Themes

Launch, satellite comms, earth observation.

As AI clusters scale, copper hits physical limits and the next bottleneck becomes optical infrastructure. Lasers, transceivers, optical engines, specialty glass, and the test/burn-in capacity behind it.

UAVs and counter-UAS - defense and commercial.

Reactors, SMRs, uranium miners and enrichment.

Server CPUs, packaging and ARM-based silicon - the non-GPU compute layer.

AI-native cloud / GPU-as-a-service operators.

Power generation, fuel cells, distribution for AI datacenters.

Foundries, memory, lithography - the AI silicon supply chain.

Enterprise software and platforms productizing AI for end-users.

Custom silicon and connectivity for inference workloads.

Edge / serverless / dev infra enabling autonomous AI agents.

Robotics, autonomy, manipulators - AI in the physical world.

Software platforms productizing AI for defense, intelligence, and battlefield decisions.

GOOGL admitted Google Cloud is leaving revenue on the table because it cannot build capacity fast enough. The bottleneck has shifted from chips to deployment - value accrues to operators with the power, real estate, and operational scale to actually deliver AI compute. These names anchor multi-billion-dollar contracted backlogs from hyperscalers (MSFT, META, GOOGL, AWS) and frontier labs (OpenAI, Anthropic), turning grid-connected MW and GPU clusters into long-dated revenue.

Primes building the next-gen US missile defense and Golden Dome architecture.

Naval platforms - surface combatants, submarines, and shipyards.

Rare-earth and strategic materials supply chain - the substrate of defense and AI hardware.

Power conversion and management silicon - the unloved bottleneck of the AI buildout.

The software layer powering next-gen robots in 2026. Generalist vision-language-action (VLA) models that train across embodiments - humanoid, quadruped, manipulation arms - are the bottleneck shifting value from hardware to the "robot brain". The investable public surface is dominated by NVDA (Isaac sim + Cosmos + GR00T) and the mega-cap diversifieds (GOOGL Gemini Robotics, MSFT/OpenAI-Figure axis, AMZN industrial fund, TSLA Optimus vertical, BIDU Apollo). The actual technological leaders - Skild AI (omni-embodied brain) and Physical Intelligence (π0/π0.5 manipulation models) - are still private, alongside the humanoid platforms (Figure, 1X, Apptronik, Agility) consuming those models. Tracked as an editorial theme rather than a bubble because the public basket would beta-drift into AI Compute / Hyperscalers.

AI-native cloud operators that buy NVIDIA (and AMD) accelerators at scale and rent GPU-hours to AI labs and enterprises — the "GPU-as-a-Service" layer sitting between the chipmakers and the model builders. Distinct from the hyperscalers: neoclouds run a thin, compute-first stack optimized for training and inference density, not a full enterprise-software cloud. The public pure-plays are CoreWeave (CRWV) and Nebius (NBIS); a second cohort of bitcoin miners has pivoted their power + real estate into HPC hosting (APLD, IREN, CORZ, WULF, CIFR). The capacity leaders driving the actual buildout are still private — Crusoe, Lambda, Together AI, Nscale — several on IPO tracks. Microsoft alone has struck ~$60B of neocloud capacity commitments across CoreWeave, Nebius and Nscale. Tracked as an editorial theme rather than a 13th bubble because the listed basket beta-drifts into semiconductors + datacenter-power. Thematic ETF: Roundhill Neocloud ETF ($NCLD).