The Twenty Minute VCThe Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha
CHAPTERS
- 0:00 – 1:23
Human alignment vs AI alignment, and setting the stage on Anj’s AI investing role
Anj opens with a provocative frame: AI alignment is hard, but “human alignment” is the deeper problem. Harry introduces Anj’s background across Anthropic, a16z, and AMP, setting up the conversation at the intersection of frontier research, investing, and compute infrastructure.
- •AI alignment framed as secondary to human/political alignment
- •Anj’s roles: early Anthropic investor, a16z AI lead, founder of AMP
- •Why compute and infrastructure sit at the center of modern AI progress
- 1:23 – 2:55
Are scaling laws dead? Why progress depends on the domain (coding vs materials science)
Harry presses on diminishing returns from scaling, referencing Demis Hassabis’s comments. Anj argues scaling laws are very much alive, but saturation differs by domain—coding may look plateaued on certain evals, while materials science is wide open.
- •Diminishing returns show up in saturated domains/evals (e.g., coding benchmarks)
- •Materials science and superconductor discovery show no saturation
- •Periodic Labs example: compute + lab automation creates rapid iteration loops
- •“Bitter lesson” still holds: more compute + better loops keep paying off
- 2:55 – 7:37
The four bottlenecks holding AI back: context, compute, capital, and culture
Anj lays out his core framework for why AI capability progress stalls: context/feedback loops (data), compute, capital, and culture. He argues algorithmic innovation is increasingly downstream of culture—great teams with mission focus will adapt architectures and find breakthroughs.
- •Four bottlenecks: context/feedback, compute, capital, culture
- •Algorithmic innovation is less the bottleneck if culture attracts top talent
- •Context feedback loops are where capability gains and commercial advantage emerge
- •Physical verification loops (labs/robots) as a way to manufacture missing data
- 7:37 – 10:37
Why “AI for science” underperformed—and how proprietary feedback loops create advantage
Anj explains that early claims about LLMs being strong at physics/chemistry didn’t match reality when benchmarked. The missing ingredient was not hype or prompting, but scarce, locked-up scientific data—prompting a push toward vertically integrated systems that generate their own ground-truth feedback.
- •Benchmarking revealed major gaps in scientific reasoning performance
- •Internet pretraining data lacks deep physics/chemistry and lab context
- •Scientific data is locked in national labs, academia, and industrial plants
- •Vertical integration (lab + robots + models + verification) can create a durable loop
- •“Claudification” risk: many software-only layers get commoditized faster
- 10:37 – 14:27
Sovereign data, the CLOUD Act, and the thesis behind Mistral’s “independent stack”
Using the CLOUD Act, Anj explains why certain workloads can’t legally or strategically run on US-managed hyperscalers—especially for European defense and mission-critical enterprises. He connects this to the investment case for Mistral: sovereignty across land/power/shell, compute, and locally trained models (with openness for deployment flexibility).
- •CLOUD Act implications: US-managed cloud can be subject to US government access
- •Mission-critical European workloads require local, trusted infrastructure
- •Opportunity: hyperscaler dominance becomes contestable in sovereign contexts
- •Mistral thesis: independence at every layer—infra + compute + models, aligned with Europe
- 14:27 – 19:35
The brutal early days of Anthropic: 21 of 22 VCs said no
Anj recounts the origin story of Anthropic as a research hypothesis turned business plan, built around scaling plus tight product-feedback loops in coding. Despite the founders’ pedigree (including GPT-3 authorship), most VCs didn’t understand the technology, forcing a rethink on fundraising scale and partners.
- •Weekly working sessions to turn “scaling” into a business hypothesis
- •Initial fundraising ambition vs re-anchoring to a smaller (still huge) seed
- •21/22 VC rejections: many didn’t know what GPT-3 was
- •Inference as dual engine: revenue + feedback to improve training over time
- •Strategic alignment with hyperscalers (e.g., Amazon) as compute/capital leverage
- 19:35 – 23:06
Public Benefit Corporations (PBCs): mission vs profit, and AMP’s governance choices
The conversation shifts to whether companies can pursue mission without being crushed by market forces or regulators. Anj defends PBC structures as a mechanism for leadership to make long-term, mission-aligned decisions (including actions that may look suboptimal to pure profit-maximizers).
- •PBCs as a governance tool to balance mission and profit over time
- •Counterargument to “win the market first, then do mission”
- •AMP example: providing significant compute at cost to support frontier innovation
- •Building standards and ecosystem health as institutional objectives
- •Why powerful AI companies inevitably interact with governments and regulation
- 23:06 – 35:30
The AMP Grid: building a ‘compute electricity grid’ and reviving hands-on venture incubation
Anj describes AMP’s model: not a hyperscaler and not a traditional VC, but a coordinator of compute capacity—analogous to an independent system operator for electricity. He ties this to a “back to the future” venture style where investors co-build with scientists (Arthur Rock, Genentech/Kleiner, Markkula/Apple) rather than just writing checks.
- •AMP Grid concept: pool capacity so teams provision for baseload, not peak
- •1885 analogy: factories running private generators at low utilization
- •Compute procurement strategy: get early, build trusted supplier relationships
- •Incubation model: deep partnership with scientists/engineers (Periodic as example)
- •Why check-writing-only venture struggles in CapEx-heavy frontier systems eras
- 35:30 – 42:16
‘GPU wastage bubble’ and why compute isn’t fungible (yet)
Anj argues we’re not in an AI capabilities bubble; we’re in an infrastructure utilization and standardization crisis. Because compute lacks fungibility across chip generations and configurations, large pools become stranded, creating inefficiency and distorted narratives about oversupply.
- •Claim: no AI bubble—capabilities are real; the bubble is “GPU wastage”
- •Compute non-fungibility across chip types (H100 vs GB200/GB300) and constraints
- •Stranded clusters: older hardware becomes memory-bound for newer training needs
- •Need for open standards and protocols for pooling and reallocating FLOPS
- •Misaligned incentives slow standardization across industry and government
- 42:16 – 47:28
China’s systems co-design advantage and distillation as a catch-up strategy
Anj outlines how China competes even without leading-edge chips by optimizing the full stack—chip, infrastructure, training, and deployment as one system. He describes adversarial distillation at scale as both a capability accelerator and a strategic threat to Western frontier labs.
- •AI race framed as full-stack systems co-design, not just a chip race
- •Performance gains via integration across the stack (Huawei + infra + training)
- •Adversarial distillation: extract capabilities from Western models via many endpoints
- •Open releases as feedback-driven iteration until frontier parity is reached
- •Security concerns: insider threats and coordinated exploitation of fragmented defenses
- 47:28 – 1:15:18
An ‘Iron Dome’ for inference, optimal competition, and what venture must become next
Anj proposes a coordinated defense layer for inference—shared signals and response to distillation attacks across labs—arguing fragmented security jeopardizes frontier leadership. He then broadens to market structure (perfect vs monopolistic vs optimal competition) and closes with a wide-ranging quick-fire touching LP diligence, founder quality (Dario), health, independence, and legacy.
- •Iron Dome concept: shared proxy/coordination layer to detect and respond to inference attacks
- •Security as a core bottleneck for scaling inference demand sustainably
- •“Optimal competition”: 3–4 serious teams per frontier; avoid both chaos and monopoly stagnation
- •Venture’s future: hands-on builders vs “bankers,” and broader public participation in upside
- •Quick-fire themes: LPs must do the work, Dario’s truth-seeking mission focus, health/time, independence, legacy
