SpaceX Prices at $1.75T With $26B AI Compute Run-Rate

The Princeton Audit That Changes the Multiple

Princeton's ICML 2026 reliability audit landed this week with the kind of finding the frontier labs would have preferred to bury in an appendix: 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 capex bought models that fail in the same ways, more fluently.

While that ceiling sits where it sat, the open-weight floor is rising into it:

• MiniMax M3: 1M token context window, open weights
• Gemma 4 QAT: multimodal in roughly 1GB on a laptop
• Ideogram 4.0: 2K native image gen on a single 24GB GPU
• Kimi K2.5 and GLM-5: frontier-adjacent agentic performance, free

The capital story reads bullish. The moat story reads complicated. The question for the next IC is which layer of the stack keeps pricing power once Anthropic prints a public multiple and open weights eat the model layer's premium.

Inference Bifurcates From Training

Google split TPU 8 into a training variant (8t) and an inference variant (8i) at Cloud Next '26, which is not a packaging decision so much as one of the largest compute buyers on earth quietly concluding that inference deserves its own SKU and, eventually, its own multiple. Inference compute is distributed, compounds with agent activity, and is contested across Google, Groq-class entrants and custom ASICs against NVIDIA's training dominance.

Separately, AI infrastructure hit 0.8% of US GDP in Q1 2026 per Epoch AI. That is an entire mid-tier sector of national output spent on the working assumption that value accrues at the model layer. Princeton just said it doesn't, or at least not on reliability.

What This Means for Multiples

Closed-model API multiples should compress from the 80-120x ARR zone toward 50-70x. The defensible moats narrow to safety tuning, enterprise distribution and governance UX of the Claude Code 7-mode permission architecture variety. Raw capability is no longer one of them.

Where the value migrates, in rough order of conviction:

1. AI FinOps and cost routing: Cloudflare shipped AI Gateway spend caps. A 10% reroute on a $10M bill saves $1M, and the bills are getting bigger.
2. Inference infrastructure: silicon, chip-to-chip networking, serving runtimes, KV-cache optimization
3. Eval-to-execution platforms: Arena pivoted from leaderboards to Agent Mode with bash recovery and tool hallucination metrics

The counter-thesis worth taking seriously: if a frontier lab posts a genuine reliability step-function in the next two quarters, the compression argument dies on the spot. Track GPT-6 and Claude 5 timelines as thesis-breaking events, not product launches.

Sources Disagree On Timing

Direction (compression) is consensus. Speed is not. The aggressive read says Anthropic's public multiple forces a repricing within 90 days of pricing. The conservative read says enterprise switching costs and distribution lock-in give closed-model names 4-6 quarters of margin buffer before it bites. Both are investable views. They size very differently.

17 Million Agent PRs and the End of Standalone

Two data points from GitHub's CPO are worth sitting with. 17 million agent-generated PRs in March 2026 alone, with the curve steepening after the December 2025 model upgrade that finally made agent-authored PRs land cleanly. And Copilot moved to usage-based billing on June 1. In the same week, OpenAI folded Codex into ChatGPT, which is the Teams-versus-Slack play applied to coding distribution.

The surge flowed to the incumbent, which is the part that matters. GitHub's 630M monthly visitors and the Microsoft channel turned the December agent-PR step-up into ~3x baseline growth. Standalone copilots that raised Series B at 2025 multiples now have to explain, with a straight face, why the same wave did not compound to GitHub instead.

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 Bundling Kill Zone

OpenAI putting Codex inside ChatGPT means the default coding surface for hundreds of millions of users is no longer a separate purchase. Cognition pitching itself as the 'Switzerland of AI Agents' tells you the agent layer is fragmenting rather than consolidating. Read together:

• Standalone coding tools face 15-30% valuation compression on bundling risk, and not in eighteen months — now
• Acceptable survivor defenses: deep workflow integration, enterprise switching costs, IDE-native distribution, agentic depth
• 'Better autocomplete' was never a moat, and the last twelve months proved it

The working window for a standalone tool to demonstrate it is a product company rather than a feature company is roughly eighteen months. Most will not clear that bar. Some of them know it.

Where the New Alpha Lives

The AI FinOps wedge is the cleanest greenfield. Enterprises running Copilot plus Cursor plus Claude Code plus an internal model want neutral-layer cost observability, budget guardrails, and routing optimization, and they want it from someone who is not also selling them the model. GitHub's Chronicle validates the demand and is platform-locked, which is the opening. The Datadog/Cloudability analog for AI dev tooling is mostly pre-Series A founders today.

Semantic Routing: The Architecture That Matters

GitHub's internal shift to semantic routing plus small-model tiers — MAI Code One Flash for trivial work, frontier models when it actually matters — quietly compresses the unit economics of anyone still pricing as if every token deserves a frontier call. This architecture is becoming default, which means the picks-and-shovels position is the routing and orchestration layer itself. Companies building cross-agent governance and cost optimization at that abstraction sit structurally well. The counter-thesis is that the hyperscalers absorb this layer too. That is possible. It is also what people said about observability in 2014.