The Jevons Paradox: Why Cheap AI Kills Margins
The cheaper the technology, the deeper the hole.
I built a simple Gemini chatbot to help my team draft emails. The goal? Efficiency. We paste an email, AI polishes the response. We stayed in the free tier for exactly three weeks.
Then cost detonated. Not because the tool failed, but because it worked too well. It was so frictionless that my team ran everything through it—routine confirms, quick updates, even internal whatsapp messages.
Query volume didn’t double; it 10x’d. I had to ration usage within a month.
This isn’t poor management. It’s physics.
The Core Misunderstanding: Training vs. Inference
Most entrepreneurs ask: “How do I use AI?” The Operator asks: “How do I use AI without going broke?”
You must understand the two costs:
- Training: The headline-grabbing, city-powering event (Meta’s Llama 3 used 2.75 TWh). You don’t pay for this.
- Inference: Using the model. Every chat, every polish, every summary. This is the silent ulcer.
Inference accounts for 96% of AI energy consumption. It is a variable cost that scales linearly with your team’s laziness.
The Jevons Paradox
You think 10x efficiency means 90% savings. You are wrong.
In 1865, economist William Stanley Jevons proved that as steam engines became more efficient, coal consumption soared. Why? Because efficiency made steam cheap enough to use everywhere.
AI follows Jevons Law. Inference costs dropped 280x in 2025. Did spending drop? No. Google’s token processing jumped from 480 trillion to 1.3 quadrillion in 5 months.
Your marketing team won’t work faster; they’ll generate 1,000 ads instead of 10. Your sales team is not going to go home early; they’ll just spam more leads.
The Trust Gap: Feature, Not Bug
AI doesn’t know facts. It predicts plausible sentences.
I sadly learned this when my AI confidently cited a revoked labor law. My staff caught it, but the lesson stuck: Hallucination isn’t a glitch. It’s the system working as designed.
Modern AI uses Connectionism—pattern recognition, not logic. It reconstructs a “plausible” poem or legal statute. It has no concept of truth, only probability.
The Rule: Treat AI as a creative assistant, never a decision-maker.
The Bubble Problem: Vendor Risk
Even if you master costs, you are building on a bubble.
The AI industry is fueled by circular financing. Nvidia invests in OpenAI, on condition OpenAI buys Nvidia chips. Oracle funds OpenAI on debt.
Depreciation is the ticking clock. Bain & Co estimates $40B in annual hardware depreciation, nearly double the industry’s revenue. When this loop breaks, prices will spike overnight.
The Team Split: Superstars vs. Commodities
AI impacts your team in two ways:
- Commoditization: Routine tasks (copy, summary) become worthless.
- Augmentation: High-judgment experts become superhumans.
MIT researchers found that top scientists doubled productivity with AI because they could filter the garbage. Low performers saw zero improvement.
This creates a barbell organization: Elite thinkers on one end, AI executors on the other. The middle manager vanishes.
The Reasonable Founder’s Strategy
Don’t be rational (chasing the best tool). Be reasonable (chasing survival).
I use a Three Path Strategy:
- Scale Path (Proprietary): Claude/Gemini (₹36k/yr). For coding and complex reasoning.
- Efficiency Path (Open Source): DeepSeek/Kimi (Free/Cheap). For generic drafts and summaries.
- Sovereignty Path (Self-Hosted): Testing local models key customer data.
The Ultimate Leverage: Walking Away
My payroll vendor hiked prices 50% because they knew I was trapped. I vowed never to let that happen with AI.
The best leverage isn’t the best prompt. It’s the ability to switch vendors in 12 hours.
- Build hybrid workflows.
- Verify every output.
- Save 3 months of “vendor failure” cash.
Don’t join the arms race. Be the one who survives the crash.
Your takeaway: AI maximizes optionality, not just capability. Build accordingly.