SpaceX Prices at $1.75T in Largest IPO Without Passive Bid

The Audit That Changes the Math

Princeton's ICML 2026 reliability audit landed this week, covering GPT 5.5, Gemini 3.1 Pro, Gemini 3.5 Flash, and Claude Opus 4.7. The finding is that the newest frontier models are not meaningfully more reliable than the ones they replaced. Another year of capex bought models that fail in the same ways, just more fluently.

This is probably wrong as a one-quarter call, but the audit reads as a pricing-power argument rather than a benchmark one. If reliability is the enterprise buying criterion and reliability is not advancing, the premium for closed-model API access erodes a little with every open-weight release that matches on capability. The counter-thesis is that Princeton captured a temporary plateau. Possible. Not what the data says.

Open Weights Are Now At the Gate

What converged this week is worth keeping in one place:

• MiniMax M3 shipped a 1 million token context window, open weights.
• Gemma 4 QAT runs full multimodal in roughly 1GB on a laptop.
• Ideogram 4.0 is a 9.3B parameter DiT, nf4 quantized, fitting on a single 24GB GPU.
• Kimi K2.5 and GLM-5 are Chinese open-weight models posting frontier-adjacent agentic scores.
• NVIDIA Nemotron 3 Ultra is now deployed by Perplexity for the Pro and Max tiers.

The frontier ceiling is sticky, the open-weight floor is rising into it, and AI capex sits at 0.8% of US GDP. Cost routing is now a first-order business problem. Structural, not a news cycle.

What This Does to Multiples

The capital-allocation implication is direct, or rather, the more interesting version of it is. Closed-model API companies currently trading at 80-120x ARR in private markets want re-underwriting against a world in which the following are simultaneously true.

1. Reliability stays flat for twelve months, which Princeton now suggests has already happened.
2. Open-weight substitutes reach roughly 80% parity on enterprise tasks, which this week's releases suggest is the current state.
3. Enterprise willingness-to-pay compresses toward commodity infrastructure margins, which is what tends to happen when the first two are true.

If the plateau holds, the target range resets to 50-70x ARR. Still premium. A 30-40% markdown from where most late-stage rounds priced. The mirror trade is that infrastructure and tooling layers benefiting from inference volume regardless of which model wins deserve a 1.5-2x multiple uplift.

Where the Value Migrates

Cloudflare's AI Gateway launch (spend caps, model fallbacks, budget enforcement) is the tell. When the network layer starts selling protection from a customer's own model bill, the bottleneck has moved one layer down. The economics they cite: rerouting 10% of a $10M AI bill saves ~$1M. That is a product with immediate ROI at enterprise scale.

Google's TPU 8t/8i separation validates the inference-compute split, the first time a hyperscaler has formally divided training and inference into distinct silicon SKUs. Inference is now its own investable sub-sector with its own unit economics and its own exit comps, which is what the rest of the buyside should care about.

Counter-thesis worth respecting: if a frontier lab posts a genuine reliability step-function in the next two quarters, the compression argument dies and the access-moat names re-rate up. The Princeton audit is one data point, not a permanent verdict. It is also the first peer-reviewed confirmation of what the market has been whispering for two quarters.

The first public comparable, and what it marks against

Anthropic filed for IPO this week. Whatever it prices at becomes the first pure-play frontier-lab public comparable, which is the number every AI app-layer company and every LP quarterly report gets marked against from now on. The private AI market has spent three years not disclosing unit economics. Quarterly disclosure begins.

There are three ways this plays out, and ranking them by probability is more useful than listing them.

1. IPO prices well, private rounds reprice upward, the capital cycle extends another year. The sell side is already writing this one.
2. IPO prices badly, private marks come under pressure, the late-stage secondary desks do the arithmetic they have been avoiding. The numbers, such as we have them, suggest this version.
3. IPO gets pulled, which tells you everything the bankers learned on the roadshow. Most informative, least likely.

The context that makes this harder

Anthropic walks into the same hostile tape as SpaceX. Rates are repricing higher, which matters more for an unprofitable name with no S&P 500 passive bid waiting behind it, and the Nasdaq dropping 4.18% in a session this week did not improve the optics. The safety branding provides some cushion, since enterprise procurement is increasingly safety-gated, but the absence of profitability means no index inclusion for 12+ months minimum.

Meanwhile Buffett's $10B Alphabet position tells you value capital has crossed over into megacap AI. That is confirmatory, not leading. When Berkshire buys the trade, the easy alpha is already booked elsewhere. The Suno mark at $5.4B says the vertical AI layer is stratifying, or rather, the data-moat winners are pulling away from the wrappers at a rate that should worry anyone who funded a wrapper at a data-moat multiple.

The portfolio repricing that follows

Within ninety days of Anthropic pricing, every AI app-layer multiple in private markets gets either validated or exposed. The discipline required is simple, which is not the same thing as easy.

The Anthropic pause call, publicly requesting a global AI freeze, is best read as IPO positioning rather than policy. Own the safe-enterprise-AI lane while OpenAI owns scale. Two go-to-market motions, two investor bases.

Quarterly disclosure of unit economics begins now. That is usually the interesting part of the cycle.

The less obvious move: the AI-security wedge. Anthropic's Mythos product being labeled a 'budget buster' in enterprise confirms AI security is becoming a separate budget line with room for cost disruptors, and the Meta Instagram breach via AI chatbot social-engineering is the first marquee AI-as-attack-surface incident. The security wedge gets repriced on the next breach, not before.