Comfort Systems USA entered July 2026 with $14.06 billion of work under contract and still unfinished. A year earlier, the backlog was $8.12 billion.
The company does not build AI models. It installs cooling, plumbing and electrical systems inside factories, hospitals and data centres. Hard hats, pipework and prefabricated modules. Not exactly what comes to mind when people picture AI fortunes.
Some of the steadiest money in the boom is landing there. The public story runs through ChatGPT, Nvidia and a small club of trillion-dollar companies. The spending reaches much further: land, power plants, transformers, cooling equipment, fibre, construction crews and consultants. New servers need all of it before they answer a single prompt.
Investors often call this the picks-and-shovels trade. The label is useful up to a point. It can also make an ordinary industrial company look like an AI stock simply because a data centre appears somewhere on its customer list.
The Power Boom did not Start with AI
Data centres used about 4.4% of U.S. electricity in 2023. A Department of Energy-backed study estimates that their share could reach between 6.7% and 12% by 2028. Nobody knows where the number will land. Construction schedules move, chips become more efficient and projections for AI adoption have a habit of aging badly.
Even 6.7% would require substantial new generation and grid capacity.
GE Vernova sells the turbines, grid equipment and power-conversion systems needed to supply it. Gas-turbine orders doubled in 2024. Quarterly net income later more than doubled to $484 million. The company reported 29 gigawatts of confirmed turbine orders and another 21 gigawatts covered by slot reservations.
About one-third of those reservations were connected to data-centre projects. The fraction is more revealing than the total. Data centres are large enough to tighten an already busy market. Electrification, factory construction and old grids account for plenty of demand on their own.
Anyone describing GE Vernova’s entire order surge as an AI windfall is borrowing more from the story than the figures allow.
Industrialization, electrification, that’s really what I would say is driving the bulk of our orders last year and really into this year.- Eric Gray
Eaton has spent heavily on the equipment that controls those electrons after they reach a site. In 2025, it agreed to pay $1.4 billion for Fibrebond, a maker of modular power enclosures. Its proposed $9.5 billion purchase of Boyd Thermal added liquid cooling. Boyd was expected to generate $1.7 billion in 2026 sales.
By the end of 2025, Eaton expected data centres and distributed IT equipment to account for around 17% of company sales. Significant exposure. Still 83% elsewhere.
A Server Hall is also a Heat Problem
AI racks pack in far more computing power than standard enterprise equipment. More electricity enters a smaller space. More heat has to leave it.
Vertiv supplies thermal-management systems, switchgear and backup power. Its revenue rose almost 28% in 2025 to $10.23 billion, with net income of $1.33 billion. Fourth-quarter organic orders jumped 252%. The company closed the year with roughly $15 billion in backlog.
Those numbers have two lives. In an earnings release, they show customers lining up for equipment. In a valuation model, they need harsher treatment. Some orders will take years to turn into revenue. Projects can slip. Customers sometimes reserve capacity early when they fear a shortage. The sale is only as good as its price, delivery cost and eventual collection.
Vertiv also earns money after installation through service and replacement work. That recurring revenue deserves separate attention from the rush to equip new facilities. It remains when the cranes leave. Comfort Systems works one layer across. It designs, fabricates and installs the mechanical and electrical systems that make high-density facilities usable. Its backlog climbed by almost $6 billion in twelve months.
Software scales with another copy. Contracting scales with skilled workers, fabrication space, permits, scheduling and the ability to finish a difficult job without an expensive failure. Comfort Systems has to earn each new dollar the slow way.
ASML Keeps More of each Sale
Nvidia’s accelerators depend on machinery produced by a far smaller group of suppliers. At the leading edge, one supplier dominates.
ASML makes the extreme-ultraviolet lithography systems used to print advanced circuits on silicon. Foundries cannot produce the most sophisticated chips at scale without them. The machines cost heavily, take time to build and have no ready substitute.
ASML reported record 2025 sales of €32.7 billion, net income of €9.6 billion and a gross margin of 52.8%. In July 2026, it raised full-year sales guidance to between €43 billion and €45 billion, with expected gross margins of 54% to 56%.
Driven by ongoing AI-related investments, our customers continue to accelerate their capacity expansion plans. -Roger Dassen
That margin is the clearest difference between ASML and a contractor carrying a huge order book. Scarcity gives ASML room to keep far more of each sale. It also attracts government scrutiny. Export controls can close off customers, semiconductor spending moves in cycles and a small number of chipmakers account for much of the market. Lithography improvements form one part of the industry’s work on cost and power consumption. Chip design, packaging, memory, networking, software and cooling carry the rest.
Two Customers Loom over Arista
Thousands of processors inside an AI facility have to exchange data at high speed. Arista Networks sells the switches and management software that connect them. Revenue rose 20% to $7 billion in 2024 as its cloud customers expanded.
Two names dominated the customer list. Microsoft supplied 20% of revenue. Meta supplied another 15%. A delayed capital programme at either company would show up quickly in Arista’s results. No industry-wide collapse required.
Equinix and Digital Realty carry a different exposure. They own facilities with power access and network connections that tenants may struggle to reproduce in the same location. Rental contracts can run for years. The capital goes out much earlier.
Equinix lowered its 2025 revenue forecast after a large campus lease took longer than expected to close. During a delay, interest, construction and operating costs keep moving. Rent waits.
The Missing Number is the Price
A record backlog can belong to an overpriced stock. Strong revenue growth can do the same.
None of the operating figures above tells an investor what return to expect at today’s share price. That calculation starts with the market value already attached to the growth. If the price assumes several years of flawless execution, an earnings beat may barely move the stock. One late project or a narrower margin can do plenty of damage.
The comparison also changes by business. For ASML, a buyer needs to weigh its earnings multiple and free-cash-flow yield against its own trading history, semiconductor cycles and export risk. Vertiv and Eaton require estimates for how long data-center orders can grow, how much margin survives as competitors add capacity and whether acquisitions earn an adequate return.
Comfort Systems calls for scrutiny of backlog conversion, labour availability and project margins. Equinix and Digital Realty have to be judged through debt costs, occupancy, development spending and funds from operations.
Current multiples belong beside those checks on the day an investment decision is made; they move too often to be treated as permanent facts in an industry article. The companies named here are businesses with evidence of demand. The figures do not amount to buy recommendations.
Consulting Revenue is Arriving before Most Companies Know What to Buy
Most businesses adopting AI will never order a turbine or build a server hall. They have old software, sensitive data, nervous compliance teams and employees who need to use new systems without creating fresh problems.
Accenture booked $5.9 billion in generative-AI work during its 2025 financial year, almost twice the previous year’s total. Revenue from generative and agentic AI reached $2.7 billion, three times the 2024 figure. It still represented less than 4% of Accenture’s $69.7 billion in annual revenue.
The smaller-business opportunity sits below that scale: preparing data, testing models, reviewing cybersecurity, documenting regulatory risk and integrating tools into a specific workflow. Lower capital requirements than a data center, obviously. Also fewer physical barriers to stop a crowd of undifferentiated “AI consultants” from turning up.
Industry knowledge is the useful filter. A firm that understands patient privacy can help a hospital deploy a model without exposing medical records. A banking specialist can test automated decisions against compliance rules. Contractors with the right certifications can install or maintain parts of a data-centre power and cooling system.
For Black entrepreneurs, these are more plausible ownership routes than a semiconductor plant or hyperscale campus. They also fit the same constraint running through the larger companies. Hospitals pay for safe deployment. Banks pay for defensible decisions. Developers pay for scarce technical labour. Nobody pays for a vague claim of being “AI-powered” for very long.
The money around AI now stretches from turbines to compliance reports. Some of it comes directly from model development. Some would have appeared through cloud growth, factory construction or grid replacement anyway. The accounts separate the two imperfectly, and investor presentations rarely rush to do so.
GE Vernova has turbine slots. Vertiv and Comfort Systems have large backlogs. ASML has a machine advanced chipmakers cannot easily source elsewhere. Arista has two customers supplying 35% of revenue. Accenture has $2.7 billion in AI revenue inside a $69.7 billion firm.
Those are the businesses. Those are also the complications.
Frequently Asked Questions
Who is making money from the physical side of AI?
Power-equipment makers, cooling suppliers, chipmaking-tool companies, electrical contractors, networking firms and data-centre landlords. Their exposure varies widely, so the AI share of revenue needs to be checked company by company.
Why do AI servers need specialized cooling?
Dense racks draw substantial power and produce heat in a small area. Many facilities need liquid cooling or more advanced air-cooling systems to operate reliably.
Will higher data-center demand guarantee utility profits?
No. Utilities may have to fund generation and grid upgrades years before recovering the cost. Regulators, financing terms and customer contracts decide much of the return.
Where can Black-owned businesses enter this market?
Specialist installation, maintenance, cybersecurity, data preparation, compliance and industry-specific implementation offer lower-cost entry points. Credible expertise and customer access carry more weight than a generic AI label.
Does the picks-and-shovels approach make these safe investments?
No. It replaces model-winner risk with valuation, customer, execution, debt and capital-spending risk.
What deserves the closest look before buying a stock in this group?
Start with price. Then examine free cash flow, margins, debt, backlog quality, customer concentration and the percentage of demand that can genuinely be traced to AI.
