AI labs spend huge sums building models that get copied cheaply and lose value fast — so why doesn't software's old payback math apply?

This explores why AI labs have less time than classic software companies to earn back what they spend building a product. The corpus has no paper that compares the two directly. Read together, though, several notes explain why the window is short. Traditional software has a well-known advantage: you build it once, it keeps selling for years, and each extra copy costs almost nothing to deliver. AI labs lose both parts of that advantage at once.

The first problem is that a frontier capability starts losing its value almost as soon as it launches. Epoch AI tracked five benchmarks and found that the price of a given level of performance falls about 66% per quarter right after it becomes state of the art. That decline slows to about 32% per quarter within two years Do AI performance prices fall faster when capabilities debut?. Their explanation is that a lab briefly holds a near-monopoly on a new capability, and competitors erode it quickly. In practice, a lab's best chance to charge premium prices lasts a few quarters, not the multi-year life of a typical software release. Efficiency gains speed this up: compact models trained on real multi-step work can now perform close to much larger ones at a fraction of the cost Does model efficiency matter more than peak capability for real work?. That makes it harder for any lab to defend a premium.

The second problem is that the costs never stop. At Anthropic, Minimax, and Z.ai, compute for both research and serving customers (inference) makes up 54–62% of operating expenses, while staff costs stay below 25% Do compute costs really dominate AI company budgets?. Software companies mostly pay people to build a product once. AI labs pay for compute every time a customer uses the model, and then pay again to train the next model before the current one's prices have collapsed. That is why these labs spend two to three times their revenue. Only Anthropic projects profitability, by 2028, and the other two expect to stay unprofitable indefinitely Do AI labs spend far more than they earn?.

The same squeeze shows up further along the supply chain. Mercor, which sells expert training data to labs, reports $2 billion in gross annualized revenue. Because 60–70% goes out to contractors, it keeps only about $600–800 million. That pressure is pushing it to build packaged products instead of acting as a pure labor broker Can AI data companies sustain margins beyond labor payout ratios?. Even training data, which you might expect to work like a reusable asset, behaves more like a recurring service cost.

The less obvious point is that "repayment window" may be the wrong way to frame it. A frontier model works less like software and more like a perishable good. It is worth the most on the day it ships and loses value quickly as competitors catch up. The labs that survive may be the ones that recover their costs within that short period at the top, or that move their revenue onto something that keeps its value longer than the model itself.