20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

What It Covers

Anj Midha — founding Anthropic investor and AMP founder — maps the four bottlenecks blocking AI progress (context feedback, compute, capital, culture), explains why compute non-fungibility creates a wastage crisis, details China's adversarial distillation strategy, and argues the industry needs an "iron dome" inference coordination protocol to protect Western frontier models.

Key Questions Answered

• •Four Bottlenecks Framework: AI progress is blocked by four specific constraints: context feedback loops (unique domain data), compute infrastructure, capital deployment, and culture. Culture ranks as the most critical because it attracts researchers who solve algorithmic problems organically. If mission-driven culture exists, algorithmic innovation follows automatically — making it no longer a standalone bottleneck as it was two to three years ago.
• •Compute Non-Fungibility Crisis: NVIDIA H100, GB200, and GB300 chips are architecturally incompatible — training runs cannot migrate between generations without purchasing entirely new clusters. This strands billions in compute across the ecosystem. Investors and operators should treat compute procurement decisions as 3-4 year infrastructure commitments, not flexible resources, and pressure the industry toward open standardization protocols analogous to TCP/IP or AC/DC electricity.
• •China's Adversarial Distillation Playbook: China compensates for chip disadvantages through full-stack systems co-design — pairing Huawei chips with custom infrastructure and training pipelines, then distilling Western frontier models at scale through open endpoints. The resulting open-source releases bootstrap domestic capability until parity is reached, at which point openness stops. Western labs should treat unusual inference traffic spikes from specific regions as active distillation attacks requiring coordinated defensive response.
• •Sovereign Data as Moat: The US Cloud Act legally requires American-managed infrastructure to grant US government data access, making it structurally impossible for European governments and enterprises with mission-critical workloads to use AWS, GCP, or Azure. This creates a genuine infrastructure sovereignty gap — the first opening in 15 years for non-hyperscaler providers. Mistral's gigawatt Paris facility, backed by Macron and Jensen Huang, is the direct commercial result of this regulatory arbitrage opportunity.
• •Optimal Competition Over Monopoly: Markets with 3-4 frontier competitors produce more innovation than either perfect competition (50+ inference companies racing to the bottom on scarce compute) or monopoly (incumbents hoarding resources instead of innovating). Current VC behavior — funding 50+ inference companies simultaneously — actively starves the 4-5 genuinely innovative teams of compute supply, which is their core product input. Capital allocators should concentrate bets on compute-secured teams, not category breadth.
• •Vertical Lab Model for Data Moats: Domain-specific AI progress requires physical data generation, not internet pre-training. Periodic Labs demonstrates the template: LLMs predict new superconductors, robots synthesize them, X-ray diffraction machines validate properties, and verification data feeds back into training runs. Any domain where critical data is locked in national labs, manufacturing plants, or physical systems — rather than the internet — represents a defensible frontier systems opportunity that general models cannot replicate through distillation.