The Investor

Three IPOs, one bad tape

Three of the most-watched private companies in the world are filing into public markets during the worst listing window in two years, which is either bad timing or — or rather, the more interesting version — exactly the timing the sellers wanted before something gets worse. SpaceX prices Friday June 12 at ~$1.75T, roughly a hundred times revenue. Anthropic has filed its S-1. OpenAI is queued behind both. The tape they are walking into is not friendly:

• May payrolls printed 172K vs. 80K consensus, with March and April revised up a combined +93K
• The three-month average is now 188K — a two-year high
• Nasdaq fell 4.18% in a single session, the worst print since April 2025
• FedWatch flipped, and a quarter-point hike by year-end is now more likely than a cut
• S&P Global confirmed on June 4 that none of these qualify for S&P 500 inclusion, on grounds of being unprofitable

The passive bid that mechanically absorbs supply in any normal mega-IPO will not be there for any of them. That structural air pocket has no precedent at this scale.

What the sources actually disagree on

Five independent analyses land on the same conclusion and split on severity. The bull case rests on SpaceX carrying $26B in annualized compute revenue from Anthropic and Google alone, a hyperscaler-tier business most investors have not bothered to model. The bear case is that this is a birthday-deadline IPO (Musk's self-imposed June 28) optimized for narrative rather than price, walking into rates that compress growth multiples. The compromise reading, which is probably the right one, is that the deal clears but the post-IPO trading band is wider and lower than where the secondaries currently mark it.

The day-one pop is the least interesting question in the room. The second-order effects are where the money is, or is not made:

1. Comp cascade: whatever multiple SpaceX prints becomes the anchor for every space, defense-tech, and AI-infra private mark in the pipeline behind it
2. SpaceX Mafia wealth event: roughly twenty years of employee paper turns liquid in a single quarter, and the pattern says it recycles into space-tech angel checks within sixty to one hundred and twenty days of lockup unlock
3. Anthropic public comp: once Anthropic prices, every AI app-layer private valuation gets re-marked against a real number rather than a deck slide

The disagreement worth taking seriously

One camp argues these names command enough strategic demand that public-market sentiment is, in their word, decorative. The other camp says at 100x revenue, you're not buying the asset — you're buying the crowd's willingness to keep paying 100x, and that crowd just watched the Fed pivot die in real time. The reconciliation is uncomfortable and probably right: SpaceX-the-business may well justify any multiple anyone cares to put on it; SpaceX-the-stock needs passive flows it is not going to get.

Lockup math

At roughly 180 days post-IPO, lockup expiration produces either a better fundamental entry for public-market longs or a 15-25% senior engineering exodus that simultaneously punctures SpaceX's execution story and funds the next cohort of space startups. Both are investable. They sit on opposite sides of the same trade.

The New Evidence

Princeton's ICML 2026 reliability audit dropped this week, covering GPT 5.5, Gemini 3.1 Pro, Gemini 3.5 Flash, and Claude Opus 4.7, and the finding is the one practitioners have been muttering about for a year: frontier models are not meaningfully more reliable than what they replaced. Another year of capex bought models that fail in the same places, just more fluently. The peer review is the new part.

Meanwhile the open-weight floor keeps rising into a ceiling that refuses to move:

The frontier ceiling is sticky and the open-weight floor is rising into it. With AI infra at 0.8% of US GDP, cost routing is now a first-order business problem — structural, not a news cycle.

What This Means for Multiples

Three independent sources land on the same arithmetic: closed-model API multiples should compress from 80-120x ARR toward 50-70x. If reliability has plateaued and open weights replicate capability on a consumer GPU, the premium is paying for distribution and brand, not technology. That is a perfectly fine business. It is not a hundred-times-revenue business.

The migration is visible if you squint. Cloudflare shipped AI Gateway spend caps, and when the network layer starts selling protection from your own model bill, the bottleneck has moved one floor down. Epoch AI sets the denominator: AI-related compute spend hit ~0.8% of US GDP in Q1 2026. A basis point shaved off that base is real money, and the recipients are cost-routing platforms rather than the labs doing the spending.

The Counter-Thesis

This is probably wrong if a frontier lab posts a genuine reliability step-change in the next two quarters — GPT-6, Claude 5 — in which case the compression argument dies and the access-moat names re-rate the other way. One source argues open-weight economics could stall on serving rather than capability; running them cheaply at scale is not solved. Both risks are live. Neither is the base case for the next twelve months given what Princeton just published.

Where Value Migrates

Three sources converge on the same shortlist: AI FinOps, inference cost-routing, and agent-execution tooling. The Cloudflare data point — a 10% reroute on a $10M bill saving roughly $1M — is the category-creation signal, or rather the version of it that already cleared a procurement department. Most founders here are still pre-Series A. Google splitting TPU into training (8t) and inference (8i) variants tells you inference is now its own sub-sector. That is where the next allocation goes.

Two Data Points That Reset the Category

GitHub's CPO disclosed that the platform processed 17 million agent-generated PRs in March 2026 alone, with record acceleration following a December 2025 model capability inflection. In the same week, OpenAI folded Codex into ChatGPT — the most direct bundling event the standalone AI coding tool category has faced. GitHub also moved Copilot to usage-based billing on June 1.

The critical insight across three independent analyses: the agent surge flowed to the incumbent, not to startups. GitHub's 630M monthly visitors and Microsoft's enterprise channel converted the December 2025 capability jump into 3x baseline growth. Standalone coding copilots pitching Series B need to explain why the same surge didn't compound to GitHub instead of them.

Generation is commoditizing into the platform layer; the alpha for the next 18 months is in verification, cost intelligence, and the agent-API ecosystem GitHub is about to open up.

The Kill Zone Map

Cross-Source Pattern

One source frames this as "OpenAI did to coding tools what Microsoft did to Slack with Teams." Another calls it the 18-month clock: standalone tools either prove they're product companies or become feature companies. A third identifies the semantic routing + small-model tiers (MAI Code One Flash) as the mechanism that compresses unit economics for frontier-only competitors. All three agree on the survivor profile: deep workflow integration, enterprise switching costs, IDE-native distribution, or agentic depth. 'Better autocomplete' is not on the list.

Where the New Alpha Lives

1. AI FinOps for engineering: Usage-based billing + token-heavy agent sessions = enterprise CFO problem that didn't exist 60 days ago. Chronicle validates demand but is GitHub-only. Multi-vendor observability is the wedge.
2. Verification layer: At 17M agent PRs/month, human code review physically breaks. Agent-native SAST/DAST, automated PR triage, and AI-aware security scanning are the new bottleneck products.
3. Agent-API ecosystem builders: GitHub explicitly signaled a shift from UI→UX→AX (Agent Experience). The 18-month platform-ecosystem window — identical to early Slack/Shopify/Stripe — just opened.

The Numbers Nobody Priced

Two disclosed contracts, one quarter, and a market tier nobody bothered to draw on the map. SpaceX is now collecting roughly $2.17 billion per month in AI compute rent.

That is roughly $26B annualized from two customers, assembled largely outside anything a public-market analyst was modeling. For scale, this puts SpaceX's compute line alone in the neighborhood of AWS's early growth years. The Google contract carries cancellation optionality after December 2026, which is a real risk and not a rhetorical one. The Anthropic deal looks more durable. This is probably wrong, but the more interesting version of the story is that even the durable one gets repriced before the cliff arrives.

When Google — a company that literally designs its own TPUs — rents 110,000 NVIDIA GPUs from SpaceX, the signal is unambiguous: even hyperscalers cannot build fast enough.

Meta's Tent Pivot Kills the DC REIT Moat

Meta is pitching five 125,000 sqft tents in Ohio because the 2-3 year data center construction cycle is too slow, and they would like to compress it to 2-3 months. If the most disciplined hyperscaler walks away from traditional construction, the traditional DC operators lose the pricing power they were charging for. The marginal dollar moves to modular DC fabricators, behind-the-meter power developers, gas turbine and SMR plays. What Meta is not doing, notably, is signing more long-dated leases with the incumbents.

SoftBank announced €75B for French data centers. The marginal AI infra dollar is now rotating toward sovereign and European capacity. The market has not fully absorbed the geography of that move.

Regulatory Crack

New York imposed a 1-year data center moratorium, which is the first state-level constraint on AI buildout worth naming. Power draw is now a voter issue in coastal states, and voter issues do not unwind quickly. The arbitrage runs to Texas, Wyoming, rural Ohio, and Tennessee. Land with power rights is the underwriting wedge.

The SpaceX Secondary Mark

Multiple sources agree the $26B compute run-rate is likely not in current secondary marks. The sum-of-the-parts for SpaceX as launch plus Starlink plus compute landlord is a different company than launch plus Starlink. The next primary round, or the IPO itself, resets the mark sharply higher on compute. The Google cancellation cliff and two-customer concentration are the things you actually have to underwrite.