The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha
Anj Midha, the prominent AI investor and founder of AMP, identifies the four critical bottlenecks limiting frontier AI progress: context feedback loops, compute, capital, and culture—arguing that culture is the most important because it attracts the talent that solves algorithmic innovation. Rather than a debate about AI alignment or capability ceilings, Midha contends the real problem is human misalignment: the failure of Silicon Valley and Washington to standardize how statistical AI systems should be procured and deployed differently from deterministic software, creating infrastructure waste and undermining Western competitiveness against China's coordinated, full-stack approach.
Key takeaways
- • The compute standardization crisis is the existential infrastructure problem: compute is non-fungible across chip types (H100s vs. GB200s), creating billions in stranded GPU capacity, and standardization following the electricity grid model is more urgent than AI safety debates
- • Frontier AI companies should be viewed as full-stack frontier systems businesses, not "foundation model companies"—they must own the entire loop from training data collection (e.g., physical labs for material science) through inference deployment to capture unique context feedback loops that competitors can't replicate
- • Vertical integration with proprietary data loops is the only defensible moat: companies like Periodic Labs (material science via physical verification loops) cannot be "claudified" because they control data generation at the point of deployment, unlike pure software models
- • Build ventures through deep partnership with scientists/founders rather than transactional check-writing—the Kleiner Perkins model of co-founding (e.g., Bob Swanson co-founding Genentech) beats the 2010s venture model where investors write checks and disappear
- • Sovereign compute infrastructure in Europe (e.g., Mistral) is competitively viable because regulatory constraints (Cloud Act) and geopolitical mission-critical workloads create genuine, defensible local demand that hyperscalers cannot serve
- • Western frontier labs need coordinated "Iron Dome" defense: a shared inference proxy that alerts all labs to distillation attacks, since China is systematically reverse-engineering U.S. models at scale via adversarial distillation and open-source bootstrapping
Recommendations (3)
"When I got to a16z as a general partner, the first thing I did is I sat down with Mark and Ben"
Anj Midha · ▶ 23:31
"I got to apprentice in that mode of venture capital because when I got to Kleiner"
Anj Midha · ▶ 27:15
Mentioned (19)
More from these creators
ElevenLabs: Building an AI Sales Machine & Why We Set a 20x Sales Quota
OpenAI Buys TBPN & Their Management Team Reboot | Mercor Hack & Why Now is the Time for Cyber
Demis Hassabis: Why AGI is Bigger than the Industrial Revolution & Where Are The Bottlenecks in AI
How Hims & Hers Reached a $4.3BN Market Cap on $2.3BN of Revenue | Andrew Dudum
OpenAI Kills Sora & Hits $100M ARR on Ads | Oura Going Public & Whoop Raises at $10BN
Why Margins Don't Matter for Early-Stage Startups | Gili Raanan