The Investor

The Setup Nobody Planned For

SpaceX prices the largest IPO in history on June 12 at approximately $1.75T — roughly 100x revenue — into a tape that just watched Nasdaq shed 4.18% in a single session after May payrolls printed 172K against 80K consensus. FedWatch now prices a hike as more likely than a cut. S&P Global confirmed it will not bend inclusion rules for unprofitable issuers, meaning no S&P 500 passive flows for SpaceX, Anthropic, or OpenAI for at least 12 months.

This is the convergence that matters: the three most-watched private names in tech are walking into the most hostile listing window in two years, without the mechanical bid that absorbed supply in every prior trillion-dollar event.

The Compute Revenue Nobody Priced

The investment case has quietly shifted. SpaceX now collects $2.17B per month in AI compute rent — $1.25B/month from Anthropic for Colossus 1 near Memphis, and $920M/month from Google for ~110,000 NVIDIA GPUs starting October 2026. That is $26B in annualized run-rate from two customers, formed largely outside public-market view.

SpaceX is no longer a launch company with a Starlink business. It is a hyperscaler with a launch business and a constellation.

The Google contract has a 90-day cancellation clause after December 2026 — real risk. The Anthropic deal appears more durable. Together, they materially change the sum-of-parts analysis and likely aren't reflected in current secondary marks.

Five Sources Disagree on What Happens Next

The intelligence splits into three camps:

• Bull case: The Google compute contract plus Starlink ARR trajectory justifies the multiple. The Space Mafia wealth-recycling thesis (modeled on Google 2004's Xoogler angel network) seeds the next generation of space-tech founders within 90 days of lockup.
• Bear case: 100x revenue is the kind of multiple that prints at cycle tops. Every smaller space competitor just got an explicit anchor they cannot reach. The IPO is deadline-driven (Musk's June 28 birthday), not pricing-optimized.
• Structural risk: Senior engineering attrition post-lockup (model 15-25% over 24 months) is bullish for downstream deal flow and bearish for the core asset. The CFO-led video signals retail-heavy distribution, which institutional allocators historically resent.

The Macro Context Is Cruel

May payrolls at 172K with March/April revised up +93K creates a three-month average of 188K — a two-year high. Inflation at 3.8% running ahead of wage growth at 3.4%. Unemployment at 4.3% with long-term unemployed at 27.5% (highest since Dec 2021). Every late-stage growth mark underwritten to 2026 rate cuts is structurally upside-down.

The Capital Allocation Playbook

The alpha is not in the IPO allocation. It's in the second-order flows:

1. Mark the late-stage book to a 'no cuts in 2026' world this week. The opportunity cost of waiting is a quarter of denial that LPs will read in the next report.
2. De-risk SpaceX secondary exposure before June 12. Without passive flows and into rising rates, the post-IPO trading band is wider and lower than secondary marks suggest.
3. Front-run the Space Mafia angel wave. Build a target list of 15-25 ex-SpaceX founders raising in the next 6-12 months. Operator angels follow within 60-120 days of lockup. The goal is cap-table presence before valuations inflate.
4. Wait on the public position. Lockup expiration (~180 days) has historically been the cleaner entry when IPOs are deadline-driven and retail-distributed.

The Princeton Verdict

Princeton's ICML 2026 reliability audit landed this week and the market has not yet read it carefully. The finding, restated honestly: GPT 5.5, Gemini 3.1 Pro, and Claude Opus 4.7 are not meaningfully more reliable than the models they replaced. Another year of frontier capex bought models that fail in the same ways with better grammar. AI infrastructure now runs at 0.8% of US GDP, spent on the working assumption that the model layer is where the value accrues.

The frontier ceiling is sticky and the open-weight floor is rising into it. The closed-model premium is now a question, not an assumption.

Open-Weight Is Closing the Gap on Hardware That Costs Less Than a Car

Four data points arrived in the same cycle, which is more interesting than any one of them:

• Gemma 4 QAT runs multimodal in roughly 1GB of VRAM
• MiniMax M3 ships a 1-million-token context window as open weights
• Ideogram 4.0 nf4 fits on a single 24GB consumer GPU and tops the Arena open-weight rankings
• Kimi K2.5 and GLM-5 (Chinese open-weight) post frontier-adjacent agentic performance

The capability gap that justified proprietary API pricing is compressing on the dimensions that actually pay for themselves: context length, multimodal quality, agentic performance, all running on silicon a mid-level engineer can expense. The live question for every portfolio company paying for frontier API access is when "good enough" at ninety percent cost savings becomes the enterprise default.

What This Means for Multiples

Closed-model API businesses currently trade at 80-120x ARR in private markets. Princeton plus open-weight pressure points toward compression to 50-70x, which is the kind of move that makes the difference between a fund-returner and a footnote. Any portfolio name whose moat reduces to "access to frontier model X" deserves a Q3 stress test.

Three layers benefit from the same compression:

Google splitting TPU 8 into training-optimized (8t) and inference-optimized (8i) variants is the tell. Inference is now a standalone capex category rather than a byproduct of training, with different unit economics, different customers, and probably different exit comps.

The Counter-Thesis Worth Taking Seriously

This is probably wrong, but only in two specific ways. A frontier lab could post a genuine reliability step-function inside two quarters, in which case the compression argument dies and access-moat names re-rate upward. Or the open-weight surge stalls on serving economics rather than capability, which is the version I find more plausible. Both risks are live. Neither is the base case.

The Bundling Event

OpenAI merged Codex into ChatGPT this week, which on its own is a product decision and not much more. Set it next to GitHub processing 17 million agent-generated PRs in March 2026 alone, and the move to usage-based Copilot billing on June 1, and the picture stops being a product decision and starts being a distribution outcome. The incumbent caught the late-2025 capability jump. The startups did not.

GitHub's 630 million monthly visitors plus the Microsoft channel converted that capability jump into 3x baseline acceleration, which is either a remarkable piece of luck or, more honestly, what happens when the surface that already owns the workflow gets handed models that can open and close a non-trivial PR end-to-end rather than finishing a line of code. The December 2025 cohort could ship multi-file changes against a brief; the November cohort could not. That capability flowed to the platform layer, the way analytics flowed into the suites a decade ago. When the platform decides a feature belongs inside, the standalone loses on distribution rather than on merit.

Standalone coding copilots pitching Series B at 2025 multiples need to defend why the December 2025 surge didn't compound to GitHub instead of them.

What Survives and What Doesn't

The category is bifurcating:

The 18-month clock started this week. Standalone coding tools that have not demonstrated they are product companies rather than feature companies by the end of it will be acquired cheaply, wound down, or kept alive by founder pride. There is a version where one of them breaks out on a vertical wedge GitHub doesn't care about. It is a real version. It is not the base case.

Three New Categories Open

Generation is commoditizing into the platform. The interesting alpha for the next 18 months sits in what the platform creates demand for but cannot own outright:

1. AI FinOps for engineering. Usage-based billing on top of token-heavy agentic sessions is, in plain terms, a CFO problem nobody has staffed for. GitHub's Chronicle proves the demand and is GitHub-locked, which is the opening. Enterprises running Copilot, Cursor, and Claude Code together need a neutral cost observability layer. Most founders here are pre-Series A, which is the right time to underwrite them.
2. Verification layer. 17 million agent PRs a month is past human review capacity by roughly any honest measure. Agent-native code review, AI-aware SAST and DAST, and automated PR triage become the actual bottleneck. The volume signal is not subtle.
3. Agent-API ecosystem. GitHub said its API is evolving toward primitives that agents, not humans, will call directly — fewer screens, more typed endpoints, schemas designed to be consumed by something without a cursor. When a platform owner signals new primitives this clearly, the 18-month ecosystem window opens, the way it did for early Shopify apps and Stripe Connect.

Cognition's Tell

Cognition repositioning as the 'Switzerland of AI Agents' tells you the agent layer is fragmenting rather than consolidating, and that enterprise buyers are pricing lock-in risk into procurement. The barbell, then: fund neutral orchestrators or fund vertical full-stack agents. The middle does not get funded.

Why This Filing Changes Everything Downstream

Anthropic filed its S-1 in the same week Buffett disclosed a $10B Alphabet position and Suno crystallized at $5.4B, which is either a coincidence or, more honestly, three different desks pricing the same idea on different timelines. The three data points rhyme. AI is being asked to clear a public-markets bar rather than a private one. The lazy read is validation. The more interesting read is that public markets will now price these names against each other every Tuesday at 9:30, which private rounds were never required to do and largely declined to.

The filing commits Anthropic to quarterly disclosure of unit economics the private AI market has spent three years not disclosing. Margins, token costs, customer concentration, capex. About to become legible.

Three Scenarios, One Action

1. IPO prices well: comparable private rounds reprice upward, the capital cycle extends another year, and secondary desks for OpenAI, xAI, and Mistral stakes move with it.
2. IPO prices badly: private marks come under pressure and the late-stage secondary market does the unpleasant arithmetic it has been politely avoiding.
3. IPO gets pulled: the most informative outcome, because it tells you what the bankers learned during the roadshow and were unwilling to put a price on.

The action does not change across scenarios, which is what makes it worth doing now rather than after the fact: build an Anthropic comp model now and re-mark every AI app-layer portco against the projected public multiple range. The multiples in current deck comps were calibrated against a private Anthropic nobody had to mark to anything. That reference point dies the moment the S-1 becomes the pricing document.

The Pause Call Is Positioning, Not Policy

This is probably wrong, but Anthropic publicly calling for a global AI freeze reads as regulatory moat-building rather than principled alarm. Calls for pauses from incumbents always disproportionately tax challengers, which is a regularity rather than a coincidence. If it gains political traction, pre-Series B foundation-model challengers absorb more compliance drag than frontier labs with existing safety teams. The enterprise safety lane is the lane Anthropic is choosing to own ahead of the roadshow.

The Buffett Signal

Berkshire putting ten billion dollars into Alphabet is not an AI call. It is Alphabet-at-this-multiple being cheap enough for value capital, which is a different sentence dressed up in the same letters. What it confirms is narrower: the easy alpha in megacap AI is gone. When Buffett buys, the consensus has already formed. The remaining alpha sits in security, sovereign infrastructure, and vertical data moats. Not in owning the hyperscalers.