Every dollar of hyperscaler capex ends up with a supplier. FactSet estimates that compute, meaning chips and servers, takes about $380 billion, roughly 60%, of the nearly $800 billion the five largest hyperscalers will spend in 2026. Nvidia’s data center revenue hit $89.0 billion in the quarter to July 2026, up 117%, and it guided to $108 billion of total revenue for the following quarter. In power and cooling, Vertiv’s Q3 2026 sales rose 61% to $4.31 billion on organic orders that more than doubled. Caterpillar’s Power & Energy sales rose 18%. Consolidation is accelerating too: GE Vernova agreed a $9.0 billion deal for Trench Group, Ecolab a $4.75 billion deal for CoolIT, and Vertiv closed a $1.45 billion acquisition of UtilityInnovation Group.
The data center supply chain is the network of companies that make the chips, servers, power gear, cooling systems and backup generators inside every AI campus. As Big Tech spending has soared, these “picks-and-shovels” suppliers have become some of the biggest winners of the AI build-out. This guide maps the supply chain layer by layer: accelerators, electrical equipment, cooling, grid equipment and on-site generation. It covers the leading companies, their latest results through October 2026, and the bottlenecks that slow construction. It is part of the Kurums Data Center Economy hub. Read it with our analysis of the AI capex supercycle and our guide to the neoclouds that buy much of this equipment.
Where does most data center money go?
To IT hardware. Microsoft says about two-thirds of its fiscal 2026 capex went to servers and other IT equipment, and FactSet estimates compute at about 60% of total 2026 hyperscaler capex.
Who are the biggest suppliers?
Nvidia dominates AI accelerators. Schneider Electric, Vertiv, Eaton, ABB and Siemens lead in power and cooling; GE Vernova, Siemens Energy and Hitachi Energy in grid and turbine equipment; Caterpillar, Cummins and Rolls-Royce in backup generation.
What is the bottleneck?
Power equipment and grid connections more than chips. Transformers, switchgear, turbines and skilled labour remain constrained, and suppliers are buying capacity and companies to keep up.
What does the data center supply chain consist of?
The supply chain has five main layers: compute (GPUs, custom accelerators, CPUs, memory, networking), power distribution inside the building (switchgear, UPS, busways), cooling (chillers, liquid-cooling loops), grid and substation equipment (transformers, high-voltage gear) and on-site generation (diesel and gas generators, turbines, batteries).
Spending is split unevenly. In Microsoft’s case, roughly two-thirds of fiscal 2026 capex of $145.3 billion went to IT hardware such as servers, GPUs and networking, and the rest to buildings, land and infrastructure, according to DatacenterDynamics’ report of its earnings call. FactSet estimates that compute will take about $380 billion, or roughly 60%, of the close to $800 billion that the five largest hyperscalers are expected to spend in calendar 2026 including leases.
The non-compute share matters more than it used to. AI racks draw several times more power than traditional racks, so each megawatt of AI capacity needs more electrical and cooling equipment. Campuses are also scaling from tens of megawatts to gigawatts, which brings grid and generation equipment directly into the data center budget. Our explainer on data center cooling covers the technical shift, and our cost-per-MW guide shows how it changes build costs.
How dominant is Nvidia in data center compute?
Nvidia is the dominant supplier of AI accelerators. Its data center revenue reached $89.0 billion in the quarter to July 2026, up 117% year over year, out of total revenue of $96.2 billion. It guided to about $108 billion of revenue for the quarter to October 2026.
For fiscal 2026, which ended in January 2026, Nvidia reported revenue of $215.9 billion and net income of $120.1 billion. Its gross margin was about 75% in the July 2026 quarter. The product cycle keeps moving: Blackwell systems dominated 2025 and early 2026 shipments, and the Vera Rubin platform, detailed at GTC in March 2026, is ramping in the second half of 2026. In late October 2026 Nvidia and Oracle announced an expansion of Vera Rubin deployments to multi-gigawatt scale. Nvidia’s outlook assumed no data center compute revenue from China because of export restrictions.
Nvidia is also investing heavily to secure its own supply chain. In October 2026 it announced plans to increase substrate capacity 2.5-fold over three years, grow interposer wafer capacity by more than 300% by the end of 2028 and raise optical-engine capacity 15-fold by mid-2027. It also made a $1.1 billion capacity-reservation advance to test partner KYEC. Those commitments show where the next bottlenecks sit: advanced packaging, substrates and optical interconnects rather than the GPU dies themselves.
Custom chips and the competition
Hyperscalers are building their own accelerators to reduce dependence on Nvidia. Google designs Tensor Processing Units, and Amazon’s Trainium powers Project Rainier, which launched with more than 500,000 Trainium2 chips. Amazon said revenue commitments for Trainium and Graviton each rose more than 20-fold year over year by Q3 2026, and it plans to offer Trainium outside AWS from late 2027. Microsoft and Meta are developing in-house chips as well. Memory makers have pricing power too. Microsoft attributed about $25 billion of its calendar-2026 capex to higher component prices, and Amazon blamed memory costs for raising its budget to about $220 billion (see our hyperscaler comparison).
| Company | Supply-chain layer | Latest reported figure | Recent strategic move (2026) |
|---|---|---|---|
| Nvidia | AI accelerators, networking | Data center revenue $89.0bn, +117% (quarter to Jul 2026) | Vera Rubin ramp; substrate/optics capacity investments |
| Vertiv | Power, thermal management | Q3 2026 net sales $4.31bn, +61%; FY guide $14.45–14.85bn | Closed ~$1.45bn UtilityInnovation Group deal |
| Schneider Electric | Power distribution, cooling | 2025 revenue €40.15bn | Agreed $22.6bn PTC acquisition; Motivair 3.5 MW CDU |
| Eaton | Electrical distribution, UPS | 2025 revenue $27.4bn | Completed Boyd Thermal; €810M COL Group deal; $242M+ Fibrebond plant |
| GE Vernova | Gas turbines, grid equipment | Closed Prolec GE transformer deal (Feb 2026) | Agreed $9.0bn Trench Group deal (Oct 2026) |
| Caterpillar | Generators, gas engines | Q3 2026 sales $20.7bn, +17%; Power & Energy +18% | Ongoing capacity investment |
| Cummins | Generators, batteries | Q2 2026 Power Systems sales $2.3bn, +19% | Gas gensets for Texas AI campuses; largest BESS deal |
Source: Nvidia Q2 FY2027 results; Vertiv Q3 2026 release; Schneider Electric and Eaton annual figures; GE Vernova, Caterpillar and Cummins releases (via StockTitan summaries); Wikipedia (Schneider/PTC).
Which companies supply power and cooling equipment?
The main power and thermal suppliers are Schneider Electric, Vertiv, Eaton, ABB, Siemens and Legrand, alongside cooling specialists such as Johnson Controls, Trane, Carrier and newly acquired liquid-cooling firms. Vertiv’s results show the demand: Q3 2026 sales rose 61%, and organic orders more than doubled.
Vertiv is the purest listed play on data center power and cooling. In Q3 2026 its net sales reached $4.313 billion, with 55% organic growth, and diluted earnings per share rose 91%. The company raised its 2026 sales guidance to $14.45–14.85 billion, up from 2025 revenue of about $10.2 billion. It is also buying capacity and capabilities: in October 2026 it completed the acquisition of UtilityInnovation Group for about $1.45 billion upfront, plus up to $1.15 billion in earnouts, and bought fluid-management firm King Environmental Services.
Schneider Electric, with 2025 revenue of about €40.15 billion, is the largest diversified player. It took a controlling stake in liquid-cooling specialist Motivair in 2024, at a reported $850 million, and in September 2026 it launched a Motivair coolant distribution unit that delivers up to 3.5 MW of cooling per unit. In October 2026 it agreed to buy industrial software firm PTC for about $22.6 billion, a deal it says will strengthen its data center business. Eaton, with 2025 revenue of $27.4 billion, completed its acquisition of Boyd Thermal to add liquid cooling. It also agreed an €810 million deal for Italian medium-voltage maker COL Group and is investing more than $242 million in an Arkansas plant to double US capacity for Fibrebond modular power enclosures.
Why are transformers and grid equipment a bottleneck?
Large power transformers, switchgear and high-voltage equipment are made by a limited number of factories, and demand from data centers, renewables and grid upgrades has outstripped capacity. Long lead times for this equipment often decide when a campus can switch on.
Suppliers are responding with acquisitions and new plants rather than short-term fixes. GE Vernova closed its acquisition of transformer maker Prolec GE in February 2026, funded partly by a $2.6 billion bond. In October 2026 it agreed to buy Trench Group, a maker of high-voltage components, from Triton Partners for $9.0 billion. Trench has about $5 billion of backlog and framework agreements and expects about $1.9 billion of revenue in 2028. Eaton’s COL Group deal and its Arkansas plant follow the same logic of adding medium-voltage and modular capacity.
The bottleneck also explains why hyperscalers increasingly build their own power. Oracle plans to run some AI campuses on natural-gas fuel cells. Amazon backed a 7.65 GW off-grid gas plant in Texas in 2026, and Microsoft signed a 20-year agreement with Chevron for a West Texas data center fed by the 2.67 GW Project Kilby, with most capacity on GE Vernova turbines and first power targeted for 2028. We cover these trends in why power is the new bottleneck.
How are generators and on-site power suppliers benefiting?
Backup and on-site generation suppliers such as Caterpillar, Cummins, Rolls-Royce, Generac and turbine makers are seeing strong demand, because every data center needs backup power and many AI campuses now need primary power on site. Caterpillar’s Power & Energy sales rose 18% in Q3 2026.
Caterpillar reported Q3 2026 sales and revenues of $20.7 billion, up 17%, with operating profit up 38% to about $4.2 billion. Its Power & Energy segment, which includes reciprocating engines, gas turbines through Solar Turbines and generator sets used at data centers, grew 18%. Cummins reported record Q2 2026 revenue of $9.5 billion, with Power Systems sales up 19% to $2.3 billion at a 24.5% EBITDA margin. It raised its full-year outlook. In June 2026 it agreed to supply natural gas generator sets to Circe Energy’s AI data center campuses in Texas, with deliveries from 2026 to 2030, and in August it won its largest battery energy storage deployment to date, for a US data center.
The shift from diesel backup to gas-fired primary power is a structural change. Behind-the-meter generation lets developers bypass multi-year grid queues, but it requires turbines, reciprocating engines and fuel infrastructure, all of which are in short supply. Turbine makers are reserving manufacturing slots years in advance. GE Vernova’s 2026 agreement with Blue Energy, for example, included slot reservations for two 7HA.02 turbines for 2029.
Who supplies networking, memory and the rest of the rack?
Beyond GPUs, AI data centers need high-bandwidth memory, networking switches and optics, CPUs, storage and server assembly. Key suppliers include memory makers SK hynix, Samsung and Micron, networking firms Broadcom, Arista, Cisco and Nvidia’s own networking unit, and server builders such as Dell, Supermicro, HPE, Foxconn and Quanta.
Memory has become a cost driver in its own right. Both Microsoft and Amazon cited higher memory and component prices for part of their 2026 capex increases. Optical interconnects are another constraint. Nvidia’s plan to raise optical-engine capacity 15-fold by mid-2027 shows how quickly demand for fast links between GPUs and racks is growing. Custom silicon also supports a separate ecosystem of chip-design partners who help hyperscalers build TPUs, Trainium and other in-house accelerators.
For colocation operators and neoclouds, these components arrive as integrated racks, increasingly whole liquid-cooled systems such as Nvidia’s NVL72 designs, assembled by original design manufacturers in Asia and Mexico. That compresses construction timelines but makes data halls depend on a single delivery schedule. Our profile of the largest colocation providers shows how landlords are redesigning halls for these racks.
How are suppliers consolidating?
Suppliers are buying capacity and technology at record prices. Major 2026 deals include Schneider Electric’s agreed $22.6 billion purchase of PTC, GE Vernova’s $9.0 billion Trench Group acquisition, Ecolab’s $4.75 billion purchase of CoolIT Systems and Vertiv’s roughly $1.45 billion UtilityInnovation Group deal.
The pattern is clear: electrical and thermal companies want to offer complete, pre-engineered power and cooling packages, because hyperscalers prefer fewer vendors and faster deployment. Ecolab, a water and hygiene company, bought CoolIT from KKR in a deal agreed in March 2026 to move into liquid cooling. Vertiv’s UtilityInnovation purchase adds grid-interconnection and power-system capabilities. Eaton’s Boyd Thermal acquisition and Schneider’s Motivair stake added liquid cooling to electrical portfolios.
Private equity is active on both sides, building and selling component businesses, as profiled in our guide to private equity’s data center platforms. For the broader deal context, see the biggest data center M&A deals.
What will shape the supply chain over the next few years?
Three forces will shape the supply chain: higher rack densities that make liquid cooling and 800-volt DC power standard, the move to on-site generation and batteries, and competition between Nvidia and custom silicon. Manufacturing capacity for power equipment will remain the gating factor.
Rack densities keep rising with each Nvidia generation, which pushes operators towards direct-to-chip liquid cooling and new electrical architectures. Schneider Electric has published research on arc-flash risk in 800 VDC data center designs, and Nvidia is promoting reference designs for AI factories that integrate power and cooling. On-site gas, batteries and eventually small nuclear reactors will make energy equipment makers central suppliers. GE Vernova, Hitachi and Samsung C&T, for example, are advancing a four-unit, 1.2 GW BWRX-300 small modular reactor project in Sweden, though without a final investment decision yet.
Demand looks secure in the near term. Hyperscalers say capacity will be short through 2026, and Alphabet has signaled that 2027 capex could top $300 billion. Synergy Research expects hyperscale capacity to triple by 2031. The risk is cyclical: suppliers expanding now could face a downturn if AI monetisation disappoints, a debate we cover in the AI data center bubble debate. For electricity demand forecasts, see our data center electricity statistics.
Frequently Asked Questions
How much of data center spending goes to Nvidia?
No exact share is published, but compute is the largest category. FactSet estimates compute at about $380 billion, roughly 60%, of 2026 hyperscaler capex, and Nvidia’s data center revenue was $89.0 billion in a single quarter to July 2026. Custom chips from Google, Amazon and others, plus memory, networking and servers, take the rest of the compute budget.
Why is Vertiv growing so fast?
Vertiv specialises in power and thermal management for data centers, the equipment AI racks need most. Its Q3 2026 net sales rose 61% to $4.31 billion, with organic orders up more than 100%, and it raised 2026 guidance to $14.45–14.85 billion. Higher rack densities mean more cooling and power equipment per megawatt, which benefits specialists like Vertiv.
What equipment is hardest to get for a new data center?
Grid and power equipment is usually the longest-lead item. Large transformers, high-voltage switchgear, gas turbines and generators are made by a limited number of factories. That is why suppliers like GE Vernova and Eaton are buying manufacturers and building plants, and why some developers install on-site power instead of waiting for grid upgrades.
Are suppliers a safer way to invest in AI data centers?
Not necessarily. Suppliers benefit when capex rises but can be hit hard if spending slows, because orders can fall faster than revenue at operators with long leases. Diversified firms like Schneider, Eaton and Caterpillar are less exposed than near-pure plays. This is general information, not investment advice.