The Investor

What's New Since Sunday

Sunday's note pegged SpaceX's $26B annualized compute revenue from Anthropic and Google, and called it interesting until proven otherwise. Three things have since proven it more interesting. A third anchor contract with Reflection AI at $6.3B / $150M per month on GB300 hardware through 2029 takes the run-rate to twenty-eight billion dollars. SpaceX is also acquiring Cursor, which is to say acquiring one of its own tenants. And a $20B bond issuance is imminent, which is the first quasi-public window anyone outside the building gets into the unit economics.

The vertical integration is the qualitative shift, or rather, the more interesting version of it. When you buy your own tenant you stop being a landlord and start being a platform with captive distribution. Cursor on SpaceX silicon has a cost-of-goods advantage no standalone dev tool can replicate. That is the deal.

Coreweave Repricing Math

Both sources land on the same conclusion, which is that Coreweave's $60B public mark is structurally exposed. The math is not subtle:

The 90-day cancellation clauses are the part nobody is pricing. SpaceX's twenty-eight billion dollars is simultaneously larger than the market understood and more fragile than the headline implies. For Coreweave-comp purposes, however, the damage is done. A private company is doing twice the revenue with cheaper capital and a tenant it can also sell.

The Bond Prospectus Is the Catalyst

When the $20B bond prints, the prospectus contains the first quasi-public disclosure of AI compute unit economics from this seller. Cross-reference the disclosed revenue against Colossus 2 capex and back into gross margin. That number re-rates every cloud-native AI infra position in the book.

SpaceX crossed from compute landlord to vertically integrated platform in one week. The $20B bond prospectus is the informational catalyst that forces every AI infra comp to reprice.

Structural Tension to Track

Anthropic now shares a compute provider with the parent of a competitor's distribution surface, namely Cursor. This is probably wrong, but the path of least resistance is that Anthropic diversifies compute, which is good for non-SpaceX infra plays. The other path is the dependency tightens and the next Anthropic round wears the narrative risk. Either way, this fact pattern will surface in diligence on every AI deal touching the SpaceX compute stack inside ninety days.

The Inflection

GLM-5.2 is the first open-weight model functioning as a credible Opus 4.8 substitute at approximately 50% cost in real agentic workloads. At $0.41 per agentic task versus Opus at $0.81, it's ranked #3 on GDPval-AA with 20+ inference providers already serving it. This isn't a benchmark curiosity — it's a gross margin event for every company in your portfolio running on closed-model APIs.

The timing matters because the inference layer just got priced. Baseten's Series F at $13B (with a notable discrepancy vs. an internal $1.5B figure worth investigating) and its customer roster — Cursor, Harvey, Notion, Abridge, Decagon — confirms inference-as-a-platform is the consensus institutional trade. That means the alpha has rotated downstream.

Where Value Accrues Now

If inference is priced and open-weight models compress API margins, the next investable layer is agent runtime infrastructure: sandboxed execution, stateful session managers, model-specific harnesses, and cost-aware orchestration. Google just blessed this layer by shipping the Interactions API to GA along with the Antigravity sandbox. When a hyperscaler standardizes a layer, Series A/B companies building there get category validation and enterprise pipeline simultaneously.

Inference is priced at $13B. Foundation models are commoditizing at 50% cost compression. The alpha vintage is agent runtime infrastructure — sandboxes, stateful sessions, and orchestration — and it's Series A/B right now.

Portfolio Diagnostic

Run a simple test: can your AI portcos drop GLM-5.2 into production at half the COGS without quality degradation? If yes, gross margins re-rate upward. If no, you've identified a genuine moat (likely fine-tuning depth, latency requirements, or compliance constraints). Either answer is valuable diligence output.

Methodological Warning

A benchmark audit across 21 LLM-as-a-Judge systems and 541K judgments shows Cohen's kappa deflates agreement by 33-41 points versus exact-match. A meaningful fraction of AI portco quality claims are likely overstated. Demand kappa-adjusted reporting in your next portco review.

From Litigation to Licensing

Four deals landed in roughly the same week, which is either coincidence or the part of the cycle where the lawyers get bored and the licensing teams take over. Hollywood and AI moved from courtroom posture to revenue-generating partnerships, and the cluster tells you where the studios actually want to spend money this year.

• Google/A24: $75M — pre-production tooling and content licensing
• Netflix/Affleck: M&A — talent-studio-AI integration
• Lionsgate/Runway: Partnership — production workflow
• Getty/OpenAI: Licensing — training data supply

Every one of these clusters at pre-production tooling and content licensing, not generative video production, which is the layer the press has narrated for two years. The Disney/OpenAI collapse around Sora is the cautionary counter-example, or rather the more useful one: model-layer entertainment bets are binary, while tooling-layer bets come with revenue attached, and the studios just signaled which one they are willing to underwrite.

Where the Entry Points Are

The investable read is layer specificity, not deal count. Storyboarding, previz (the unglamorous middle of the stack), and virtual scouting are the wedges being validated by major studio spend, priced today at Series A/B before the multiple expansion that tends to follow A24-scale validation. The generative-video model layer, meaning Runway, Sora, Kling and the rest, stays high execution risk until unit economics derate, which so far they have not. This is probably wrong if the next Sora-class model ships with margins attached, but nothing in the current filings suggests it will, and the studios are voting with their checkbooks one layer above.

Model-layer entertainment bets are binary. Tooling-layer bets have revenue. The studios just told you which layer they're buying.

The Structural Dynamic

Hollywood's embrace of AI tooling follows a pattern that should not surprise anyone who watched the VFX rollouts of the last two decades: studios license technology that reduces pre-production cost without threatening the creative workforce narrative, which is politically safe and margin-accretive in a way that scales quietly across the content portfolio. The deals cluster where they cluster because the alternative — generative production at studio scale — is still radioactive, and the studios know it better than the model labs do.