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

Culture is the primary constraint

Algorithmic innovation is a function of culture; mission-driven environments attract flexible researchers who solve architectural challenges naturally, making culture more limiting than technical specifications.

Context feedback loops create defensibility

Unique data generation and verification systems—such as Periodic Labs' physical robots validating superconductor predictions—provide greater moats than generic pre-training data.

Scaling laws remain domain-dependent

While coding shows diminishing returns, domains like material science exhibit super-exponential gains from increased compute, with no saturation in superconductor discovery.

Infrastructure requires structured capital

Beyond algorithms, progress requires coordinated deployment of land, power, shells, and diverse financing vehicles to sustain RL loops at scale.

Data sovereignty reshapes cloud competition

The US Cloud Act prevents mission-critical European workloads from running on American infrastructure, creating openings for sovereign providers like Mistral to capture enterprise and government demand.

Physical vertical integration beats general models

Companies like Periodic Labs avoid being 'Claudified' by owning proprietary data generation through physical labs, whereas horizontal tools without unique data moats face commoditization.

Scientific reasoning requires physical verification

Current LLMs lack physics and chemistry capabilities because the data is locked in national labs and manufacturing plants, necessitating physical infrastructure to unlock scientific frontiers.

21 of 22 VCs rejected the seed pitch

Despite the team inventing GPT-3, most investors in 2021 lacked technical literacy to evaluate the scaling hypothesis, asking 'What's GPT-3?' during pitches.

The inference flywheel strategy

Anthropic's early operational model focused on deploying coding assistants to generate both revenue and context feedback simultaneously, creating a sustainable capability improvement loop.

Bridging research and business timing

Success required converting the research hypothesis of scaling laws into a business hypothesis through methodical operationalization over 12-24 months.