The AI Margin Window Is Open; Course Hoarders Are Stuck at n = 0
The cost of a product attempt has fallen so far that a solo programmer can run a portfolio of tries on free tiers and small subscriptions; the scarce skill is asking people for money, not learning more AI tools. Most tries fail, but a non-zero success probability times many attempts beats a guaranteed zero from the sidelines.
AI's current margin window is a time lag: production costs are collapsing while customer price anchors stay stuck at pre-AI labor rates. Model capability that cost $60 per million tokens in late 2021 costs $0.06 by the end of 2024, and a16z measured inference costs falling about 10x per year. That spread is what solo developers can capture; payment data shows it is being captured. Stripe's top 100 AI companies reached $1M ARR in a median 11.5 months, one-person products such as HeadshotPro run around $3.6M in annual recurring revenue, and China's Kling took in 340 million yuan in Q4 2025 on the same day OpenAI shut down Sora.
The same data is harsher. TrustMRR's verified sample of 10,150 products shows 26.1% never earned a dollar, 49.1% had revenue in the last 30 days at a median of $169, and just 2.4% reached $100K annualized revenue. AI products are not more likely to sell than non-AI products; AI merely changes how cheaply a product can be produced. Translated into repeated-try math, at roughly 2.4% success per attempt, ten tries give a 21.6% chance of at least one hit, thirty tries push past 50%, and zero tries is exactly zero.
The operational advice is simple: pick a narrow problem people already pay for, build it on free tiers and cheap subscriptions, charge from day one, and check price anchor, repeat purchase, and crowding before starting. Course hoarding is the documented losing alternative, with only 3.13% of edX registrations producing certificates while paid cohorts finish at 44-56%.
The price of a failed attempt has dropped far more than the odds of success have risen; the winning behavior is to multiply attempts, not to raise skill first. Repeated-attempt math turns a 2.4% per-try hit rate into a coin flip over 30 tries.
Course buying is closer to paying for the feeling of progress than for progress. The edX completion gap between unpaid and paid cohorts suggests having a financial stake drives follow-through more than course quality does.
The Stanford gap for workers aged 22-25 appears alongside no overall unemployment, so the damage is concentrated in entry-level hiring, not layoff waves. AI lowers the price of codifiable labor while the market for judged, accountable work holds.
Zero-revenue rates for developer tools (69%) are the strongest signal in the data: the crowd of makers is largest where the willingness to pay is weakest, so distribution and niche access matter more than building ability.