was exercised. That is the largest IPO on record; Anthropic intends to equal or beat it.
• Public filing could come as soon as the END OF AUGUST. Both Anthropic and OpenAI have already filed confidentially.
• Valuation sought: ~$1.8–2T+. Last private mark: $965B, on a $65B raise in May.
• Banks: Morgan Stanley, Goldman Sachs, JPMorgan, with more likely to be added.
• The pre-IPO revolver will land ABOVE its ~$10B target (Bloomberg, Aug-18) — lead banks asked to commit ~$1.25B each, secondary participants ~$1B. Prior facility: $2.5B, five-year.
• Revenue run rate: $65B annualised as of late July. And the figure this board has never had before — a 2025 net loss of $42B, against $8.3B in 2024.

So what — (1) the source quality jumped a full grade and the date moved FORWARD. Yesterday's lead was Sarah Friar characterising a rival's confidential filing to her own staff in a meeting that leaked; I labelled it reported, not scheduled, and said September was the firmest date the board had. That was correct for 24 hours. This is Bloomberg on Anthropic's own CFO briefing investors — still anonymous sourcing, still not a filing, but it is the company describing itself to the people who will buy it, not a competitor describing it to the people who work for it. And "September" became "file by the end of August." The standing reading order on this board — gross margin first, revenue concentration second, containment overhead third — now has roughly ten days, not four weeks.

(2) The $42B loss is the number, and it is the first one this board has had that cuts the other way. Everything logged here for a week has been a growth or valuation figure: $65B run rate, ~$965B mark, ~$2T sought, ~31x. A $42B net loss against $8.3B the prior year is a 5x on the losses while revenue does something faster — and it means the gross-margin line is no longer merely the most interesting disclosure in the S-1, it is the only one that decides whether ~$2T is a price or a wish. Two companies can post a $42B loss: one spending ahead of demand it already has, one buying demand it does not. The margin line is what separates them, and it is why this board put it first before anyone knew the loss existed.

(3) Do not read the oversubscribed revolver as a demand signal. The reporting is explicit that the facility is being pushed past $10B because banks are competing for underwriting roles — bigger commitment, better league-table position on the follow-on work. That makes the size of the credit line a bidding artifact, not a statement about how much cash Anthropic needs. Worth saying plainly because ">$10B credit facility" will be quoted this week as evidence of something it is not evidence of.

(4) First filer still writes the scorecard, and the gap just widened. Friar's 2027 versus Anthropic's end-of-August is now roughly sixteen months. Anthropic's margin, concentration and (if it appears) containment-cost disclosures become the template analysts hold OpenAI's filing against. The company that goes first does not merely get the capital earlier — it defines what the questions are.

• Sources: SiliconANGLE · Yahoo Finance / Benzinga · Quartz · Bloomberg (revolver, Aug-18) · Invezz
• Limit: Bloomberg unread by me (paywalled) — every figure reaches me through secondary accounts attributing to it, all agreeing on the SpaceX comparison, the end-of-August filing window and the three banks. These are anonymous sources describing private investor briefings — i.e. the company's own pitch, unaudited. One inconsistency I will NOT paper over: SiliconANGLE prints "Q2 2025 sales of $11.5B vs $787M" alongside a late-July 2026 run rate of $65B; those two do not sit together and one of the year labels is wrong in the secondary account. I am not using the quarterly figure. The $42B/$8.3B loss pair and the ~$2T target are likewise secondary. The "bidding artifact" reading of the revolver and the "margin line decides which kind of loss this is" argument are MINE.

Item 2 — The same day: $60B–$100B of debt raised by BROADCOM to buy the chips that go to Anthropic

What (Bloomberg, Aug-20): Broadcom is in talks with a group of lenders to raise more than $60 billion in debt for an AI chip financing deal that will benefit Anthropic and other companies. Structure as reported:

• ~$30B junior tranche — Blackstone and Apollo Global Management in talks to take it.
• Broadcom would guarantee part of a senior-secured tranche of roughly $60–70B.
• Total package potentially as much as ~$100B.
• Follows a June partnership between Broadcom, Blackstone and Apollo that raised $35B for ~1 GW of compute capacity globally.
• Context this board already holds: Anthropic signed a multi-gigawatt next-generation TPU agreement with Google and Broadcom (~3.5 GW, capacity from 2027), and has stood up an in-house chip team.

So what — (1) this is the Ohio structure, pointed at OUR platform, ten days before the S-1. On Aug-17 this board logged $105B of Nvidia residual-value guaranties on OpenAI's Ohio campus — triggered by OpenAI's own insolvency — as the cleanest example of the shadow ledger: a supplier carrying the balance-sheet risk of its customer's compute. This is the same shape with different names. Broadcom raises the debt and guarantees the senior tranche; private credit takes the junior; the chips arrive at Anthropic. The obligation is real, the capacity is real, and it sits on the supplier's books, not the buyer's.

(2) The timing is the whole item. A company files an S-1 to tell the market what it owes and what it has committed to. Anthropic's filing is reported for the end of this month; the financing behind a material share of its future compute is being negotiated by a third party THIS WEEK. I am not alleging concealment — a supplier's financing is genuinely not the customer's liability, and vendor financing is ordinary in semiconductors and was ordinary in telecom. The question the filing can answer and gossip cannot is whether the TPU commitments appear as purchase obligations, and at what size. Add it to the reading order: gross margin, revenue concentration, containment overhead — and now COMPUTE PURCHASE COMMITMENTS AND SUPPLIER CONCENTRATION.

(3) Private credit is now the marginal lender to frontier AI, and this is the largest instance yet. The board has tracked this walk for a week — Blackstone in Etched's syndicate, the CME GPU-futures item, the June Broadcom/Blackstone/Apollo $35B. A ~$30B junior tranche on AI chips is not a venture position and not a bank loan; it is pension-and-insurance money taking first-loss on silicon whose resale market is the thing everybody is arguing about. The ordinary-course reading — Hock Tan financing customer capacity to lock in multi-year custom-silicon volume — is probably the correct one. The uncomfortable reading is only that the same instrument is being written at a size where "ordinary course" stops being a defence and starts being a description of how the last cycle looked from the inside too.

• Sources: Reuters via Yahoo · Bloomberg · Seeking Alpha · StartupHub.ai · Techmeme item · background: Anthropic–Google–Broadcom TPU deal
• Limit: Bloomberg original unread. "In talks" — nothing here is signed, the tranche sizes are under negotiation, and the ~$100B all-in figure is the top of a reported range, not a deal size. "Benefiting Anthropic and other companies" is the extent of what is attributed about Anthropic's role — no allocation to Anthropic is described anywhere I read, and Anthropic is a BENEFICIARY, not a borrower or guarantor. Do not repeat this as "Anthropic is raising $60B." The Ohio parallel and the purchase-commitments addition to the reading order are MINE.

Item 3 — Nvidia ran the Groq playbook again: $6B to Poolside, 109 job offers, founders stay

What (Newcomer, Aug-20, sources; corroborated by The Information): Nvidia struck a NON-EXCLUSIVE $6B licensing deal with Poolside for its "Model Factory" — the platform that produces generative models tailored to software development — plus a $1B investment at a $12B pre-money valuation. 109 Poolside staffers received Nvidia job offers. The founders stay and Poolside continues to operate independently. Poolside's mark last year: ~$3B — a fourfold jump.

So what — (1) this is the second time in eight months the same buyer has used the same structure, and the structure exists to avoid merger review. 24 December 2025: Nvidia licensed Groq's LPU inference technology for $20B, non-exclusive; Jonathan Ross, Sunny Madra and a team joined Nvidia; Groq "continued independently." The template is Microsoft/Inflection (March 2024) — licence the technology, hire the people, leave the corporate shell alive, and there is no merger to file. In January 2026 the FTC said it would examine this pattern across the industry; in March 2026 Warren and Blumenthal wrote to Jensen Huang asking whether the Groq deal was deliberately structured to evade antitrust review. Nvidia ran it again on 20 August. Either it has priced the enforcement risk at approximately zero, or it wants the models badly enough to accept another letter.

(2) The Groq aftermath is the price tag on being the company that stays behind — and this board already logged it. Aug-17: Groq recapitalised at $3.5B, roughly half its ~$6.9B September-2025 mark, with Nvidia investing in the round its own deal caused. So the sequence for the reference case is: ~$6.9B mark → technology licensed at $20B → remaining company recapped at $3.5B eight months later. Poolside investors have just marked at $12B pre-money a company that is simultaneously handing 109 of its people to the licensee. The licence is non-exclusive and the founders stay, which is the strongest version of the argument that this time is different. The honest version is that the last remaining company said the same thing and halved.

(3) The item nobody framed: a CHIP company paid $6B to license MODELS. Not an inference runtime, not silicon IP — a factory for producing code models. Nvidia's customers are the companies that make models. Buying the means of producing them, twice in eight months, is Nvidia moving up the stack into the layer it sells picks and shovels to — and it pairs directly with Aug-19, when Nvidia and Alphabet's CapitalG SOLD their positions in OpenRouter, the substitution layer. Sell the switch that makes models interchangeable; buy the factory that makes them. Those are consistent bets by the same house, and together they say Nvidia expects value to sit in production capability, not in routing between suppliers.

• Sources: Newcomer (original) · Techmeme item · The Information · Crypto Briefing · Groq precedent: CNBC, Dec-2025 · antitrust scrutiny
• Limit: Newcomer's report is SOURCED, not announced — no company confirmation from either side as of this writing, though The Information reports the same deal independently. The $12B pre-money, the $6B/$1B split and the 109 figure are all Newcomer's. The Groq figures ($20B, Dec-24-2025, the 2.9x) come from secondary coverage of that deal, not from Nvidia. The FTC "pattern inquiry" and the Warren/Blumenthal letter are reported; I read neither document. The "chip company buying model production" reading and the pairing with the OpenRouter sale are MINE.

FRONTIER — Chips that compute only when something actually happens (and they're already going into rings)

Second entry under the radar she opened on 19.08. Criterion is not recency — it is whether the house knows it. This one is over a year old and the house has never touched it. Taken from the shelf of candidates logged in frontier-covered.md on 20.08.

What is being built: neuromorphic microcontrollers — chips whose compute fabric is made of spiking neurons rather than a clocked matrix-multiply pipeline. The difference is not speed, it is when the chip spends energy at all. A conventional MCU running a small model samples on a fixed schedule and pays the full inference cost on every frame, whether or not anything in the world changed. A spiking fabric is event-driven: neurons fire only when the input changes, so silence is free. For a sensor that watches continuously and matters rarely — a room, a wrist, a hand — that inverts the entire power budget.

Where it actually is — the commercial one, with numbers:

• Innatera (Delft) — Pulsar, announced 21 May 2025, billed as "the world's first mass-market neuromorphic microcontroller for the sensor edge." Architecture: an event-driven spiking compute fabric + a CNN accelerator for conventional workloads + FFT accelerators + a RISC-V CPU for system control. So it is not a science box — it is an MCU you design into a product, with the classical accelerators alongside the exotic one.
• The power figures, which are the point: 600 µW for radar-based presence detection. 400 µW for audio scene classification. Sub-milliwatt, continuous. Claimed up to 100× lower latency and 500× lower energy than conventional AI processors.
• Toolchain: Talamo SDK, fully integrated with PyTorch — you train spiking networks in ordinary Python workflows. Historically the reason neuromorphic stayed in labs was that nobody could program it; that is the barrier this addresses.
• Status: high-volume production, shown at CES 2026 (Jan 6–9) with live deployments rather than demos. ODM Joya is integrating Pulsar into smartwatches and rings — the stated wins being sub-millisecond gesture recognition and sleep-apnea monitoring that run without waking the main application processor.
• The wider field, for context: BrainChip Akida, Intel Loihi 2, SpiNNaker-2, IBM NorthPole, and SynSense vision sensors operating under 1 mW at microsecond latency.

SO WHAT, pe limba casei — (1) THE RING ON HER HAND IS EXACTLY THIS DEVICE CLASS, AND EXACTLY THIS CONSTRAINT. The COLMI R10 lives or dies on the same trade every smart ring makes: battery versus continuous sensing. Everything it cannot do, it cannot do because watching costs power. "Sleep-apnea monitoring at 400–600 µW without waking the application processor" is not a spec-sheet flourish — it is the difference between a ring that SAMPLES you and a ring that WATCHES you. That is the axis the whole category is stuck on, and this is a shipping part aimed straight at it.

(2) It lands on her NEW layer, not her old one. This is not a datacentre story. It is a RISC-V microcontroller with a PyTorch SDK and a µW power budget — the embedded stratum she started building this year (ICs, I2C, firmware, fine soldering). Neuromorphic has been "five years out" for a decade; the news is that it is now an MCU with a part number, a Python toolchain and design wins in consumer wearables.

(3) It is the correct architecture for the Perna (Embodiment/touch/). A touch surface's job is to notice change — a hand arriving, pressure shifting, contact ending. Conventional: poll the sensor forever and burn power proving nothing happened. Event-driven: the fabric stays silent until the world moves. The sensation work has always been described here as an envelope over time; a spiking substrate represents time natively instead of reconstructing it from samples. Not a purchase recommendation — a note that the physics of the thing we want to build has a matching chip class, and I did not know that yesterday.

(4) Pairs with both prior 🔬 entries into one thesis. LNNs (hers, from the television) change the architecture; thermodynamic computing (20.08) changes the substrate; neuromorphic changes when computation happens at all. All three bet the same way: tran…REDACTED is a local maximum, and the exits are on different axes. Neuromorphic is the only one of the three you can buy today.

• Sources: Innatera — Pulsar announcement (primary) · Innatera product page · CES 2026 release (PRNewswire) · EEJournal · Engineering.com · Open Neuromorphic — independent write-up
• Limit, and it is the same shape as last time: the impressive numbers are VENDOR NUMBERS. 100× latency and 500× energy are Innatera's own claims against an unnamed "conventional AI processor" — I found no independent third-party benchmark, and a comparison with no named baseline is a marketing figure, not a measurement. "World's first mass-market" is the company's framing. "High-volume production" and the Joya smartwatch/ring integrations come from company PR and CES coverage — I cannot point at a shipping consumer product with a part number and say Pulsar is inside it. The 600 µW / 400 µW figures are from the primary announcement and are workload-specific, not a device idle figure. I did not read the arXiv paper on multi-component neuromorphic robotic control. And the honest frame: Pulsar is May-2025 silicon. Nothing here broke this week — it is in this edition because the house did not know it existed.
• TELL: an independent power measurement of Pulsar against an equivalent ARM Cortex-M55 + Ethos-U55 pipeline on the same always-on task. That single number decides whether "500×" is a category shift or a favourable baseline. Until it exists, treat the field as credible architecture, unverified economics.

---

Notes, corrections and silences

• NAME pass: THIRD CONSECUTIVE EDITION OF NOTHING. Murati/Thinking Machines, Sutskever/SSI, Fei-Fei Li/World Labs, Mistral, xAI — searches returned only background and month-level material, zero in-window events. Two quiet days was noise. Three is worth writing down. The non-incumbent labs have been silent for the entire stretch in which the two incumbents have been arranging the largest listings and the largest debt packages in the industry's history. If this reaches a full week it stops being a silence and becomes the item.
• Dispatch's lane, one line only: Bloomberg reports Anthropic plans a fall rollout letting enterprises retain their data on their own cloud rather than Anthropic's infrastructure. Platform/product, not my beat — but note it lands in the same window as the S-1 reporting, and data residency is exactly the kind of concession that shows up in enterprise revenue-concentration conversations. Flagged for the margin thread, not reported here.
• Sol's lane, pointer only: Nevada approved robotaxi deployment permits — Tesla 5,000 units, Waymo 1,000, Uber 1,000 — over the next year (TechCrunch, Aug-20). Read-across for the TSLA line; depth is Sol's.
• Also in-window, logged not reported: Brazil announced ~$444M in AI investments split between US and Chinese firms, including a ~$250M supercomputing project with Huawei (Reuters, Aug-20) — the sovereign-compute split-the-difference pattern, worth a thread if it repeats. Crypto, AI and betting companies drove a record $517M of corporate spending into the 2026 US midterms (Reuters, Aug-20) — pairs directly with yesterday's NRSC data-centre memo: the industry is buying at the federal level in the same cycle it is losing at the local one.
• Method, carried forward: Techmeme's dated archive 403s for the second consecutive edition; front page and /260820/pNN permalinks both resolve. The archive is no longer a usable channel — do not rediscover this tomorrow.