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

Key points

• Four bottlenecks for AI progress: compute, context (data/feedback loops), capital, culture. Culture is upstream of everything — get mission-driven research talent and the algorithmic work takes care of itself.
• Bitter Lesson is emphatically alive in under-explored domains. At Periodic Labs (Midha's material-science incubation), throwing more compute at superconductor discovery is producing 'super-exponential' per-iteration gains. Coding evals are saturating; material science isn't close.
• Anthropic's seed round was a near-failure. 21 of 22 VCs Midha introduced them to passed — most asked 'what's GPT-3?' The $100m seed (re-anchored down from a $500m ambition) eventually closed because ML-literate people (SBF, some EA-adjacent) and Amazon got it. Amazon's $4B compute-for-equity was the unlock.
• Sovereign AI is real and commercially consequential. The US Cloud Act forces any workload on US-company-managed infrastructure to be reachable by the US government — which rules out European defence, logistics (ASML, CMA CGM), and critical infrastructure. That's why Macron and Jensen stood with Arthur Mensch in Paris unveiling a gigawatt facility: full independent stack, locally trained models, locally controlled compute.
• AMP is positioning as the 'independent system operator' for compute — not a cloud, not a VC firm, a coordinator. Analogy: we're in 1885 Industrial Revolution England. Every AI lab is running its own generator at half capacity. Pool them and output scales.
• 'We are not in an AI bubble. We're in a GPU wastage bubble.' Compute is not fungible — H100 clusters can't be used for GB200-era workloads, so billions of dollars of compute sit stranded. Standardisation (think TCP/IP or AC/DC) is the missing institutional layer.
• 'Iron dome for inference': Midha's proposal for a coordinated defensive layer where all Western frontier inference is served through a shared proxy that can detect and respond to adversarial distillation attacks (largely from China). Today coordination is informal group chats with founders comparing notes.
• VC model critique: 'Perfect competition is for losers. So is monopolistic competition. What you want is optimal competition — 3-4 teams per frontier.' Fifty inference companies is a race to the bottom that starves the actually-innovative 4-5. Back-to-the-future era: Arthur Rock / Bob Swanson / Mike Markkula style co-founding + capex-heavy incubation is returning.
• On labels: 'foundation model company' is misleading. Anthropic, Mistral etc. are Frontier Systems companies. Cloud Code was always part of the Anthropic plan; the VCs who missed it were working from associate market maps, not first principles.
• Wealth-creation problem: almost no venture firm was in the Anthropic seed. LPs (endowments, pension funds, teachers) are therefore mostly miss the largest value-creation event of the cycle. Fix: educate LPs on the bottlenecks and invest in the bottleneck-solvers.

Notable quotes

Themes

• The four bottlenecks of the AI frontier
• Sovereign AI and the end of hyperscaler dominance
• GPU wastage bubble, not AI bubble
• Iron dome for inference — defending the Western frontier
• Back-to-the-future VC: incubation over check-writing