Google's 2025 capex came in north of $75B — up 43% from $52.5B in 2024, and revised toward $85B mid-year as AI demand outran capacity. That's the short answer. The longer answer is more interesting.
Almost none of that increase is offices, vehicles, or the moonshots people associate with Alphabet. It is data centers, servers, and custom silicon — the physical plumbing of AI. The question worth asking isn't whether $75B is a big number. It obviously is. It's what a company gets for spending it, and whether the revenue compounds faster than the depreciation it creates.
Google capex 2025 AI: the headline numbers
Google's 2025 AI capex started at a guided $75B and was raised toward $85B during the year, almost entirely directed at AI infrastructure: data centers, servers, and custom TPU chips. That is a 43% jump over the $52.5B Alphabet spent in 2024, which itself was up roughly 63% over 2023. In three years Alphabet's annual capital spend more than doubled, and the slope keeps steepening.
For context, Alphabet generated about $350B in revenue in 2024 and was tracking well above that in 2025. Spending $75-85B of capex against that base means roughly one in every five dollars of revenue is being plowed back into physical infrastructure — a ratio that looks more like a utility or a telecom than a software company. That shift is the whole story.
Google capex 2025 by quarter
Capex didn't arrive evenly. It ramped through the year as data centers came online and chip orders cleared. Here is the approximate quarterly shape against 2024, based on Alphabet's reported and guided figures.
| Period | 2024 Capex | 2025 Capex | YoY Change |
|---|---|---|---|
| Q1 | $12.0B | $17.2B | +43% |
| Q2 | $13.2B | $18.5B | +40% |
| Q3 | $13.1B | $20.0B | +53% |
| Q4 | $14.2B | $22.0B+ | +55% |
| Full year | $52.5B | $75–85B | +43–62% |
| 2023 (ref) | $32.3B | — | — |
Quarterly 2025 figures are approximate, reflecting reported actuals early in the year and management guidance later. The direction is unambiguous: every quarter ran 40-55% above the prior year, and the back half accelerated rather than tapered. Track the live numbers on the Big Tech Earnings dashboard.
Where the AI capex actually goes
"$75B on AI" is a slogan, not a budget. The money splits into a few buckets, and the proportions tell you what Alphabet believes about the next five years.
Servers & TPU silicon (~55-65%)
The single biggest line. Google designs its own TPU accelerators, so a large share of compute avoids Nvidia's 70%+ gross margin.
Data center construction (~25-30%)
Land, shells, power, and cooling. Lead times stretch 18-36 months, which is why capex is committed years before revenue lands.
Networking & fiber (~5-10%)
AI training clusters need enormous east-west bandwidth; interconnect is a rising share of every dollar.
Power & energy deals (rising)
Google signed nuclear and renewable PPAs in 2025 — power availability, not chips, is becoming the real constraint.
The TPU strategy is the part most analysts underweight. Microsoft and Amazon buy the bulk of their accelerators from Nvidia, paying a margin that runs well above 70% gross. Google designs TPUs in-house and fabricates them through partners, which means a meaningful slice of its $75B+ doesn't flow to Nvidia at all. On a like-for-like compute basis, Google's effective cost per FLOP can run materially below a Nvidia-only buyer's. That's a structural cost advantage that compounds every year capex grows.
How Google capex compares to other hyperscalers in 2025
Google isn't spending in a vacuum. All four US hyperscalers ramped simultaneously, and the combined number crossed $300B for the first time. The comparison matters because compute is partly a relative game — falling behind on capacity means losing AI workloads to whoever has GPUs available.
| Company | 2025 Capex | 2024 Capex | Chip Strategy |
|---|---|---|---|
| Amazon (AWS) | $100B+ | $77B | Nvidia + custom Trainium/Inferentia |
| Microsoft | ~$80B | $55.7B | Mostly Nvidia + Maia in-house |
| Google (Alphabet) | $75–85B | $52.5B | Custom TPU + some Nvidia |
| Meta | $65–72B | $39.2B | Nvidia + MTIA in-house |
| Oracle | ~$25B | $8.7B | Mostly Nvidia (OCI) |
| Four-hyperscaler total | $320B+ | $224B | — |
Amazon leads on raw dollars, but Google's position is arguably the most efficient because of the TPU offset. Meta's capex nearly doubled year over year — the steepest acceleration of the group — while Oracle's tripled off a small base as it landed AI hosting contracts. The takeaway: this is no longer a question of whether to spend. Every player concluded that being short on compute is more dangerous than overbuilding.
What Alphabet is actually building with it
Strip away the accounting and the $75B+ funds four concrete things, each tied to a revenue line.
Revenue-generating buildout
- ✓ Gemini training & inference compute
- ✓ Google Cloud capacity (growing 28-35% YoY)
- ✓ TPU capacity it rents to Cloud customers
- ✓ AI-driven Search and ad ranking infrastructure
The cost it creates
- ✕ Depreciation on $75B+ of assets over ~6 years
- ✕ Power and operating costs that scale with usage
- ✕ Margin drag until AI revenue catches the spend
- ✕ Stranded-asset risk if a chip generation lags
Google Cloud is the proof point. It turned its first full-year operating profit in 2023 and expanded margins through 2025, with revenue growing roughly 28-35% year over year and a backlog measured in the tens of billions. That backlog is the justification for the capex — it's contracted demand the company can't serve without more data centers. When the spend funds a backlog, it's investment. When it funds speculative capacity, it's a gamble. Alphabet's leadership has been explicit that they'd rather risk overbuilding than miss the demand.
What it means for founders and investors
The hyperscaler capex wave reshapes the ground startups operate on. Three implications stand out.
Compute is getting cheaper to rent, not buy
Google, Amazon, and Microsoft are absorbing the capital cost so startups don't have to. Renting TPU or GPU capacity beats raising a Series A to buy your own cluster — the unit economics favor renting until you're at serious scale.
The infrastructure layer is consolidating
When four companies control $320B+ of annual compute spend, the picks-and-shovels opportunities narrow to what they can't or won't build: vertical data, specialized inference, and tooling that sits above the cloud.
Depreciation is the next earnings story
Watch operating margins, not just capex headlines. As $75B+ of 2025 assets begin depreciating, the question for public investors shifts from 'how much are they spending' to 'is AI revenue outrunning the depreciation it created.'
You can see the same dynamic across the sector on the AI Spending dashboard and in how the market prices it on AI Valuations. The capex race is the clearest signal yet that the incumbents believe AI is a generational platform shift — they don't spend $75B a year hedging.
$75B isn't the bet. The TPU offset and the Cloud backlog are.
Google is the only hyperscaler spending at this scale while owning the chip — and that's the edge that compounds.
Track hyperscaler capex and AI spending on the Big Tech Earnings dashboard at Value Add VC. Originally published in the Trace Cohen newsletter.