Leaked IPO docs: Anthropic tempts investors with existential risk warning

Anthropic's leaked IPO prospectus reportedly projects $518 billion of cloud, compute and infrastructure obligations, while the company recorded an operating loss exceeding $8 billion last year despite revenue rising 12x to roughly $4.6 billion. The article highlights concerns that AI investment economics are unsustainable, citing Oracle's reported force-majeure notice related to power delivery for a New Mexico datacenter tied to its planned $300 billion OpenAI capacity buildout. With AI capital expenditure projected at $31.6 trillion through 2050 and current investment estimated at about one-third of US economic growth, a reversal could create broad economic and technology-sector downside.

Analysis

The investable issue is not AI demand alone but the conversion of contracted demand into energized, financeable capacity. ORCL's valuation now carries disproportionate execution sensitivity to power delivery, construction milestones and customer credit quality; a delayed campus can defer revenue while fixed equipment, lease and financing costs continue to accrue. The market should treat any contractual escape clause as a warning that signed AI capacity is not equivalent to near-term recognized cloud revenue.

Over the next 1-3 months, ORCL is vulnerable if quarterly RPO growth remains strong but management cannot provide a clear split between contracted capacity, powered capacity and billable capacity. That distinction could widen the valuation gap versus MSFT, AMZN and GOOGL, whose broader cash flows and existing infrastructure reduce dependence on a small number of megaprojects. OWL's direct exposure is not yet quantifiable: the relevant variables are the size of its capital commitment, remedies available under the contract, collateral and whether any impairment affects fee-bearing AUM or realizations.

The second-order beneficiary of buildout friction is not necessarily GPU vendors but incumbent power owners and transmission-linked infrastructure. CEG, VST and utilities with available generation/interconnection optionality can retain pricing power even if a developer postpones a single site; conversely, VRT and ETN face timing risk if data-center starts slip, though a broad cancellation cycle—not an isolated power delay—would be required to impair their 6-18 month demand outlook. Consensus is likely too quick to extrapolate a project-level constraint into an AI-demand collapse, but too complacent about the capital-intensity discount that private model developers may face at eventual public-market valuation.

A material de-risking would be independently verified energization dates, customer prepayments, and evidence that cloud gross margin scales despite rising depreciation and power costs. Falsifiers for the bearish ORCL relative thesis are sustained cloud revenue acceleration accompanied by stable or expanding operating margin, plus disclosure that delayed capacity has been redeployed to other customers without incremental capital needs.

Key Decisions for Investors

• Initiate a 3-6 month relative-value position: short ORCL / long MSFT in equal dollar amounts. The thesis is that ORCL's multiple is more exposed to construction and capacity-conversion slippage; cover if ORCL reports cloud growth acceleration with no deterioration in operating margin or gives firm energized-capacity milestones.
• Do not establish an OWL directional position until contract terms, commitment size and balance-sheet/fee-AUM exposure are disclosed. Set an alert for a disclosed impairment, litigation, redemption impact or fee-related earnings revision; those would justify reassessing a short.
• Maintain a selective 6-12 month long basket in CEG and VST rather than broad AI hardware exposure, sized modestly because power-price and regulatory risk remain high. Add only on evidence of incremental long-duration data-center contracting; exit if power-market pricing weakens materially or new capacity/interconnection approvals erode scarcity.
• Avoid chasing VRT and ETN on AI-capex headlines until construction starts and power availability are reconciled. A broad supplier short is premature absent evidence that multiple hyperscalers are cutting orders, rather than shifting project timing.