Build moats, not demos, in AI
AI rewards strategy, not flashy features: costs stay real, commoditization moves fast, and hesitation gets punished. This playbook lays out a 4D framework—direction, differentiation, design, deployment—and the 2Ps of pricing and positioning to turn AI from a costly demo into a compounding moat. Case studies from Chegg, Jasper, Duolingo, Perplexity, Midjourney, and OpenAI show how winners model unit economics early, embed AI where users already work, and price to align value with inference spend.
Points clés
- Chegg lost about 90% of its valuation after moving too slowly on AI as students shifted to ChatGPT.
- Jasper raised $125M at a $1.2B valuation but stalled as “AI wrapper” economics and commoditization drove churn and price cuts.
- Duolingo’s rushed AI tutor rollout triggered reputational damage, user churn, and a loss of 300,000 followers within weeks.
- ChatGPT hit 100M users in two months, then introduced ChatGPT Plus at $20/month to contain compute costs.
- AI unit economics differ from SaaS: each query incurs inference costs (tokens/GPUs), so scale can increase costs without design efficiencies.
- Example cost model: at $29/user/month and 500 queries costing $0.002 each, per-user inference is ~$1; at 100,000 users and 50M queries/month, inference hits ~$10M/year.
- Perplexity reduced token spend with retrieval-augmented generation, cutting costs and improving speed and trust via citations.
- Midjourney scaled with hybrid tiers ($10–$60/month) and GPU minute caps, ending free trials to prevent runaway compute burn.
- The 4D framework: direction (choose data, distribution, or trust moat), differentiation (survive commoditization), design (adoption plus cost efficiency), deployment (pricing, infra, team).
- The 2Ps of pricing and positioning: usage-based (OpenAI API), outcome-based (pay per result), seat-based (risk without caps), and hybrid (e.g., Midjourney) to align revenue with inference costs.
À retenir
Start with a moat, not a magic trick: pick data, distribution, or trust and double down like your GPU bill depends on it—because it does. Price like a grown‑up (caps, usage, or outcomes), design for caching and model routing, and build evals before Twitter does it for you. Do these and you’ll scale; skip them and you’ve built a very generous charity for your users—and your cloud provider—congratulations.
Sources
