The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha

Anjney Midha reveals Anthropic’s brutal seed round — 21 of 22 VCs said no — and argues compute non-fungibility, not AI hype, is the real bubble.

• 21 of 22 VCs passed on Anthropic’s seed; they originally targeted $500M, reanchored to $100M, still couldn’t explain GPT-3 to most investors.
• Amazon’s initial $4B compute-for-equity Anthropic deal was driven by watching Azure’s OpenAI advantage, not mission alignment.
• AMP has secured 1.3 gigawatts of compute (~$40B cloud spend over 4 years), financed ~80% debt, 20% equity (~$10B equity required).
• Europe needs 12–15 GW of sovereign AI infrastructure to match Alphabet’s internal pool; currently nowhere close.
• Compute is not fungible: H100, GB200, and GB300 clusters cannot share workloads, creating billions in stranded, unutilized GPU capacity.
• China’s strategy is adversarial distillation at scale — extract Western frontier model performance via open endpoints, release as open source, iterate until frontier-competitive, then stop open-sourcing.
• Midha argues for an inference Iron Dome: a shared proxy across Western labs that detects and coordinates response to distillation attacks in real time.
• Perfect competition (50 inference companies) and monopoly are both losing structures; optimal is 3–4 well-resourced teams per frontier category.

2026-04-14 · Watch on YouTube