The 30,000-ft map of AI infrastructure stocks across nine layers — compute, memory, HBM packaging, lithography, photonics, networking, cooling, power, and neocloud operators — with the publicly-listed leaders in each layer and why a focused stack beats the BOTZ / AIQ / ROBO / IRBO ETFs.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
AI infrastructure stocks are publicly traded companies that build the physical layers beneath every GPU — the custom silicon, memory, advanced packaging, lithography equipment, photonics, networking, cooling, power, and neocloud capacity that turn a hyperscaler's capex budget into a running AI cluster. The investable universe spans nine distinct layers, each with its own bottleneck and a small handful of publicly-listed beneficiaries. This guide is the 30,000-ft map — who gets paid as hyperscaler capex climbs from roughly $150B in 2023 toward $300B+ in 2026E.
👉 See live prices across the whole stack: track the Data Center Buildout trend on Macroplane — the meta-page that aggregates every basket below into one performance view.
Each layer below maps to a focused Macroplane basket. Cumulatively they form the picks-and-shovels universe.
The accelerator itself. $NVDA is the anchor, but the real edge is in the second tier: $AMD's MI300/MI400 ramping into hyperscaler diversification budgets, $AVGO and $MRVL designing custom ASICs for Google, Meta, and ByteDance, $ARM collecting royalties on every CPU core in a hyperscaler rack. Below that: $INTC (Gaudi + foundry optionality), $QCOM (edge AI), $LSCC (low-power FPGAs in smartNICs), $CBRS (Ceva IP cores).
The bottleneck here isn't silicon design — it's advanced packaging capacity downstream (Layer 3). Custom ASIC revenue at $AVGO and $MRVL is growing faster than the merchant GPU market because each hyperscaler wants its own accelerator for inference. That re-rates the whole Custom Silicon basket.
You cannot feed a GPU without high-bandwidth memory, and HBM is the most supply-constrained component in the AI stack. $MU is the only US-listed pure HBM play, ramping HBM3E and HBM4 into $NVDA, $AMD, and custom ASIC sockets. $SNDK (the spun-off Western Digital flash business) covers NAND for AI storage tiers. $SIMO and $RMBS sit on the controller-IP side, monetizing every memory-rich rack.
HBM gross margins are running 50%+ vs sub-30% for commodity DRAM. HBM die area is roughly 2x a standard DDR5 die, yields are lower, and every wafer is pre-sold 12-18 months in advance. The pure-play exposure is the Memory Supercycle basket.
This is where the actual bottleneck lives. CoWoS (TSMC's chip-on-wafer-on-substrate) capacity is the gating constraint on global GPU supply. $AMKR is building US-based 2.5D packaging for AI customers. $ASX provides outsourced packaging. The picks-and-shovels are in inspection and tooling: $CAMT, $ONTO, $ENTG (specialty chemicals), $COHU, $FORM (probe cards), $BESIY (hybrid bonders — the single most supply-constrained tool in advanced packaging), $KLIC, and $TER.
If AI capex is structural, advanced packaging capacity has to at least double by 2028. Every name in the HBM/Packaging basket sees their TAM rerate.
The semiconductor capital equipment names. $ASML is the only EUV lithography supplier on Earth — every AI chip at 3nm and below passes through one of their machines. $TSM is the foundry no other can match for AI volumes. Around them: $LRCX, $KLAC, $AMAT (the WFE big three), $ASMI (atomic layer deposition), $MKSI, $UCTT, $ACLS (ion implantation), and $PLAB (photomasks).
This is the cleanest "5-year visibility" trade in the stack. Fab capex commitments lock in tool orders 2-3 years ahead. See the Lithography & Fab Tools basket.
As cluster sizes scale past 100,000 GPUs, copper interconnects hit a wall. The industry's answer is co-packaged optics — moving optical engines onto the same package as the switch ASIC. $COHR is the leading laser and CPO module supplier. $LITE plays the same role with InP lasers and EML chips. $FN is the contract manufacturer behind virtually every high-speed optical transceiver. $CIEN, $POET, $AAOI, $VIAV, and $NOK round out the Photonics/CPO basket.
Private names matter here too — Lightmatter, Ayar Labs, Celestial AI — but they're not investable. The publicly-listed proxies are the laser and module suppliers, because every CPO design needs an External Light Source. The macro lens is the Silicon Photonics & Optical Interconnects trend.
The fabric between racks. $ANET is the merchant networking leader for AI clusters; $CSCO is the legacy incumbent rebuilding for AI fabrics. The differentiated picks are $ALAB (PCIe retimers and CXL controllers — the unsung hero of GPU-to-GPU communication) and $CRDO (Active Electrical Cables, bridging passive copper and optics). $MXL covers SerDes and physical-layer chips.
Retimers are the most under-appreciated investment in the AI stack — every NVL72 rack uses dozens. See the Networking/Retimers basket.
Air-cooled racks max out around 50-60 kW. Blackwell racks run 130 kW. Rubin will exceed 250 kW. Direct liquid cooling is no longer optional. $VRT leads pre-fab modular data center and CDU (coolant distribution unit) integration. $MOD builds heat exchangers. $NVT, $JCI, $TT, and $CARR round out the building thermal layer. The Cooling basket is the most under-owned layer in the AI infra stack relative to its capex pull-through.
A single AI campus can pull 1 GW. Texas, Arizona, and the Midwest are running out of grid headroom. $GEV builds the gas turbines and grid equipment. $CEG, $VST, and $TLN are merchant nuclear operators that have signed PPAs directly with hyperscalers. $ETN, $HUBB, and $VICR cover power distribution, switchgear, and rack-level conversion. $BWXT, $OKLO, and $SMR are the small modular reactor optionality plays. See the Power & Grid basket.
Cloud operators that only exist to rent GPUs. $CRWV is the largest pure-play GPU cloud, with multi-billion-dollar contracted backlog from Microsoft. $NBIS is the European entrant. $IREN and $APLD pivoted from bitcoin mining into AI hosting. $SMCI and $CLS aren't operators but build the AI-optimised servers and racks that fill those data centers. This is where the highest leverage to capex sits — and the highest balance-sheet risk. The AI Cloud/Neoclouds basket lets you size that exposure deliberately.
Three forces are reinforcing each other.
2023 combined hyperscaler capex (Microsoft, Google, Meta, Amazon, Oracle) was roughly $150B. 2024 ran $230B. 2025 cleared $280B. 2026 guidance is $300B+ and rising. This isn't a single-product cycle — it's an infrastructure buildout with 5-7 year amortisation horizons. Every capex dollar locks in years of tool orders, memory wafer starts, packaging slots, and grid contracts.
Saudi Arabia's HUMAIN, UAE's G42, India's IndiaAI Mission, Japan's AI Bridging Cloud, France's Mistral cluster — none existed at scale 18 months ago. Each needs the same nine layers. Sovereign demand layers on top of, not instead of, hyperscaler demand. The implication for $NVDA, $AVGO, $MU, $TSM, $ASML, $COHR, and $VRT is straight TAM expansion.
HBM is sold out through 2026E. CoWoS capacity needs to double. $BESIY hybrid bonders are on 12-month lead times. $ASML EUV machines are pre-allocated for years. Gas turbines are sold out through 2027. Each bottleneck is its own equity story — as long as it holds, pricing power holds.
| Layer | Bottleneck | Publicly-listed leader |
|---|---|---|
| 1. Compute | Advanced packaging slots (not silicon design) | $NVDA, $AVGO, $MRVL |
| 2. Memory (HBM) | HBM wafer capacity, sold out through 2026 | $MU |
| 3. HBM Packaging | Hybrid bonders, CoWoS capacity | $BESIY, $AMKR, $ONTO |
| 4. Lithography / Fab Tools | EUV machine output (single-supplier) | $ASML, $LRCX, $AMAT, $KLAC |
| 5. Photonics / CPO | InP laser yields, ELS supply | $COHR, $LITE, $FN |
| 6. Networking / Retimers | PCIe retimer supply per rack | $ALAB, $CRDO, $ANET |
| 7. Cooling | Liquid CDU integration capacity | $VRT, $MOD |
| 8. Power & Grid | Grid interconnects + gas turbines + nuclear PPAs | $GEV, $CEG, $VST, $ETN |
| 9. Neoclouds | Capital + GPU allocation + power contracts | $CRWV, $NBIS, $SMCI |
Framework for picking inside a layer: prioritise (a) single-supplier choke points over competitive ones; (b) highest revenue concentration in AI infra vs legacy; (c) customer concentration in hyperscalers signing multi-year POs, not spot resellers.
The dominant retail "AI" ETFs — BOTZ, AIQ, ROBO, IRBO — dilute AI infrastructure with adjacent themes that don't scale with hyperscaler capex.
| BOTZ / AIQ / ROBO / IRBO | Macroplane AI-infra stack | |
|---|---|---|
| AI-infra pure-play | ~30-40% — diluted with industrial automation, robotics, and legacy software | 100% — every name maps to a specific AI-infra layer |
| Sector mix | Mixes factory automation with capex beneficiaries | Pure capex-pull-through across 9 layers |
| Index methodology | Backward-looking — names added after re-rating | Forward-looking — bottleneck-driven |
| Concentration | Diluted across 30-50 names | Concentrated in highest-conviction bottleneck names |
| Expense ratio | 0.40-0.68% per year | Free to track on Macroplane |
If your thesis is "hyperscalers and sovereigns are building a $1.5T physical infrastructure over 5 years," BOTZ and friends are the wrong vehicle — they hold industrial automation names whose revenues don't inflect on AI capex.
👉 Build your own: combine constituents across the nine Macroplane baskets — Custom Silicon, Memory Supercycle, HBM/Packaging, Lithography & Fab Tools, Photonics/CPO, Networking/Retimers, Cooling, Power & Grid, AI Cloud/Neoclouds.
There's no single answer — there are nine layers, each with its own bottleneck depth. Default megacap exposures: $NVDA (compute), $MU (memory), $ASML and $TSM (fab tools), $AVGO (custom ASIC). Asymmetric mid-caps: $ALAB and $CRDO in retimers, $VRT in cooling, $GEV in power, $COHR in photonics, $BESIY in hybrid bonders, $CRWV in neoclouds. A diversified portfolio across the nine layers is the standard institutional approach.
"AI stocks" lumps software (Microsoft, Salesforce, Palantir), hyperscalers, and infrastructure together. AI infrastructure is specifically the physical buildout — silicon, memory, packaging, networking, cooling, power, and the data centers that house them. It's what monetises hyperscaler capex regardless of which AI application wins.
No — and that's the central insight here. $NVDA captures the compute layer; the other 8 layers have their own publicly-listed leaders. $MU, $ASML, $AVGO, $COHR, $VRT, $GEV, and $CRWV are all materially exposed to the same capex super-cycle without taking single-architecture risk on Blackwell vs Rubin vs MI400.
A genuine demand pause from one or more hyperscalers. A Microsoft, Google, or Meta capex cut would re-price the entire stack. Mitigants: (a) sovereign demand layering on top, (b) 12-24 month booking visibility, (c) bottleneck depth — when you cannot get hybrid bonders for 18 months, pricing power holds even if one buyer pauses.
Macroplane organises the universe into the nine baskets above plus the Data Center Buildout trend and the Silicon Photonics & Optical Interconnects trend. Each company page links to live financials, news, and the supplier/buyer relationship map — so when $NVDA announces a Rubin design win at $AVGO, you can trace the impact to packaging tool orders at $BESIY and laser shipments at $COHR.
There's no single answer — there are nine layers, each with its own bottleneck depth. Default megacap exposures: $NVDA (compute), $MU (memory), $ASML and $TSM (fab tools), $AVGO (custom ASIC). Asymmetric mid-caps: $ALAB and $CRDO in retimers, $VRT in cooling, $GEV in power, $COHR in photonics, $BESIY in hybrid bonders, $CRWV in neoclouds. A diversified portfolio across the nine layers is the standard institutional approach.
"AI stocks" lumps software (Microsoft, Salesforce, Palantir), hyperscalers, and infrastructure together. AI infrastructure is specifically the physical buildout — silicon, memory, packaging, networking, cooling, power, and the data centers that house them. It's what monetises hyperscaler capex regardless of which AI application wins.
No — and that's the central insight here. $NVDA captures the compute layer; the other 8 layers have their own publicly-listed leaders. $MU, $ASML, $AVGO, $COHR, $VRT, $GEV, and $CRWV are all materially exposed to the same capex super-cycle without taking single-architecture risk on Blackwell vs Rubin vs MI400.
A genuine demand pause from one or more hyperscalers. A Microsoft, Google, or Meta capex cut would re-price the entire stack. Mitigants: (a) sovereign demand layering on top, (b) 12-24 month booking visibility, (c) bottleneck depth — when you cannot get hybrid bonders for 18 months, pricing power holds even if one buyer pauses.
Macroplane organises the universe into the nine baskets above plus the Data Center Buildout trend and the Silicon Photonics & Optical Interconnects trend. Each company page links to live financials, news, and the supplier/buyer relationship map — so when $NVDA announces a Rubin design win at $AVGO, you can trace the impact to packaging tool orders at $BESIY and laser shipments at $COHR.