AI Infrastructure Investment in United States 2026
AI infrastructure investment in the United States has reached a scale that no longer maps cleanly onto normal corporate spending categories. The four largest hyperscalers, Amazon, Google, Meta and Microsoft, are on track to spend a combined $725 billion on capital expenditure in 2026, up 77% from roughly $410 billion in 2025, and almost all of that increase is going toward GPU clusters, custom silicon, data centers, and the power needed to run them. This is not a single company placing a bold bet; it is four of the most cash-rich businesses on earth arriving at the same conclusion at the same time and racing to build faster than their competitors.
Behind the capex headlines sits a physical supply chain straining at every link. NVIDIA’s data center revenue hit $75.2 billion in a single quarter this year, GPU lead times stretch past 36 to 52 weeks, and hyperscalers are now signing 20-year nuclear power contracts just to guarantee electricity for facilities that haven’t broken ground yet. Cloud backlogs measured in the hundreds of billions of dollars show that demand still outstrips what the industry can physically deliver, even at this unprecedented pace of spending. The statistics below cover where this capital is going, what is bottlenecking it, and how far the current buildout is projected to run.
Interesting Facts About AI Infrastructure Investment in US 2026
| Fact | Figure |
|---|---|
| Combined 2026 hyperscaler capex (Big Four) | $725 billion |
| Increase over 2025’s $410 billion | 77% |
| NVIDIA data center revenue, Q1 FY2027 (May 2026) | $75.2 billion (+92% YoY) |
| AI chip market size, 2025 | $120 billion (triple 2023 levels) |
| NVIDIA’s share of the AI training chip market | ~80% |
| NVIDIA’s confirmed AI chip demand through 2027 | ~$1 trillion |
| AWS cloud backlog | $244 billion (+40% YoY) |
| Oracle’s infrastructure backlog | $523 billion |
| Microsoft’s undeliverable Azure backlog | $80 billion |
| Nuclear power committed to US AI data centers | 9.8 GW across 13 deals |
Source: PitchBook, Tom’s Hardware, Mordor Intelligence, ValueAdd VC
Big Four Hyperscaler Capex: 2025 vs 2026
2025 ████████████████ $410B
2026 ████████████████████████ $725B
Every one of these figures describes the same underlying reality: demand for AI compute is outrunning supply at nearly every point in the chain. The jump from $410 billion to $725 billion in combined hyperscaler capital spending is a 77% increase in a single year, and NVIDIA’s $75.2 billion quarterly data center revenue, up 92% year over year, shows just how directly that spending is flowing into chip demand. With NVIDIA controlling roughly 80% of the AI training chip market and holding $1 trillion in confirmed orders through 2027, the company sits at the center of nearly every hyperscaler’s infrastructure plan.
The backlog numbers are just as telling as the spending totals. AWS’s $244 billion backlog and Oracle’s $523 billion figure mean customers have already committed to buying capacity that doesn’t exist yet, while Microsoft’s $80 billion in undeliverable Azure orders is stuck specifically because GPUs are sitting in warehouses waiting for power, not because chips are unavailable. That distinction, chips waiting on power rather than power waiting on chips, is why 9.8 GW of nuclear capacity has already been committed to US AI data centers even though none of it will generate a single watt for AI workloads before 2027.
Hyperscaler Capital Expenditure Statistics in US 2026
| Company | 2026 Capex |
|---|---|
| Amazon | ~$200 billion |
| Google/Alphabet | ~$175-185 billion |
| Meta | ~$115-135 billion |
| Microsoft | ~$110-190 billion (guidance varies by report) |
| Oracle | ~$50 billion (5-company total target) |
| Combined Big Four total | ~$725 billion |
| 2025 combined total, for comparison | ~$410 billion |
| Projected 2027 combined total | Over $1 trillion |
Source: ValueAdd VC, Futurum, CNBC, Tom’s Hardware
2026 Capex by Company (USD Billions, approx.)
Amazon ██████████████████████████ $200B
Google ████████████████████████ $180B
Meta ███████████████████ $125B
Microsoft ██████████████████ $120B+
Amazon leads all hyperscalers in 2026 spending at roughly $200 billion, narrowly ahead of Google at $175 to 185 billion, with Meta and Microsoft each committing well over $100 billion of their own. Every one of these companies blew past its own early guidance in 2025: Google guided $75 billion and actually spent $91.4 billion, Amazon guided “about $100 billion” and hit $131.8 billion, and Meta guided $60 to 65 billion but landed at $72.2 billion. That consistent pattern of overshooting guidance is why Goldman Sachs has publicly noted that “consensus capex estimates have proven too low for two years running.”
When Oracle’s roughly $50 billion target is added to the Big Four, the five largest US cloud and AI infrastructure providers are collectively committing between $660 billion and $690 billion for 2026 alone, nearly double what the same group spent in 2025. Combined annual capex has effectively tripled since 2024, when the group spent roughly $226 billion, and with analysts already projecting the Big Four alone will top $1 trillion in 2027, there is little sign the spending curve is close to flattening.
AI Chip and GPU Market Statistics in US 2026
| Metric | Figure |
|---|---|
| AI chip market revenue, 2025 | $120 billion |
| Growth since 2023 | Tripled |
| NVIDIA’s AI training chip market share | ~80% |
| NVIDIA data center revenue, Q1 FY2027 | $75.2 billion (+92% YoY) |
| NVIDIA FY2026 total revenue | ~$215.9 billion (+65% YoY) |
| AI GPUs shipped globally, 2025 | 1.7 million units |
| Data-center GPU lead times, 2026 | 36 to 52 weeks |
| H100 GPU peak wait times, 2024 | 6 to 12 months |
Source: AI Cloudbase, Intellectia.ai, Mordor Intelligence
NVIDIA Data Center Revenue: Growth Trajectory
2022 baseline (quarterly) ██ ~$3.6B
Q1 FY2027 (May 2026) ██████████████████████████ $75.2B
The chip layer of AI infrastructure has produced some of the fastest revenue growth in semiconductor history. A $120 billion AI chip market in 2025, roughly triple its 2023 size, sits almost entirely on NVIDIA’s shoulders, whose data center revenue has grown from around $3.6 billion per quarter in 2022 to $75.2 billion in a single quarter by May 2026, a jump of roughly twentyfold in under four years. The company’s full fiscal year 2026 revenue of approximately $215.9 billion, up 65% year over year, confirms this is not one exceptional quarter but a sustained growth trajectory backed by $1 trillion in confirmed demand through 2027.
Supply has not kept pace with any of this. Even with 1.7 million high-end AI GPUs shipped globally in 2025, data-center GPU lead times still stretch 36 to 52 weeks in 2026, an improvement from the brutal 6-to-12-month waits customers faced for H100 orders at peak demand in 2024, but still long enough to shape how hyperscalers plan capacity years in advance. That persistent gap between chip demand and chip availability is the single biggest reason infrastructure spending keeps outrunning even aggressive prior-year forecasts.
Cloud Backlog and Compute Demand Statistics in US 2026
| Metric | Figure |
|---|---|
| AWS cloud backlog | $244 billion (+40% YoY) |
| Oracle infrastructure backlog | $523 billion |
| Microsoft Azure undeliverable backlog | $80 billion |
| Google 2025 capex: guided vs. actual | $75B guided → $91.4B actual |
| Amazon 2025 capex: guided vs. actual | ~$100B guided → $131.8B actual |
| Meta 2025 capex: guided vs. actual | $60-65B guided → $72.2B actual |
Source: ValueAdd VC, BuildMVPFast, CNBC
Cloud Provider Backlogs (USD Billions)
Oracle ██████████████████████████ $523B
AWS ████████████ $244B
Microsoft ████ $80B
Backlog data is arguably the clearest evidence that AI infrastructure demand is real rather than speculative. Oracle’s $523 billion backlog, which includes signed commitments from companies like Meta and Nvidia, and AWS’s $244 billion figure, up 40% year over year, represent capacity customers have already agreed to pay for once it exists. Microsoft’s $80 billion in Azure orders it cannot yet deliver is the starkest example: the company has GPUs sitting in inventory that simply cannot be deployed because there isn’t enough power available to run them, a constraint that no amount of chip supply alone can fix.
The pattern of every hyperscaler blowing through its own capex guidance tells the same demand story from a different angle. Google’s actual 2025 spending of $91.4 billion landed 22% above its $75 billion guidance, Amazon’s $131.8 billion came in roughly 32% above its “about $100 billion” target, and Meta overshot its guidance range by nearly 12%. When every major player in an industry consistently underestimates its own future spending for two consecutive years, it signals that demand visibility, not investor caution, is now the limiting factor on how fast AI infrastructure gets built.
Data Center Construction Statistics in US 2026
| Metric | Figure |
|---|---|
| US data center construction starts, 2025 | $77.7 billion (+190% YoY) |
| January 2026 monthly starts record | $25.2 billion |
| Average cost per square foot, 2026 | ~$746 (+45% vs. 2025) |
| Total US data centers, March 2026 | 4,011 |
| Combined US data center capacity, end 2025 | Over 50 GW |
| Southeast US projected 2030 capacity share | 35% |
Source: ConstructConnect, FERC, Brightlio
US Data Centers vs Nearest Competitor (March 2026)
United States ██████████████████████████ 4,011
United Kingdom ███ 511
The physical construction backing this capital wave has grown just as fast as the spending behind it. US data center construction starts hit $77.7 billion in 2025, a 190% increase year over year, and January 2026 alone brought a record $25.2 billion in new groundbreakings. Costs have climbed right alongside volume: the average price per square foot reached roughly $746 in 2026, a 45% jump from the prior year, driven by AI-specific rack density, cooling, and power distribution requirements that didn’t exist in general-purpose cloud facilities.
The US now hosts 4,011 data centers as of March 2026, nearly eight times the 511 operating in the UK, its closest competitor by count, with combined US capacity exceeding 50 GW by the end of 2025. Development is also actively spreading beyond its original hubs: the Southeast, anchored by Georgia, the Carolinas, and Texas, is projected to control 35% of national data center capacity by 2030 as traditional markets like Northern Virginia hit power and permitting ceilings that newer regions haven’t yet reached.
Power and Energy Procurement Statistics in US 2026
| Metric | Figure |
|---|---|
| New utility-scale generating capacity planned, 2026 | 86 GW (record) |
| Data centers’ share of US electricity demand growth | 55% |
| PJM grid power supply cost increase, one year | $2.2 billion to $14.7 billion |
| US interconnection queue, early 2026 | Over 2,600 GW |
| Average interconnection wait time | ~5 years |
| Project withdrawal rate from that queue | ~80% |
| Memory’s share of hyperscaler data center spending | 30% (4x increase vs. 2023) |
Source: EIA, Brookings, Tom’s Hardware
US Interconnection Queue: Projects Waiting for Grid Power
Total queued ██████████████████████████ 2,600+ GW
Average wait ~5 years
Power has overtaken almost every other constraint as the defining bottleneck of AI infrastructure. The Energy Information Administration expects a record 86 GW of new utility-scale generating capacity in 2026, driven substantially by the fact that data centers now account for 55% of all US electricity demand growth. Even with that record build-out, the national interconnection queue has ballooned to over 2,600 GW of proposed projects, with average wait times near 5 years and roughly 80% of queued projects eventually withdrawn, a bottleneck severe enough that it is now shaping where hyperscalers choose to build as much as land cost or tax incentives do.
The strain is visible in market pricing too. In the PJM grid region alone, power supply costs jumped from $2.2 billion to $14.7 billion in a single year, a shift regulators trace substantially to AI data center demand. Memory has quietly become part of the same power-and-cost equation: it now consumes 30% of hyperscaler data center spending, a fourfold increase since 2023, as high-bandwidth memory becomes as tightly constrained a resource as electricity itself.
Nuclear and Alternative Power Statistics for AI Infrastructure in US 2026
| Metric | Figure |
|---|---|
| Total nuclear capacity committed to AI data centers | 9.8 GW across 13 deals |
| Microsoft’s Three Mile Island restart (Crane Energy Center) | 835 MW, online ~2027 |
| Amazon-Talen Susquehanna PPA | 1.92 GW, 17-year deal to 2042 |
| Amazon investment in X-energy SMRs | $700 million, up to 12 reactors |
| Google-Kairos Power SMR deal | 500 MW |
| Meta’s combined nuclear commitments | Up to 6.6 GW |
Source: SMR Intel, Carnegie Endowment, Build Insights
Hyperscaler Nuclear Commitments by Company (GW)
Meta ██████████████████████████ 6.6 GW
Amazon ████████ 1.92 GW
Microsoft ███ 0.835 GW
Google ██ 0.5 GW
Every major hyperscaler has now signed at least one nuclear power deal specifically to secure electricity for AI infrastructure, something that would have seemed implausible just a few years ago. Meta leads with commitments of up to 6.6 GW spread across TerraPower, Oklo, Vistra and Constellation, while Microsoft’s deal to restart Three Mile Island’s Unit 1 reactor, rebranded the Crane Clean Energy Center, will deliver 835 MW once it comes online around 2027. Amazon’s approach has been the most diversified, combining a 1.92 GW, 17-year power purchase agreement with Talen Energy’s Susquehanna plant through 2042 alongside a $700 million investment in X-energy for as many as 12 small modular reactors.
The logic behind all of this is straightforward: nuclear offers 24/7 carbon-free baseload power without competing for space in the same overloaded interconnection queues straining the rest of the grid. Google’s agreement with Kairos Power for 500 MW of small modular reactor capacity follows the same pattern, though none of these facilities, including the more advanced restart projects, are expected to deliver power before 2027 at the earliest. That timeline gap means hyperscalers will lean on natural gas generation to bridge the years between today’s construction boom and nuclear’s eventual contribution, even as the long-term nuclear commitments keep growing.
Semiconductor Supply Chain Statistics in US 2026
| Metric | Figure |
|---|---|
| Data centers’ share of global memory chip consumption | 70% |
| Memory order backlog created by that demand | 18 months |
| Micron’s DRAM revenue growth, year over year | 69% |
| AI data center semiconductor revenue, 2026 | $670 billion |
| North America’s share of the GPU semiconductor market, 2025 | 49.51% |
| Combined hyperscaler capex directed at AI infrastructure | ~75% of total spend |
Source: Mordor Intelligence, Semiconductor Industry Association, Programs.com
Data Centers' Share of Global Memory Chip Demand
Data centers ██████████████████████████ 70%
All other uses ████████████ 30%
The chip shortage driving AI infrastructure spending extends well beyond GPUs into the memory supply chain underneath them. Data centers now consume 70% of all memory chips produced globally, creating an 18-month backlog for new orders and pushing Micron’s DRAM revenue up 69% year over year as manufacturers scramble to keep pace. The Semiconductor Industry Association now estimates total AI data center semiconductor revenue reached $670 billion in 2026, confirming that AI has become the single largest demand driver across the entire chip industry, not just within the GPU segment specifically associated with training large models.
North America’s 49.51% share of the global GPU semiconductor market reflects how much of this demand originates domestically, even though the underlying manufacturing, particularly high-bandwidth memory, remains concentrated among a small number of Asian suppliers racing to expand capacity. With roughly 75% of combined hyperscaler capex now flowing directly into AI infrastructure rather than general-purpose cloud capacity, the entire semiconductor supply chain, from GPUs to memory to advanced packaging, has effectively been reorganized around a single application. For a deeper look at how this compute buildout translates into physical facilities, the AI data center statistics in US report breaks down the site-level detail behind these national totals.
Long-Term AI Infrastructure Investment Forecast Statistics in US 2026
| Metric | Figure |
|---|---|
| Goldman Sachs Big Four capex forecast, FY2025-FY2030 | $5.3 trillion |
| Baseline aggregate capex forecast, 2026-2031 | $7.6 trillion |
| McKinsey global data center capex forecast by 2030 | $7 trillion |
| Global AI data center market, current to 2032 | $344B → over $2 trillion |
| Anthropic’s annualized revenue run rate, May 2026 | $47 billion |
| New US billionaires tied to the AI infrastructure boom, 2025 | 50+ |
Source: Goldman Sachs, McKinsey, Forbes/IPS
Big Four Hyperscaler Capex Forecast: FY2025-FY2030
Prior estimate ██████████████████ $4.5T
Current estimate ██████████████████████ $5.3T
Wall Street’s own forecasts keep moving upward alongside the actual spending. Goldman Sachs now projects $5.3 trillion in combined capex from the Big Four hyperscalers between fiscal 2025 and fiscal 2030, revised up from $4.5 trillion before this year’s first-quarter earnings calls, while a broader baseline estimate covering compute, data centers, and power puts the 2026-to-2031 total at $7.6 trillion. That roughly matches McKinsey’s independent $7 trillion forecast for global data center capital spending by 2030, and the global AI data center market itself is projected to grow from $344 billion today to over $2 trillion by 2032.
The capital flowing through frontier labs illustrates just how fast the revenue side of this equation is catching up to the spending side. Anthropic’s annualized revenue run rate reached $47 billion by May 2026, up from roughly $1 billion less than eighteen months earlier, a pace of growth that helps explain why infrastructure providers keep signing multi-year, multi-billion-dollar contracts with confidence, a trajectory covered in more detail in the Anthropic IPO statistics report. That same capital wave has already reshaped personal fortunes as much as corporate balance sheets: more than 50 new US billionaires emerged from the AI infrastructure boom in 2025 alone, a wealth-creation pattern detailed further in the list of billionaires in US report.
Disclaimer: The data research report we present here is based on information found from various sources. We are not liable for any financial loss, errors, or damages of any kind that may result from the use of the information herein. We acknowledge that though we try to report accurately, we cannot verify the absolute facts of everything that has been represented.
