Replit CEO: Why the SaaS Apocalypse is Justified & Why Coding Models are Plateauing | Amjad Masad

The Viral Reversal on Coding Education

In March 2025, Masad publicly declared that people should no longer learn to code but instead learn to create, acknowledging that agentic AI enables non-technical founders to build multi-million dollar software businesses without developers.

The 2024 Agentic Unlock

The transformative breakthrough was not generative AI itself—which existed since GPT-3 in 2022—but the emergence of agentic AI capable of maintaining coherent actions over long horizons, requiring extensive infrastructure to manage autonomous execution.

From Billion Developers to Billion Creators

Replit's original 2016 mission to create a billion developers has evolved into enabling a billion creators, removing the coding bottleneck that previously prevented people from turning ideas into software.

Approaching the Asymptotic Plateau

Masad argues coding models are reaching an asymptotic plateau on the S-curve, making it increasingly viable to fine-tune open-source models for specific domains rather than relying exclusively on frontier models that improve rapidly.

The Society of Models Strategy

Replit orchestrates multiple providers simultaneously, using Anthropic as the 'workhorse' for core agent loops while leveraging Google's Gemini for price-sensitive subtasks, maintaining strict optionality across the ecosystem.

The Build-Versus-Buy Calculus

Replit trained models in 2023 that outperformed GPT-3.5, but the gap closed within months when Claude launched, demonstrating that building proprietary models is only viable during specific plateau windows.

Dancing With Model Capabilities

AI product development requires continuous infrastructure adaptation—building guardrails for Agent v2 in September 2024, deleting them for v2 in March 2025, then adding parallel processing for Agent 4—depending on whether autonomy is built into the underlying model.

Performance Trumps Cost Optimization

Companies should prioritize product performance over margins during rapid advancement phases, as premature optimization leads to wrong decisions; cost rationalization only becomes viable when approaching an asymptotic plateau where improvements slow.

Inference Becomes Customer Acquisition

Free token allowances and inference credits have replaced traditional marketing spend, with companies using the 'creative addiction' of agentic development as a viral loop to acquire users, though retention remains the critical challenge.