20VC: Lovable CEO Anton Osika on $120M in ARR in 7 Months | The Honest Truth About Defensibility and Unit Economics for AI Startups | The State of Foundation Models: Long Grok, Short OpenAI, Why | Replit vs Lovable vs Bolt: What Happens

What It Covers

Anton Osika, CEO of Lovable, discusses scaling from zero to $120M ARR in seven months, AI startup defensibility challenges, foundation model competition dynamics, and building a generational European tech company through extreme execution velocity.

Key Questions Answered

• •Revenue composition: Lovable's $120M ARR splits 80% complex application builders, 10% enterprise prototyping, 10% hobbyist websites. Enterprise segment grows fastest as product leaders use Lovable to build working demos instead of documents, fundamentally changing how companies validate product ideas before engineering investment.
• •Model provider economics: Majority of paid usage revenue passes through to Anthropic and OpenAI today, but margin expansion comes through platform lock-in as users accumulate value. Future revenue shifts from build-time compute costs to subscription retention once users establish their technical infrastructure on the platform.
• •Foundation model strategy: Lovable uses complex agentic chains mixing fast small models with Anthropic for code writing and GPT-5 for hard debugging. Building for tomorrow's model capabilities rather than optimizing current ones enables faster product iteration as AI advances monthly with completely different capabilities.
• •Defensibility framework: Early-stage AI startups should ignore defensibility and execute like chickens shot from cannons, flapping faster than competitors. Defensibility emerges later through platform value accumulation where users create so much on the system they cannot leave, not through initial technical moats or model optimization.
• •Talent assessment methodology: Hire for slope over current capability by evaluating conversation dynamism and learning rate. Seek candidates who demonstrate extreme trauma or masochism, indicating resilience for startup intensity. Video camera test asks what their actual past work performance looked like, not resume achievements.