a16z: what AI apps can still charge for as models get cheaper and better
rseroter · x · 2026-11-05
An a16z piece examines pricing and moats for AI application companies as inference costs fall. Key points:
- Cheaper, stronger models improve margins but also make app features easier to replace with general-purpose models; the core question is what customers keep paying for.
- Four pricing levers: domain expertise, deep integrations, reliable execution, and hard-to-access context.
- Build-vs-buy framework: customers assembling the stack vs vendors absorbing workflow orchestration, context, evaluation and ongoing optimization — compare finished work on quality, speed, total cost, risk.
- Case study: Repsol's legal team blind-tested Harvey against Microsoft Copilot and its internal prototype Lexia; Harvey won, then outperformed in an 8-week pilot with 50 lawyers across 12 jurisdictions, leading to full rollout, with lawyers reportedly saving 4–6 hours weekly and 96% adoption.
More from Venture
- Nvidia down, Amazon up: traders may be betting existing datacenters gain value as construction slows — Darpinian · 2026-11-05
- Qualcomm Q4 revenue falls 8% y/y, non-GAAP earnings down 28% — firstadopter · 2026-11-05
- Agentic commerce splits two ways: network tokens vs. agents with their own stablecoin wallets — schwentker · 2026-11-05
- RicursiveAI team hailed by investor Striker for string of AI breakthroughs — annadgoldie · 2026-11-05
- a16z's Andrew Chen: You can't go direct by copy-pasting AI — andrewchen · 2026-11-05
- NYC startups raised $4.94B in October, the city's biggest funding month ever — moonsandhues · 2026-11-05