The short seller’s argument nobody on the coming mega IPO roadshow wants you to make

Anthropic has confidentially filed for an IPO, while OpenAI is reportedly preparing a draft, alongside a cited $965 billion Anthropic valuation and $852 billion OpenAI valuation. The article argues that frontier AI revenue is overstated versus actual demand, pointing instead to monetization in infrastructure, COBOL modernization, fraud detection, and AI crop models, with Bain warning AI needs $2 trillion in annual revenue by 2030 to justify compute spend. It also cites Oracle's $248 billion in data-center leases and falling inference prices as margin headwinds for model-layer businesses.

Analysis

The market is still pricing the AI value chain as if model ownership is the toll booth, but the more durable monetization usually shifts one layer lower once pricing gets competitive. If inference continues to deflate 30-50% annually, the enterprise software layer becomes a margin pass-through business while infrastructure, workflow integration, data governance, and compliance remain sticky. That is structurally bearish for pure model economics and selectively bullish for the picks-and-shovels names that sit inside regulated workflows.

The biggest second-order winner is not the flashy frontier lab acquirer; it is the software and data vendors that can embed AI into legacy systems where switching costs are high and ROI is measurable. IBM is the clearest near-term beneficiary of modernization demand, but the reaction risk is that the market extrapolates too much from a one-off rerating and ignores how slowly mainframe replacement actually converts into revenue. Oracle has a different vulnerability: long-duration data-center commitments look attractive only if utilization and pricing hold; if model margins compress faster than expected, lease obligations become leverage rather than moat.

The underappreciated upside is in enterprise data plumbing and workflow orchestration where every AI deployment must connect to messy internal data. That favors names like ServiceNow, Informatica, Salesforce, and even Cisco/Akamai-style infrastructure providers more than the labs themselves. Emerging-market AI winners are likely to be fintech rails and fraud/identity vendors, but that thesis is still several quarters from showing up in public equity fundamentals; the better trade is to own the enablers before revenue inflects, not the end-demand story.

Catalyst timing matters: the IPO window can sustain the narrative for weeks, but the real check comes over the next 2-3 earnings seasons when customers quantify payback. If enterprise CFOs keep treating AI as discretionary opex, the re-rating in frontier labs could reverse quickly; if modernization and fraud use cases show up in budget line items, the market will rotate down the stack rather than abandon AI altogether.

Key Decisions for Investors

• Long IBM vs short ORCL, 3-6 month horizon: IBM benefits from legacy modernization and consulting-led AI adoption, while ORCL is more exposed to valuation risk from long-dated capacity commitments if pricing compresses faster than utilization improves.
• Buy INFA on weakness into the IPO tape, 1-2 quarters: the market should reward whichever layer becomes indispensable to AI deployment; risk/reward is favorable because data integration is a toll road, not a model-cycle bet.
• Add NOW / CRM basket against frontier-lab sentiment, 6-12 months: workflow ownership and seat expansion are more durable than pure model economics; use a relative-long structure to avoid beta to the AI narrative unwinding.
• Long CSCO and AKAM as infrastructure hedges, 3-6 months: if AI revenue disappoints at the model layer, traffic, delivery, and network optimization remain monetizable; these names offer lower multiple risk than software-heavy beneficiaries.
• Avoid chasing fresh AI IPOs at filing/roadshow stage; if you want exposure, use call spreads only after first 1-2 quarters of public data, since the setup is likely to be narrative-rich but execution-poor.