20VC: Four bottlenecks in AI — Anj Midha

Category: Expert Interviews · Duration: 75 min · ▶ Watch

Speakers: Anjney Midha · Harry Stebbings

Switch language → 中文

Segments (13)

  • 00:00 · Introduction
    • Introduction of Anjney Midha, his background with Anthropic, Mistral, and Amp.
  • 00:58 · Scaling Laws and Diminishing Returns
    • Anjney argues against diminishing returns in AI scaling, citing super-exponential gains in material science via Periodic Labs.
  • 02:55 · The Four Bottlenecks of AI
    • Discussion on the primary bottlenecks: culture, context feedback (data), compute, and capital.
  • 10:06 · Sovereign Data and Local Infrastructure
    • The necessity of local AI infrastructure in Europe due to regulations like the US Cloud Act, highlighting Mistral’s strategy.
  • 14:27 · The Early Days of Anthropic
    • Anjney recounts the struggles of raising Anthropic’s seed round, facing 21 rejections from VCs who didn’t understand the technology.
  • 19:35 · Public Benefit Corporations and Mission Alignment
    • The role of PBCs in balancing long-term mission with profit motives, using Amp and Anthropic as examples.
  • 25:21 · The Back to the Future Era of Venture Capital
    • A comparison of current frontier tech investing to the early days of Intel and Genentech, requiring deep GP involvement.
  • 35:30 · The GPU Wastage Bubble
    • Anjney explains that we are in a GPU wastage bubble due to non-fungible compute, not an AI capabilities bubble.
  • 40:06 · China’s AI Strategy and Adversarial Distillation
    • How China uses systems co-design and adversarial distillation of Western open models to catch up in the AI race.
  • 44:26 · An Iron Dome for Inference
    • The proposal for a coordinated defense mechanism across Western AI labs to protect against state-sponsored distillation attacks.
  • 47:04 · Optimal Competition in the Inference Market
    • Why having 50 inference companies is a race to the bottom, and the need for optimal competition among 3-4 strong players.
  • 55:24 · Capital Requirements for Frontier AI
    • Estimating the massive capital and power (gigawatts) needed to compete with hyperscalers like Google.
  • 59:15 · Advice to LPs and the Future of VC
    • Anjney advises LPs to do their own research and criticizes GPs who invest in AI without building with it.

Specific Prices (4)

Timestamp Item Value Context
14:27 Anthropic Seed Round $100 million The re-anchored seed round size for Anthropic after initially trying to raise $500 million.
14:27 Amazon Partnership with Anthropic $4 billion The size of the compute and capital partnership between Amazon and Anthropic.
22:30 Amp Cloud Spend $4 billion The estimated cloud spend over the next four years for the 1.3 gigawatts of compute infrastructure Amp is securing.
22:30 Amp Equity Capital $10 billion The estimated equity capital portion required to finance Amp’s compute infrastructure build-out.

Memory Facts (4)

  • [00:58] Periodic Labs operates a facility for AI-driven material science.
    • 30,000 square feet
  • [14:27] Number of rejections Anjney received when pitching Anthropic’s seed round to Sand Hill Road VCs.
    • 21 nos
  • [22:30] Amount of compute infrastructure Amp has started securing.
    • 1.3 gigawatts
  • [55:24] Estimated infrastructure capacity of Google for internal and external AI workloads.
    • 12 to 15 gigawatts

Bottleneck Claims (3)

  • [02:55] Context feedback (domain-specific data) is the primary bottleneck for advancing AI capabilities in new fields.
    • Evidence: Early AI models failed at physics and chemistry because that data is locked in national labs and physical plants, not available on the public internet.
  • [35:30] The lack of compute standardization and fungibility is a massive infrastructure bottleneck.
    • Evidence: You cannot easily move workloads between different generations of chips (e.g., H100 to GB200) without buying a entirely new cluster, leading to wasted resources.
  • [47:04] Venture capital misallocation is a bottleneck for the inference market.
    • Evidence: Funding 50 different inference companies creates a race to the bottom and starves the truly innovative teams of the scarce compute resources they need.

Predictions (4)

  • [07:36, Near to medium term] We will see a generation of vertically integrated foundation model companies that generate their own proprietary physical data to build moats.
  • [10:06, Medium term] Europe will develop a fully independent, sovereign AI infrastructure stack (land, power, compute, models) to avoid reliance on US hyperscalers.
  • [29:20, Long term] The AI compute market will transition from a pre-standardization era to a standardized grid model, similar to the evolution of the electricity grid.
  • [47:04, Medium term] The inference market will consolidate from dozens of competitors to an ‘optimal competition’ state of 3 to 4 dominant, highly profitable players.

Key Technologies (4)

  • LLMs (Large Language Models): Used in Periodic Labs to predict new materials and superconductors.
  • Transformers vs Diffusion Models: Different underlying neural network architectures for AI models; the guest argues culture is more important than being tied to one specific architecture.
  • H100 / GB200 GPUs: Different generations of Nvidia AI accelerator chips; the lack of interoperability between them causes compute wastage.
  • Adversarial Distillation: A technique where a smaller or competing model is trained using the outputs of a larger, state-of-the-art model to rapidly catch up in capabilities.

Companies Mentioned (8)

Anthropic · Mistral · Amp · Periodic Labs · Amazon (AWS) · Google · Nvidia · Huawei

Notable Quotes (4)

There’s no saturation in superconductor discovery for example at all. The bitter lesson is holding is well and alive. — Anjney Midha @ 00:58

We are not in an AI crisis. We are not in an AI bubble… We are definitely in a GPU wastage bubble. — Anjney Midha @ 35:30

If we don’t secure frontier model inference… behind a coordinated Iron Dome, I don’t think we have a sustainable shot at staying at the frontier over the next decade. — Anjney Midha @ 44:26

Perfect competition is for losers. I also think monopolies are mafias. What we need is optimal competition. — Anjney Midha @ 47:04

Key Topics

AI Scaling Laws and Diminishing Returns · Bottlenecks in AI Development (Culture, Data, Compute, Capital) · Sovereign AI Infrastructure and Data Privacy · The Evolution of Venture Capital in Frontier Tech · Compute Standardization and the GPU Wastage Bubble · Geopolitics of AI: China's Strategy and Western Defense · Market Dynamics: Inference Competition and Monopolies

Takeaways

  • AI scaling is not slowing down; it is accelerating rapidly in specific physical domains when paired with proprietary data generation.
  • The biggest threat to AI progress isn’t a lack of capability, but a lack of standardized, fungible compute infrastructure, leading to massive resource waste.
  • Geopolitical regulations like the US Cloud Act are forcing regions like Europe to build entirely independent, sovereign AI stacks.
  • To defend against state-sponsored model theft via distillation, Western AI companies must collaborate to build a shared ‘Iron Dome’ for inference.
  • Venture capital must return to its roots of deep, co-founding partnerships to successfully fund the massive capital requirements of frontier AI systems.