Your cloud provider stops quoting on-demand rates and starts mapping a three-year reservation, which tells you something about how much capacity it has already committed to build. Alphabet (Google), Amazon, Meta and Microsoft are putting hundreds of billions of dollars into artificial intelligence (AI) capital expenditure (capex) this year, and your compute terms come out of those decisions.

Founders raising a seed or Series A have to price margins against a buildout they don't control, and compete with whatever the hyperscalers build to fill that capacity. This article covers what AI capex is, where the money goes and how founders should read the cycle.

What AI Capex Is

AI capex is capital expenditure directed at future compute capacity. It covers graphics processing units (GPUs) and custom accelerators, data center shells, networking equipment, power systems and cooling. The purchase creates a long-lived asset, so the buyer records it on the balance sheet and depreciates it over years rather than expensing it up front. That timing difference explains how a company can commit tens of billions in a single quarter while its reported earnings barely move.

None of the hyperscalers reports AI capex as a separate line item. The cash sits inside "purchases of property and equipment" on the cash flow statement, mixed in with warehouses, office buildings and satellites. Pulling the AI portion out means reading finance lease disclosures next to management commentary, and because each company draws that boundary differently, published totals for the same company vary. Any AI capex figure you read is a reconstruction, not a reported number.

How Much AI Capex the Largest Hyperscalers Committed in 2026

Combined hyperscaler commitments now total roughly $725 billion, up about 77 percent from around $410 billion in 2025. Guidance rose at Alphabet, Amazon and Meta during the year, while Microsoft's figure came in below expectations for reasons that have more to do with accounting than with capacity:

• Alphabet: Two upward revisions during the year took guidance to $195 to $205 billion, well above the initial $175 to $185 billion. The money funds tensor processing unit (TPU) clusters and delivery of the cloud backlog.
• Amazon: The largest single commitment sits at roughly $220 billion, after memory chip costs pushed the figure up by $20 billion mid-year. Most of it goes to Amazon Web Services (AWS) capacity and Trainium silicon.
• Meta: Component prices climbed through the year, lifting the range to $130 to $145 billion from an initial $115 to $135 billion. Spending targets training runs and ranking infrastructure.
• Microsoft: Calendar 2026 spending comes in at roughly $175 billion, below the $190 billion the market expected, because extending building useful life from 15 to 25 years moved some leases out of reported capex. The underlying build plans haven't changed, and Azure still holds an $80 billion backlog of orders it can't fill for lack of power.

Beyond the hyperscalers: Oracle spent $55.66 billion in its fiscal 2026 and guided up to $95 billion for fiscal 2027, while AI infrastructure companies like CoreWeave and Crusoe add further capacity. Counting those investments and sovereign programs in the Gulf and the European Union, global AI investment for 2026 exceeds $1 trillion.

Memory and component prices drove most of the mid-year raises at Amazon and Meta, so part of that increase buys the same capacity at a higher price. Anyone reading the totals as pure capacity growth overstates what the money delivers.

Where AI Capex Dollars Go

Chips are the largest single line, while construction lead times and grid interconnection now constrain new AI capacity. Spending splits unevenly across silicon, buildings, power and networking, and each category runs into a different bottleneck. The composition tells you more about delivery timing than the headline total does.

Accelerators and Custom Silicon

GPUs and custom accelerators account for roughly one third of total data center capex, the largest category by a wide margin. Custom parts are the fastest-growing slice within it, as Alphabet's TPUs, Amazon's Trainium and the design programs at Broadcom and Marvell take share from merchant GPUs. Advanced packaging capacity at Taiwan Semiconductor Manufacturing Company (TSMC) is the binding constraint here, and lead times on the most advanced parts remain long.

Sites and Construction

Construction takes a smaller share of spend than the hardware that goes inside the building, but it sets the delivery schedule. Shell and core construction averages about $11.3 million per megawatt this year, and the AI tenant fit-out can add up to $25 million per megawatt on top. Getting a hyperscale project from groundbreaking to commissioning can take years. A meaningful share of the capacity announced for 2026 hadn't broken ground by mid-year, which is why announced commitments and online capacity keep diverging.

Power and Cooling

Delays concentrate in the electrical build, and speed to power is now the first thing developers screen for when they pick a site. Substation transformer lead times exceed 160 weeks in 2026, up from around 140 weeks in 2023. Grid interconnection queues in Texas and Virginia stretch out for years, and power prices in data center-heavy regions have risen sharply. Regulators in several of those markets have started screening queued projects before they approve new large-load connections.

Networking and Interconnect

After the accelerators themselves, networking is the largest information technology expense inside the facility. Ethernet data center switch revenue topped $8 billion in a single quarter of 2025, more than double its level three years earlier, as clusters scale to tens of thousands of interconnected GPUs.

Every additional GPU needs high-bandwidth connections to the rest of the cluster, so spending here tracks the accelerator deployment schedule that also drives how the buildout gets financed.

How the AI Capex Buildout Gets Financed

Hyperscalers capitalize the hardware and buildings, then depreciate them over years. The power bills and rented capacity to run them hit the income statement right away as operating expense, and that division determines how much of the spend reaches earnings and when.

Funding comes from retained earnings or debt, and the mix changes what founders and investors should watch. Hyperscalers and their suppliers issued $225 billion in bonds through the first half of 2026, and a growing share of the buildout no longer comes out of operating cash flow. Negative free cash flow shows a company leaning less on the cash its operations generate, and slower buybacks reinforce that reading. Signed data center leases that haven't begun yet may sit outside reported capex, so uncommenced lease commitments in the footnotes deserve a look.

Depreciation charges continue for years after the cash goes out, since accelerator costs get spread over several years even though the hardware cycle can run faster. Shortening that schedule lowers operating income, while extending it lifts near-term margins and pushes the charge into later years. When purchases of property and equipment run ahead of reported depreciation, that margin drag keeps building. Concentration compounds the risk when a small set of AI labs underwrites a large share of a provider's backlog growth.

The Early Signs of an AI Capex Slowdown

Observable indicators would appear before headlines confirm a slowdown, although outsiders can't determine whether the spending will produce adequate returns. A real pullback would likely register across several of the indicators below at the same time:

• Accelerator lead times compressing: A sustained move toward materially shorter delivery times would mean supply had caught demand.
• A guidance cut from a major hyperscaler: A year-over-year capex reduction from Alphabet, Amazon, Meta or Microsoft would carry more weight than any outside estimate.
• Cloud backlog growth decelerating faster than committed capacity: Backlogs are the demand evidence behind the spending, and bookings slowing while capacity keeps landing would break utilization economics first.
• Power purchase agreement pricing easing: Softening rates in Iowa, Texas and Virginia would suggest developers expect less demand than the announcements imply.
• Useful-life assumptions shortening: Recent moves have run the other way, toward longer depreciation schedules that flatter near-term margins. A shift back toward shorter lives would concede that hardware ages faster than the books assume.

None of those signals has fired, since accounting treatment explains Microsoft's shortfall against expectations more than any decision to build less. Guidance rose everywhere else and the queues that gate delivery kept lengthening, which points to continued buildout rather than a pullback.

How Founders Should Read the AI Capex Cycle

CRV is an early stage venture capital firm, and early stage AI companies rent this capacity and build on the models rather than own the hardware. Those same companies compete with the products the hyperscalers ship to fill their data centers. Both effects shape investor diligence, and you don't need a view on whether the spending pays off to assess them.

Compute Access and Pricing

When a startup rents GPU capacity by the hour, the provider carries the depreciation risk on the underlying hardware. If accelerator generations turn over faster than the provider's depreciation schedule assumes, the write-down lands on the provider's books, and your exposure ends with the rental period. Providers can raise on-demand GPU prices while available capacity stays scarce, and neocloud providers can move smaller customers from hourly tiers into multi-year reservations. The lock-in risk to watch is a multi-year contract signed at today's rates right before the next hardware generation makes today's hardware cheaper.

Bundling and Vertical Integration

Hyperscalers that commit hundreds of billions to capacity have a structural incentive to fill it with their own application-layer products. Application-layer launches from hyperscalers can overlap directly with startups building on their models, which puts pressure on companies without a differentiated product or distribution advantage. Thin application companies absorb the worst of it, while anything that makes committed capacity more productive gains from the same incentive.

Vercel took the second path: it expanded from web deployment into AI agent tooling that helps teams use the buildout. CRV led Vercel's Series A and backed the company through its B, C, D and E rounds.

Gross Margins and Diligence

Fundraising conversations for AI companies now cover compute dependency and gross margin trajectory alongside the product itself. AI companies can justify structurally lower gross margins than typical software businesses, because infrastructure costs scale with usage. A credible path to margin improvement accounts for those costs instead of assuming scale erases them.

Single-provider concentration gets its own line of questions. Credits and discounts that pull a team onto one cloud early turn into renewal-time pricing power once the workloads settle there. We ask how hard it would be to move a meaningful share of those workloads elsewhere.

AI Capex Has Become a Power Story

In 2024, accelerator supply constrained data center builds. In 2026, the bottlenecks are power infrastructure and the financing needed to carry projects until power comes online. A shell without power remains unusable inventory, which is why transformer queues and interconnection waits now set the timeline for when committed spending becomes usable compute. Financing shapes that timeline too, since debt and lease structures determine which announced projects reach energization.

From our seat, the early stage founders positioned best in this cycle are the ones making committed capacity more productive. That can mean tools for inference efficiency and utilization, or the infrastructure layers that let application teams ship faster.

Competing with the hyperscalers on raw compute or thin applications means fighting the balance sheet math laid out above. If you're an early stage founder looking for a lead investor who understands how compute costs shape your margins, reach out to us to see if we'd be a good fit.

Frequently Asked Questions About AI Capex

Is AI capex the same as total AI spending?

AI capex covers capitalized assets: chips, buildings, networking and power infrastructure that depreciate over years. Total AI spending adds electricity costs. It also includes cloud rental fees and model usage charges, so aggregate estimates for the same year can differ by hundreds of billions of dollars based on scope.

Who benefits most from the AI capex cycle?

Suppliers of the constrained inputs capture the most direct gains: accelerator makers, construction firms, electrical equipment manufacturers and power developers. Startups that rent compute benefit indirectly, because the depreciation risk on the hardware belongs to the provider, which competes to fill capacity.

How does AI capex affect United States economic growth?

AI capex reaches economic growth through domestic investment in construction, power equipment and data center fit-out. Imported servers and chips pull in the other direction, because gross domestic product (GDP) accounting subtracts imports, and the headline spending totals don't separate the domestic portion from the imported one. The domestic share, not the headline total, is what shows up in growth figures.

Should a startup treat its own AI infrastructure as capex or opex?

Hardware you own goes on the balance sheet as capex and depreciates over its useful life. Compute you rent by the hour or month counts as operating expense (opex) in the period you use it. Implementation costs for cloud services depend on the project stage, and you can capitalize eligible application development work. Most early stage teams rent managed inference instead of owning hardware, so nearly all of their AI infrastructure spending lands in opex.