The AI Capex Boom Is Turning Into a Depreciation and Free-Cash-Flow Test

Bottom line

The capex story has moved from optional to contractual.

Two years ago, AI spend was still being sold as a growth option. In mid-2026, it looks more like a contracted industrial program: server purchases, data-center shells, networking, power, and lease obligations that arrive before the monetization does.

That shift changes how you should underwrite the megacap AI complex. The right comparison is no longer software margin alone. It is a hybrid of telecom-style buildout, utility-like depreciation, and platform economics that may take years to normalize.

My view: the AI winners are now the companies that can fund infrastructure without destroying free cash flow. The losers are the names that only look good while capex is still outrunning depreciation.

What the numbers say

The spend is broad, large, and still accelerating.

2026 AI infrastructure spend scale

This chart mixes 2026 guidance with Microsoft's FY2025 actual additions to property and equipment. It is intentionally a scale comparison, not a like-for-like quarterly series.

Unidad: USD billions

Amazon

2026 capex plan

200

Alphabet

2026 guidance midpoint

185

Meta

2026 guidance midpoint

135

Microsoft

FY2025 additions to PPE

64.6

What each company is really saying
Company	Published signal	Interpretation
Alphabet	Q1 capex was $35.7B; full-year 2026 guidance is $180-190B; 60% of the quarter's spend was servers.	Alphabet is still building the compute stack aggressively, but it is also showing that spending is tied to technical infrastructure, not vanity projects.
Amazon	Management said it is not investing about $200B in capex in 2026 on a hunch and that much of AWS capex will monetize in 2027-2028.	Amazon is explicitly framing capex as future FCF production, not just current expense.
Meta	2026 capex guidance moved to $125-145B, up from $115-135B, because of higher component pricing and future data-center capacity.	Meta is admitting the buildout is getting more expensive even before the next monetization leg arrives.
Microsoft	FY2025 additions to property and equipment were $64.551B; depreciation expense was $22.0B; committed datacenter construction obligations were $32.1B.	Microsoft is already carrying a much bigger depreciation base, which means the cash conversion test is going to get harder, not easier.

Cash flow and depreciation

The market is underestimating how fast the expense side will catch up.

Alphabet said Q1 2026 free cash flow was $10.1 billion and trailing-12-month free cash flow was $64.4 billion. Meta reported Q1 2026 free cash flow of $12.39 billion and cash from operating activities of $32.23 billion. Microsoft's FY2025 annual report shows that the company used $72.6 billion in investing cash flow and that depreciation expense reached $22.0 billion.

Amazon is the clearest version of the same argument from the other side. Its shareholder letter says the company expects much of the AWS capex spent in 2026 to monetize in 2027-2028. That is a good business statement, but it also means the market is still financing a gap between cash outflow and monetization.

Cash-flow lens
Company	Free-cash-flow signal	Why it matters
Alphabet	$10.1B in Q1 and $64.4B TTM	Still very cash generative, but the reinvestment bar is rising.
Meta	$12.39B in Q1 FCF	CF is positive, but component pricing and capacity spending are dragging on flexibility.
Amazon	AWS capex is being pre-funded for 2027-2028 monetization	The payback window is long enough that execution risk matters.
Microsoft	$22.0B depreciation in FY2025	More assets now have to earn back their cost every quarter.

The main analytical mistake is to treat capex as if it stays in the future forever. In reality, capex becomes depreciation, and depreciation becomes a margin ceiling unless revenue per watt keeps rising.

Who wins

The winners are the firms that control utilization, silicon, and customer commitments.

Cloud platforms with sticky demand can use scale to lower unit economics.
Custom silicon reduces dependence on external accelerators and can compress inference cost.
Companies with committed demand can front-load buildout without taking pure speculation risk.
The market should reward high utilization and disciplined deployment, not just the biggest absolute spend number.

Where the pressure lands first
Layer	Likely effect	Investor lens
Server vendors	Higher unit volume, but more pricing pressure	Revenue can rise even when hardware economics get tighter.
GPU and networking suppliers	Strong demand, but more procurement scrutiny	The better operators will lock in the best mix and terms.
Utilities and power equipment	More load and larger interconnection needs	Energy becomes a gating factor, not a background cost.
Public cloud providers	Higher depreciation and lease obligations	FCF quality matters more than headline growth.