• Two questions about any software product.

• The maker spends $1M on inference. A competitor, or a customer with a weekend and an API key, spends $100. Does the $1M version become meaningfully better? Or does $100 get you 95% of the way there?

• Is the product already past what its users need, or is there still valuable territory left to build into?

• Where $1M compounds: Microsoft Copilot sits on the entire Office graph. Decades of email, calendar, documents, org charts. A dollar of inference against that data produces something a dollar against a cold start can’t replicate. Salesforce, where the AI that predicts which deal closes next week only works if you have years of deals that didn’t. The intelligence gets better with data. The data took a decade to accumulate. These companies charged for features and gave away data accumulation for free. Turns out they were underpricing the thing that actually mattered.

• Where $100 is enough: Canva. Basic email tools. Simple dashboards. The product does more than most users touch. The moat was complexity and switching costs, not genuine capability. AI doesn’t need to match the full product. It needs to match the 20% that delivers the value.

• Most SaaS is overshooting. The feature bloat that justified pricing tiers for a decade is now the surface area that gets cloned for free.

• The mistake is treating AI as uniform pressure on software. It strengthens where intelligence compounds with proprietary data. It dissolves where the product was already past the point users cared about. Every software company spent a decade calling itself a data company. AI is the audit.