SpaceX Buys Cursor, Absorbing Its Own Compute Tenant

The Stack Is Pricing Layer by Layer

Three data points today confirm the AI value chain is pricing sequentially — and the next investable layer is clear:

• GPU Compute: Priced. SpaceX at $28B, Coreweave at $60B market cap. Blackwell at $10+/hr. Arbitrage rents, not venture returns.
• Inference Control Plane: Priced. Baseten at $13B Series F, serving Cursor, Harvey, Notion, Abridge, Decagon. Category de-risked but returns capped at late-stage multiples.
• Agent Runtime Infrastructure: Underpriced. Google's Interactions API GA + Antigravity sandbox just standardized the category. Hermes at 200K GitHub stars. Series A/B entry points still available.

What Is Agent Runtime Infrastructure?

The specific companies and capabilities to source: sandboxed execution environments, stateful session managers, model-specific harnesses, and cost-aware orchestration layers. Google blessing this with Antigravity is the category-legitimizing signal — the same way AWS Lambda legitimized serverless.

Inference was the platform layer of 2024. Agent runtime is the platform layer of 2025. The gap between category legitimization and multiple expansion is your entry window.

Open-Weight Acceleration Compresses the Model Premium

GLM-5.2 sitting at #3 on GDPval-AA and costing $0.41/agentic task vs. Opus at $0.81 is the substrate fact making runtime infrastructure more valuable than model access. If any enterprise can drop in an open-weight model at half the COGS, the moat moves to how you run models, not which model you run.

The practical implication for portfolio construction: audit every AI portco for proprietary-model dependency. If GLM-5.2-class substitution drops COGS 40-50% without quality degradation, gross margins re-rate upward. If it can't substitute cleanly, you've identified a genuine closed-model moat. Either answer is investable intelligence.

Methodological Warning

A cross-source caution on quality claims: an LLM-as-a-Judge audit across 21 judges and 541K judgments shows Cohen's kappa deflates agreement by 33-41 points vs. exact-match. Translation: a meaningful fraction of AI portco quality benchmarks are overstated. Demand kappa-adjusted reporting in diligence.

The Deal Cluster

Four deals printed in rapid succession, all clustering at the same layer of the stack:

The Pattern

Every deal that closed lives at pre-production or licensing — storyboarding, previsualization, virtual scouting, content libraries. None sit at the generative production layer, which is interesting given how much of the press cycle has been spent there. The cautionary data point is the Disney/OpenAI collapse when Sora was shut down: model-layer entertainment bets pay off once or not at all, while tooling-layer bets are doing the much less glamorous work of generating revenue every quarter.

Hollywood is buying AI tools that augment the $15B pre-production workflow, not AI models that replace the $200B production workflow. The investable surface is narrower and more capital-efficient than the hype suggests.

Entry Points

The Google/A24 deal at $75M validates budget-level commitment from a studio that does not write checks casually. Series A entry points still exist in pre-production tooling — storyboarding automation in particular, with virtual scouting and AI-assisted previz behind it — before the multiple expansion catches up to the deal flow. The diligence question is narrow: does the tool shorten the pre-production timeline, which is investable, or does it generate final-frame content, which is binary risk dressed up as a product.

Deprioritize generative-video model-layer plays until execution risk de-rates. This thesis is probably wrong in the obvious ways — a model that actually ships and integrates would re-rate the layer overnight, or a partnership structure designed to survive a model shutdown would do the same more quietly — and neither has happened yet. The Sora shutdown took the Disney partnership with it. That is what platform dependency looks like before it shows up in a deck.