AI's Economics Don't Make Sense

The article argues that AI subscription and compute economics are structurally broken, highlighting Microsoft’s GitHub Copilot shift to usage-based pricing on June 1, 2026 and claiming Anthropic users can burn $13-$30 per day, or $150-$250 per developer per month in enterprise use. It says OpenAI may need $852 billion in revenue and funding through 2030 to meet compute obligations, while Oracle’s Stargate buildout requires roughly $150 billion more to finish and carries significant debt and cash flow pressure. The piece is highly critical of AI business models, pricing transparency, and management execution across OpenAI, Anthropic, Microsoft, and Oracle.

Analysis

The market is starting to reprice AI from a software-margin story into a working-capital and financing story. Once usage-based billing becomes the norm, demand elasticity matters far more than top-line logo count: heavy users will self-ration, casual users will churn, and enterprise procurement will push back on open-ended spend. That shifts value away from user-interface wrappers and toward the capital providers, power owners, and the few vendors with enough balance-sheet strength to survive lower utilization and more volatile take rates.

The second-order loser is not just the AI labs; it is the entire financing stack behind them. If token pricing resets lower utilization assumptions, the marginal economics of new data-center builds deteriorate fastest for highly levered colocation and single-tenant assets, where even a short delay can wipe out a year of equity returns. That creates a reflexive loop: tighter billing reduces demand, weaker demand slows take-up, slower take-up worsens debt coverage, and weaker coverage raises the cost of capital for every participant downstream.

The most interesting near-term catalyst is not an AI demand collapse, but a disclosure event: guidance, covenant language, or a quarterly update that forces investors to see token burn as opex rather than “innovation.” That would matter most for Oracle and Microsoft because they sit closest to the revenue-recognition illusion and the infrastructure commitments. On a 3-6 month horizon, the risk is less that AI usage vanishes than that budgets get capped, creating a roll-over problem for names whose valuations assume perpetual capacity expansion.

Contrarianly, the selloff in the most obvious beneficiaries may be too linear. Hyperscalers and chip vendors can survive a pricing reset because they have multiple end markets and negotiating leverage; the fragile part is the private-credit-funded edge of the ecosystem and the single-customer projects built on heroic growth assumptions. The cleaner short is therefore not ‘AI’ broadly, but the capital-intensive enablers whose equity is pricing in full utilization, rapid commissioning, and unbroken customer solvency.

Key Decisions for Investors

• Short ORCL on rallies over the next 1-3 months; use a defined-risk put spread into the next earnings cycle. Thesis: the market is underestimating how quickly AI capex turns from growth engine into balance-sheet drag if utilization or customer payments slip.
• Buy MSFT 6-12 month put spreads or put LEAPS as a hedge against Copilot monetization friction and enterprise AI budget pushback. Risk/reward improves if management has to acknowledge usage-based pricing pressure or slower attach rates.
• Pair long NVDA / short CRWV for 3-6 months. NVDA keeps pricing power across customers, while the neocloud model is most exposed to utilization, funding, and contract-renewal risk if demand normalizes.
• Short APLD on strength, with a 6-9 month horizon. The market is likely too sanguine on levered data-center builds whose returns depend on perfect occupancy and cheap refinancing.
• Hedge a broad AI basket with a small short in CRM or UBER against token-based pricing pushback. These are the names most likely to face ROIscrutiny from customers and boards once AI spend moves from experiment to audited expense.