The Investor

Menlo Ventures turned a roughly $1B Anthropic position into a $14B mark, a stake ten times larger than any investment in its 50-year history, and posted 40%+ IRR on its 2023 and 2026 vintages, a number that excludes the May Series H. One position rehabilitated a firm whose 2015 and 2018 funds nobody remembered, and earned Menlo a $3B raise. That is the benchmark LPs are about to apply.

Anthropic and OpenAI are filing because private capital ran out of appetite for $60B+ rounds, not because either business needed the spotlight. Anthropic closed a $65B Series H at $965B four days before filing, above OpenAI's $852B, with Goldman, Morgan Stanley and JPMorgan leading. The debut becomes the comp anchor for the entire late-stage stack.

The sequencing is the risk. It prints into a cooling tape. SpaceX broke below its $135 IPO price to $131.11 inside a month, the Nasdaq-100 fell 1.6%, and the marginal buyer is discriminating rather than euphoric.

Where the sources diverge is the interesting part. The bulls argue late-stage private marks lead public comps and will hold. The credit-and-repricing camp notes marks are running a quarter or two behind, which is a different claim entirely. Oracle's CDS quintupled to 198bps, SpaceX's bonds trade like junk, and the auditor resets the mark for you if you don't do it first. When the bond desk disagrees with the equity story, the bond desk usually wins.

Anthropic's range will settle whether late-stage AI marks are conservative or fiction. That is information LPs get to price, whether or not the holders wanted it priced.

One diligence line the correction just made mandatory: Commerce disabled Anthropic's frontier models globally within 90 minutes and kept them dark for 14 days. Any diligence memo that skips the regulatory kill-switch clause and billing-meter integrity is incomplete now.

The signal isn't any single round — it's who shares the cap tables. Sequoia, Founders Fund, a16z, General Catalyst, Lowercarbon and 8VC all sat in Senra's $65M Series B, funds that normally compete rather than co-invest. When rivals cluster like that, they're marking a category to market: 'software-defined skilled assembly' just moved from contrarian to consensus. The economics that justify it — 99% first-pass yield vs 75% incumbents, a 20%→50%+ margin curve, 4-week vs 4-month lead times — sit behind an ITAR moat, against a $1T+ defense budget and launches up 64% YoY.

The same pattern repeats up the froth curve. Walden Robotics took $300M at a $1.1B seed (Nvidia, Boeing, Samsung, CoreWeave Ventures, Toyota); Humanoid's $150M Series A first tranche priced at $1.2B. Radiant is running full-power reactor tests toward a hard 2028 delivery, and Musk quietly bought APR Energy, a ~$1B gas-turbine firm, to feed data centers.

Two structural reads matter more than any valuation. First, compute-rich strategics are converting GPU leverage into equity access — CoreWeave and Nvidia crowding seed tables squeezes pure-financial VCs on both price and allocation. Second, the binding constraint has moved from chips to power: no domestic HALEU supply exists, and behind-the-meter generation is the design-agnostic chokepoint every reactor depends on.

Where the evidence is thinnest: these are top-of-cycle prices — a seed marked like a Series C — and Senra's 50%+ margin is booked, not yet realized. The froth risk is real even where the demand tailwind is.

When rival top-tier firms share one hardtech cap table, they're pricing a category — and the cheap entry is one component-vertical over, not in the headline round.

Underwrite the composition, not the multiple. Kalshi is raising at ~$40B — 8x its mark 18 months ago, roughly 20x its $2B annualized revenue, about 2x Polymarket's $15B. Which is either a sensible price for a fast-growing venue or a wager on nine justices, because roughly 65% of the volume is sports contracts a dozen states are trying to ban. The comp treats Kalshi and Polymarket as the same animal. They aren't. Kalshi is CFTC-regulated and takes real economics on flow; Polymarket ran years at zero fees on a token that doesn't yet exist.

The interesting thing the round buries — or rather the durable thing — is the compute forward curve: prediction contracts on future GPU rental prices, the first genuine financialization of AI compute. The value there accrues one layer down, at the indices and hedging platforms and brokerages where entry multiples are still sane, and not at the headline exchange everyone is pricing.

The disruption gets corroborated from the short side, which is the tell. Kalshi has cleared $100B cumulative and runs ~$10B in weekly notional, and DraftKings is down ~40% since September to a $12.5B cap. The Bear Cave's point isn't that the threat exists — the chart already conceded that — it's that consensus drifted from 'this won't matter' to 'this is priced in,' which is the more comfortable belief and, conveniently, the more fadeable one, while the flow keeps migrating.

The 40% DraftKings drop discounts the fear; it has not yet discounted the flow migrating to the exchange.

The clean expression is a paired read — the structurally challenged incumbent against the exchange collecting the flow — not a naked directional bet that punishes you on timing, which is a subtler way of saying you can be right about the migration and still lose money waiting for the market to agree.

The target inventory nearly doubled year-over-year to 160+ privately held companies worth over $1B that haven't raised since June 2024, which is not a cyclical dip so much as a liquidity plumbing problem: the demand that used to find these names now gets vacuumed up by SpaceX, Anthropic and OpenAI, while the public enterprise-software comps sit 30%+ off. So the consensus buyer target becomes data infrastructure, and roughly a dozen data-software names on the list are already fielding approaches from both big tech and PE. Watch the structure, not the headline. The Meta/Scale template — 49% for talent plus data — is the one everyone will copy.

Alongside it, a fresh category just got reference-priced. Oak emerged with a $60M seed from Accel, CRV and Greylock for AI-agent identity, and a syndicate that size only assembles when three tier-1 firms have quietly agreed a category exists. The corroboration is unusually broad: attackers are systematically scanning for MCP servers and LLM endpoints in the wild — the same pre-signal that ran ahead of Wiz's cloud-security category — an autonomous hackbot cleared net-profitability at an 89% confirmation rate, and multiple sources converge that the harness layer is commoditizing while agent identity is the Okta-shaped position with the switching costs attached.

The divergence worth pricing, and this is where late buyers historically lose money: 'non-replicable data' is a claim, not a fact. Microsoft and Databricks sit on agent-data tooling too, and either could consolidate the layer, foreclosing the independent exits the whole thesis rests on. Paying ahead of confirmed transaction comps is precisely how the last cycle's late buyers lost money.

Sellers hold the leverage now, not later — the window is the interval between high buyer interest and the first closed deal, and no meaningful data-software deal has closed yet.