AI Infrastructure Capex

AI infrastructure capex refers to the wave of hyperscaler, neocloud, and frontier-lab capital spending on data centers, chips, and the electricity, cooling, and transmission infrastructure those data centers require. As of mid-2026 it had become the dominant feature of the US AI compute economy. It is distinguished from generic enterprise IT spending by its scale (single hyperscalers reporting $80B–$145B in 2026 capex), the concentration of its physical assets in a small number of US states and overseas hubs, and the upstream demand it places on electric utilities, chip foundries, and private-credit markets. Its effects connect to questions of whether the spend is justified (AI Bubble Debate), where it is sited (AI Data Centers), who supplies its power (Dominion Energy, NextEra Energy, Constellation Energy), and who finances it (Private Credit & AI Infrastructure).

Snapshot

Reported 2026 capex by spender

Numbers are reported figures from each company's most recent guidance; subject to revision on the next earnings cycle.

Memory-supply commitments

Capex commitments have extended upstream to the memory supply the accelerator buildout depends on. Kioxia Corporation and Sandisk Corporation announced on August 27, 2026 anticipated investments in Japan totaling over $31 billion, approximately 5 trillion yen, contingent upon government support. The investments run through 2032 and fund continued buildout at the Yokkaichi Plant and the Kitakami Plant along with related infrastructure and technology. The two companies said their joint venture has invested over $50 billion, approximately 9 trillion yen, in Japan over the past 25 years, and framed the new commitment as supporting multi-year bit growth and stable flash-memory supply. The figure is contingent and the government support is not committed (Source: kioxia.com).

The August 28, 2026 close of Andreessen Horowitz's $1.1 billion Machine Age fund, its first dedicated hardware-infrastructure vehicle covering AI processors, memory chips, networking equipment, data storage and robotics, is a venture-scale counterpart to the same supply-side concern (Source: wsj.com). See Andreessen Horowitz (a16z).

Non-hyperscaler cloud providers have made the same kind of pre-purchase against specific customer contracts. On September 24, 2026 Akamai authorized its contract manufacturer Jabil to buy approximately $1.7 billion of memory components, which Jabil holds on consignment and repurchases at cost as they are used. Akamai estimated total capital expenditure related to Anthropic's approximately $11.6 billion commitment at approximately $5.5 billion, and an increase of approximately $1.7 billion in its 2026 capital expenditure "to secure and pre-purchase critical supply chain components, including memory," with no change to its 2026 revenue guidance (Akamai Technologies Form 8-K: Anthropic Project Plans and Warrant Agreement (September 2026)).

Construction put in place

The U.S. Census Bureau's construction-spending release of September 1, 2026 put data-center construction put in place in July 2026 at an annual pace of more than $75 billion, rising at an annualized rate of nearly 60% from July 2025. The series measures building shells only, which Cushman & Wakefield estimates at roughly 20% of all-in data-center cost, so it tracks the physical buildout rather than total capex (Source: axios.com).

Demand and capacity

Epoch AI's May 25, 2026 compute-crunch model puts the demand side at roughly 10×/year token growth against roughly 3.4×/year inference-capacity growth, implying a binding capacity constraint for long-context agentic workloads through 2027 even at the prevailing capex run-rate (Epoch AI; Source: epoch.ai).

Off-balance-sheet commitments

A Wall Street Journal analysis of company filings published August 16, 2026 reported that Alphabet, Amazon, Meta, and Microsoft carry $1.52 trillion in purchase commitments and $904 billion in leases not yet started off their balance sheets, against $248 billion in lease liabilities and $356 billion in long-term debt on them. The piece is headlined as putting big tech's AI spending $3 trillion higher than reported capital-expenditure figures suggest; only the deck and the underlying chart were retrievable, and the claims here are confined to those figures (Source: wsj.com). Stijn Van Nieuwerburgh's Brookings Papers on Economic Activity paper (Fall 2026) cites the same Journal analysis as totalling $2.4 trillion of off-balance-sheet liabilities for the four companies against $604 billion on balance sheet, so that only one-fifth of about $3 trillion of data-center-related obligations is currently recognized, and cites a February 2026 Moody's estimate that hyperscalers hold roughly $970 billion of lease commitments, about $660 billion of them future leases not yet on balance sheet (Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)). See Private Credit & AI Infrastructure.

Structural drivers

Three structural pressures distinguish AI-infrastructure capex from prior IT-capex cycles.

The first is a reweighting from training toward inference. Inference is forecast to reach roughly two-thirds of AI compute demand by end-2026 (Baseten pitch deck, May 2026). That shifts the capex mix from a small number of very large training clusters to a much larger number of geographically distributed inference clusters, raising power, cooling, and latency constraints.

The second is the shift toward agentic workloads. Per Epoch AI's model, long-context agentic workloads (multi-step tool-using agents holding large state) are the binding driver, implying that capacity allocation increasingly favors agent-style inference, with downstream consequences for the Agentic Economy (Epoch AI).

The third is the electricity bottleneck. Northern Virginia's "Data Center Alley" — Loudoun and Prince William counties — is the world's largest concentration of AI/cloud data-center load, and Dominion projects 5.5%/year demand growth there through 2030. The May 18, 2026 NextEra–Dominion $67B all-stock merger was framed as a response to AI-data-center demand and was the first such major utility-consolidation event explicitly cast in those terms; Constellation–Calpine ($27B) and BlackRock–AES/Allete ($39B) are the comparable utility-sector transactions (How the AI revolution has turbocharged M&A).

Aggregate scale, funding mix and depreciation coverage

Exponential View's June 2026 model puts cumulative hyperscaler and neocloud capex at $2 trillion through 2026E on a property-plant-and-equipment-plus-leases basis across seven stack-builders, of which $535 billion is above the pre-AI trend. It cautions that announced totals are not all AI: they include pre-planned spending for existing cloud and SaaS businesses, Meta's metaverse work and Amazon's logistics (The State of the AI Economy (Exponential View, June 2026)).

The same BPEA paper reports that aggregate capital expenditures of Oracle, Microsoft, Amazon, Meta and Alphabet rose from roughly $97 billion in 2020 to more than $400 billion in 2025 and are projected to exceed $800 billion in 2026, surpassing the five firms' combined operating cash flow for the first time (the series includes non-AI capex and excludes data centers the hyperscalers lease; 2026 values combine reported results, guidance and Wall Street estimates). Building up from a representative 200 MW campus costing about $8.2 billion, roughly two-thirds IT equipment, and Cleanview project data, its central scenario has 182.7 GW of U.S. capacity completed in 2025–2032 and $10.3 trillion of investment over the period, 3.63% of GDP, peaking at $2,058 billion (5.1% of GDP) in 2032; the estimate covers initial construction and equipping only, not hardware refreshes. The paper states that financing is shifting toward leases, joint ventures, project debt, private credit, securitization and special-purpose vehicles as capex outgrows internal cash flow (Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026)).

A narrower cash-flow comparison was presented at the AI Infra Summit in Santa Clara on September 21, 2026. A technology equity analyst at the London-based research firm Arete Research said Alphabet, Amazon, Meta and Microsoft are expected to spend $1.8 trillion on capital investments through 2028, about $600 billion more than their underlying businesses generate in cash over the same period; the coverage's headline describes capex as consuming nearly all of the four companies' operating cash flow in 2026. The retrievable account gives neither the analyst's name nor the underlying model (Source: broadbandbreakfast.com).

The composition of a build dollar has shifted toward silicon. Between 2021 and 2026E, chips' share of total data-center build cost rose from 40% to 60%, and memory moved from about 2% to about 18% — the single largest mover in the mix. Nvidia supply commitments grew from $31 billion to $95 billion over the year to June 2026, and the extra electricity the US grid is expected to need by 2030 rose from 24 GW to 166 GW since 2022, with data centers accounting for about 55% of that growth (The State of the AI Economy (Exponential View, June 2026)). See AI Data Centers.

Funding mix is diverging by operator type: neocloud capex is primarily debt-funded while hyperscaler capex remains primarily cash-funded, so the marginal AI-infrastructure dollar is increasingly externally financed. The report frames the distinction as one about where a bad bet lands — external funding "moves risk outside firms, as third-parties expect repayment," while paying cash "keeps a bad bet inside the firm, only denting profits" — and notes that hyperscalers' market-capitalization weight transmits risk to the wider market regardless (The State of the AI Economy (Exponential View, June 2026)). See Circular Financing in AI.

Against a 2026E depreciation charge approaching $111 billion (IT equipment over six years, buildings over fourteen), quarterly AI revenues first exceeded quarterly capex depreciation in Q4 2025. Q1 2026 headroom, defined as (revenue − depreciation) ÷ revenue, was 19% for hyperscaler and neocloud revenues and 32% across all GenAI revenue; cumulative revenue had nearly covered cumulative depreciation but not the expected headroom. The coverage is highly sensitive to the assumed chip life — 19% at six years, 36% at eight — and the report's evidence for a longer life is that gross rental yields at 50% utilization exceed the 17% six-year depreciation charge for newer GPUs and persist beyond six years for older ones (The State of the AI Economy (Exponential View, June 2026)).

A per-unit cost model in the same report puts the annual cost to own and operate 1 GW of AI capacity at $7.9 billion, 89% capital and 11% operating: servers $4.6 billion (480,000 GPUs in 6,700 systems, six-year life), facility $1.4 billion (fourteen-year life), data-center network $1.0 billion, land and utility $33 million, energy $594 million and other operating cost $308 million. Energy is 66% of operating cost but under 8% of the all-in annual figure — the buildout is capital-dominated rather than electricity-dominated at current prices, even where electricity is the binding physical constraint on siting (The State of the AI Economy (Exponential View, June 2026)). See Inference Economics and Token Pricing.

Measured contribution to GDP

An Epoch AI report of August 24, 2026 addresses why spending of this size has shown a modest measured effect on US GDP growth. It puts US investment in computing equipment at roughly $400 billion per year, nearly triple its 2023 level, and shows net imports of computers and peripherals rising nearly in lockstep, so that the two largely cancel in expenditure-side GDP. It accepts that much of the equipment is imported but argues the explanation is incomplete: the value-add of US fabless chipmakers, chiefly Nvidia, is counted in the import price of servers entering the US but never recorded as a US export, because the chips are manufactured and sold abroad. Using Nvidia's US-booked operating income as a proxy, it estimates that annualized US GDP growth over the prior year was underestimated by about 0.3 percentage points, and that the gap could reach almost two percentage points of growth per year by 2028 if Nvidia's operating income keeps growing at its current exponential rate (The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026)). See AI and Productivity.

M&A and private-credit ramifications

How the AI revolution has turbocharged M&A (FT, 2026-05-22) is the anchor source for the case that AI capex is the central driver of the 2025-2026 M&A boom. Transactions tied to the AI-infrastructure capex thesis include:

KKR's CoolIT exit returned 15× on a liquid-cooling investment, an indication of how rapidly private capital was crystallized by the data-center cooling and power-management subsegments (How the AI revolution has turbocharged M&A).

Debates and positions

Sources disagree on whether the spend is justified. Gary Marcus's May 26, 2026 Substack post argued that if Uber's COO's disclosure of "not seeing proportional productivity gains from AI" is confirmed across more enterprises, "the AI bubble pops" (Source: garymarcus.substack.com). Cited counter-evidence includes Cloudflare's greater-than-30% revenue growth alongside the Matthew Prince layoff thesis; OpenAI's Pennsylvania ChatGPT public-sector deployment averaging 95 minutes saved per user per day; and Anthropic's reported $14B Claude Code ARR.

A second debate frames the capex as either a moat or a stranded-asset risk. The moat case is associated with Nvidia & TSMC — AI Compute Infrastructure, Oracle, and SoftBank Group; the stranded-asset case appears on AI Bubble Debate and in the FT M&A piece's "discretionary state capitalism" framing (How the AI revolution has turbocharged M&A).

A third concerns public-sector exposure. OpenAI's May 27, 2026 call for a "Strategic Compute Reserve," modeled on the Strategic Petroleum Reserve, would federalize a slice of the capex stack for national-security AI use (Strategic Compute Reserve).

Microsoft's May 2026 wind-down of some Claude Code licenses, which Fortune attributed to token-cost pressure, has been read as an early sign of a hyperscaler rationalizing AI-token spend (Source: theverge.com).

Relationships

• related: The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026) — argues measured GDP omits most of fabless chipmakers' value-add from the hardware buildout
• supports: AI Data Centers; Data Center Siting / AI Power Politics; AI Environmental Impact; Private Credit & AI Infrastructure
• contradicts: the strict version of AI Bubble Debate (which treats the capex as the bubble itself)
• depends-on: Agentic Economy (workload mix), Inference Economics and Token Pricing
• related: Akamai Technologies, Akamai Technologies Form 8-K: Anthropic Project Plans and Warrant Agreement (September 2026) — capex and memory pre-purchase tied to the Anthropic contract
• related: Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026); Stijn Van Nieuwerburgh; Dominion Energy; NextEra Energy; Constellation Energy; Meta AI; Microsoft; Google DeepMind; Amazon; OpenAI; Baseten; Epoch AI; Strategic Compute Reserve; How the AI revolution has turbocharged M&A

Sources

• How the AI revolution has turbocharged M&A — consolidated long-form survey of the 2025-2026 M&A wave through the AI-capex lens
• Epoch AI compute-crunch model (2026-05-25) (Source: epoch.ai)
• AWS–Snowflake $6B agentic-CPU deal (2026-05-27) (Source: wsj.com)
• Marcus on AI-spend rationalization (2026-05-26) (Source: garymarcus.substack.com)
• The State of the AI Economy (Exponential View, June 2026) — deduplicated capex, funding-mix and depreciation-coverage model, June 2026
• Financing the AI Buildout (Van Nieuwerburgh, BPEA Fall 2026) — BPEA campus-cost, capacity and investment scenario (2025–2032); hyperscaler capex versus operating cash flow