What it measures
Hyperscaler AI capex aggregates the announced capital expenditure from the four largest cloud providers — Microsoft Azure, Alphabet/Google Cloud, Meta, and Amazon AWS — for calendar year 2025. Reuters Breakingviews, drawing on individual company earnings disclosures and guidance, put the combined figure at approximately $320B.
This is the best available public proxy for infrastructure commitment to AI. Unlike revenue metrics (which measure what customers pay) or market-size estimates (which include services and software), capex measures money actually committed to servers, networking, power infrastructure, and data center construction. It is upstream, hard to fake, and large enough to have macroeconomic consequences.
Why it matters
When four companies collectively deploy $320B in a single year into physical infrastructure, the spending becomes a macro event. Construction labor markets tighten. Power utilities must plan grid capacity years in advance. Equipment manufacturers — GPUs, networking switches, fiber — face demand spikes they cannot quickly satisfy. Commodity prices for copper, rare earth metals, and specialty cooling fluids respond.
The $320B figure also reveals the competitive dynamic between the four companies. None of them can afford to invest significantly less than the others: a capex gap today translates into a compute gap tomorrow, which translates into a model capability gap the year after. The result is a race to spend that has taken on a self-reinforcing character.
Microsoft: ~$80B | Alphabet: ~$75B | Meta: ~$65B | Amazon: ~$100B+. Amazon's figure carries the most uncertainty — some analyses of reported and guided figures suggest 2025 capex was closer to $131B when all data center and infrastructure categories are included. The Reuters Breakingviews aggregate of ~$320B is the most-cited combined figure.
What it misses
Capex commitment is not the same as revenue generation. The $320B being deployed in 2025 will not fully translate into AI services revenue for two to four years — the typical time horizon between groundbreaking and operational data center capacity. Several risks compound this lag:
- Overbuild risk: If demand grows more slowly than models project, hyperscalers face stranded assets and margin pressure. The telecom sector's fiber overbuild in the late 1990s is the historical analogue most often cited by bears on hyperscaler capex.
- Concentration risk: The $320B is controlled by four companies. Geopolitical restrictions, antitrust action, or a single catastrophic infrastructure failure could disrupt supply in ways that a distributed capex base would not.
- Non-hyperscaler spend excluded: The figure does not include independent data center operators, Chinese cloud providers (Alibaba, Tencent, Baidu, ByteDance), sovereign AI infrastructure programs, or enterprise on-premises AI hardware investment.
- Announced vs. actual: Capital expenditure guidance can be revised downward. Treat the $320B as a commitment signal, not a certainty.
Collectively, the four hyperscalers generated approximately $150B in cloud revenue in 2025. They are deploying more than twice that in capex. The bet is that AI will expand the addressable market dramatically enough to justify this ratio. Whether that expansion materializes on the timeline the models assume is the central uncertainty in the AI infrastructure story.
What happens next
$320.0B in a single year from four companies is a number without historical precedent in technology infrastructure investment. The nearest analogue — the telecom fiber buildout of the late 1990s — ended in a bust when revenue assumptions proved too optimistic. The key difference: AI has demonstrated real, growing demand (NVIDIA revenue confirms it), where the telecom buildout preceded demand. But the 27:1 ratio of hyperscaler capex to consumer AI subscription revenue still demands explanation — and that explanation will arrive in 2026–2028 earnings.
Pros — Benefits
- Best public proxy for infrastructure commitment — audited capex, not forecast
- At $320B, creates macro spillovers: construction labor, power, equipment manufacturing
- Competitive dynamics mean all four companies must invest at similar scale — durable commitment
- Hard infrastructure creates durable competitive moat: capacity built today generates cloud revenue for 10+ years
Cons — Risks
- Revenue from 2025 capex may not materialize for 2–4 years — overbuild risk is real
- Concentrated in four companies — geopolitical or regulatory risk is a single-point-of-failure scenario
- Does not include Chinese hyperscalers (Alibaba, Tencent, Baidu) — global total is higher
- Announced guidance can be revised; treat as commitment signal not certainty
What to watch for
- Quarterly capex guidance from the four hyperscalers (Microsoft, Alphabet, Meta, Amazon earnings calls)
- Power utility demand growth in data center-dense regions (PJM, ERCOT)
- Data center REIT occupancy and pricing (Equinix, Digital Realty) — proxy for hyperscaler demand
- NVIDIA, AMD, Broadcom data center revenue — upstream of hyperscaler capex commitment
- Hyperscaler cloud AI revenue growth: must accelerate to justify capex trajectory
Most critical tipping point
What you can do
- Track quarterly earnings from the four hyperscalers — capex revisions are the earliest signal of demand expectations changing
- Follow power utility earnings and data center REIT filings — downstream indicators of capex flowing through
- Monitor construction labor market tightness in Virginia, Texas, and Phoenix — hyperscaler geography signal
- If your business depends on cloud AI services, lock in pricing contracts before capacity tightens further
- Evaluate on-premises vs cloud cost trajectories — hyperscaler capex will eventually lower cloud AI prices
- Track hyperscaler AI service announcements: capex precedes product launches by 18–24 months
- Develop power grid capacity planning frameworks that account for hyperscaler demand projections
- Require hyperscaler environmental impact disclosures: water, land, and power consumption at the project level
- Fund antitrust research on concentration dynamics when four companies control AI infrastructure at this scale
Data & methodology
- Source
- Reuters Breakingviews; individual company earnings disclosures
- Companies
- Microsoft Azure + Alphabet/GCP + Meta + Amazon AWS
- Excludes
- Chinese hyperscalers, independent data centers, enterprise on-premises AI hardware
- Amazon note
- Amazon 2025 capex may exceed $100B; some analyses put the figure at $131B
- Update cadence
- Quarterly earnings; annual summary in Reuters Breakingviews
- Dashboard anchor
- AI Spending section on dashboard