Bubbles

GPUs, accelerators, and the AI-specific silicon stack that moves as one under capital flows.

Picks and shovels of chip manufacturing - the machines that make the machines.

DRAM, HBM, memory IP, advanced packaging, and storage - the memory bottleneck of AI compute scaling. The Samsung+SK Hynix+Micron trio is forecast to print ~$454B in FY26 operating income (Samsung +413% / SK Hynix +593% / Micron +684%) on HBM allocation tightness through 2026. **The cycle framework.** Memory is the most cyclical sub-sector in semis - commodity output, multi-year capex lag against quarter-by-quarter demand, three-player oligopoly with prisoner's-dilemma supply discipline. Every up-cycle in the last decade has ended in a 50-75% drawdown within 12-18 months of the peak (2015 -75%, 2018 -57%, 2022 -51%). The current cycle is +1,851% off the 2023 lows - larger than the prior three combined. AI dampens the cycle (HBM LTAs, healthier customer concentration, slower supply response) but does not eliminate it; roughly 70% of Micron's bit output is still commodity DRAM/NAND. **Where the cycle sits now (2026-05-28).** 8 of 10 structural signals read late or peak: weekly RSI 86, +1,873% off the low, op margin 67.6% (above any prior cycle peak), forward PE 8.76× (canonical signature of peak earnings about to roll over), analyst price-target lagging price by -50%, peak-cycle capex committed ($100B+ across the three players). Two signals have NOT triggered: spot DRAM has not rolled over, and Micron has not yet issued equity at ATH (the classic peak confirmation - they did at $60 in 2018 and $90 in 2021). The bloc is in the late phase of an up-cycle that may have one more leg. **Eight leading indicators to watch** (in rough order of lead time): DXI/DRAM spot price (3-6 months ahead), HBM allocation tightness (~3 months), memory lead times (~3 months), hyperscaler inventory days (1-2 quarters), fab utilization (~1 quarter), capex/sales ratio (late-cycle confirm), MU equity issuance at ATH (peak confirm), MU relative performance vs SOXX (~6 weeks). Two cycle-turn triggers within 30 days historically precede the price peak by 4-8 weeks. Full framework + position-management template: /articles/memory-cyclicality-the-supercycle-that-still-has-a-cycle.

The subset of mega-caps whose AI capex narrative dominates price action.

Workflow and data-platform SaaS monetizing AI - the legacy-software AI trade.

Moving bits between tens of thousands of GPUs - optical is non-negotiable.

AI eats electricity. Utilities and independent power are the unlock.

Clean baseload for hyperscaler campuses - SMRs and uranium miners.

Physical plant of AI - cooling, switchgear, prefab datacenter components.

Physical AI — industrial, surgical, and humanoid robotics. The cluster spans mature automation (ISRG surgical, ABB/ROK industrial arms, TER test + humanoid exposure) and the emerging humanoid build-out. **Why now.** Two structural unlocks moved roughly together in 2025-2026. First, the software bottleneck broke: NVIDIA's GR00T foundation models, Google DeepMind's RT-2, and the open OpenVLA stack gave robots general-purpose vision-language-action intelligence instead of narrow task-specific programming — the same "scale + foundation model" pattern that unlocked LLMs is now being applied to embodied agents. Second, the economic crossover has started: humanoid lease/rental costs are now reported below blended human-labor costs in some warehouse and manufacturing deployments, which is the trigger that turns a lab demo into a procurement line item. **The public-market read-through is thin.** Most of the credentialed humanoid builders — Figure AI, Apptronik, 1X, Sanctuary AI, Unitree — remain private. CCXI (Churchill Capital Corp XI) is the first public on-ramp: a SPAC merger with Agility Robotics announced 2026-06-24 at a $2.5B pre-money valuation, Foxconn-led PIPE, NVIDIA/Amazon/SoftBank-backed, with Digit already contracted at Schaeffler, GXO, Toyota, and Mercado Libre. It renames to AGLT on deal close. Supply-chain proxies (AMBA vision SoCs, OUST LiDAR, VPG force/torque sensors) capture the sensing + perception layer that every humanoid platform needs regardless of which builder wins. **Bull:** foundation-model-driven generality removes the "one robot, one task" cost structure that killed prior robotics cycles; deployed paid pilots (not demos) at real logistics/manufacturing customers; NVIDIA building a dedicated platform (Isaac, GR00T, Halos safety) the way it built CUDA for GPUs. **Bear:** thin public-market exposure means most of the upside accrues to private cap tables until more humanoid names IPO; unit economics claims are early and self-reported; SPAC/IPO mechanics add deal-completion risk on top of the operating thesis; industrial robotics incumbents (ABB, ROK) face margin pressure if humanoid generalists commoditize fixed-purpose automation.

Edge compute, global data pipes, earth observation - orbital layer of AI.

Post-GPU compute optionality - long-dated but asymmetric.