AI Infrastructure Companies: The Stack, Layer by Layer
Mark aistrategy.guide as a preferred source Our AI strategy guides show up more often, with a Preferred badgeA strategy map of the AI infrastructure companies an enterprise depends on, whether or not it buys from them directly: chips, neoclouds, power, cooling, and data center developers and REITs. Where the constraints sit in each layer, and what to ask your providers. Not investment advice.
Most executives buy AI capacity from one or two providers and never look below them. That worked when cloud capacity was elastic. In 2026 it is not, and the delivery date on any AI capacity contract depends on a chain of companies the buyer never signs with: the chip vendor, the power supplier, the cooling and electrical equipment makers, and the developer that owns the building. A slip in any of them becomes a slip in your roadmap.
This page maps those layers from the buyer’s side. It names companies to show where the dependencies are. It is not investment advice, and nothing here is a view on any company’s shares.
The stack at a glance
| Layer | What it supplies | Representative companies | Current constraint |
|---|---|---|---|
| Chips and systems | GPUs, accelerators, networking, fabrication | NVIDIA, AMD, Google (TPU), AWS (Trainium), TSMC | Allocation of the newest generation |
| Compute providers | Rented GPU capacity | AWS, Microsoft Azure, Google Cloud, Oracle; neoclouds such as CoreWeave, Nebius, Lambda, Crusoe | Energised data center space |
| Power | Electricity, firm and clean | Utilities, independent power producers, Constellation, new-nuclear developers | Grid interconnection and new generation |
| Cooling and electrical | Liquid cooling, power distribution, backup | Vertiv, Schneider Electric and others | Lead times for high-density equipment |
| Buildings | Land, shells, halls, interconnection | Equinix, Digital Realty, private developers | Sites with power available |
Layer 1: chips and systems
NVIDIA remains the reference supplier. Its data center segment reported $89.0 billion of revenue in the quarter ended 26 July 2026, up 117% year on year, and the company said its Vera Rubin platform was in full production. AMD sells competing Instinct accelerators, and the largest clouds design their own chips (Google’s TPU, AWS Trainium) to reduce their dependence on a single vendor. Most leading-edge accelerators, whoever designs them, are manufactured by TSMC.
What it means for a buyer: you rarely buy chips directly, but the generation you rent determines your cost per token for the life of the contract. When a new generation ships, the previous one gets cheaper to rent. Avoid long commitments to hardware that is about to be superseded, and ask providers which generation your capacity will run on and when.
Layer 2: compute providers, hyperscalers and neoclouds
Hyperscalers sell GPUs inside a full cloud platform. Neoclouds sell GPU capacity as their core product. The neocloud layer has grown on the back of very large contracts: CoreWeave announced in September 2025 that its contracts with OpenAI totalled up to about $22.4 billion, and Nebius signed a five-year agreement with Microsoft worth $17.4 billion in September 2025. In other words, hyperscalers and AI labs are themselves major neocloud customers.
What it means for a buyer: a neocloud’s largest customers get priority when capacity is short. Check customer concentration and the energisation status of the facility your capacity sits in. The full placement framework is in hyperscaler vs neocloud vs on-prem, and the platform-by-platform view is in GPU and inference platforms.
Layer 3: power
Power is the slowest layer to expand, and it now sets the pace for the others. The US Department of Energy’s summary of the 2024 Lawrence Berkeley National Laboratory report put data centers at about 4.4% of US electricity in 2023 and projected 6.7% to 12% by 2028.
The largest buyers are signing long-term deals for firm, low-carbon power, much of it nuclear:
- Microsoft and Constellation signed a 20-year power purchase agreement in September 2024 to restart Three Mile Island Unit 1 as the Crane Clean Energy Center, about 835 MW.
- Google and Kairos Power agreed in October 2024 to buy power from multiple small modular reactors, up to 500 MW, with the first unit targeted for 2030.
- Amazon and X-energy announced plans for small modular reactors, with an initial Washington state project of 320 MW and an option to expand to 960 MW.
- Microreactor developers such as Valar Atomics and Radiant reached test milestones in 2026; the buyer’s view of what that means is in Valar Atomics vs Radiant.
What it means for a buyer: nearly all of the new nuclear capacity arrives in 2027 or later. For the next two planning cycles, power availability is a site and provider selection criterion. Ask every provider where the power for your capacity comes from and when it is energised.
Layer 4: cooling and electrical equipment
The newest AI systems are liquid-cooled and dense. Vertiv and NVIDIA’s reference architecture for the GB200 NVL72 platform is built around racks of up to 132 kW, many times the density older data center halls were designed for. That makes cooling distribution units, chillers, switchgear, transformers and backup power part of the critical path. Vertiv and Schneider Electric are among the large suppliers; there are many others.
What it means for a buyer: most enterprise server rooms cannot host current high-end AI systems without a retrofit. If an on-prem plan assumes the existing room, check density and cooling before buying hardware. For colocation, ask whether the hall you are offered is already built for liquid cooling or is waiting on equipment.
Layer 5: data center developers and REITs
The buildings are owned by publicly listed data center REITs such as Equinix and Digital Realty, by privately held developers, and increasingly by the hyperscalers themselves. Developers compete on one thing above all in 2026: sites with power available or contracted.
What it means for a buyer: if you lease colocation directly, the developer’s power and interconnection position is your capacity risk. If you buy from a neocloud, the same risk sits one step removed. Either way, ask who owns the building and what the energisation date is.
How should an executive use this map?
- Trace one critical workload down the stack. For the AI workload your roadmap depends on most, name the provider, the facility, the GPU generation and the power source. Gaps in that chain are your risks.
- Ask the bottleneck question in every renewal. “Which layer is constraining your delivery to us this year?” A provider that cannot answer it does not know.
- Separate long-dated bets from current supply. New nuclear, new chip generations and new campuses are real but mostly 2027 and later. Plan this year’s capacity on what is energised now.
- Keep a second path. Whatever layer is tightest, a second provider in a different facility on different power is the cheapest hedge available.
The engineering view of the same stack, from rack design to interconnects, is on ctaio.dev’s AI infrastructure hub.
The verdict
Your AI capacity is only as reliable as the weakest layer under your provider, and in 2026 that layer is usually power. Buy accordingly.
Sources & methodology
- NVIDIA: financial results for Q2 fiscal 2027 (26 Aug 2026)
- CoreWeave: CoreWeave expands agreement with OpenAI by up to $6.5B (25 Sep 2025)
- TechRadar: Nebius and Microsoft multi-billion AI infrastructure agreement (Sep 2025)
- US Department of Energy: report evaluating increase in electricity demand from data centers (LBNL 2024)
- Constellation: Constellation to launch Crane Clean Energy Center (20 Sep 2024)
- Google: advanced nuclear clean energy agreement with Kairos Power (Oct 2024)
- Amazon: small modular reactors and X-energy
- Vertiv: complete power and cooling blueprint for NVIDIA GB200 NVL72
- Equinix and Digital Realty company sites
- Methodology: company facts from company announcements and filings, checked 20 October 2026. Companies are named as examples of each layer, not as a complete or ranked list. Not investment advice.
