The Investor

Run the SpaceX arithmetic slowly

Andrew Ryan, who runs venture fund share distribution at Alex. Brown (Raymond James) and therefore sits downstream of essentially every tender and listing, supplies the cleanest mechanism study available. SpaceX's last reported private mark was roughly $800B, from the December 2025 insider sale. The June 12 IPO at $135 implied about $1.75T. At the early-August low close of $108.27, implied valuation was still above $1.4T. The private marks held; the IPO premium on top of them is what went away.

The reason is participation, not business quality. Company-run tenders exclude short sellers, index rebalancers and margin-called holders, and the company typically sets timing, approves buyers and influences price. Every mark derived from that mechanism inherits the same one-way bias. Ryan's field diagnostic on the unlisted backlog is the line to keep: "They're carrying themselves at eight billion and they don't want to go public at four billion."

Where the three stories converge

Anthropic is the live version of mechanism one. Fortune's Term Sheet records secondaries clearing at $1.3-1.4T against a $25M floor ticket, generally $50M+, with Caplight fielding roughly $1.5B of interest since the start of Q2 2026 and Anthropic deliberately narrowing who may join the cap table after a May demand surge. Underline Capital's Clara Vydyanath is on the record that Anthropic "feels overpriced" while OpenAI "feels underpriced." A three-to-five-times demand ratio measures interest, not willingness to pay at $1.4T.

Newcomer's leaked India Quotient fund documents show the other end of the same problem. A roughly $1M seed check into ShareChat in 2015 was marked at 77x by March 2022 — about $77M gross against a $20M Fund II, or nearly four times the whole fund on one line. Ten years after entry the firm has realized nothing beyond small secondaries, ShareChat's fair value is described in the documents as uncertain, and the firm took an across-the-board write-down. A mark that never converts is an option on someone else's listing window.

The power law is a distribution of exits, not a distribution of marks.

The trigger nobody has calendared

Ryan's second point is the one that transmits across your whole book: as real-time private pricing from providers like Caplight becomes observable, private positions move "from a level three asset to a level two asset pretty quickly." If observable secondary prints become the valuation input, the markdown trigger stops being a down round and becomes the auditor — arriving on an audit schedule, hitting the portfolio simultaneously, and removing the discretion that has smoothed venture NAV for a decade. It transmits further because NAV facilities, continuation-vehicle pricing and LP co-invest decisions all reference marks that just became contestable.

Two caveats belong in the memo. Every SpaceX private figure here comes from public reporting rather than filings, and the publisher repeatedly flags such indications as potentially unaudited or incomplete while also holding an Anthropic position and operating a competing secondary platform. One company in one summer is a mechanism, not a comp set. But the mechanism is what you can act on before December.

Read the argument, not the headline

The Justice Department's September 2 filing in the consolidated New York Times case does something more interesting than pick a side: it splits copying from output. Entire works go in, nothing comes out published, and outputs "often if not always lack substantial similarity," so market harm cannot be assumed from training alone. The filing also waves off the Copyright Office's contrary report as carrying no binding legal authority, a report whose author, Register of Copyrights Shira Perlmutter, was fired shortly after publication and is now challenging her dismissal. Techpresso notes the administration paired this with pressure at the G20 to permit training on creators' work. The executive branch volunteered a position in a case where it is not a party.

Consequences land on a specific cohort. Rights marketplaces, publisher-data brokers and clearance-infrastructure companies raised over the last 24 months on one implicit premise: labs will be forced to pay. If fair use holds, licensing becomes a reputational and quality purchase rather than a compliance necessity, attach rates fall, and discretionary TAMs clear at multiples that look nothing like compulsory ones. Publisher plaintiffs lost settlement leverage in the same motion, which compresses the observable comps for every licensing deal signed from here.

The channel that fair use does not touch

Same day, Jason Isbell, David Lowery, Guy Forsyth and Eduardo Calle sued Suno in federal court in Massachusetts, and not for reproducing their songs. The claim is misappropriation of identity and style, which runs orthogonal to fair use. A total developer victory on copyright resolves none of it.

The third row compounds. MIT Technology Review's reporting on the ad-tech ruling carries Forrester's Nikhil Lai: "I think this will be the end of an era of the US trying to break up platforms this complex." Two of the largest non-compute liabilities on the platform and AI stack got cheaper inside the same window. Anything underwritten on those overhangs persisting is now positioned against stated federal policy.

An AI legal-risk model with one variable is underwriting the wrong half of the market.

Where the sources disagree

Pivot 5 takes the counter-thesis seriously, and it deserves that. This is executive-branch advocacy rather than settled law, politically compromised by the Perlmutter firing, and dozens of cases remain live against OpenAI, Microsoft, Anthropic and Meta. Fair use could hold broadly and kill the compulsory-licensing thesis outright; a court rejection or an administration change reinstates the prior regime; or copyright resolves for developers while identity claims keep the liability alive. The third is probably where the money sits, which argues for repricing rather than exiting the licensing cohort. Voluntary licensing for provenance, quality and enterprise indemnification survives, and EU exposure plus procurement preference is real residual optionality.

The policy record cuts the same way on tail risks. California AB 412, which would have required copyright training-data disclosure on rightsholder request, died in session, deferred rather than dead, since the framework can be reintroduced. And Anthropic won summary judgment against the Department of War, with the court calling the supply-chain designation "illegal and baseless." Federal exclusion is litigable now. That lowers the discount on defense and federal revenue in govtech AI models.

The durable wedge is the one nobody funded at scale: companies selling consent and identity clearance rather than training rights. This may be wrong on timing, but artists' deals signed early look like a defensible counter-position sellable to labels and enterprises. Everyone else in generative media is carrying an unpriced liability.

The reference price is the asset

The CFTC has opened a comment file on the Listing of Compute Derivatives Contracts, comments closing October 20, 2026, framed around transparency, liquidity and responsible innovation. Read past the civics and this is the point at which compute stops being a procurement line and starts being a hedgeable commodity with a published benchmark. In commodity markets the economics have a habit of settling on whoever owns the reference price rather than on the physical seller, which means index and benchmark providers, exchanges, brokers, risk analytics.

The second-order effect is the one that reaches an infrastructure book. Forward compute contracts that can be hedged make neocloud revenue financeable against a forward curve instead of an announced backlog, and that changes the cost of capital for the entire GPU-fleet asset class. This is probably wrong on timing, or rather the timing is the only part worth arguing about, because the funding structure for these companies is being redesigned in a docket that closes in weeks.

Assurance becomes a licensed profession

Third-party AI audit stopped being voluntary this session in three separate jurisdictions, each with a hard statutory date attached:

Demand becomes compulsory and supply at scale does not exist. The moat is a registration you can only win by being early, which is a strange sort of moat, and a durable one for as long as the dates hold. The honest caveat is that the sponsor of this policy brief is also publicly building the federal preemption case; the August 24 McKinley and Perault piece argues state AI laws diverge across states and shift within them. If you back a compliance company whose value proposition is navigating a 50-state patchwork, you are implicitly long the patchwork surviving. The more interesting version of the trade is assets whose demand persists under harmonization: evaluation, provenance, security, age assurance.

Siting flipped from incentive to gate

Infrastructure moved the other way in the same period. Pennsylvania's EO 2026-05 conditions permit review on legally enforceable GRID commitments, strips projects from the Permit Fast Track Program, and ties tax exemptions to compliance. On August 3, Governor Abbott ordered PUCT and ERCOT to audit every pending interconnection request, including cost-shifting, self-supplied power and water, and ownership and controlling interests.

A queue position now buys a disclosure obligation rather than a megawatt, which is not what it was underwritten as.

Speculative queue positions get culled and ownership opacity becomes disqualifying, which produces dispersion rather than sector-wide impairment. Developers who turn up with self-supplied generation, water plans, community commitments and a transparent cap table clear faster and inherit the scarcity value. Treat that as a screen on positions already held this quarter and leave the sector call alone.

One deadline in the same brief does not wait for 2028. Newsom has until September 30 to sign or veto California's end-of-session AI bills, and SB 1000 carries an urgency clause, effective on enactment with no phase-in. It would remove the one-million-monthly-user carve-out and require latent disclosure whenever generative AI creates or alters an output. Bloomberg's reporting on the same regulatory split is the useful frame: governance loosening internationally while tightening at state and municipal level. A fragmented compliance surface is a moat for incumbents and a tax on sub-scale consumer AI. It can land three ways, and the ordering matters: he signs and sub-scale consumer AI absorbs the cost on enactment, he vetoes and the patchwork thesis loses a year, or preemption arrives first and the compliance names reprice. Signature is the base case.