The Investor

The Setup Nobody Expected

SpaceX prices Friday, June 12, at roughly $1.75T, the largest IPO ever printed, and the headline writers will tell you it is an aerospace story. The more interesting version is that SpaceX has quietly become a hyperscaler. Filings in the last two weeks show $2.17 billion a month in AI compute rent — $1.25B from Anthropic for Colossus 1 near Memphis, $920M from Google for roughly 110,000 NVIDIA GPUs starting October 2026 — which annualizes to $26B from exactly two customers. None of that existed on a public tape until two weeks ago.

The secondaries are stale. Most holders are still marking a launch-plus-Starlink SOTP and ignoring the compute line entirely, which is either an oversight or a gift, depending on whether the listing prints clean.

The Tape Is Not Friendly

May payrolls came in at 172K versus 80K consensus, March and April were revised up by a combined +93K, and FedWatch quietly flipped to a hike being more likely than a cut. The Nasdaq dropped 4.18% in a session, its worst day since April 2025. On June 4, S&P Global confirmed it would not bend the inclusion rules: SpaceX, Anthropic, and OpenAI all sit out the S&P 500 for at least twelve months plus four profitable quarters.

Three of the most-watched private names in the world are walking into the most hostile listing window in two years without the indexers behind them.

No passive bid means day one is entirely active managers and retail. SpaceX's CFO video, which is doing a Brin-and-Page 2004 direct-to-public homage whether it admits it or not, says the management team has decided to route around the institutional gatekeepers. Institutions tend to remember being routed around. Price discovery gets bumpier.

The Space Mafia Liquidity Wave

The second-order trade matters more than the first. A decade of restricted SpaceX paper turns liquid inside a quarter, which means several hundred senior engineers and operators now hold enough to seed the next cohort of space companies themselves. The 2004 Google IPO did the same thing — that cap table funded YouTube, LinkedIn, Yelp, and Palantir before the lockup had really expired in spirit.

The recycled capital lands in the obvious places: propulsion, in-space manufacturing, satcom infrastructure, lunar logistics. Anyone with a deck can list those four. The alpha is being the first call when a propulsion lead actually puts in notice, which is a relationship business, not a thesis.

Risk: The Google Contract Has a Kill Switch

This is probably wrong, but: the Google deal carries a 90-day cancellation right after December 2026 and a September 30 GPU delivery cliff, and that is real concentration. The Anthropic line is durable, the Google line is optional. Sizing should haircut the Google revenue by 30 to 40% on a probability-weighted basis. Maybe more if the cliff slips.

What This Means for the Book

Three repricings land at the same time, and the book has to absorb all of them:

1. Pre-IPO space secondaries get marked up or down on the print — model both a 20% upside and a 30% compression and stop pretending you know which
2. Late-stage growth at 2025 marks is structurally underwater in a no-cuts-2026 tape, and the opportunity cost of holding the denial is a quarter of returns
3. Meta's tent data centers — five 125K sqft units stood up in two to three months versus the two to three year industry norm — say the DC REIT moat is gone, and the modular off-grid operators are what replaces it

The Ceiling and the Floor Are Converging

Princeton's ICML 2026 reliability audit is the cleanest empirical reading we have so far: GPT 5.5, Gemini 3.1 Pro, and Claude Opus 4.7 are not meaningfully more reliable than the versions they replaced. Another year of frontier capex bought models that fail in the same ways, only more fluently. The floor, meanwhile, keeps rising into the ceiling:

• MiniMax M3: one million token context window, open weights.
• Gemma 4 QAT: multimodal on a laptop at roughly 1GB.
• Ideogram 4.0: 9.3 billion parameters fitting a single 24GB GPU, top of Arena open-weight.
• Kimi K2.5 / GLM-5: Chinese open weights posting agentic scores competitive with Opus 4.7.

Sticky frontier ceiling, rising open-weight floor. Closed-model API names move to underweight pending a reliability step the audit says has not arrived.

Anthropic's IPO Creates the Public Ceiling

Anthropic filing for IPO is not really a capital markets event in the usual sense; it is the first time private AI app-layer multiples will face a public comp. Every deck calibrated against a private Anthropic that nobody had to mark to anything gets repriced inside 90 days of listing. The three branches worth pricing:

1. IPO prices well. Private comps reprice upward and the capital cycle extends another year.
2. IPO prices badly. Private marks compress and the secondary market finally does the arithmetic it has been avoiding.
3. IPO gets pulled. The most informative outcome of the three, because it tells you what the bankers learned on the roadshow.

The sell side is already writing scenario one. The numbers suggest scenario two is more likely, though this is probably wrong in some way I have not figured out yet. Buffett's $10B Alphabet position, in the meantime, is confirmation that value capital has crossed into AI megacaps, which means the easy alpha in megacap AI exposure has already been collected. By the time Berkshire shows up, the consensus is already in the room.

Where Value Migrates

If model quality commoditizes, value accrues to whichever layers keep their pricing power regardless of which model serves the token:

Google splitting TPU 8 into a training variant (8t) and an inference variant (8i) is the hardware-layer tell. Inference is now a standalone capex category with its own SKU and, eventually, its own multiple. AI infrastructure spend is running at 0.8% of US GDP, which is the size of a mid-tier sector, and it is running there on the premise that value stays in the model layer. That premise is breaking.

The Counter-Thesis Worth Holding

The version of this where I am wrong: a frontier lab posts a genuine reliability step inside the next two quarters, an actual GPT-6 moment, the compression argument dies on contact, and access-moat names re-rate upward. This has been the bear's graveyard before. What makes this audit different is that it was peer-reviewed across all three majors, and it is the first credible empirical evidence the plateau is real rather than a Twitter narrative. Weight it accordingly.

Two Data Points That Reset the Category

GitHub's CPO disclosed that the platform processed 17 million agent-generated pull requests in March 2026 alone, with record acceleration following a December 2025 capability inflection. Separately, Copilot moved to usage-based billing on June 1. In the same week, OpenAI folded Codex directly into ChatGPT — absorbing the standalone coding tool's core value proposition into the platform's free distribution.

The surge flowed to the incumbent. GitHub's 630M monthly visitors and Microsoft channel converted the December capability inflection at roughly 3x baseline expectations. The question standalone coding tools need to answer: why didn't this compound to you instead of GitHub?

When the platform absorbs the feature, the standalone doesn't get a softer landing for being earlier in the category. It gets a worse one, because the bundler is not pricing for margin.

What Dies and What Emerges

The category is bifurcating into platform consolidators (capturing distribution) and adjacent layers (where new bottlenecks live). Generation is no longer the scarce resource — verification, cost predictability, and routing intelligence are.

Kill zone (15-30% multiple compression):

• Standalone coding copilots competing on completion quality
• Any tool whose moat thesis doesn't survive Copilot at usage-based pricing
• Companies reliant on undifferentiated autocomplete without workflow data ownership

Greenfield opportunities:

• AI FinOps for engineering — Usage-based billing creates an enterprise CFO problem overnight. Chronicle validates demand but is GitHub-locked. Multi-vendor environments (Copilot + Cursor + Claude Code + internal models) need neutral-layer cost observability.
• Verification layer — 17M agent PRs/month exceeds human review capacity. Agent-native code review, AI-aware SAST/DAST, and automated PR triage are structurally underfunded vs. demand.
• Agent-API ecosystem — GitHub explicitly framed APIs evolving to 'agent-centric' and the design paradigm shifting from UX to AX (Agent Experience). First movers compound default distribution before saturation, as with Slack apps and early Stripe Connect.

The Contradiction Sources Surface

There's a tension between two readings of this data. One source frames Cognition's 'Switzerland of AI Agents' pivot as validation that the agent layer is fragmenting into neutral orchestrators vs. vertically-integrated stacks — with the middle compressed. Another frames the same signal as bundling making everything converge to platforms. Both are simultaneously true at different layers: the generation layer is consolidating while the orchestration layer is fragmenting. The barbell thesis funds neutral orchestrators OR vertical full-stack agents and treats the middle as a structural short.

Anthropic's Mythos being labeled a 'budget buster' in enterprise AI security adds a parallel signal: even frontier labs face pricing pushback when enterprise buyers have alternatives. The cost-disruptor wedge is opening in every AI tool category simultaneously.