The More AI Costs Fall, The More You'll Spend
Inference costs are falling but don't bet on that to save SaaS margins
The numbers coming out of the SaaS industry tell a troubling story. According to the 2025 SaaS Benchmarks report, gross margins for early-stage companies dropped nearly 10 percentage points year-over-year, likely driven by AI costs. Traditional SaaS companies targeting 75-80% gross margins are now seeing those numbers compress to 60-65% as they integrate AI features.
Silicon Valley’s prevailing wisdom goes something like this: Yes, AI features are expensive now, but inference costs are falling dramatically. OpenAI’s GPT-4 launched at $30 per million input tokens in March 2023, and by 2024, GPT-4o cost just $5 per million tokens—an 83% reduction in 18 months. Just wait. The unit economics will work themselves out.
They won’t. And the reason has everything to do with a principle discovered in 1865 by an English economist studying coal consumption.
The Jevons Paradox Primer
William Stanley Jevons observed something counterintuitive about steam engines in his 1865 book “The Coal Question.” As James Watt’s improvements made steam engines dramatically more efficient at burning coal, Britain’s coal consumption didn’t fall—it exploded. British coal consumption tripled by 1900 even as steam engines became vastly more efficient.
The mechanism is straightforward: when something becomes cheaper and more efficient, we don’t just do the same amount of it for less money. We find entirely new uses for it, expanding consumption faster than efficiency improvements can offset.
This isn’t mere economic theory. We’ve watched it play out across technology: computing power became millions of times cheaper, yet we spend more than ever on computing. Data storage costs fell 99.9%, consumption exploded. Faster internet didn’t reduce internet bills. It enabled streaming 4K video.
Now it’s happening with AI inference.
