The AI Infrastructure Map.

AI training and inference consume electricity at industrial scale. Hyperscaler contracts are reshaping utility P&Ls, restarting shuttered nuclear plants, and pulling gas turbines and transmission into multi-year backlog.

• 01CEGConstellation EnergyLargest US nuclear fleet; Three Mile Island restart for Microsoft — the purest nuclear-for-AI trade.
• 02VSTVistraTexas nuclear + gas generation signing hyperscaler PPAs; recent top-performing S&P utility.
• 03TLNTalen EnergyOwns the Susquehanna nuclear and Cumulus data-center campus feeding AWS; post-bankruptcy compound.
• 04NEENextEra EnergyLargest US renewables developer plus regulated Florida utility; the boring way to own AI load growth.
• 05GEVGE VernovaGas turbines, grid equipment, offshore wind; multi-year backlog constraints are the bull case.
• 06ETNEatonSwitchgear, UPS, data-center electrical distribution; the quiet compounder of electrification.
• 07PWRQuanta ServicesThe contractor actually building transmission; labor-constrained grid-buildout beneficiary.
• 08CCJCamecoLargest Western uranium miner; tier-one fuel supplier to the nuclear renaissance.
• 09OKLOOkloSmall modular reactor pre-revenue moonshot; Sam Altman-chaired — speculative but on-thesis.
• 10SMRNuScale PowerFirst NRC-approved SMR design; the pure SMR bet if the technology actually scales.
• 11BWXTBWX TechnologiesNaval nuclear components and SMR supplier; the picks-and-shovels of the SMR trade.
• 12WMBWilliams CompaniesNatural-gas midstream; pipelines are what actually fuel new data-center gas plants.
• 13URNMSprott Uranium Miners ETFBasket exposure to uranium miners for investors who don't want single-name risk.

The chip stack is the binding constraint on AI progress. GPU supply is gated by TSMC advanced packaging, which is gated by ASML EUV, which is gated by HBM memory from Micron, SK hynix, and Samsung.

• 01NVDANvidiaThe king: GPUs, the CUDA moat, Mellanox networking, and the expanding software stack.
• 02AVGOBroadcomCustom silicon partner to Google (TPU) and Meta (MTIA); dominant AI networking ASICs.
• 03AMDAMDMI300 and MI350 accelerators; the only credible GPU alternative buyers want to fund.
• 04TSMTSMCFabs every leading-edge AI chip on earth — the real bottleneck of the entire stack.
• 05ASMLASMLEUV lithography monopoly; no EUV means no sub-5nm chips.
• 06AMATApplied MaterialsBroadest wafer-fab-equipment exposure; benefits from every fab buildout.
• 07LRCXLam ResearchEtch and deposition leader; critical to HBM memory stacking.
• 08KLACKLAProcess-control and metrology monopoly; yields collapse without it.
• 09MUMicron TechnologyHBM3E memory for Nvidia; memory bandwidth is now the binding constraint on AI performance.
• 10ARMArm HoldingsCPU IP inside every hyperscaler's custom silicon — Graviton, Axion, Cobalt.
• 11MRVLMarvell TechnologyCustom silicon (AWS Trainium) and optical DSPs for data-center interconnect.
• 12ALABAstera LabsPCIe retimers and CXL connectivity — the glue chips inside every AI rack.
• 13CRDOCredo TechnologyActive electrical cables and SerDes for AI-rack interconnect; pure-play AI networking IC.

Data centers are the physical factories. Roughly $300B in global construction is planned. Colocation REITs, cooling specialists, networking, and systems integrators build the box around the chips.

• 01EQIXEquinixLargest global colo REIT; interconnection hub between clouds, carriers, and enterprises.
• 02DLRDigital RealtyNumber two colo REIT; hyperscale wholesale exposure across six continents.
• 03IRMIron MountainPaper-to-data-center pivot is working; under-appreciated REIT re-rating.
• 04AMTAmerican TowerTower REIT with data-center assets via CoreSite; 5G and edge-compute crossover.
• 05VRTVertiv HoldingsPower and cooling infrastructure; leader in liquid cooling for dense AI racks.
• 06MODModine ManufacturingCooling systems; the cheaper, less-covered Vertiv alternative.
• 07ANETArista NetworksHigh-performance Ethernet switching for AI clusters; Nvidia-networking challenger.
• 08CIENCienaOptical networking between data centers; benefits as AI workloads go multi-site.
• 09SMCISuper Micro ComputerServer integrator; Nvidia's preferred rack-scale system builder (governance risk noted).
• 10DELLDell TechnologiesPowerEdge AI servers; enterprise AI infrastructure beneficiary.
• 11HPEHewlett Packard EnterpriseCray-derived AI systems plus Juniper (pending close) for AI networking.

Twenty-four terms decoded in plain English. Keep this open the next time you read a SemiAnalysis post or a hyperscaler earnings call.

Hyperscaler

The biggest cloud and AI operators — AWS, Azure, Google Cloud, Meta, Oracle. They set AI infrastructure demand.

Colocation (colo)

A data-center business model where tenants rent space, power, and cooling rather than own the facility.

HBM (High-Bandwidth Memory)

Stacked DRAM packaged next to GPUs. HBM3E is the current generation; HBM4 ships in 2026. Capacity is the binding constraint on AI chips.

CoWoS

TSMC's advanced-packaging process that bonds HBM stacks to GPUs. Limited capacity here limits Nvidia shipments.

SMR (Small Modular Reactor)

Factory-built nuclear reactors in the 50–300 MW range. Pitched as the answer to data-center power demand.

PPA (Power Purchase Agreement)

Long-term contract to buy power at a fixed price. Microsoft, Amazon, and Google are signing multi-decade PPAs with nuclear and renewables operators.

Interconnect

The networking that moves data between GPUs, racks, and data centers. Bottlenecks here stall AI training runs.

SerDes

Serializer/deserializer — the chips that convert parallel data to serial for high-speed transmission. Every link in an AI cluster contains them.

REIT

Real Estate Investment Trust. Data-center and tower REITs are the landlords of the AI buildout.

Megawatt (MW)

The common unit of data-center capacity. A large AI campus is 1,000+ MW — roughly the power draw of a mid-sized US city.

EUV

Extreme ultraviolet lithography. ASML is the sole supplier of EUV scanners, the machines that pattern leading-edge chips.

Foundry

A contract chip manufacturer. TSMC is the dominant foundry; Samsung and Intel Foundry compete for share.