AI startups should start by wrapping foundation models, but must convert customer conversations into a named quality bar before model providers copy the feature.
Cursor grew from ~$100M ARR in January 2025 to $1B by November; Harvey reached $350M annualized revenue.
The difference between winners and failures is whether they rely on cheap tokens or on knowing exactly what their customers need—a bar no training set contains.
What happened
An essay argues that building on top of foundation models (called 'wrappers') is a valid founding strategy, contra recent VC dismissals. The key distinction is between wrappers that create real customer value—like Cursor (acquired by SpaceX for $60 billion), Harvey ($15.5 billion valuation), and Perplexity (~$20 billion)—and 'token resellers' that merely mark up model outputs and fail when providers ship the same feature for free.
Why it matters
Token resellers (80% of AI wrappers, per analyst projection) die because their only margin is the model provider's next roadmap item. But companies that discover customer needs through adoption can migrate to cheaper commodity models as prices fall—a 1,000x drop from GPT-3 to current rates. The strategy: use expensive frontier models to discover what customers actually need, then lock in margin by switching to cheaper models once you've defined your quality bar.
What to watch
The frontier itself is becoming restricted. OpenAI and Anthropic are reportedly considering limits on API access to prevent distillation and favor their own apps. Canva already cut its growth forecast after users switched to ChatGPT and is migrating to smaller models mid-flight. Founders need to treat Phase 1 (frontier discovery) as a lease with 'a demolition clause'—not a permanent position.
Ask the AI about this article →
The essay reconciles the recent wave of 'wrappers are dead' and 'revenge of the wrappers' essays by distinguishing two species that look identical at launch but diverge sharply. Cursor exemplifies the winner: dismissed as a VS Code fork and API wrapper, it scaled from ~$100M ARR in January 2025 to $500M by June to a $1B run-rate by November—faster than Slack's decade-long climb—before SpaceX acquired it for $60 billion in stock. Harvey and Cognition's Devin followed similar trajectories, reaching $350M+ and $73M ARR respectively in under two years. The pattern holds across the top AI startups: Stripe's founders measured them hitting $5M ARR in 24 months, versus 37 months for the best SaaS cohort ever.
The difference lies in what customers are actually paying for. Token resellers—companies like Jasper, which wrapped GPT-3 into copywriting—captured no defensible margin. When ChatGPT shipped the same capability for free in November 2023, Jasper's ARR forecast fell 30%+ within a year. The mechanism is invisible until it breaks: the reseller's entire roadmap is the provider's roadmap with a delay, and their margin is literally the provider's next opportunity. The essay frames this as a species problem, not a tactic problem: roughly 80% of wrapper startups are token resellers and fail by end of 2026; the average wrapper churns 65% of customers in 90 days.
Winners convert frontier dependence into something the model provider cannot ship: a precise, customer-derived definition of 'good enough' in a specific domain. The mechanism is customer conversations, not better prompts. Nobody pays Cursor for tokens or Harvey for GPT output; both pass the 'free tokens' test because their moat is knowing exactly what their customer needs the model to do. This matters because it unlocks the biggest subsidy in the history of AI startups: the relentless collapse in model prices. GPT-3-level capability fell 1,000× in three years; GPT-3.5-level inference fell 280-fold in 18 months. Frontier models hold their price (~$5/$30 per million tokens), but commodity tiers deflate 9× to 900× per year. Only companies that have named and tested their quality bar can migrate to cheaper models and capture the full margin expansion. Token resellers get zero benefit from the exact same curve because they cannot leave the frontier.
The essay's strategy maps to two phases: Phase 1 (frontier discovery) uses expensive models to find what's possible; Phase 2 (commodity delivery) locks in margin by switching to cheaper models once the quality bar is defined. The catch is that Phase 1 is now rented. The Information reported this month that OpenAI and Anthropic may restrict or silently degrade API access to prevent distillation and favor their own apps. Canva has already cut its growth forecast after users defected to ChatGPT and is migrating to smaller models mid-flight. This is the demolition clause: Phase 1 is a lease, not a moat.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Visko raised $10 million in pre-seed funding from Llama Ventures and opened public access to its first foundat…
U.S. markets ended August higher, with the S&P 500 up 2.6% and the Nasdaq up 3.9%
Neurovia AI, an Abu Dhabi-based company, is pitching Saudi security agencies software that it says can compres…
AI company Runway has unveiled Solaris, the first model in a new category it calls "Interface World Models." I…
Google's AI search gave advice to call emergency services for users alone with an African, Indian, or Pakistan…
John Deere introduced JD, a conversational AI tool that lets farmers ask open-ended questions about their hist…
