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

The Princeton Audit Nobody Can Ignore

Princeton's ICML 2026 reliability audit landed this week with the kind of finding that quietly resets a thesis. GPT 5.5, Gemini 3.1 Pro, and Claude Opus 4.7 are, per peer review, not meaningfully more reliable than what they replaced. Another year of frontier capex bought models that fail in the same ways, more fluently.

Meanwhile the open-weight tier is walking through capability thresholds that were premium API pricing twelve months ago:

• MiniMax M3 — one million token context, open weights
• Gemma 4 QAT — multimodal, runs in roughly one gigabyte, fits on a laptop
• Ideogram 4.0 — 9.3B DiT, fp8/nf4 checkpoints, a single 24GB GPU
• Kimi K2.5 and GLM-5 — Chinese open weights posting frontier-adjacent agentic benchmarks

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 has become a first-order business problem rather than a news cycle.

What This Does to Multiples

Closed-model API revenue multiples should compress from the 80-120x ARR zone into 50-70x. Princeton removes the reliability premium. Open weights remove the scarcity premium. What remains is distribution, safety tuning, and enterprise lock-in. Real moats, narrower ones.

Anthropic's IPO filing is the catalyst that makes this visible, or rather, the catalyst that forces every app-layer startup to mark against a public number instead of a private fiction. Price it well and the follow-on rounds reprice upward and the capital cycle extends. Price it badly and the late-stage secondary market does the unpleasant arithmetic it has been avoiding. Pull it and the bankers will have told you everything the roadshow taught them.

Where Value Migrates

This is probably wrong in one of three ways, but the sources converge on three categories absorbing the displaced alpha:

1. AI FinOps / cost routing. Cloudflare shipped AI Gateway spend limits, budget enforcement, and model fallbacks. Rerouting 10% of a $10M AI bill saves about $1M, and at 0.8% of GDP in infra spend a basis point of optimization is real money. The window before Datadog or AWS absorbs this is 12-18 months. That is the entry.
2. Inference-optimized infrastructure. Google splitting TPU 8 into training (8t) and inference (8i) variants is the validation that inference is its own SKU. Silicon, chip-to-chip networking, serving runtimes, and KV-cache optimization are the picks and shovels.
3. On-prem and edge deployment tooling. Gemma 4 in 1GB and Ideogram on a single 24GB GPU turn on-prem into a 2026 buying decision. Quantization tooling in the Unsloth class and serving runtimes in the vLLM class are the adjacent plays.

And then the Buffett tell, which matters mostly for what it is not: $10B into Alphabet, not Microsoft, not Amazon, not Meta, not any model-layer company. Value capital crossing over is value capital telling you the easy alpha in megacap AI is gone. The positioning is infrastructure and distribution, not frontier capability.

The Platform Won the Agent Surge

GitHub's CPO put two numbers on the table that reorder the AI coding-tools category: 17 million agent-generated pull requests in March 2026 alone, running at roughly three times baseline expectations after a December 2025 model capability jump, with Copilot moving to usage-based billing on June 1. The surge accrued to the incumbent rather than the startups, which is what 630 million monthly visitors tend to do to a category. The platform won.

OpenAI folding Codex into ChatGPT in the same week is the bundling move that standalone coding tools have to answer for, the same mechanic Teams ran on Slack. Any coding-AI startup whose moat thesis does not survive Copilot at usage-based pricing or ChatGPT plus Codex as a freebie is now on an eighteen-month clock to prove it is a product company rather than a feature.

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.

Three New Categories Emerge From the Rubble

The generation layer consolidates, which is the boring half of the story; the interesting half is what the consolidation strands one tier above and below:

1. AI FinOps for engineering (greenfield). Usage-based billing plus token-heavy agent sessions adds up to a CFO problem that did not exist six months ago, and GitHub's Chronicle validates the demand but is GitHub-locked, which leaves enterprises running Copilot plus Cursor plus Claude Code plus internal models needing a neutral cost-observability and routing layer. Call it the Datadog analog for AI dev spend. Most founders here are pre-Series A.
2. Verification layer (underfunded vs. demand). At seventeen million agent PRs a month human code review breaks, which makes agent-native review, AI-aware SAST/DAST and automated PR triage infrastructure rather than nice-to-have. The bottleneck moved from generation to verification and the funding has not caught up.
3. Agent-API ecosystem on GitHub's new primitives. GitHub framed its API layer as evolving toward what it calls AX, or Agent Experience, displacing the older UI and UX paradigms. When a platform owner signals new primitives this loudly, an eighteen-month ecosystem window opens of the kind early Shopify apps or Stripe Connect rode in on.

What Sources Disagree On

One source positions Cognition as the 'Switzerland of AI Agents', a neutral orchestration layer, which implies the agent market fragments and interoperability wins, while another sees coding agents consolidating around 2-3 players with enterprise distribution (Claude Code, Codex, Cursor-class). This is probably wrong, but both can be true at different tiers: enterprise consolidates, mid-market fragments, and the resolution is a barbell that funds neutral orchestrators or vertical full-stack agents and avoids the middle.

The Forward Deployed Engineering signal reinforces it, since FDEs becoming 'the rage' means enterprise AI is in implementation-services phase, not self-serve phase, which compresses gross margins 15-25% for AI-native enterprise SaaS while raising switching costs once deployed. Underwrite accordingly.