How the AI Boom Turns Into a Stock Boom: Following the Capex Chain

Definition

Artificial intelligence runs in buildings. Training and serving large models requires warehouses of accelerators, the electricity to run them, the cooling to stop them overheating and the fibre to connect them. Every dollar committed to that infrastructure leaves the buyer’s balance sheet and lands on a supplier’s income statement.

That transfer is the entire link between “AI is important” and “these stocks went up.” It is not a story about model capability. It is an accounting identity: capex out, revenue in.

How the Capex Chain Works

Compute. The largest single share of a data centre’s cost is the accelerator, but the chain runs deeper than one supplier. Each accelerator requires high-bandwidth memory, advanced packaging capacity, foundry wafer starts and substrate supply. The bottleneck moves around: when memory is scarce, memory producers capture the margin. Microsoft disclosed that roughly $25 billion of its approximately $190 billion 2026 capital expenditure is component price inflation — meaning a material share of hyperscaler spending is now being transferred to component suppliers through price rather than volume.

Power. A large AI campus draws power at the scale of a small city. Meta’s Hyperion campus in Louisiana is designed to draw up to 5 GW at peak. That demand supports generators, utilities, transformer and switchgear manufacturers, and cooling suppliers — but power is where the chain is slowest and least elastic, because grid interconnection queues and transmission build-out run on multi-year regulatory timelines that no amount of capital compresses.

Networking. Training a large model requires thousands of accelerators to exchange data continuously, which makes interconnect a performance constraint rather than a commodity. Optical transceivers, switching silicon and cabling scale with the size and topology of clusters.

Where the Spending Stands in 2026

For scale on the trajectory: Amazon’s approximately $200 billion for 2026 compares with roughly $96.5 billion spent in 2025 and $83 billion in 2024. Meta raised its full-year 2026 guidance mid-year, citing higher component pricing and additional data centre costs.

The capital intensity figures are the ones that matter for risk. A company spending 54% of its revenue on physical assets has committed to a future in which that capacity earns a return. If the return does not materialise, the depreciation still does — it flows through the income statement for years regardless.

How to Use Capex Data in Practice

1. Read capex guidance as the leading indicator for the whole chain. Rising guidance signals supplier revenue growth ahead; a flattening or cut signals the reverse, and it reaches supplier share prices long before it reaches supplier revenue.

2. Separate revenue recognised from revenue promised. A supplier’s backlog, book-to-bill ratio and remaining performance obligations tell you what has been contracted. A customer’s capex guidance tells you what has been announced. Announcements are revised; backlog is harder to cancel without penalty.

3. Distinguish price inflation from volume growth. Microsoft’s disclosure that roughly $25 billion of its 2026 capex is component price inflation means a meaningful part of the spending increase buys no additional capacity. For a component supplier that is margin; for the buyer it is cost with no return attached.

4. Check where a supplier sits on the lag curve. Compute suppliers convert orders in quarters. Power infrastructure converts over years, constrained by interconnection queues. Two companies exposed to the same theme can have completely different revenue timing, and the market frequently prices them as though they do not.

5. Measure your own concentration across the chain. A chip company, a utility with data centre exposure and a networking supplier look like three separate positions. If all three depend on the same capex assumption, they are one position held three times. Cluenex covers the top 1,000+ US-listed stocks, which allows that shared dependency to be examined name by name.

6. Watch the constraints, not the ambitions. Announced project pipelines exceed what the grid, the permitting process and local politics will permit. The binding constraint on AI buildout in 2026 is not capital.

Common Mistakes and Misconceptions

Example: When the Chain Runs in Reverse

The Virginia Digital Gateway in Prince William County, Virginia, was planned as approximately 2,100 acres and 22 million square feet of gigawatt-scale data centre capacity — the largest such campus proposed anywhere. Blackstone-owned QTS abandoned it after years of litigation.

Trace what that removes from the chain:

How Cluenex Uses Capex and Sector Data

Cluenex AI ingests capital expenditure guidance, sector conditions and company financials across the top 1,000+ US-listed stocks when calculating predicted short-term and long-term price movement. Those inputs feed the sentiment scores and forecasts shown on the platform rather than appearing as separate indicators.

The practical use for a capex-chain position is examining the shared dependency directly. Cluenex’s discounted cash flow and owner earnings tools value each supplier from the cash its operations generate, and the moat analysis distinguishes a company whose advantage is structural from one whose recent results reflect a temporary bottleneck. That distinction is what separates a durable business from a cycle beneficiary, and it is invisible in a single quarter of revenue growth.

Frequently Asked Questions

• What is capex and why does it matter for stock prices? Capital expenditure is money spent on long-lived physical assets, recorded on the balance sheet and depreciated over years rather than expensed at once. It matters twice: for the spender, it is a commitment that becomes depreciation whether or not the asset earns a return; for suppliers, it is the source of future revenue. Because share prices discount expected future profits, a change in a large buyer’s capex guidance moves supplier share prices immediately.

• How much are the hyperscalers actually spending in 2026? Amazon, Microsoft, Alphabet and Meta plan roughly $725 billion combined, up about 77% from a record $410 billion in 2025. Individually: Amazon approximately $200 billion, Microsoft approximately $190 billion, Alphabet $180–190 billion, and Meta $125–145 billion after a mid-year increase. As a share of sales this represents roughly 54% at Meta, 47% at Microsoft, 46% at Alphabet and 25% at Amazon.

• Which companies benefit most from AI capital spending? The spending divides across three links: compute — accelerators, memory, storage, packaging and foundry capacity; power — generation, transmission, transformers, switchgear and cooling; and networking — optics, switches and interconnect. Compute converts orders to revenue fastest, typically within quarters. Power is slowest, constrained by multi-year grid interconnection queues that capital cannot accelerate.

• What happens to supplier stocks when a data centre project is cancelled? The expected future revenue is removed from the forecast, and because share prices discount expected profits rather than reported ones, the repricing can happen immediately even though no order was ever booked. The market frequently cannot distinguish a local cause from a demand signal in the moment, so an individual project failure can move the whole chain before the reason is understood.

• Is the AI capex boom sustainable? There is real disagreement and no settled answer. Supporting continuation: cloud revenue growth remains strong and the largest buyers have raised rather than cut guidance. Against it: capex at 46–54% of sales converts into depreciation regardless of return, investors began scrutinising hyperscaler capex disclosures during the July 2026 reporting season, and physical constraints — grid interconnection, permitting, local opposition — are binding in ways capital cannot solve.

• Why did the Virginia Digital Gateway project fail? Blackstone-owned QTS abandoned the roughly 2,100-acre, 22-million-square-foot campus in Prince William County, Virginia, after years of lawsuits, with the project ultimately failing over a procedural newspaper-notice technicality. The cause was land-use litigation rather than financing difficulty, weak demand or power availability. That distinction matters, because it means the cancellation carries no information about aggregate AI demand.

• How do I avoid over-concentrating in the AI capex chain? Check whether your holdings share a single dependency rather than counting them as separate positions. A chip designer, a memory producer, a utility with data centre load and a networking supplier are four tickers exposed to one assumption about hyperscaler spending. If that assumption changes, they reprice together. Examine each company’s revenue exposure to the theme directly rather than relying on sector labels.

Related Concepts

• How to Evaluate AI Stocks: Metrics That Actually Matter Beyond the Hype — assessing an individual company in the chain
• The Memory Chip Cycle: Why the Same Stocks Boom and Bust on Repeat — the component bottleneck currently capturing the margin
• How to Analyze Semiconductor Stocks: Wafer Starts, Lead Times, and Inventory Cycles — reading supply signals in the compute link
• What is the Magnificent Seven and Are They Still Worth Owning in 2026 — the companies doing the spending
• How to Diversify a Stock Portfolio: Sector Allocation and Correlation Explained — measuring hidden shared exposure
• Concentrated Stock Positions: The Risk Hiding in Your Best Winner — when several holdings are really one bet