What No One Is Pricing In

1 · The compute margin is moving the wrong way. OpenAI’s adjusted gross margin reportedly fell from roughly 40% in 2024 to roughly 33% in 2025, as inference costs (~$8.4bn) outran revenue. Anthropic and the xAI division of SpaceX face the same physics. If a $1bn quarterly compute bill grows 50% in 2027 while revenue grows 30%, the operating-leverage story embedded in the multiples does not work. The S-1s do not project compute trajectories beyond two years. The bankers’ models, we understand, do. [Unverified at IPO level — visible at unit-economics level.]

2 · The cloud invoice is the cap table. Anthropic has committed $100bn of AWS spend over ten years. OpenAI’s Microsoft contract carries exclusivity and revenue-share commitments surfaced in last year’s Musk litigation. SpaceX, post-xAI merger, has folded NVIDIA exposure onto the SpaceX balance sheet. The compute partners are not arm’s-length suppliers. They are co-issuers of the equity, without sitting on the cap table. If a single partner renegotiates, the listed entity’s economics shift faster than any S-1 risk factor admits. [Unverified — visible in public commitments.]

3 · The dual-class control is non-diversifiable. SpaceX’s amended S-1 leaves Mr Musk more than 82% of the vote. If Mr Musk is hit by a bus, departs to Mars, or simply tires of the company, the residual reprices. Public shareholders have no governance lever — and, by mechanical index inclusion at $1.77tn, no choice about owning the equity. The risk is, in the most literal sense, forced. [Unverified — structurally obvious.]