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

Identified Bottlenecks for Frontier Progress: Anj Midha categorizes the primary constraints on advancing AI capabilities into four specific pillars:

• Context Feedback Loops: The scarcity of real-time data from specific domains (e.g., physics, chemistry, materials science) where internet pre-training data is insufficient.
• Compute Supply: The critical need for coordinated, scalable infrastructure rather than fragmented, siloed data centers.
• Capital Availability: The requirement for diverse financial structures (equity, debt, balance sheets) to fund land, power, and shell infrastructure.
• Culture: The most significant long-term bottleneck, defined as the ability to attract mission-driven talent that prioritizes problem-solving over architectural dogma.

The "Iron Dome" Inference Proposal: Midha argues that Western leadership in AI requires a coordinated defense mechanism for state-of-the-art model inference to prevent adversarial attacks.

• This involves a shared proxy system where companies coordinate to detect and respond to distillation attacks across the Western ecosystem.
• Without this coordination, the West risks losing its competitive edge over the next decade due to uncoordinated vulnerabilities.

Compute Standardization and the "Grid" Analogy: Midha posits that the current AI infrastructure crisis mirrors the pre-standardization electricity era of 1885.

• Compute is currently non-fungible across chip types (e.g., H100 vs. GB200) and manufacturers, leading to significant stranded assets and inefficiency.
• AMP Grid: The goal is to function as an independent system operator to pool and coordinate compute capacity, similar to how the electrical grid standardized power delivery, reducing waste and enabling better resource allocation.

Critique of Current Market Structures:

• Avoidance of Perfect Competition: Midha argues that "perfect competition" (e.g., 50+ inference companies) is destructive, as it fragments resources and drives a "race to the bottom" on price, preventing innovation.
• Optimal Competition: He advocates for a landscape of "three or four" dominant, well-funded teams per frontier that maintain sufficient competition to innovate without becoming stagnant monopolies.
• Rejection of "Foundation Model" Labels: He contends that successful AI companies are "frontier systems" engaging in full-stack co-design (data, compute, application), not just model providers.

Sovereign Infrastructure and Geopolitics:

• European Independence: Midha highlights Mistral's strategy to build a fully sovereign, open-stack AI infrastructure in Europe (land, power, shell, compute, models) to bypass U.S. jurisdiction laws like the Cloud Act.
• Global Tension: He warns that China's strategy involves systems co-design and adversarial distillation of Western models to catch up rapidly, necessitating a unified Western response.

Anthropic's Genesis and Investment Thesis:

• Midha led the seed round for Anthropic (raising $100M after initial rejections) by pitching the "scaling recipe" as a business hypothesis: deploy models, gather context feedback, and reinvest revenue into more compute.
• He notes that early investors failed to understand the "compute multiplier" effect, where a dollar of VC capital could yield more intelligence if compute access was secured early.

Periodic Labs and Physical-Data Loops:

• Vertical Integration: Periodic Labs is building a 30,000 sq. ft. facility where LLMs predict new superconductors, robots synthesize them, and physical machines (X-ray diffraction) validate results.
• Data Moat: The unique, proprietary physical data generated by this loop creates a barrier to entry that prevents commoditization by generalist models.

Venture Capital Evolution:

• Return to "Back to the Future" Model: Midha compares the current need to the era of Arthur Rock (Intel) and Mike Markkula (Apple), where investors acted as active co-founders and operational partners rather than passive capital allocators.
• Public Benefit Governance: As the founder of AMP, a Public Benefit Corporation, Midha prioritizes mission alignment over short-term profit, even if it means offering compute at cost to support the independent ecosystem.
• LP Education: He advises investors to stop outsourcing due diligence, insisting that LPs must personally understand the technology (e.g., by building agents) to identify true defensible advantages.

Financial Scale and Capital Allocation:

• AMP Infrastructure: The organization is currently securing approximately 1.3 gigawatts of compute capacity, representing roughly $40 billion in cloud spend over four years.
• Capital Structure: This infrastructure is financed with approximately 20% equity ($10 billion) and 80% debt, utilizing mission-aligned balance sheets to de-risk large-scale procurement.

Personal Insights and Cultural Shifts:

• Time Scarcity: Midha cites personal health experiences as a catalyst for valuing time and relationships over financial accumulation, shifting his focus from the "money treadmill" to mission execution.
• Dario Amodei's Leadership: Attributes Amodei's success to a physicist's mindset (empiricism, truth-seeking) and an unwavering mission alignment that forces hard trade-offs.
• Legacy: Midha aspires to be remembered for accurately predicting the future trajectory of AI and establishing the institutional standards required to sustain it.

Forward-Looking Predictions:

• Inference Growth: Demand for inference is expected to grow combinatorially over the next 3–5 years if scaling laws continue to hold.
• Market Consolidation: The number of viable inference providers will likely contract to a few key players due to compute supply constraints, rather than fragmenting indefinitely.
• Model Specialization: While general models will remain accessible, specialized frontier models will likely be segmented, with access restricted to enterprise clients capable of absorbing development costs.