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

Key points

• Coding-model performance is approaching plateau on the S-curve. Implication: 'when you focus on cost is when you reach a certain asymptotic plateau.' At plateau, fine-tuned/distilled models for specific use cases become economically rational. Intercom's new customer-support model better than the frontier is Exhibit A.
• Society of Models thesis (Masad's 2022 framing, now operational): use whatever model is best per task. Anthropic = core long-running agent workhorse. Gemini = best price/performance for tasks like code search → spawn cheaper subagents. In-house = where data flywheel + plateau intersect.
• 'Agent labs' as a category: Replit, Cursor, Lovable, base44 — start from the user problem, walk back to the technology. Different from frontier labs (which start from training capability and walk forward).
• On Cursor's decision to build their own model: 'In 2023 we trained models that beat GPT-3.5 at coding. But Anthropic was spending $100B on agents — would have been a dumb fight to pick.' Now (2026) the opportunity reopens because plateau + good open-source + your own data.
• Optionality > commitment: model strategy must change every 3-6 months. The cost of being locked in is bigger than the cost of switching.
• Margins are NOT 80% to the model providers. Replit's % is 'way less,' and management mode oscillates: build the best product first, optimise margins later. 'Premature optimisation is the root of all evil' — applied to gross margins.
• On Lemkin's 'inference is the new sales and marketing': agreed. Free Claude Code / Codex tokens drove the late-2025 → early-2026 hype cycle. Agentic development is addictive ('better than social-media addiction — it's creative'), making free-token onboarding work as a customer-acquisition lever.
• Future of product organisations: still engineers (handling AI/ML, infrastructure, embedded, low-level), still product builders (some technical-leaning, some design-leaning, some product-leaning) — but the role names blur. The skill is figuring out what to build next.
• Replit's next ICP: **operations teams.** Not product teams (already won). Ops sit at the nexus of data flows, buy lots of SaaS that silos data, run on Excel + manual workarounds. ROI on ops automation is as good or better than ROI on product-team automation. Examples cited: quote configurators, deal-desk automation, support operations.
• On Chinese open-source models: not a moral issue, but a security one for enterprise customers. 'I'd love to see a US corporation invest in national open-source. The US government could start a consortium to keep the market competitive.' Without it, an AI oligopoly will collude on price and limit API access.
• On Mythos / cyber: defensive-offensive leapfrog, no permanent advantage. Companies updating their codebase against new vulnerabilities is 'a multi-year effort, every time.'

Notable quotes

Themes

• Coding-model plateau and the rise of fine-tuned vertical models
• Operations as the next vibe-coding ICP
• Inference as the new sales and marketing
• Anti-oligopoly case for US open-source