The 7-layer data center stock map — compute, memory, networking, storage, cooling, power, construction — with the publicly-listed leader in each layer (from $NVDA and $MU to $VRT, $GEV, and $PWR) and why the SRVR / DTCR / AIQ ETFs underweight the most asymmetric layers.
More Macroplane supply-chain guides · Baskets · Investment theses · Macro trends
Data center stocks are publicly traded companies that build, equip, or power the physical infrastructure behind the AI boom — the silicon, memory, networking, storage, cooling, electrical, and construction layers that turn megawatts of grid power and trillions of dollars of capex into trained models and inferenced tokens. These are the picks-and-shovels of the AI buildout: most are not pure AI plays, but every one sells a critical component into a hyperscaler order book that has expanded 60–80% year-over-year heading into 2026.
This guide walks through the seven layers of a modern AI data center, the public names that matter at each stage, and why the 2026 trade is structurally different from anything before it.
👉 See live prices and performance: Data Center Buildout trend on Macroplane — live tracking of the categories, baskets, and companies powering the AI buildout.
A modern AI campus is not "a building full of servers." It is a stack of seven interdependent supply chains, each with its own public-equity expression.
$NVDA — NVIDIA ships most merchant AI training silicon (Blackwell, Rubin in 2026–2027) plus the NVLink/InfiniBand networking that ties racks together.
$AMD — Advanced Micro Devices is the credible #2 in AI accelerators (MI300X, MI350); EPYC remains the workhorse general-compute CPU in AI servers.
$AVGO — Broadcom co-designs Google's TPUs and Meta's MTIA accelerators and ships the Tomahawk switching silicon — the cleanest read on hyperscaler in-house-silicon strategy.
$MRVL — Marvell is the other custom-silicon house (Amazon Trainium, Microsoft Maia) and owns dominant share of electro-optical DSPs inside every 800G and 1.6T transceiver.
$ARM — Arm Holdings is the architecture under hyperscaler in-house CPUs (AWS Graviton, NVIDIA Grace, Microsoft Cobalt, Google Axion). Every custom CPU shipped into a hyperscaler pays Arm a royalty.
$INTC — Intel has the largest installed base of data center CPUs plus optionality on the US foundry ramp and Gaudi/Falcon Shores accelerators.
👉 Custom Silicon basket — merchant and hyperscaler-custom AI silicon as one watchlist.
$MU — Micron is the American HBM (high-bandwidth memory) play. Every Blackwell, every MI300, every TPU needs 6–12 stacks of HBM3E/HBM4. One of three HBM producers globally and the only US-listed pure-play. HBM is structurally tighter than commodity DRAM — capacity is contracted years ahead and pricing has decoupled from the broader memory cycle.
$SNDK — SanDisk is the NAND side; AI workloads write huge volumes of intermediate state and checkpoints to flash. Re-listed in 2025 after the Western Digital spinout — the cleanest US-listed enterprise NAND exposure.
👉 Memory Supercycle basket — HBM and enterprise NAND benefiting from AI memory tightness.
$ANET — Arista Networks is the dominant merchant switch vendor for hyperscale Ethernet fabrics. AI clusters spend two to four times more on networking per GPU than traditional cloud workloads.
$CSCO — Cisco is the incumbent enterprise switching name, repositioning around the AI campus — less pure than Arista but a credible vehicle for enterprise AI infrastructure spend.
$ALAB — Astera Labs is the PCIe and CXL retimer pure-play. Every Blackwell rack uses dozens of Astera retimers to hold PCIe Gen 5/6 signals together.
$CRDO — Credo ships active electrical cables (AECs) and retimers for short rack-to-rack copper links. Hyperscaler-concentrated revenue, which cuts both ways.
👉 Networking & Retimers basket — switches, optical DSPs, and retimers in one view.
$STX — Seagate and $WDC — Western Digital are the nearline HDD duopoly — the bulk-storage layer holding training datasets and model artifacts. HAMR and ePMR are pushing per-drive density past 30TB.
$NTAP — NetApp and $PSTG — Pure Storage are the enterprise flash array names. Pure in particular has won share in AI training storage with FlashBlade — checkpoints need very high parallel read bandwidth.
$DELL — Dell and $HPE — Hewlett Packard Enterprise are the AI server integrators. Thinner margins than component vendors, but the revenue scale and enterprise channel access are unique.
👉 Storage basket — disk, flash, and AI server OEMs as a single exposure.
$VRT — Vertiv is the pure-play cooling and power-management vendor — the most direct exposure to the air-to-liquid transition. Order book essentially sold out into 2026.
$MOD — Modine ships specialty heat exchangers and CDUs — smaller and more cyclical than Vertiv, levered to the same shift.
$NVT — nVent, $JCI — Johnson Controls, $TT — Trane, $AAON, $CARR — Carrier, and $SBGSY — Schneider Electric are the broader thermal/HVAC names with meaningful data-center exposure. Data center is the fastest-growing segment for all of them; Schneider bundles cooling with power and prefab modular data centers.
👉 Cooling basket — pure-play and broader-play thermal management.
$GEV — GE Vernova is the largest pure-play power-generation equipment vendor. Gas-turbine lead times have stretched to 2028–2029.
$CEG — Constellation, $VST — Vistra, and $TLN — Talen Energy are the nuclear and gas-generation owners signing direct PPAs with hyperscalers. Constellation's Three Mile Island restart for Microsoft and Talen's Susquehanna PPA with Amazon are the templates — baseload that bypasses 5–7 year interconnect queues.
$ETN — Eaton, $HUBB — Hubbell, $POWL — Powell, and $VICR — Vicor cover switchgear, busways, transformers, and power-conversion modules. Eaton is the bellwether; Powell is a smaller pure-play and one of the largest percentage beneficiaries of the buildout.
$GNRC — Generac and $FLNC — Fluence cover backup generation and grid-scale battery storage. Even campuses with PPAs need on-site backup, and batteries are now part of the grid-firming story behind every new contract.
👉 Power & Grid basket — generation, distribution, and on-site power.
$PWR — Quanta, $EME — EMCOR, $FIX — Comfort Systems, $MTZ — MasTec, $DY — Dycom, $MYRG — MYR Group, and $PRIM — Primoris are the electrical, mechanical, and fiber contractors assembling data centers in the field. Quanta and MasTec dominate utility-scale grid work; EMCOR and Comfort Systems lead MEP contracting inside the building; Dycom strings the fiber. Backlogs at all seven are at multi-year highs.
👉 Construction & MEP basket — the contractors converting capex into operating capacity.
$APH — Amphenol, $TEL — TE Connectivity, $GLW — Corning, $BDC — Belden, and $COMM — CommScope ship every cable, connector, and passive optical patch panel in a hyperscale fabric. Amphenol is the most concentrated AI play in the group; Corning's fiber is a direct read on long-haul and intra-campus optics.
👉 Connectors basket — the physical interconnect layer underneath everything else.
Three forces are driving the trade right now.
$MSFT, $AMZN, $GOOGL, $META, and $ORCL have collectively guided to $500B+ of 2026 capex versus ~$320B in 2025. Every dollar flows into one of the seven layers above. Microsoft and Meta have publicly committed to triple-digit GW campus plans. Forecasters now expect 100+ GW of new capacity online globally by 2030 — roughly doubling the installed base. There is no historical analog for this rate of single-end-market capex acceleration.
The real shortage is not GPUs — it is grid interconnects and gas turbines. Utility-scale connection queues now run five to seven years in PJM, ERCOT, MISO, and CAISO. $GEV turbines are sold out into 2029. That's why hyperscalers are signing direct nuclear PPAs ($CEG with Microsoft, $TLN with Amazon), restarting decommissioned plants, and standing up behind-the-meter gas at the campus. Pricing power is concentrated in $CEG, $VST, $TLN, $GEV, $ETN, $POWL, $GNRC, and $FLNC.
Blackwell racks dissipate 120+ kW. Rubin will push that to 200+ kW. Air cooling is physically impossible above roughly 50 kW per rack, so every new AI campus is being built with direct-to-chip liquid, immersion, or rear-door heat exchangers from day one. Retrofitting existing air-cooled campuses is the second wave. $VRT, $MOD, and the HVAC majors ($JCI, $TT, $CARR, $SBGSY) are the public expressions. Backlogs at $VRT have grown faster than revenue every quarter for two years.
Most investors think "data center = NVIDIA." That's wrong for portfolio construction — the dollar breakdown is far more distributed than the headlines suggest.
| Layer | Approximate $/MW of build | Listed names | Margin profile |
|---|---|---|---|
| Compute silicon | $25–40M | $NVDA, $AMD, $AVGO, $MRVL, $ARM, $INTC | High gross margin, customer-concentrated |
| Memory (HBM + NAND) | $4–8M | $MU, $SNDK | Cyclical; HBM tighter than commodity DRAM |
| Networking & optics | $3–6M | $ANET, $CSCO, $ALAB, $CRDO, $AVGO, $MRVL | High margin at silicon, thinner at systems |
| Storage | $1–3M | $STX, $WDC, $NTAP, $PSTG, $DELL, $HPE | Mixed — components high, systems thin |
| Cooling | $3–6M | $VRT, $MOD, $NVT, $JCI, $TT, $AAON, $CARR, $SBGSY | Improving with liquid-cooling mix shift |
| Power & grid | $8–15M | $GEV, $CEG, $VST, $TLN, $ETN, $HUBB, $POWL, $VICR, $GNRC, $FLNC | Strong — capacity-constrained, multi-year backlogs |
| Construction & MEP | $6–12M | $PWR, $EME, $FIX, $MTZ, $DY, $MYRG, $PRIM | Improving — backlogs at all-time highs |
Silicon is roughly a third of total build cost. The other two-thirds — memory, networking, storage, cooling, power, construction — are spread across forty-plus public companies, many carrying valuations a fraction of NVIDIA's despite being structurally short of capacity.
Two ETF families are worth knowing.
Data-center REIT ETFs — DTCR (Global X) and SRVR (Pacer) — own primarily the campus landlords: $EQIX, $DLR, and international peers. That's real-estate exposure, not infrastructure-equipment exposure: it benefits from leasing rates and occupancy but not directly from the silicon, cooling, or power capex cycle.
AI infrastructure ETFs — AIQ (Global X) and BOTZ (Global X) — are software- and applications-heavy. AIQ's largest holdings are megacap software ($MSFT, $GOOGL, $META); BOTZ is robotics-tilted. Neither delivers the pure picks-and-shovels exposure that the cooling, power, and construction names provide.
| DTCR / SRVR | AIQ / BOTZ | Macroplane data-center baskets | |
|---|---|---|---|
| REIT exposure | Yes — primary | No | No |
| Compute silicon | Minimal | Yes — megacap-tilted | Yes — Custom Silicon |
| Cooling | No | No | Yes — Cooling |
| Power & grid | No | No | Yes — Power & Grid |
| Construction & MEP | No | No | Yes — Construction & MEP |
| Memory (HBM/NAND) | No | Limited | Yes — Memory Supercycle |
| Pure picks-and-shovels | Partial (REITs) | No | Yes |
| Expense ratio | 0.50% / 0.55% | 0.68% / 0.68% | Free to track on Macroplane |
If you want REIT-style yield exposure to data-center landlords ($EQIX, $DLR), DTCR or SRVR are fine. If you want exposure to the capex — the silicon, memory, networking, cooling, power, and construction inside the buildings — no single ETF gives you that cleanly. The Macroplane baskets isolate one layer at a time so you can size cooling separately from silicon, separately from power.
👉 Build it your way: clone any of the Data Center Buildout baskets on Macroplane, mix layers, set weightings — live performance updated continuously.
The buildout is a stacked supply chain — the answer depends on which layer you're underweight. The cleanest pure-AI plays are $NVDA, $AMD, $AVGO, and $MRVL on silicon; $MU on HBM; $ANET and $ALAB on networking; $VRT on cooling; and $GEV, $CEG, $ETN, and $POWL on power. Broader-tailwind names like $JCI, $TT, $CARR, $ETN, and $HUBB give buildout exposure without single-end-market concentration. Most institutional books hold a mix across layers.
GPU supply is constrained but scaling — $NVDA, $AMD, $AVGO, and $MRVL keep adding capacity at TSMC and Samsung. Power is constrained by physics and regulation: a new gas turbine takes $GEV three to four years; a grid interconnect in PJM or ERCOT runs five to seven years; a new nuclear plant is a decade-plus. AI campuses can't be built faster than power can be delivered, which is why nuclear PPAs, on-site gas, batteries, and behind-the-meter generation have become primary investment themes.
Some are pure AI plays — $NVDA, $AMD, $MRVL, $ALAB, $CRDO, $VRT — where the buildout is essentially the entire growth story. Others are broader-tailwind names where AI is the incremental driver: $JCI, $TT, $CARR, $ETN, $HUBB, $APH, $TEL. Pure-plays have higher beta to a hyperscaler capex slowdown; broader-tailwind names give up some upside but absorb a soft patch better.
Roughly 70–80%. The remainder is land, internal labor, software, and a small share of private suppliers. The $500B+ of guided 2026 capex translates into roughly $350–400B of addressable spend across the public companies on this page.
The Data Center Buildout trend on Macroplane shows the categories, baskets, and companies tied to the buildout in one place, with continuous price tracking and per-company news. The individual baskets — Custom Silicon, Memory Supercycle, Networking & Retimers, Storage, Cooling, Power & Grid, Construction & MEP, Connectors — let you isolate one layer at a time.
The buildout is a stacked supply chain — the answer depends on which layer you're underweight. The cleanest pure-AI plays are $NVDA, $AMD, $AVGO, and $MRVL on silicon; $MU on HBM; $ANET and $ALAB on networking; $VRT on cooling; and $GEV, $CEG, $ETN, and $POWL on power. Broader-tailwind names like $JCI, $TT, $CARR, $ETN, and $HUBB give buildout exposure without single-end-market concentration. Most institutional books hold a mix across layers.
GPU supply is constrained but scaling — $NVDA, $AMD, $AVGO, and $MRVL keep adding capacity at TSMC and Samsung. Power is constrained by physics and regulation: a new gas turbine takes $GEV three to four years; a grid interconnect in PJM or ERCOT runs five to seven years; a new nuclear plant is a decade-plus. AI campuses can't be built faster than power can be delivered, which is why nuclear PPAs, on-site gas, batteries, and behind-the-meter generation have become primary investment themes.
Some are pure AI plays — $NVDA, $AMD, $MRVL, $ALAB, $CRDO, $VRT — where the buildout is essentially the entire growth story. Others are broader-tailwind names where AI is the incremental driver: $JCI, $TT, $CARR, $ETN, $HUBB, $APH, $TEL. Pure-plays have higher beta to a hyperscaler capex slowdown; broader-tailwind names give up some upside but absorb a soft patch better.
Roughly 70–80%. The remainder is land, internal labor, software, and a small share of private suppliers. The $500B+ of guided 2026 capex translates into roughly $350–400B of addressable spend across the public companies on this page.
The Data Center Buildout trend on Macroplane shows the categories, baskets, and companies tied to the buildout in one place, with continuous price tracking and per-company news. The individual baskets — Custom Silicon, Memory Supercycle, Networking & Retimers, Storage, Cooling, Power & Grid, Construction & MEP, Connectors — let you isolate one layer at a time.