AI Infrastructure Capex: What 863 SEC Filings Reveal

Amazon spent $128.3 billion on capital expenditures in 2025 — up 65% in a single year, recorded in an SEC 10-K filed February 6, 2026. Alphabet spent $91.4 billion in the same period, up 74% year-over-year. Meta committed $115 to $135 billion for 2026 before that year had barely begun. Together these three hyperscalers are deploying more capital into physical computing infrastructure than the United States spent building the interstate highway system. Searching SEC EDGAR for "AI infrastructure" or "AI capex" returns 863 operational filings across five form types — evidence that this spending has moved from strategic announcement to required regulatory disclosure. Analyzing five 10-K and 10-Q filings across the value chain reveals three structural patterns the headline numbers alone do not capture.

The Value Chain Hiding in Plain Sight

These five filings do not tell separate stories. They describe different positions in the same transaction. Amazon and Alphabet record the capital as deployed cash. Meta records it as a forward commitment. Microsoft records it as cost pressure compressing current margins. And NVIDIA — the primary hardware supplier receiving these purchase orders as revenue — names the same infrastructure being built as the binding constraint on its own future growth.

Reading all five together reveals a self-reinforcing loop invisible from any single annual report: hyperscaler demand funds NVIDIA's GPU revenue, GPU supply enables more AI model deployment, more deployment drives more hyperscaler infrastructure spend, and the capacity of data centers and power grids to absorb continued GPU deployment becomes NVIDIA's growth ceiling. The cycle is now documented on five separate SEC filings.

Amazon and Alphabet: The Record-Breaking Deployment

Amazon's FY 2025 10-K reports the largest single-year capital expenditure in corporate history. Cash capex reached $128.3 billion, up from $77.7 billion in 2024, with the filing specifying this "primarily reflect[s] investments in technology infrastructure (the majority of which is to support AWS business growth)." The technology and infrastructure operating expense line grew 23% year-over-year — from $88.5 billion to $108.5 billion — the fastest-growing major cost line in the company. The 10-K explicitly guides that capex will increase again in 2026.

We expect spending in technology and infrastructure to increase over time as we continue to add infrastructure and employees, including to support our artificial intelligence and machine learning initiatives.

Alphabet's picture is similar but with a faster acceleration rate. Capital expenditures reached $91.4 billion in 2025, up 74% from $52.5 billion in 2024 — the fastest single-year infrastructure acceleration ever disclosed by a technology company at this scale.

Capital expenditures, which primarily reflected investments in technical infrastructure, were $91.4 billion for the year ended December 31, 2025.

Alphabet's 10-K also discloses a $4.8 billion acquisition of Intersect, a provider of data center and energy infrastructure solutions — a signal that Alphabet is moving up the physical infrastructure stack to control the underlying energy and facility layer, not just buying GPU capacity. Its 2026 guidance is explicit:

In 2026, we expect to significantly increase, relative to 2025, our investment in our technical infrastructure, including servers and network equipment, and data centers. The costs associated with operating our technical infrastructure — depreciation, energy, equipment, and network capacity — are expected to significantly increase as developing and serving AI offerings require more compute power than our historical consumer and enterprise offerings.

Combined, Amazon and Alphabet deployed $219.7 billion in capital expenditures in 2025 — and both guided those figures would increase again in 2026.

Meta: The $125 Billion Commitment Without a Comparable Baseline

Meta's FY 2025 10-K, filed January 29, 2026, presents a different kind of data point: not a reported figure but a forward commitment. The $115-135 billion range for 2026 represents what Meta intends to spend on infrastructure over a single calendar year — with the midpoint ($125 billion) sitting between Amazon's actual 2025 record and Alphabet's $91.4 billion actual.

We anticipate making capital expenditures of approximately $115 billion to $135 billion in 2026 to support our AI efforts and core business.

Meta generated $200.97 billion in revenue in FY 2025 at a 41% operating margin. A $125 billion midpoint capex commitment would represent more than 62% of revenue — an extraordinary capital intensity ratio for a software-driven business that historically ran on thin physical infrastructure. The filing frames this investment as supporting near-term ad targeting monetization through AI-driven ranking, longer-term bets including "superintelligence," and the Reality Labs hardware business — all drawing from the same infrastructure budget.

Microsoft: Where Infrastructure Costs Surface in Reported Margins

Microsoft's 10-Q for the quarter ending September 30, 2025 is the most actionable data point in this analysis — not because of its dollar size, but because it is the only filing in this group that shows AI infrastructure costs compressing reported margins in a closed financial period, not projecting them into future periods.

Azure revenue grew 40% in that quarter. Intelligent Cloud cost of revenue grew 43%. That 300-basis-point gap — revenue growing slower than the costs to deliver it — is attributed directly to AI infrastructure in the filing.

Gross margin percentage decreased driven by the impact of scaling our AI infrastructure, offset in part by efficiency gains in Azure.

We will continue to invest in capital expenditures to support growth in our cloud offerings and our investments in AI infrastructure and training.

The other hyperscalers describe this dynamic in the future tense. Microsoft is reporting it in the past tense, for a quarter already closed and audited. That distinction matters for investors timing positions around whether AI infrastructure costs represent a temporary investment phase or a structural margin drag on cloud businesses.

NVIDIA: The Supply-Side Mirror

NVIDIA's FY 2026 10-K, filed February 25, 2026 for the period ending January 25, 2026, closes the loop. The company's revenue growth "was driven by data center compute and networking platforms for accelerated computing and AI solutions" — meaning the hyperscaler capex described in the four filings above flows directly through NVIDIA's income statement as revenue. But what makes NVIDIA's filing structurally distinct is its characterization of the same infrastructure being built as its own primary business risk.

We expect to increase capital expenditures in fiscal year 2027 relative to fiscal year 2026 to support the future growth of our business.

Availability of data centers, energy, and capital to support AI infrastructure buildout.

The company receiving the most revenue from AI infrastructure spending identifies the same physical infrastructure as its growth ceiling. NVIDIA's demand is not constrained by chip manufacturing capacity but by whether enough data center space, power grid capacity, and deployment capital exist to absorb continued GPU output. Hyperscalers racing to lock in data center capacity are not building only for their own workloads — they are creating the absorption infrastructure that NVIDIA's revenue trajectory depends on.

NVIDIA also absorbed a $4.5 billion charge during the Blackwell product transition for H20 excess inventory and purchase obligations — evidence that even the primary GPU beneficiary of this cycle faces real supply-demand mismatch risk when hyperscaler purchasing timing shifts.

The Pattern

Three structural patterns emerge from reading these five filings together, and none is visible from any single annual report alone.

The spending is accelerating, with no plateau in sight. Every company in this dataset is guiding higher capex for 2026 versus a 2025 that was already a record year. Alphabet went from $52.5 billion to $91.4 billion in a single year and guided "significantly increase" again. Amazon went from $77.7 billion to $128.3 billion and guided higher. Meta committed to a range whose low end ($115 billion) exceeds either company's 2025 actual. No filing in this dataset contains language that signals a spending ceiling.

Infrastructure costs are appearing in reported margins now, not just future guidance. Microsoft's Q1 FY 2026 10-Q is the leading indicator for what Alphabet and Amazon will begin reporting as they close out 2026 and 2027 quarters: infrastructure depreciation, energy, and networking costs growing faster than the cloud revenue they enable. Alphabet's 10-K already flags this explicitly: "depreciation, energy, equipment, and network capacity - are expected to significantly increase as developing and serving AI offerings require more compute power than our historical consumer and enterprise offerings." The question for investors is not whether this will compress cloud margins — it will — but when it becomes visible in reported numbers for each hyperscaler.

NVIDIA's growth constraint is the hyperscalers' construction agenda. The GPU supplier's identification of data centers, energy, and capital as key uncertainties creates the feedback loop: hyperscaler demand funds GPU revenue, GPU supply enables AI model deployment, more deployment drives more infrastructure spend, and the rate at which physical infrastructure absorbs deployed GPUs becomes NVIDIA's revenue ceiling. Investors tracking only the hyperscaler capex figures are watching one side of a two-sided market. The other side — physical data center construction timelines, power interconnection queues, and water cooling constraints — is moving at a materially slower pace than either side of the transaction has publicly priced.

Three specific, observable signals to watch: (1) Microsoft Azure gross margin direction in coming quarters — its trajectory from the Q1 FY 2026 data point will signal whether AI infrastructure is temporarily or permanently dilutive across the peer group. (2) NVIDIA data center revenue relative to hyperscaler capex commitments — divergence between stated spend and GPU purchase rates would signal either demand softness or improving compute efficiency per dollar of infrastructure. (3) The Alphabet Intersect acquisition close and any resulting disclosures about owned data center and energy cost structure — the first hyperscaler to disclose unit economics on fully-owned physical infrastructure changes the industry's cost baseline.

Frequently Asked Questions

What is AI infrastructure capex and why does it matter for investors?

AI infrastructure capex is capital spending on data centers, servers, networking equipment, and physical computing infrastructure required to train and serve AI models at scale. It matters because the numbers have reached utility-scale: Amazon alone deployed $128.3 billion in capex in 2025, Alphabet $91.4 billion, and both guided higher for 2026. At this scale, the spending reshapes free cash flow profiles, cloud gross margins, energy demand, and real estate requirements across multiple sectors — not just technology.

How much are hyperscalers spending on AI infrastructure capex?

Based on FY 2025 10-K filings filed in early 2026: Amazon reported $128.3 billion in cash capital expenditures (up 65% from $77.7 billion in 2024), and Alphabet reported $91.4 billion (up 74% from $52.5 billion). Meta guided $115-135 billion for 2026. Combined, Amazon and Alphabet alone deployed over $219 billion in capital expenditures during 2025 — and that figure does not include Microsoft or Meta's actual spend.

How does AI infrastructure capex show up in Microsoft's current financial results?

Microsoft's 10-Q for the quarter ending September 30, 2025 is the clearest current example: Azure revenue grew 40% while Intelligent Cloud cost of revenue grew 43%, with the filing explicitly attributing gross margin compression to "scaling our AI infrastructure." This makes the infrastructure cost dilution observable in a closed reporting period — not just forward guidance, which characterizes every other hyperscaler filing in this analysis.

What does NVIDIA's 10-K say about the AI infrastructure buildout?

NVIDIA's FY 2026 10-K identifies "availability of data centers, energy, and capital to support AI infrastructure buildout" as a key uncertainty for its own business growth. The GPU supplier views infrastructure capacity absorption as the binding constraint on its revenue trajectory — the same data centers and power grids that hyperscalers are racing to build also determine how quickly NVIDIA can convert its GPU supply into revenue.

Is AI infrastructure capex accelerating or reaching a plateau?

Accelerating, based on all five filings reviewed. Amazon's 2025 capex was up 65% from 2024 and its 10-K guides higher for 2026. Alphabet's was up 74% and guided to "significantly increase" again. Meta committed $115-135 billion for 2026. Microsoft stated it will "continue to invest in capital expenditures to support growth in our cloud offerings and our investments in AI infrastructure and training." No filing in this dataset contains language describing a plateau or a reduction in spend pace.

What should investors watch to track AI infrastructure capex going forward?

Three observable signals: (1) Azure and Google Cloud gross margin trends — Microsoft already shows the compression in reported quarters, and Alphabet explicitly flags the cost growth dynamic in its 10-K. (2) NVIDIA data center revenue relative to hyperscaler capex commitments — a gap between stated spend and GPU purchase orders signals either demand softness or improving compute efficiency per dollar of infrastructure. (3) Whether Amazon's actual 2026 capex exceeds its 2025 baseline of $128.3 billion, which the 10-K explicitly guided would increase.

Methodology

This analysis used MetricDuck's SEC filing intelligence tools to search 863 operational filings across five form types (8-K, 10-Q, 10-K, S-1, DEF 14A) for entities matching "AI infrastructure," "AI capex," or "artificial intelligence capital expenditure." We identified five companies with substantive capex disclosures (META, GOOGL, AMZN, MSFT, NVDA) and analyzed their MD&A, liquidity, and capital allocation sections for patterns across both demand (hyperscaler spending) and supply (GPU hardware) sides of the AI infrastructure transaction.

Tools used: SEC EDGAR Full-Text Search (EFTS) for discovery across all EDGAR filings from March 2025 through March 2026, MetricDuck filing intelligence for company-level signal extraction and triage, MetricDuck filing section reader for evidence extraction from specific MD&A and liquidity sections.

Limitations: EFTS keyword matching returns filings that mention the search terms but does not distinguish between companies making direct AI infrastructure investments and those citing AI capex in the context of market conditions or competitor risk. The 863 operational filings include many S-1 and DEF 14A filings that reference AI capex trends without company-specific spending data. Additionally, this analysis covers disclosures through March 2026; actual 2026 capex figures will not be available until FY 2026 annual reports are filed in early 2027, meaning Meta's $115-135 billion commitment remains guidance, not a reported figure. Microsoft's 10-Q data reflects one fiscal quarter (July-September 2025) and may not represent full-year 2025 trends.

Disclaimer: This analysis is for informational purposes only and does not constitute investment advice. All figures are sourced directly from SEC filings and represent reported or guided financial data as disclosed by the companies at the time of filing. Past capital expenditure patterns do not guarantee future spending levels or investment returns.