AI Capital Spending Is Becoming a Macro Variable
· Updated
In a single quarter, GE Vernova booked $2.4 billion of data-center equipment orders in its Electrification business, more than it booked in all of 2025.
Picture GE Vernova's data-center equipment orders in its Electrification business through 2025, one quarter at a time.
Pour the four quarters together into a single pool: everything it booked in 2025.
Now a single quarter of 2026: $2.4 billion of orders.
One quarter outweighs the whole of 2025. AI spending has reached the grid's equipment makers.
AI spending has spread well beyond tech, into electricity demand, utility regulation, imports, 
inflation and credit. At its June meeting, the Federal Reserve discussed AI infrastructure as a source of both real growth and near-term price pressure. Below: where the money goes, why power grids are the tightest bottleneck, a simple compounding exercise that shows how fast one grid's demand could grow, and the signals we'd watch.
The short version
- The scale: the
Bank for International Settlements estimates the five largest hyperscalers will spend more than $1 trillion on AI-related capital across 2025 and 2026.
- Power is the pinch point: data centers used about 4.4% of US electricity in 2023; projections for 2030 run 9.5% to 15.3%.
Picture all the electricity the US uses in a year as one pool.
Zoom in. In 2023, data centers used about 4.4% of it.
By 2030, projections start at 9.5%.
And run as high as 15.3%. Power is the pinch point of the AI build.
- Our read: the build can push prices up before any productivity payoff pulls them down, and markets tend to underrate that.
Figures are as of our research cutoff: July 14, 2026, 5:49 a.m. Eastern Time, before June CPI was released. Hyperscaler capex guidance is updated through the July 22–30 second-quarter reports.
- Capital (Cash, bonds, credit) to Compute (Chips, memory, servers)
- Compute (Chips, memory, servers) to Sites (Buildings, cooling)
- Sites (Buildings, cooling) to Power (Grid, generation)
- Power (Grid, generation) to Economy (Prices, imports, rates)
The spending has reached industrial scale
At our July 14 cutoff, Microsoft expected roughly $190 billion of calendar 2026 capex (capital expenditure, money spent on long-lived assets like data centers). Alphabet guided to $175 billion to $185 billion and Meta to $125 billion to $145 billion. Amazon spent $43.2 billion of cash capex in the first quarter, net of equipment sales and incentives ($44.2 billion before them), up from $24.3 billion on the same net basis a year earlier.
Tap the phrase you think is the tell.
Spotted it.Not that phrase. Same quarter, two definitions. That's why the four companies' capex figures never line up exactly, and why our tracker labels each one.
- “Amazon spent $43.2 billion of cash capex” This is the smaller of the two numbers, not the reason it's smaller.
- “net of equipment sales and incentives” Amazon subtracts money from selling equipment and from incentives. Before those, the quarter was $44.2 billion.
- “up from $24.3 billion on the same net basis a year earlier” A comparison on the same net basis, so it doesn't explain the gap.
The July 22–30 second-quarter reports pushed the plans higher. Alphabet now guides to $195 billion to $205 billion, Amazon to about $220 billion and Meta to $130 billion to $145 billion. Microsoft's figure fell to about $175 billion only because some leases moved out of capex, not because it is building less. Combined, the four now guide to $720 billion to $745 billion. Our hyperscaler capex tracker has each company's number and sources.
| 2026 capex | At our July 14 cutoff | After Q2 reports |
|---|---|---|
| Microsoft | About $190B | About $175B (lease change) |
| Alphabet | $175B–$185B | $195B–$205B |
| Amazon | $43.2B in Q1, net | About $220B |
| Meta | $125B–$145B | $130B–$145B |
| Combined | — | $720B–$745B |
The money flows through four layers:
| Layer | What it buys | Why it matters now |
|---|---|---|
| Compute | Accelerators, high-bandwidth memory, networking, servers | Where scarcity first appeared and much of the profit went; as chip supply improves, the bottleneck moves |
| The data-center site | Buildings, cooling, backup power, switchgear, busways | Decides whether chips can run reliably |
| The grid | Substations, transformers, transmission | Chips arrive in quarters; big power gear can take years |
| Generation | Power plants, often natural gas as a bridge | Data centers need dependable power every hour |
Suppliers' order books show the shift
| Company | What it reported |
|---|---|
| 44% organic sales growth in the Americas in Q1; backlog at the end of 2025 more than doubled from a year earlier | |
| Rolling orders in Electrical Americas up 42% organically in Q1, driven by data centers; Electrical segment backlog up 48% | |
| $2.4 billion of data-center equipment orders in Electrification in the quarter, more than in all of 2025 |
High-density systems need liquid cooling and power distribution, not just racks. The numbers also signal a bottleneck: a backlog only pays if a company can deliver it and protect margins.
The Department of Energy says distribution-transformer lead times went from three to six months in 2019 to 12 to 30 months in 2023. Some large substation transformers take three to four years, industry participants say.
Solar and storage are growing fast, but where new transmission and clean, always-on power aren't available, natural gas often fills the gap. The Energy Information Administration's July outlook expects record gas use in the power sector next year. In a high-data-center-demand scenario, EIA found gas generation could rise 7.3% between 2025 and 2027, against 1.7% in its base case.
Growth in US gas-fired generation, 2025 to 2027.
Electricity is where it becomes macro
The 9.5% to 15.3% projection for 2030, up from about 4.4% in 2023, comes from a Department of Energy summary of Lawrence Berkeley National Laboratory work. Even the low end is a big shift for a slow system.
Locally, it's sharper. In Dominion's Virginia zone, the summer peak hit 23,905 megawatts in 2025, 23% above 2019. PJM, the regional grid operator, expects summer peak demand there to rise about 5.4% a year over the next decade. Texas is handling a similar wave of large-load requests.
Worked example: what 5.4% a year does to a grid
A growth rate sounds modest until you compound it. Using only the Dominion figures above:
Spot on.Close, a little low.Close, a little high.Way off, too low.Way off, too high. It was 1.69×.
About 1.69×. From 23,905 megawatts, that's roughly 40,450 megawatts, our arithmetic rather than a PJM forecast.
| Step | Math | Result |
|---|---|---|
| 1. Implied 2019 peak | 23,905 MW ÷ 1.23 | About 19,435 MW |
| 2. Annual growth, 2019 to 2025 | 1.23^(1/6) − 1 | About 3.5% a year |
| 3. Ten years at PJM's 5.4% | 1.054^10 | About 1.69x |
| 4. Illustrative peak after ten years | 23,905 MW × 1.69 | About 40,450 MW |
| 5. Years to double at 5.4% | 72 ÷ 5.4 (rule of 72) | About 13 years |
Two lessons. First, PJM's expected pace of about 5.4% is well above the roughly 3.5% compound annual growth rate of the past six years, so the forecast assumes acceleration. Second, ten years at that pace adds roughly 16,500 megawatts (about 40,450 minus 23,905), about 3.7 times the rise since 2019 (about 23,905 minus 19,435, or roughly 4,470 megawatts). Step 4 is our arithmetic, not a PJM forecast of a specific number, and it assumes growth starts from the 2025 peak.
Equipment timing makes this harder. Distribution-transformer lead times went from three to six months to 12 to 30 months, four to five times longer (12 ÷ 3 = 4; 30 ÷ 6 = 5). A utility planning for that load has to order gear years ahead.
For utility investors, that concentration cuts three ways:
- Growth: utilities can grow their rate base (the assets regulators let them earn a return on) if regulators approve.
- Better use of assets: a steady, round-the-clock load appeals to generators and grid owners.
- Backlash: if a data center requests costly upgrades, cancels, and leaves ordinary customers paying, regulators step in.
That third risk is already live. In June, the Federal Energy Regulatory Commission ordered six regional grid operators to justify or reform their rules for connecting large loads, including how to prevent cost shifting.
So load growth isn't automatically good for every utility. Ask whether the project is funded or only announced, whether the customer has made a real financial commitment, who pays if it's cancelled, whether the utility earns a regulated return, whether the load can be cut during grid stress, and whether new generation is arriving with it. More in our AI data-center power demand guide.
Member research: Transformers, turbines and gas plants are where the power bottleneck turns into revenue for suppliers. Our
It also reaches trade and inflation
A Federal Reserve analysis estimated US data-center spending would top $500 billion in 2025, lifting exports from economies that supply servers and components. Because those parts cross borders, one supplier's disruption can delay a multibillion-dollar facility.
The build raises demand for electrical gear, construction labor, turbines and power before it delivers productivity gains. Microsoft said higher component prices added about $25 billion to its calendar 2026 capex estimate. Meta also cited component and data-center costs when it raised guidance.
Fed officials are weighing two clocks: a buildout clock that can push prices up because capacity is scarce, and a productivity clock that may pull them down later if AI lets firms do more with less. Markets often price the second and underrate the first.
- Demand for electrical gear, labor, turbines and power
- Scarce capacity pushes prices up
- Component prices added about $25B to Microsoft's 2026 estimate
What Fed officials are weighing
- AI lets firms do more with less
- Could pull prices down
- Arrives only after the build delivers
What markets tend to price
The hidden financing layer
Hyperscalers also fund the build with bonds, leases, special-purpose vehicles and private credit. The BIS analysis of AI financing says hyperscaler gross bond issuance topped $100 billion in 2025. It also describes off-balance-sheet deals where a separate entity funds a facility and the tech company signs long-term lease or capacity commitments.
The risk is opacity. Capex can look lower at the hyperscaler while the obligation sits elsewhere as rent, an offtake agreement or project debt. If usage disappoints, the burden stays.
- Lenders (Bonds, private credit) to Separate entity (Funds the facility): Project debt
- Separate entity (Funds the facility) to Hyperscaler (Uses the capacity): Capacity
- Hyperscaler (Uses the capacity) to Separate entity (Funds the facility): Lease payments
Amazon shows the pressure. Its trailing 12-month purchases of property and equipment, net of incentives, reached $147.3 billion by the end of the first quarter. Free cash flow (operating cash minus capex) over that period fell to about $1.2 billion from $25.9 billion a year earlier. It isn't in distress, but the cycle absorbs cash fast.
The counterweight: adoption trails construction
The strongest bear case: infrastructure may be outrunning business adoption. The Census Bureau found AI use among US businesses hovered around 17% to 20% between December and May, much higher at large firms and in information and finance. A
Federal Reserve adoption review found roughly 18% of firms had adopted AI by the end of 2025, though employee use of generative tools was wider.
In the healthy version, better apps and cheaper inference fill the capacity. In the unhealthy one, duplicate grid requests and weak app economics lead to cancellations. Efficiency cuts both ways: cheaper tasks can spur more use, or mean fewer servers are needed.
Try it yourself
A 10-minute exercise on the power bottleneck:
- Open the free data-center PUE and electricity cost calculator. Keep its default 100 MW campus and note the yearly electricity cost.
- Raise the electricity price by a third and note how much the cost changes. That's the exposure a tight grid creates.
- Open the free CAGR calculator and enter Dominion's implied 2019 peak (about 19,435 MW) and its 2025 peak (23,905 MW) over six years. Check that you get about 3.5%.
- Open
Eaton's first-quarter release and find the orders and backlog figures in the supplier table above.
- Check the free 2026 AI capex dataset to see how hyperscaler guidance has moved since our July 14 cutoff.
You should now be able to answer: is the grid growing fast enough for the demand planned on it, and which suppliers are turning that shortage into delivered revenue rather than just backlog?
What we'd watch
Our read: this is a capital-flow map, not one industry. A supplier can own a real bottleneck and still be a poor investment if its price assumes years of perfect execution. A hyperscaler can have huge demand and still destroy value if depreciation and financing outrun revenue. A utility can win load growth and lose the fight over who pays.
Signals that separate a lasting expansion from a temporary shortage:
- Energized megawatts, not announced ones.
- Supplier backlog conversion into revenue and cash, not just orders.
- Cloud backlog conversion into recognized sales.
- Depreciation and free cash flow, the cost of keeping the fleet current.
- Transformer and turbine lead times, an early sign of easing scarcity.
- Large-load deposits and cancellation rules, which firm up utility forecasts.
- Business adoption: recurring workloads, not demos.
- Electricity prices and gas burn, the sign the build has become a visible macro input.
The key question: where does each new AI dollar go, what scarce asset does it hit, and who keeps the profit once the constraint clears? Member issue AI Capex Shock adds company scorecards, three 2026–2028 scenarios and catalysts (membership). For a single stock, start with our stock-research checklist and how we test a thesis.
Questions readers ask
Is AI spending good or bad for inflation?
Both, on different clocks. The buildout raises demand for scarce equipment, labor and power now, which can push prices up. Productivity gains may pull prices down later if AI lets firms do more with less. Fed officials are watching the first clock; markets tend to focus on the second.
Why does natural gas keep coming up in an AI story?
Data centers need power every hour. Where new transmission and clean, always-on power aren't available, gas plants are often the generation that can respond. That's why EIA's high-data-center scenario shows gas generation rising 7.3% between 2025 and 2027, against 1.7% in its base case.
Does rising electricity demand make utilities a good investment?
Not automatically. Load growth helps only if regulators approve the spending, the customer has made a real commitment, and ordinary ratepayers aren't left with the bill if a project is cancelled. FERC's June order shows regulators are already scrutinizing those rules.
Why do off-balance-sheet deals matter if the hyperscaler isn't borrowing?
Because the obligation still exists. It just sits somewhere else, as rent, a capacity agreement or project debt. If data-center use disappoints, those payments are still due, which is why capex alone can understate a company's commitments.
Go deeper with member research
The macro view shows where the money goes. Member research follows it to the balance sheets and the companies that supply the power:
- Issue 05: NVIDIA Is Building the Credit Market for AI: How NVIDIA turns its roadmap into cheaper capital for customers, and why AI infrastructure could move into private credit, insurance balance sheets and securitized debt.
- GE Vernova and the value of delivering dependable power: Gas-power and grid economics as of January 29, 2026, and why it was already a recognized trade by then.
Limits of this data
- Hyperscaler capex guidance is updated through the July 22–30 second-quarter reports; energy forecasts and other figures are as of July 14, 2026.
- The ten-year grid figure in the worked example is our compounding arithmetic, not a utility or PJM forecast.
- Other sources used:
June 2026 FOMC minutes,
EIA Annual Energy Outlook 2026,
DOE 2026 transmission needs study,
Microsoft FY26 Q3 earnings,
Meta Q1 2026 results and
Amazon Q1 2026 10-Q.
Companies in this article
Keep reading
Members-only research
Members are reading this
- Issue 05: NVIDIA Is Building the Credit Market for AI
NVIDIA has found a way to turn knowledge of its own roadmap into a lower cost of capital for its customers. The result could move AI infrastructure from venture equity into private credit, insurance balance sheets and securitized debt.
- Meta jumped 11%. What must happen next to justify the price?
A member research report on Meta’s AI rally: Muse, subscription economics, the cash-flow burden, legal costs and the earnings assumptions behind the new price.
The thesis, the numbers behind it, and what would break it. Full access is $39 a month.
Read the full research