2026-10-10 补充校核：此文财务数字属于所引分析的估算，未经本次财报逐项核验；“40–70倍”须保留计算量/成本口径，“补贴游戏结束”属于预测。

AI’s Affordability Crisis

Core Thesis

AI platforms run massive subsidies (40-70x) to create dependence, then switch to token-based pricing causing 7x+ cost increases.

Key Evidence

• OpenAI 2025: $13.07Brevenue,$34B costs, $20.92Bloss(netloss$38.53B including conversion costs)
• OpenAI spent 44% of revenue on Sales & Marketing ($5.73B)
• Anthropic burn-through: $200/mosubscriptionbuys$8K worth of compute tokens
• Enterprise customer quoted: “Our spend went up 7x the first day” after token pricing
• Microsoft shifting from Claude Code to Copilot CLI by June 30, 2026 to cut costs
• Nvidia VP: “For my team, cost of compute is far beyond cost of employees”
• Hyperscaler AI investment returns (FT/Panmure Liberum, zero costs assumed): Microsoft -9.2%, Alphabet -15.7%, Meta -28.8%, Oracle -35.6%, Amazon +7.2%

Implications for Hermes

1. Local-first architecture insulates from token pricing volatility
2. Token efficiency (context compression, attention innovations) = competitive advantage
3. Chinese model alternatives (DeepSeek, Qwen) at fraction of cost — evaluate for non-critical paths
4. The subsidy game is ending — plan for realistic token costs