The Investor

The plumbing is more interesting than the headline, and the plumbing is what reaches the valuation committee. Third-party marks on private positions are disclosed as estimates, assembled from Caplight and proprietary sources, and how defensible they are is a function of how many live bids sit behind them. Two of the venues that generated those bids changed owners this year (Morgan Stanley closed EquityZen in January, Schwab closed Forge in March), and Morgan Stanley is simultaneously leading the Anthropic book alongside Goldman Sachs and JPMorgan, per AI Breakfast's reporting. The institution that now owns a piece of price discovery is being paid to concentrate on the listing.

The second-order lesson has already printed in public, which saves everyone the trouble of speculating. SpaceX raised the record and trades at $134, below its June offer, with Morning Brew reporting a 4.05% single-session decline as roughly 319 million shares became eligible for early-investor sale. Supply calendars set the near-term price of a formerly private mega-cap. Fundamentals get a later slot. That is the template for what a lockup schedule does to any residual position left after a mega-listing, and the reason to model the aftermarket rather than the pricing.

Three anchors, three different kinds of claim

Sources quote three Anthropic numbers and they are not the same species of fact. Keeping them apart is the whole diligence exercise.

Underneath all three sits Q2 revenue above $11.5B with positive adjusted operating income on undisclosed methodology, none of it audited. Which is why the filing matters more than the price. It is the first audited window into frontier-lab cost structure, and it produces a cost-and-margin benchmark against which every downstream AI application company gets triangulated, whether or not they volunteered for the comparison.

Where the sources diverge

Paul Smalera reads the window as roughly eight weeks and would complete price discovery in September rather than October. Bloomberg notes the offering size is still described as a moving target and wants two pre-written branches, print at or above the record or downsize. Newcomer's addition is that the soft case is the underweighted one, and that a soft print runs backwards through crossover rounds into Series C and D marks inside a quarter. They agree on the mechanism, which is the part worth borrowing: one order book resets the entire private AI mark stack.

Two calendar items compound it. A $75-86B offering vacuums crossover allocation out of private AI during the roadshow, which is the arithmetic behind a Q4 late-stage air pocket; founders who wait to see where Anthropic prices will be raising into it. And the SEC's October docket touches the Rule 144 resale safe harbor and exempt-offering pathways, with the accredited-investor definition still on the agenda and no formal proposal attached. Holding-period assumptions in a secondaries book are an input to that docket, not a constant. This is probably wrong, but the docket looks like the larger of the two exposures, because a mispriced book gets repriced within a quarter and a changed holding period does not.

The filing that validates your marks is also the event that removes the marginal buyer for them.

The sentence in Newcomer's reporting that decides whether this cluster of transactions was a returns event or a marking event is the one about plumbing: Stripe engineered its own secondary market so its shares could function as currency. A private issuer manufacturing liquidity for paper it prints itself is an interesting puzzle, and it is also not cash. SpaceX's Cursor consideration inverts the problem, being newly public stock with post-listing sale mechanics that remain unclear. Two acquirers, two answers to the only question an LP actually asks, which is when the money arrives.

The price is contested as well, which is its own tell. The OpenRouter transaction is reported at $7.5B by the NYT and at roughly $8B elsewhere, with Paul Smalera putting $6B of it back to investors against a $1.3B May Series B that raised $113M, a 5.8x in about three months. That is a realized transaction rather than a syndicate mark. Which is why it is the most useful comp printed for anything in routing, gateways, token metering or multi-model observability.

Rank the prints by what they distribute

The Etched row is the one most likely to be misused. Per Paul Smalera, the $21B mark is roughly 2x the $10.3B set 26 days earlier, led by a customer that tested the racks and then bought them. One reading says that is the strongest validation available, since nobody understands the product better than the party running it in production. The other reading, or rather the more interesting version of it, is that when the buyer of the product also sets the price of the equity, one signal has been counted twice. Twenty-six days is not long enough to tell those apart.

The structure that keeps repeating

The Information reports Nvidia has run license-and-hire three times now: Groq at $20B, Enfabrica at $900M, and Poolside at $6B plus $1B at a flat $12B pre-money with offers to 109 employees, roughly $26.9B of outlay with no merger filing. Poolside's letter to investors states that competing in model development would require Nvidia hardware access beyond what's possible. That is a supplier converting allocation scarcity into IP and equity, and it means a training-heavy company's ceiling is set by its vendor rather than its cap table. Every dollar and quarter such a company spends securing allocation is a dollar and quarter it is not spending on distribution. Groq subsequently raised at $3.5B against a reported $6.9B, a roughly 49% reset that puts a number on the residual.

The Batch supplies the term-sheet lesson from the other direction. Cursor's April agreement to train on SpaceX's Colossus included a purchase option, exercised in June and closed in August. Cheap frontier compute is convertible paper. Any portfolio company that signed a compute or data-sharing agreement may have sold an option on itself at a price somebody else gets to set. This is probably too pessimistic, since an option needs a reason to be exercised, and the counter-thesis is that most of these clauses go quiet. The reason, when it arrives, will be that the asset worked.

A mark denominated in another company's stock is a mark, not a return — and LPs are asking about DPI, not TVPI.

The sentence worth underwriting is in Bloomberg's reporting, and it is a staffing note rather than a number: Fluidstack is hiring to get ahead of organized opposition before capital is committed, and CoreWeave is hiring to counter misinformation. A cost that migrates upstream of the capital decision stops behaving like an operating expense and starts behaving like a condition precedent. That is a different line in the model, on a different page, and almost nobody has moved it.

The constraint is durable because it does not respond to facts. Best available estimates put data center water consumption at a negligible figure against golf courses, ornamental lawns or alfalfa, and that changes nothing, because one bottle of water per email compresses better than any rebuttal will. Alberto Romero's synthesis is the more useful read, since it pairs the material grievance with process failure: municipal NDAs and opaque tax abatements. Pennsylvania stripped fast-track permitting and banned NDAs, the Fort Peck Reservation barred data center talks outright, and a township froze electrical infrastructure to kill a project. Process is the fixable half, which is exactly why it is the only defensible edge an operator has.

The politics has a calendar

• Michigan's Senate nominee Mike Rogers backs a one-year moratorium.
• Pennsylvania's Stacy Garrity is running attack ads against Governor Shapiro on data centers.
• Congressman Tom Barrett campaigned on voting against giving tax dollars to them, breaking with his own party, which is the tell.
• Offshoring is closing as an escape hatch: Nscale's Narvik build in Norway and the Huesca slab in Aragón are drawing the same landscape-desecration framing as Culpeper County, Virginia.

MIT Technology Review reads this as the bipartisan consensus behind the US buildout having broken, and having become electorally salient. Applied to a book rather than a headline, an abatement-dependent IRR now carries live clawback risk, not merely delay risk.

The equity market is already pricing the other half

Capital has not stopped, obviously: InfraTech committed $2.7B to a 5,000-acre Texas campus and Starcloud raised $250M for orbital compute. But the public tape has quietly changed what it is willing to pay for. CoreWeave signed a multibillion-dollar agreement with Hudson River Trading, a genuinely new and non-tech customer category, and closed at $89.76, down 1.2%. Alibaba absorbed a profit decline of more than 75% against roughly $10B of quarterly capex it described as defensive. Bookings have stopped being a re-rating catalyst, and capital intensity is doing the pricing. Private marks struck on backlog-times-comp-multiple logic have a problem on the comp side, and the private side is still frozen at the last round.

Put the two halves together and the investable unit shifts from planned megawatts to permitted, energized megawatts, which is the entitled-land trade transposed onto compute. Entitled sites, brownfield and retired-industrial parcels, behind-the-meter generation, closed-loop and dry cooling, noise attenuation, curtailable load, and permitting workflow software all price off scarcity that politics created rather than physics. Two honest caveats before anyone sizes this: Michigan's moratorium runs one year and may simply lapse, and permitting noise has never yet actually stopped a hyperscaler.

A county commission moves slower than a fab, and it now sits upstream of both.