The AI CapEx Arms Race: Who Is Spending What

The AI capex arms race in 2026 is the largest synchronized capital build in corporate history. Four hyperscalers, Amazon, Microsoft, Alphabet, and Meta, have guided to a combined figure near $725 billion this year, up roughly 77% from 2025 (analyst consensus per CreditSights). That is not a budget. It is a coordinated bet that the cost of being short on compute exceeds the cost of overbuilding.

The headline number is the easy part. The story the filings actually tell sits one line down: depreciation. GPUs lose their economic value in three to four years, but the companies buying them book five-to-six-year useful lives. That gap quietly smooths near-term earnings, and it is the variable that decides whether this build looks like genius or excess in 2027.

This piece reads the spend through the disclosures, not the keynotes. It lays out who is spending what, why none of them can stop, and where the accounting hides the real risk. The framing is analytical: how to locate the pressure in the build, not what to do about any stock.

---

Key takeaways

• The spend is the demand signal. Amazon, Microsoft, Alphabet, and Meta have guided to roughly $725 billion in combined 2026 capex, up about 77% from ~$410 billion in 2025 (CreditSights consensus). Around 75% of it, near $450 billion, funds AI-specific infrastructure.
• Two companies lead in absolute terms. Amazon at ~$200B (Q4 FY2025 call, Feb 5, 2026) and Microsoft at ~$190B (Q3 FY2026, Apr 29, 2026) top the list, with Alphabet at $175-185B and Meta at $115-135B.
• Depreciation is the real story. GPUs run a three-to-four-year economic life but are depreciated over five to six years (SiliconANGLE/theCUBE Research, Nov 2025), spreading cost thin and flattering current margin.
• The accounting is already moving. Amazon shortened server and networking life back to five years effective January 1, 2025, and Microsoft’s gross margin fell to 67.6% partly on accelerated depreciation.
• Capex is outrunning revenue. Capex-to-sales is tracking near 54% at Meta, 47% at Microsoft, 46% at Alphabet (Forbes, June 2, 2026), all historically elevated. Revenue has to catch the curve.

---

Why are hyperscalers spending $725 billion in the AI capex arms race 2026?

Because the risk is asymmetric. Being short on compute when demand spikes is catastrophic, costing lost revenue, share, and positioning. Overbuilding only depreciates over five to six years. So the four hyperscalers build ahead of demand certainty rather than chase it, treating excess capacity as the cheaper mistake.

That asymmetry is the entire logic of the build. A hyperscaler that cannot serve an AI workload loses the customer to a rival who can, and likely loses the rest of that customer’s cloud spend with it. A hyperscaler that builds too much carries depreciation on idle racks for a few years. One of those failures is permanent; the other is temporary and amortizable.

The combined guidance is unprecedented. Per analyst consensus aggregated by CreditSights, 2026 capex across Amazon, Microsoft, Alphabet, and Meta (plus Oracle in the wider basket) lands near $725 billion, up about 77% from roughly $410 billion in 2025. CreditSights and other analysts put AI-related infrastructure at roughly 75% of that, near $450 billion of AI-specific spend.

This is the strongest demand read the public disclosures produce. It is also why the spend reads as a distribution war as much as an infrastructure one: owning the capacity is owning the surface every model and app runs on, the same logic dissected in Google’s AI strategy as a distribution war.

---

The CapEx Scorecard: who is spending what in 2026

Here is the central framework, which I will call the CapEx Scorecard. It is the analytical asset this piece is built around, and it is meant to be citeable on its own: a one-screen read of the four hyperscalers, their 2026 guidance, the year-over-year move, where the money lands, and the depreciation overhang each one is taking on. How that spend translates into competitive position across the big three clouds is the subject of AWS vs Azure vs Google Cloud.

Hyperscaler	2026 capex (guidance)	YoY move	Where it lands	Depreciation overhang
Amazon	~$200B	Up sharply from $131.8B (2025)	AWS data centers, chips, networking; AI for internal + third-party workloads	Server/networking life shortened back to 5 years, effective Jan 1, 2025
Microsoft	~$190B	Up ~61%	Azure capacity, GPUs; ~$25B of guidance tied to component price inflation	Accelerated depreciation already in COGS; gross margin fell to 67.6%
Alphabet	$175-185B	Up from ~$91B (2025)	Google Cloud, TPUs + GPUs, internal AI	Heavy GPU/TPU base on 5-6 year schedules
Meta	$115-135B	Up from ~$72B (2025)	Internal recommendation/ranking AI, training clusters	Largest capex-to-sales ratio (~54%) of the four

The CapEx Scorecard. Guidance figures: Amazon Q4 FY2025 call (Feb 5, 2026); Microsoft Q3 FY2026 (Apr 29, 2026); Alphabet Q1 2026 guidance; Meta Q4 2025 guidance. Combined consensus ~$725B per CreditSights, “Tech: Raising Hyperscaler Capex 2026 Estimates.” The framework itself is original analysis.

Read the scorecard left to right and a pattern appears. Three of the four roughly doubled or more their prior-year spend. Amazon and Microsoft lead in absolute dollars; Meta carries the heaviest capex-to-sales weight. And every one of them is taking on a depreciation schedule that will define its margin for years, regardless of how utilization tracks.

One detail inside Microsoft’s number deserves a flag. Management attributed roughly $25 billion of the $190 billion guidance to higher component prices, primarily memory and storage, rather than added capacity (Q3 FY2026 commentary, CFO Amy Hood). Part of the headline spend is supply-chain inflation, not incremental compute. That spend can be cut without cutting capacity if prices normalize, which is exactly what softening demand would look like in the filings.

---

The combined number: $725B, and where it actually goes

The aggregate is easy to quote and easy to misread. The $725 billion figure is a consensus estimate, revised up repeatedly through late 2025 and early 2026 (from ~$600B in November 2025 to ~$620B in January 2026 to ~$725B), not a single audited disclosure. It is built from individual company guidance, each of which is itself a range.

Of that total, CreditSights and other analysts attribute roughly 75% to AI-related infrastructure, near $450 billion. The rest is the maintenance-and-growth capex hyperscalers would spend regardless: general cloud capacity, real estate, networking refresh. The AI slice is the part growing at triple-digit rates.

Two facts about the destination matter more than the headline. First, a large share flows straight to one supplier layer, the chip vendors, which is why the same dollars that compress hyperscaler margin expand semiconductor margin. The mechanics of that transfer are the whole subject of the AI Infrastructure Market Map. Second, a meaningful portion lands in fixed assets that cannot be redeployed: substations, transmission interconnects, cooling, and long-term power-purchase agreements. Those amortize on schedules that often run longer than the chips inside.

The combined-capex framing is useful as a demand signal. It is misleading as a precision figure. Treat $725 billion as a directionally correct consensus, not a number you can tie to a single filing line.

---

Why depreciation is the real story

The headline is capex. The earnings risk is depreciation. This is the methodology that separates the two, and it is the part most coverage skips.

When a hyperscaler buys a GPU, the cash leaves immediately but the cost hits the income statement gradually, spread across the asset’s assumed useful life. Assume a longer life and each quarter’s depreciation charge is smaller, which makes current margin look healthier. Assume a shorter life and the charge is larger, compressing margin now.

Here is the gap. Industry analysis puts a GPU’s true economic useful life near three to four years, because a new generation obsoletes the prior one and the economics of running old silicon collapse (SiliconANGLE/theCUBE Research, November 2025; corroborated across multiple independent analyses). But hyperscalers depreciate these assets over five to six years on the financial statements. That spread of one-and-a-half to two-and-a-half years understates the true cost of the build in every quarter it persists.

The accounting has started catching up to the physics. Amazon moved its server and networking useful life back to five years from six, effective January 1, 2025, an explicit acknowledgment that AI hardware obsolesces faster. Microsoft’s gross margin fell to 67.6% in its most recent quarter, down from 68.7% and its lowest since 2022, with accelerated depreciation cited as a driver.

Methodology: estimating the depreciation overhang

• Inputs: combined 2023-2026 hyperscaler capex; reported useful-life assumptions of five to six years; independent estimates of true economic life at three to four years (SiliconANGLE/theCUBE Research).
• Assumption: the gap between booked life and economic life means depreciation expense is recognized more slowly than the assets actually lose value, deferring cost into later periods.
• The illustrative scale: across the heavy 2023-2026 capex years, the cumulative understatement of depreciation has been estimated in the $200 billion-plus range industry-wide. This is a reported analyst estimate built on assumed lives, not a figure disclosed in any single filing; treat it as a directional flag, not a measured number.
• What this misses: some accelerators do retain useful secondary lives in inference or lower-tier workloads, which would extend the effective life and shrink the gap. The filings do not disclose retirement curves, so the true overhang is not measurable from public data alone.

The point is not a precise dollar figure. It is the mechanism: the longer the assumed life, the better today’s margin looks, and the larger the catch-up when reality intervenes. Margin behaves much the way it does across the SaaS stack, where the cost structure below the revenue line quietly sets the ceiling, the argument in why gross margin is destiny in SaaS.

---

Can AI revenue catch the spend curve?

Not yet. Capex is currently growing far faster than the AI revenue meant to absorb it. Per Forbes (June 2, 2026), capex-to-sales is tracking near 54% at Meta, 47% at Microsoft, and 46% at Alphabet, all historically elevated levels for these businesses.

Those ratios are the heart of the absorption question. A company spending nearly half its revenue on capital assets is making a bet that the revenue base will be much larger by the time the depreciation lands. If revenue accelerates, the ratio normalizes and the build pays for itself. If revenue merely grows at today’s pace, the depreciation arrives against a base too small to carry it, and margin compresses.

The mismatch is structural for now. Capital spending growth across the group is estimated to have run above 50% in 2026. Several analysts expect that growth to decelerate sharply, toward roughly 13% in 2027, as the build matures (this 2027 deceleration is a forward analyst projection, not a guided figure, and should be read as illustrative of trajectory rather than a committed number).

A decelerating capex growth rate is not the same as falling capex. It means the absolute dollars stay enormous while the year-over-year increase shrinks. The depreciation from the 2025-2026 build keeps flowing through the income statement either way. The revenue has to show up to meet it. The same cloud-economics reset, read at the segment level, is the subject of AWS margin pressure and the cloud reset.

---

The accounting question: five-year lives for three-year assets

Why book a five-to-six-year life for an asset that loses its edge in three? Because the longer schedule smooths reported earnings, and because, between 2021 and 2023, extending lives was a defensible response to genuinely longer-lasting general-purpose servers.

The history matters. Hyperscalers extended server useful life from three-to-four years toward six years across 2021-2023, a move analysts estimate deferred roughly $18 billion in annual depreciation expense industry-wide (Breaking Into Wall Street analysis; a reported estimate, not a single disclosure). At the time, the logic held: cloud servers genuinely lasted longer as workloads got more efficient.

AI hardware broke that logic. Frontier-model training churns through accelerator generations far faster than general compute, and the gap between a current and a prior-generation GPU is large enough to strand the old hardware economically. The asset that justified six-year lives is not the asset filling the new data centers.

That is why Amazon’s reversal to five years, effective January 1, 2025, is a tell worth watching across the group. When a company shortens a useful-life assumption, it is conceding that its assets wear out faster than its prior accounting implied, and it is accepting higher depreciation charges as the price of that honesty. The question for every hyperscaler is whether five years is still too generous for a GPU in a training cluster.

---

What operators should take from this

The scorecard does not tell you which hyperscaler wins. It tells you that the binding constraint is not the cash spent, but the depreciation booked against it. Here is how to act on that if you build software rather than data centers.

1. Watch capex-to-revenue, not headline capex. A big number is noise; the ratio is signal. When capital spend climbs toward half of sales, the company is betting hard on future revenue. Track whether revenue is closing the gap or the gap is widening.
2. Read the useful-life footnote. Depreciation-schedule changes are buried in the filings and they move margin directly. A shortened useful life is a quiet admission that earnings were flattered before. Read the footnote, not the headline.
3. Separate price inflation from real capacity. Microsoft attributed ~$25B of its guidance to component prices, not added compute. When you read any capex number, ask how much is volume and how much is the supply chain charging more. They have opposite implications.
4. Match your cost commitment to your revenue visibility. Amortization punishes whoever commits capital against demand that has not arrived. Commit fixed cost only where you can defend the utilization that pays it down, the same unit-economics discipline that separates usage-based from seat-based pricing.
5. Treat the supplier layer as the place the spend lands first. The hyperscalers compress their own margin to fund the chip layer’s. If you are mapping where pricing power sits in any supply chain, find the scarce input and follow it.

As a small, illustrative analog from running an AI feature inside a product like PDF9to5: committing to a fixed block of reserved compute is the operator-scale version of the hyperscaler bet. Reserve too little and you throttle at peak; reserve too much and you depreciate idle capacity against revenue that has not materialized. The product does not change. Only the match between your cost commitment and your demand curve does.

---

The bear case: what the skeptics get right

A build this large deserves the strongest argument against it. The bear case is not that any single number is wrong. It is that the entire structure rests on a demand forecast that has to keep materializing, and that the accounting is currently hiding how exposed that forecast is.

Start with the depreciation gap. If GPUs really do wear out economically in three to four years and are booked over five to six, then current margins across the group are overstated, and the correction is not optional. It arrives mechanically as the assets age. Amazon’s move back to five years is the first crack; a move toward four would compress margins across the sector at once. The bear reads the $200 billion-plus cumulative-overhang estimate as deferred pain, not avoided pain.

Then the absorption problem. Capex at 46-54% of sales only works if revenue accelerates dramatically. If AI workload adoption settles below the implied trajectory, the hyperscalers are left with fixed depreciation on underutilized capacity. New capex can be cut in a quarter; depreciation already in the ground cannot. That asymmetry, fast-to-cut spend against slow-to-clear depreciation, is precisely where a demand miss turns into multi-year margin damage.

And there is the concentration objection. The same supplier dependence that makes the chip layer rich makes the whole build fragile to one disruption, a single point of failure that mirrors the customer-concentration risk every analyst already checks on a single filing.

Here is the honest weighing. The bear case is right that the depreciation overhang is real, that absorption is unproven, and that a demand miss would be expensive and slow to unwind. What it has to assume is that demand misses materially, and the strongest counter is the asymmetry the hyperscalers themselves cite: the cost of being short on compute in a real AI demand wave is worse than the cost of carrying idle capacity for a few years. The bear case is a forecast that demand disappoints. The build is a bet that it does not. Both are statements about a curve that has not finished forming.

---

Where this breaks: the stranded-asset scenario

A credible map names where it fails. The clearest failure mode is utilization stalling while depreciation keeps landing.

Picture the mechanism with numbers used only to show the shape, not drawn from any filing. A hyperscaler builds capacity assuming sustained 40% cloud growth, and growth instead settles toward 20%. The revenue base it depreciates against is smaller than planned, but the five-to-six-year depreciation schedule does not shrink to match. The gap shows up directly in gross margin and persists for the life of the assets. That is the debt-service shape of the risk: not literal default, but a multi-year stretch where committed spend outruns the revenue it was built to serve.

The second crack is the supply chain itself. The accelerator supply narrows to a dominant design vendor, a small set of advanced foundries, and a thin layer of memory and packaging suppliers. That concentration is what hands the chip layer its pricing power. It is also what makes the entire stack above it fragile to one shock: a foundry constraint, a packaging shortage, or an export-control change can starve every layer at once, with no second source at scale.

The honest caveat for the whole scorecard: the public filings do not separate revenue-serving capacity from internal capacity, nor power and real estate from server capex, nor disclose accelerator retirement curves. So per-dollar return on this spend, and the precise depreciation overhang, cannot be cleanly attributed from public data alone. What the filings do show is the direction: enormous fixed cost, committed now, against a demand curve that has to keep compounding to justify it.

---

Analysis, not investment advice. Figures are drawn from disclosed capex guidance and earnings commentary cited inline by company and period (Amazon Q4 FY2025; Microsoft Q3 FY2026; Alphabet Q1 2026; Meta Q4 2025), aggregated by CreditSights, with depreciation analysis from SiliconANGLE/theCUBE Research and others. Combined and overhang figures are reported analyst estimates, not single audited disclosures. Frameworks here, including the CapEx Scorecard, are for understanding business structure and tradeoffs, not for making buy or sell decisions.

Want the full toolkit for reading filings like this, the CapEx Scorecard, the depreciation-overhang methodology, and the capex-to-revenue worksheet used above? It’s in the Tech Business Analysis Playbook.