AI startups can still build durable businesses against ChatGPT, Claude, and Gemini, but only in categories the labs will never prioritize. Olivia Moore (partner at a16z) maps where the gaps are, why OpenClaw matters, and what the next wave of consumer AI looks like.
1. Public Sentiment vs. Actual Adoption
A new NBC News poll shows 57% of American voters believe AI risks outweigh benefits, ranking AI below ICE, Marco Rubio, and the Republican Party in favorability. Moore attributes this to media coverage of resource consumption fears and anxiety about creative-sector job displacement. Her counter: a Wharton study of 800 enterprise leaders found heavy AI adopters expected to hire more humans to meet increased demand. An Anthropic labor report found no significant unemployment increase, with the most-impacted roles being engineers, researchers, and finance workers. The gap between average and power AI users is 8-9x in utilization, meaning early adopters compound advantages fast.
2. The Lab Constraint Thesis
Moore's core investment framework: the major labs are resource-constrained despite their scale. Every hour building a creative model is an hour not spent on coding agents or AGI. This creates durable gaps. ChatGPT and Gemini absorbed image generation because OpenAI had DALL-E and Google had YouTube training data. Video and audio have not been similarly consolidated. The ChatGPT and Claude app stores each have 200-plus apps with only 11% overlap, reflecting genuine divergence: ChatGPT targets fashion, retail, and transport; Claude targets finance, science, and medicine data sets. Gemini usage spikes almost entirely on new model drops, not organic retention.
3. Where Startups Can Still Win
Moore avoids horizontal categories like AI email, calendar, and docs, where Google and the labs already own distribution and user data. Winning startup profiles: (1) verticalized workflows where the last 1-2% of accuracy is most of the value, such as investment banking models requiring specific formatting and audit trails; (2) painful legacy integrations that require dedicated engineering and which agents cannot yet self-complete; (3) head-start quality moats, with ElevenLabs as the example: founders consistently try cheaper audio alternatives and switch back because voice quality is categorically better. The labs find it uneconomic to close that gap when bigger priorities exist.
4. OpenClaw: Architecture Unlock, Not Consumer Product
Moore calls OpenClaw the most important architecture unlock of 2026. Roughly half the founders she meets cite it as inspiration for vertical agent products. OpenClaw's February web traffic is flat-to-down from launch, meaning it has not reached mainstream consumers. Current power users are developers running 8-9 hours per day. For non-technical users, Claude's scheduled tasks deliver 99% of the value without setup friction. The real impact is structural: companies like Pulsia wrap OpenClaw and Claude Code so non-technical founders can prompt a business into existence. Pulsia reportedly hit $3 million ARR in under two weeks.
5. Sora, Social AI, and the Limits of Virality
Sora reached 1 million downloads faster than ChatGPT and spent 20 days at the top of the U.S. App Store, driven by the Cameo feature that let users make Jake Paul memes. The structural problem: exported Sora videos competed directly on TikTok and Reels against best-in-class human-made content, making the native feed experience inferior. Downloads have fallen sharply. Sora retains 3 million daily active users as a creative tool but failed as a social graph. Moore's conclusion: nobody has cracked AI social yet, and it will be very hard.
6. Incumbents and the SaaS Risk
Google now has four standalone products on the a16z top-100 list: Gemini, NotebookLM, AI Studio, and Google Labs. Moore would not have predicted that 24 months ago when Bard launched. Incumbents face a genuine but slow-moving threat: AI-native startups founded today will default to AI-native tooling rather than 25-year-old legacy software. Sam Altman's view, which Moore largely endorses, is that next-era winning software will be built ground-up for AI, not bolted on. The timeline is years, not months, particularly where incumbents hold data lock-in and integration switching costs.
Key Takeaways
- The major AI labs are genuinely constrained: compute, inference, and headcount force prioritization, which means large categories like vertical finance software, audio quality, and complex enterprise workflows remain open for startups to own.
- Horizontal AI apps (email, calendar, docs) are high-risk investments because Google and the labs already own the distribution and data needed to dominate them; the defensible plays are opinionated, verticalized products where the last 1-2% of accuracy carries most of the business value.
- OpenClaw matters not as a consumer product (traffic is flat, setup is too hard) but as an architectural blueprint: it proved AI can run async, multi-step tasks across applications, and roughly half of new founders Moore meets are building vertical versions of that idea.
- AI intensifies work rather than reducing it: Moore's own report output grew in length and density because AI gave her more analytical leverage, consistent with a Harvard Business Review finding that heavy AI users do more work, not less.
- The SaaS disruption is real but slow: incumbents with data lock-in and integration depth have years, not months, but startups founded today will choose AI-native tools by default, and that compounding selection effect is the actual long-term threat to legacy software.