COMPUTE

---

 Compute is industrial capacity for calculation. A data center turns electricity into answers: forecasts, drug candidates, fraud checks, factory schedules, and everyday AI tools.

Two kinds matter. AI already runs at national scale in buildings packed with specialized chips that need power, cooling, and fast links to other sites. Quantum computing is smaller and earlier. It uses quantum physics on a narrower set of hard problems — molecules, grid planning, materials, some optimization — that ordinary computers struggle with. Useful large-scale quantum machines are still a mid-term bet, not a 2026 mass product, but investment and early commercial use are already large enough to count.

The two are linked. AI is buying power, land, fiber, and talent now. Quantum is being built to attack some of the same bottlenecks, including energy planning and new materials for batteries, grids, and chips.

The International Energy Agency (IEA) reports that capital spending by the largest technology companies, much of it on data centers, exceeded $400 billion in 2025 and is expected to rise another 75 percent in 2026. McKinsey estimates global data-center investment could reach about $7 trillion by 2030. U.S. Census figures cited in 2026 put data-center construction above a $50 billion annual rate, larger than general office building for the first time.

That money buys concrete, steel, transformers, turbines, fiber, cooling, and labor. S&P Global research cited in 2026 found AI data centers and related high-tech spending accounted for a large share of U.S. private-demand growth in early 2025. McKinsey’s 2025 AI survey found about one-third of firms already scaling AI and 64 percent reporting innovation benefits — one reason planners treat the power and construction pipeline as durable through 2030.

Growth Rates and Spend:

Global   data-center electricity use: 485 TWh in  2025 (IEA): Projected: About 950 TWh  by 2030

Share of world   electricity: About 1.5% in 2024-25: Projected About 3% by   2030 (IEA)

U.S.   data-center electricity: 192 TWh in 2024, ~4.7% of U.S. use (DOE 2026): Projected 9.5%-15.3% by 2030

Hyperscaler   capex: Over $400   billion in 2025 (IEA) Projected+75% expected   in 2026:

Quantum   economic value (est.): $1B+ company  revenue in 2025  Up to $2.7 trillion value by 2035:

The most visible effect is on electricity. Data-center use grew 17 percent in 2025; AI-focused sites grew about 50 percent — far faster than the 3 percent rise in world electricity demand (IEA). A typical AI hall already uses as much power as about 100,000 homes. The largest sites now under construction will use many times that.

That load is also a buyer with long contracts and a high willingness to pay for reliable power. That combination is changing what gets built.

IEA figures put electricity generated for data centers at about 460 TWh in 2024, rising to more than 1,000 TWh in 2030. Renewables — wind, solar, and hydro — are expected to meet nearly half the extra demand through 2030, growing about 22 percent a year in this use. Gas and coal together still cover more than 40 percent of the added load in that window, because they run when sun and wind do not. Nuclear’s share grows later in the decade.

Long-term power contracts make new plants easier to finance. Microsoft signed with Constellation to restart Three Mile Island Unit 1 (about 835 megawatts), targeted for 2027. Amazon Web Services contracted with Talen Energy for up to 1,920 megawatts from Susquehanna. Conditional offtake deals between data-center operators and small modular reactor (SMR) projects grew from 25 gigawatts at the end of 2024 to 45 gigawatts by early 2026 (IEA). Those reactors are not all built. Compute is nonetheless one of the first large commercial buyers for a nuclear class that had struggled to find customers.

Gas turbines tell the same story. Global orders jumped about 70 percent in 2025, a 25-year high. AI is not the only reason, but it is a major one. Tens of gigawatts of on-site gas are proposed or in development for U.S. data centers. Once built, that capacity can also support nearby homes and factories when a campus does not need every megawatt.

A Congressional Research Service brief notes ways campuses can help the grid, not only strain it. Sites with their own generation can, where rules allow, send spare power out. Large customers that can briefly cut or shift load act as a shock absorber. McKinsey argues that plants, storage, and wires built for AI will still serve the wider economy after this construction wave cools. 

AI used on the grid itself is the second energy benefit. The IEA estimates that if existing AI tools were widely adopted in electricity, they could save up to $110 billion a year and unlock about 175 gigawatts of extra transmission capacity — more power on lines that already exist. Proven AI uses in energy-heavy industry can cut energy costs by 3 to 10 percentage points. Scaled up, existing AI energy applications could save on the order of 300 TWh, roughly the annual electricity use of Australia and New Zealand.

Compute uses a lot of power. It is also a tool that can make power systems cheaper and less wasteful. Both are true.

A modern AI data center is closer to a factory than an office: a building designed around power density, cooling, and uptime. AI server power density rose about 11-fold from 2020 to 2025 and is set to rise again by 2027 (IEA). One rack can draw as much peak power as dozens of homes.

Location now follows electricity and land. McKinsey notes some future campuses may draw 5 to 10 gigawatts — in the range of a large city’s power use — so more sites will sit next to generation rather than at the end of a long wire. Hybrid models are appearing: on-site plants, batteries (IEA sees 20–25 GW of storage that could sit inside data centers by 2030), microgrids, and direct deals with plant owners.

U.S. data-center construction is now its own Census category. Electrical work is commonly 45 to 70 percent of the construction budget. That is why electricians, not software engineers, are the binding constraint on many sites.

Brookings found that labor markets getting their first large data center see data-processing jobs rise about 56 percent over the first decade and telecom jobs about 43 percent — typically 100–200 jobs in those sectors in a treated county. Hyperscale campuses produce more telecom-ecosystem work than smaller halls. Construction employment jumps first; operations staff stay lean (a few dozen operators per 100 megawatts is a common rule; Microsoft’s Quincy site had hundreds of construction workers and on the order of 50 permanent staff). Lasting local gains are tax base, vendor contracts, and the fiber and substations other businesses can use.

  Chips in one building are useless if they cannot talk to chips in another, or to users. Fiber-optic cable — glass threads that carry light — is the road system. AI has turned that system from a telecom story into an industrial one.

The Fiber Broadband Association and RVA estimate the United States may need to nearly double long-haul route miles, from about 95,000 to about 187,000, and raise total fiber miles from about 159 million to about 373 million by 2029. Each new hyperscale campus needs, on average, about 135 route miles of new connectivity. Inside one large Meta campus planned for about a million graphics processors, a Corning executive put the internal fiber need on the order of 8 million miles.

Private capital is following. Prysmian committed $1.25 billion to three U.S. plants and up to 600 manufacturing jobs, more than doubling U.S. fiber output. Meta signed a Corning deal worth up to $6 billion. Lumen has reported nearly $13 billion of private connectivity contracts and plans to grow its intercity network from 17 million fiber miles at the end of 2025 toward 47 million by 2028. Verizon disclosed a dark-fiber agreement with Google worth more than $1 billion. U.S. providers passed a record 11.8 million homes with fiber in 2025.

  Households gain indirectly. Fiber built so two AI campuses can swap training data also carries calls, remote work, telemedicine, and school traffic.

 Public debate often treats AI as a threat to office work. The physical build is doing the opposite in the trades. Listings for data-center electricians rose more than tenfold from 2023 to 2026. Skillit put average pay on data-center construction around $81,800 in early 2026, about 32 percent above other commercial work. Indeed found installation and maintenance listings at data centers paid about 42 percent more than similar jobs elsewhere. Overtime in tight markets has pushed some electrician pay well into six figures 

The premium reflects a shortage. The Fiber Broadband Association and the Power & Communication Contractors Association have estimated a U.S. shortfall on the order of 58,000 tradespeople for the fiber build tied to this wave — about 28,000 construction workers and 30,000 technicians — with retirements making the longer-run gap larger. Industry comments put extra electrician demand over the next decade in the hundreds of thousands when AI load is included.

Tech firms are funding trade training because they cannot open buildings without it. Reporting in 2026 described more than $265 million in commitments from Meta ($115 million), Google ($50 million), and BlackRock ($100 million). Google is working with contractors and the International Brotherhood of Electrical Workers to lift apprenticeship enrollment. Meta launched a short fiber-technician pathway tied to site interviews. Short courses will not replace multi-year apprenticeships overnight. They do show compute demand pulling money into skills that have been scarce for years.

  The footprint spreads past the fence. Cushman & Wakefield, studying six major U.S. data-center markets from 2022 to 2025, found ecosystem firms accounted for 10.4 percent of new industrial leasing, rising to 14.4 percent in 2025. That supported an estimated 33,000 to 50,000 initial industrial jobs and, with suppliers and household spending, 81,000 to 124,000 jobs and about $11.6 billion a year in output. Every 100 megawatts of new capacity was associated with about 1,285 jobs across the chain and $344 million in output.

Two cautions. Permanent on-site jobs per dollar are low compared with a factory; Virginia and campus announcements often show one lasting operations job per tens of millions of dollars invested. Brookings found little overall local wage lift and a 2 to 5 percent rise in home prices, which helps owners and can pinch renters. Hyperscale sites locate first for power, land, and fiber, not tax breaks. Compute is a construction and supply-chain employer first, a long-run local payroll second, and a national productivity tool third.

Infrastructure is the means. The reason societies tolerate the power and land use is the work the machines do.McKinsey has estimated generative AI use cases could add up to $4.4 trillion a year across the functions it studied. A later McKinsey Global Institute analysis put about $2.9 trillion a year of U.S. value by 2030 if firms redesign whole workflows. Goldman Sachs Research has described a 15 percent cumulative lift to U.S. productivity and GDP after widespread adoption, while its nearer-term view is modest: AI’s current net addition to measured U.S. GDP growth is about 0.1 percentage point, with potential growth rising toward 2.3 percent in the early 2030s. McKinsey’s 2026 survey found 80 percent of respondents saying AI improved their own productivity and 50 percent saying it improved decisions, while only 37 percent of firms yet traced a clear profit impact. Tools show up in individual work first, in national accounts later.

Concrete uses are easier than headline dollars. The IEA notes energy use per AI task has been falling by an order of magnitude per year as models and chips improve, even as total use rises because more people use the tools. McKinsey has reported drug-discovery timelines cut by as much as 80 percent in some AI-assisted programs. Utilities are testing AI for outage prediction, equipment health, and renewable integration. Those are field trials and early deployments, not slogans.

Quantum computing is not a replacement for AI data centers. Ordinary computers use bits that are 0 or 1. Quantum machines use qubits that can hold more complex states, which helps on problems that explode when solved by brute force — how a molecule behaves, how to schedule a power system, how to search a huge set of options.

McKinsey’s 2026 Quantum Technology Monitor put potential economic value at up to $2.7 trillion worldwide by 2035. More than 300 companies, including Airbus, JPMorgan Chase, and utility E.ON, were working with quantum vendors. Firms booked more than $1 billion of revenue in 2025, with a path toward as much as $4.4 billion by 2028. Startup investment reached about $12.6 billion in 2025. The Quantum Economic Development Consortium counted 7,420 quantum-engaged organizations and 16,482 workers at pure-play firms. Governments announced on the order of $12.7 billion in new public funding in 2025.

Energy is an early customer. A World Economic Forum paper describes hybrid uses — quantum plus ordinary computers — for grid planning, asset scheduling, and materials research. Utilities including ComEd have started building the unusual support gear these machines need, such as deep cooling and reinforced power feeds. If hardware matures, payoffs include better batteries and catalysts, tighter renewable integration, and faster answers to dispatch problems today’s computers only approximate.

Quantum will also use energy and scarce materials. A 2026 ScienceDirect review warned that water and helium-3 could constrain a large fleet of fault-tolerant machines even if electricity stays inside previously modeled high-demand cases. Treat quantum as a high-upside research and early-commercial sector already creating specialized jobs, lab construction, and utility partnerships, with its largest economy-wide effects still ahead of 2030.

Compute is physical. AI is buying turbines, restarting reactors, pouring slabs, and stringing glass across states in 2026. Quantum is hiring specialists and signing utility pilots for problems current machines still cannot finish. Together they concentrate demand in power, construction, electrical gear, fiber, and cooling, and spread possible benefits across medicine, materials, grids, software work, and household tools that already save people hours.

The numbers that can be checked today are the infrastructure numbers: global data-center electricity on track to roughly double by 2030; U.S. data-center load on track for a high-single-digit to mid-teens share of national electricity use; hundreds of billions of dollars a year in tech capital spending; fiber route miles that may need to nearly double; construction wages already about a third higher on these sites; and a nuclear and gas order book that did not exist at this scale before AI campuses needed firm power. The larger productivity prize is still being earned, not banked. That is normal. Steel mills were built before the skyline they made possible was finished.

Sources Here