Cost is not the same as defensibility
Stanford’s 2025 AI Index documents a dramatic historical fall in the inference price for GPT-3.5-level benchmark performance. That supports testing cost and capability assumptions. It does not establish that all software, distribution or customer service is free.
Our analytical view: improvements in models can help a business and weaken its differentiation at the same time. Lower inference costs may improve margins while making an answer-generation feature easier for competitors to reproduce.
Run the substitution test
Imagine a capable customer rebuilding your headline feature over a weekend. Which parts of the customer outcome would remain difficult? Consider access to customers, integration with their operations, permissioned feedback, reliable delivery and the cost of switching.
“We have proprietary data” is not a complete answer. Document the right to use it, whether those rights transfer, whether the data improves outcomes and whether the same useful information is readily available elsewhere.
Run the survival test
Model three shocks: inference costs triple; the dominant acquisition channel disappears; the founder becomes unavailable for two weeks. Do not assign invented probabilities. Write down the operational consequence and the cheapest experiment that could reduce uncertainty.
The strongest result may be an unglamorous improvement: diversified lead sources, an export path, a tested fallback model or an exception-handling procedure. Buyers inherit those capabilities, not your enthusiasm about the technology.
Turn assessment into work
The free AI Moat Test separates an opinion from documented evidence. Its score is a prioritization aid, not a valuation or a prediction that someone will acquire the company. Complete one missing evidence task before adding more features.
- Stanford HAI · AI Index 2025: research and development ↗Historical inference-cost evidence; not a prediction that complete software businesses cost nothing to run. Checked 2026-09-15.