Free research
AI Capex Tracker 2026: $695B–$720B in Spending
- Research desk
- Yield Theory Research
- Reviewed
- Evidence
- 5 external references · Method
The 2026 hyperscaler capex number
Direct answer: Microsoft, Alphabet, Amazon, and Meta have outlined $695 billion to $720 billion of combined 2026 capital expenditure. The midpoint is approximately $707.5 billion. It is the largest disclosed spending envelope in this research set, but it is not a comparable pure-AI total.
How much are hyperscalers spending on AI capex in 2026?
The companies do not disclose one comparable AI-capex line. The best public proxy is their combined capital-expenditure guidance, which covers AI infrastructure alongside other servers, data centers, networking, fulfillment, and corporate investment. The defensible conclusion is that hyperscaler infrastructure spending is enormous and still growing; the indefensible conclusion is that every dollar in the combined figure is AI spending.
That figure is useful for scale, but it is not a clean measure of AI spending. Each company defines and reports capital expenditure differently. Amazon's plan covers the whole company, including non-AI investments. Meta includes principal payments on finance leases in its guidance. Microsoft discusses calendar-year spending even though its fiscal year ends in June. Alphabet's range includes servers, data centers, networking, and investments across several businesses.
The honest conclusion is not that these companies will spend exactly $707.5 billion on GPUs. It is that the largest technology platforms have entered an infrastructure cycle large enough to affect semiconductor supply, electricity demand, construction, financing, and reported margins.
Research cutoff: July 30, 2026. This page incorporates Meta's July 29 narrowed outlook and Amazon's July 30 second-quarter update. It will be refreshed after any material guidance change.
2026 capex guidance by company
| Company | 2026 capex guidance | Midpoint used here | What the number includes |
|---|---|---|---|
| Microsoft | About $190bn | $190bn | Calendar-year capital expenditure, including short-lived compute and long-lived data-center assets |
| Alphabet | $175bn to $185bn | $180bn | Servers, data centers, networking, and other property and equipment |
| Amazon | About $200bn | $200bn | Companywide capital expenditure, including AI, AWS, chips, robotics, fulfillment, and satellites |
| Meta | $130bn to $145bn | $137.5bn | Capital expenditure including principal payments on finance leases |
| Combined | $695bn to $720bn | $707.5bn | Not directly comparable and not all AI-specific |
Microsoft: about $190 billion
Microsoft expects to invest roughly $190 billion in calendar 2026 capital expenditure. The company said approximately $25 billion of that total reflects higher component pricing.
In its fiscal third quarter, Microsoft spent $31.9 billion. Roughly two-thirds went to short-lived assets, primarily GPUs and CPUs, while the remainder went to assets expected to support monetization over much longer periods. Microsoft also said customer demand continued to exceed available capacity.
The source is Microsoft's fiscal Q3 2026 earnings call.
Open the Microsoft AI capex profile →
Alphabet: $175 billion to $185 billion
Alphabet expects $175 billion to $185 billion of 2026 capital expenditure, up from $91.4 billion in 2025. Management said the mix should remain roughly similar to 2025, when about 60% of investment went to servers and 40% to data centers and networking equipment.
Alphabet also warned that depreciation and data-center operating costs, including energy, would pressure the income statement as the new infrastructure enters service.
The source is Alphabet's fourth-quarter 2025 earnings call.
Open the Google and Alphabet AI capex profile →
Amazon: about $200 billion
Amazon expects to invest approximately $200 billion across the company in 2026. Management named AI, chips, robotics, and low-earth-orbit satellites among the opportunities behind the plan.
This is the least AI-specific number in the table. Amazon also invests heavily in fulfillment centers, transportation, devices, and other businesses. Investors should not label the full amount as AWS or AI capex.
The source is Amazon's fourth-quarter 2025 results.
Amazon's second-quarter 2026 update kept the cash-flow cost visible: trailing free cash flow moved to a $7.6 billion outflow, driven primarily by a $66.1 billion year-over-year increase in property-and-equipment purchases net of incentives.
Open the Amazon AI capex profile →
Meta: $130 billion to $145 billion
Meta narrowed its 2026 capital-expenditure outlook to $130 billion to $145 billion on July 29, raising the lower end from $125 billion. The measure includes principal payments on finance leases. The spending supports data centers, servers, networking, and the compute required for advertising, recommendation systems, generative AI, and research.
Meta is different from a cloud seller. It can monetize infrastructure indirectly through engagement and advertising improvements rather than charging an external customer for every unit of compute.
The source is Meta's second-quarter 2026 results.
Open the Meta AI capex profile →
Why the totals are not directly comparable
Four accounting differences matter.
- Cash purchases versus finance leases: A company can receive an asset through a lease without the same immediate cash-flow presentation as an outright purchase.
- Fiscal versus calendar year: Microsoft's fiscal calendar does not match the other three companies' reporting years.
- AI versus companywide investment: None of the four provides a perfectly comparable audited "AI capex" line.
- Short-lived versus long-lived assets: GPUs and servers may be depreciated over several years, while buildings and site infrastructure can remain in service for decades.
This is why a single combined number should be presented as an infrastructure-spending envelope, not an accounting identity.
Where the spending flows
The first-order beneficiaries are not limited to accelerator designers.
| Layer | What hyperscalers buy | Useful operating signals |
|---|---|---|
| Compute | GPUs, custom accelerators, CPUs | Unit shipments, utilization, performance per dollar |
| Memory | HBM, server DRAM, storage | Capacity, bandwidth, pricing, qualification, yield |
| Networking | Switches, optical links, interconnects | Backlog, port speed, cluster scale |
| Data-center systems | Cooling, power distribution, racks | Orders, backlog, deployment time |
| Grid and generation | Transformers, substations, transmission, power | Interconnection queues, contracted capacity, energization dates |
| Construction and financing | Campuses, leases, debt, private credit | Commitments, borrowing, construction in progress |
Read our supporting guides to AI data-center power demand, high-bandwidth memory, and custom AI chips versus GPUs.
If an accelerator count is disclosed, the free HBM demand calculator translates it into installed memory, aggregate peak bandwidth, and raw model-weight capacity without pretending that those physical quantities are a supplier-revenue forecast.
The five numbers investors should track
1. Capex growth versus cloud and AI revenue growth
Spending can create value when productive capacity comes online into strong demand. It becomes more concerning when capital investment rises while cloud growth, advertising monetization, or contracted demand weakens.
2. Gross margin
Revenue growth alone does not prove attractive economics. AI products can consume far more infrastructure per user than traditional software. Gross margin shows whether pricing and efficiency are keeping pace with the cost of serving demand.
3. Depreciation
Cash leaves before the full expense reaches earnings. New servers, network equipment, and buildings begin depreciating only after they are placed in service. Yield Theory's member research explains why the AI depreciation clock arrives late.
4. Free cash flow after capex
Operating cash flow can rise while free cash flow falls if infrastructure investment grows faster. The relevant question is whether the current build produces enough future operating cash to justify the outlay.
5. Utilization and supply constraints
Management comments about capacity constraints, deployment times, customer backlog, and utilization help distinguish productive scarcity from speculative overbuilding.
What would weaken the AI capex thesis?
The combined spending plan is not automatically bullish for every company in the supply chain. Warning signals would include:
- Cloud or AI revenue growth slowing while capacity comes online
- Gross margins deteriorating without a credible efficiency bridge
- Customers renegotiating or delaying large commitments
- Construction in progress growing without corresponding utilization
- Useful-life extensions masking faster technological obsolescence
- More debt financing without recovery in free cash flow
- Power or permitting delays leaving expensive equipment idle
Bottom line
The 2026 capex cycle is large enough to reshape several industries, but the total is not a forecast of guaranteed AI demand or investment returns. Treat the guidance as a map of where capital is being committed. Then test whether revenue, utilization, margins, and cash flow follow.
This tracker will be updated when the companies change their guidance or report material new infrastructure data.
This article is for educational purposes only and is not investment advice. Company guidance is forward-looking and can change. Verify all figures against the linked primary sources before making investment decisions.
Continue the research
Know where the money is going. Decide where to invest.
The companies we would own, the catalysts that can move them, the risks the market is missing, and the breakpoint that changes the call. New evidence is checked against the written thesis, so every update tells you whether the call is stronger, weaker, or broken.