AI has a subsidization problem

Theo (T3.gg) argues AI’s subsidization era is ending as Google cuts Gemini free tiers and Anthropic’s $200/month plan now covers up to $5,000 in compute.

• Claude Max plan subsidy grew from 10x ($2,000 compute for $200) last year to 25x ($5,000 compute for $200) — Anthropic is accelerating subsidies to pull users away from Cursor.
• GitHub Copilot student plan is dropping GPT-4, Claude Opus, and Sonnet; Google is locking free users off Gemini Pro models entirely.
• Token consumption per prompt has risen ~10x since 2023 due to tool calls, reasoning tokens, and large context — cost-per-session outpaced per-token price drops.
• The 10th-to-90th percentile gap in prompt cost is roughly 400x; a single T3 Chat message can cost $1 against an $8/month subscription.
• YouTube ad revenue for a channel with 3M monthly views from senior developers is ~$9k/month — Theo uses this to show ads cannot subsidize inference at any realistic scale.
• Chinese companies reportedly ran cheap Claude wrappers (~$10/month) specifically to intercept prompt/response pairs for model training.
• Anthropic sent lawyers to block Open Code’s Claude Max plugin — a deliberate move to eliminate harnesses that let users switch models, since subsidies only pay off if users stay lifetime customers.
• Google subsidized so aggressively that internal dev teams were denied access to Gemini Pro 3.1 because capacity was reserved for free-tier consumer users.

2026-03-25 · Watch on YouTube