One AI Supercycle -
10 Layers. 10 Tickers.
Layer 1 — Power |
$BE
A single AI data center can consume 100–500MW. The US grid wasn’t built for this. Bloom Energy sits at the intersection of distributed power generation and the insatiable energy appetite of hyperscalers. Power is the foundational constraint — before chips, before cooling, before anything else.
Layer 2 — Substrates |
$AXTI
InP (Indium Phosphide) and GaAs (Gallium Arsenide) wafers are the raw material for photonic components — lasers, modulators, detectors. AXT Inc supplies these specialty substrates to the photonics supply chain. Demand is structurally rising as Co-Packaged Optics (CPO) and 1.6T transceivers scale. Tight supply, long qualification cycles, few alternatives.
Layer 3 — Chips |
$NVDA
H100. H200. Blackwell. Each generation widens the moat rather than narrowing it. CUDA lock-in is one of the deepest competitive advantages in tech history.
$NVDA isn’t just a chipmaker — it’s the operating system of the AI era.
Layer 4 — Memory |
$MU
HBM3E (High Bandwidth Memory) is the bandwidth interface between the GPU and data. Without it, the most powerful chips in the world are throttled. Micron is one of only three companies globally that can produce HBM at scale — alongside SK Hynix and Samsung. Supply is tight. ASPs are rising. The AI upgrade cycle is a multi-year HBM demand wave.
Layer 5 — Photonics |
$AAOI
As data centers scale from 400G to 800G to 1.6T optical speeds, the components inside transceivers — lasers, modulators, detectors — face an exponential demand surge. Applied Optoelectronics is a pure-play photonics manufacturer benefiting directly from this cycle. Margin expansion volume ramp = a powerful setup.
Layer 6 — Optics |
$LITE
If photonics makes the components, optics assembles them into the interconnect.
Lumentum is a leader in optical networking — coherent transceivers, 3D sensing, EML lasers. The 800G → 1.6T transition is a hardware replacement cycle that touches every hyperscaler and co-lo data center globally. This isn’t incremental demand. It’s a full network overhaul.
Layer 7 — Cooling |
$VRT
Vertiv designs and manufactures liquid cooling, immersion systems, and thermal management infrastructure for high-density AI racks. As GPU power density climbs past 1kW per chip, traditional air cooling fails. Vertiv is already embedded with the largest hyperscalers. Backlog is growing. Lead times are extending.
Layer 8 — Networking |
$ANET
Arista Networks builds the high-speed Ethernet switching fabric that connects thousands of GPUs inside AI training clusters. Their software-defined architecture and 400G/800G switching platforms are designed for exactly the traffic patterns AI workloads generate. AI networking is a separate, incremental growth vector on top of their already dominant enterprise business.
Layer 9 — Data Centers |
$NBIS
Nebius Group is building AI-native data centers — purpose-built for GPU density, liquid cooling, and low-latency networking. Unlike legacy co-los retrofitting old facilities, Nebius is starting from scratch for the AI era. Backed by Yandex’s original infrastructure DNA, they’re scaling fast in a market where capacity is chronically constrained.
Layer 10 — Hyperscalers |
$GOOG
Google has committed $75B in capex for 2025 alone. Their TPU buildout, data center expansion, and AI product integration (Gemini, Search, Cloud) make them both a consumer and a builder across the stack. Every dollar they spend flows down through layers 1–9.
The AI supercycle isn’t a software story — it’s a physical infrastructure buildout that rivals the railroad era. Every layer of this stack is capacity-constrained, capital-intensive, and structurally undersupplied relative to where demand is heading. Most investors own one or two names at the top of the stack. The opportunity is in understanding all 10 layers — and sizing accordingly.
Not financial advice .