The Future of Compute
is Supply-Constrained.
Eleven research reports spanning semiconductor bottlenecks, robotics, inference scaling, cross-border equity signals, and AI capital-allocation maps.
The bottleneck shifts every 2–3 years
Memory & packaging (2026–27) → Power (2028) → Fabs & EUV (2029–30) → Geopolitics (2031+). Each phase creates different winners.
HBM, storage, and packaging have been discovered
June refresh: HBM +237.4% median, storage/data +200.0%, advanced packaging +164.0%. The highest current alpha now shifts toward power, cooling, and selected packaging names that still have bottleneck exposure without the same rerating.
ASML is the most asymmetric trade
A $400M EUV tool enables $14.3B downstream value. ASML captures <3% of what it creates. Pricing power inflects 2028–2032.
Robotics is escaping the teleop trap
EgoScale: 20,854h of human video with R²=0.9983 scaling law. DreamDojo: 44,711h. That’s 89.4x Figure’s teleop data. The scaling substrate is shifting from robot demos to human video.
Structure beats brute-force thinking
Structured test-time scaling — recursion, context isolation, verification — outperforms naive chain-of-thought. Value shifts to verifier infrastructure and recursive training flywheels.
880+ assets and episodes mapped across the research stack
100 GPU buildout equities, 100 passives residual-alpha names, 120 unified semiconductor/CPO alpha names, 100 semiconductor AI node/connection names, 100 latent AI company nodes, 50 robotics companies, 100 test-time scaling names, 50 cross-border signal names, and 162 podcast episodes. Each is scored on chokepoint exposure, mispricing, scaling alignment, method fit, network centrality, or hidden alpha.
Eleven Reports, One Thesis
Report I maps the physical constraints — EUV tools, power grids, memory fabs — that limit how fast AI can scale. This is the macro framework.
Report II zooms into 100 public companies across 10 sectors. Filter by sector, sort by current alpha, and explore bull/bear theses for each name.
Report III asks: can robots learn from human video instead of expensive teleop? 10 methods scored, 5 labs profiled, 50 companies ranked.
Report IV examines whether AI can get smarter by structuring its own thinking. 5 novel RLM methods, 16 scenarios, 100 companies.