A developer recently generated buzz on Hacker News by setting AI agents loose to decompile a first-person shooter, burning through an astonishing 500 billion tokens. At true API market rates, a workload that size would cost anywhere between $100,000 and $750,000, yet power users are pulling off these massive compute feats for a tiny fraction of the price thanks to heavily subsidized subscription plans.
For early-stage founders and small-scale backers, this isn't just a fun technical experiment—it's a critical lesson in artificial unit economics. Big AI labs are currently happy to absorb massive compute losses to inflate usage metrics and hook power users, mirroring the early, heavily subsidized days of ride-hailing platforms.
If you're building or funding a startup today, you cannot mistake subsidized compute for a sustainable business model. The moment AI providers turn off the tap and demand real prices, products relying on brute-force token usage will see their margins evaporate unless they deliver undeniable, high-margin value that users are genuinely willing to pay for.