Amazon, Alphabet, Microsoft and Meta now point to roughly $733 billion of capital spending in 2026. The number is real. The interpretation is usually too loose.
That total mixes land, power, buildings, cooling, networking, chips, servers, leases and non-AI projects. It also mixes assets already earning revenue with sites that may still be waiting on an interconnection, a transformer or a completed data hall.
The investable question is how quickly each dollar moves from approved budget to energized, utilized compute. That conversion rate decides which suppliers get paid, when cloud revenue shows up and whether the spending earns an acceptable return.
Start with the denominator
The current headline is built from four large guidance figures.
| Company | Approx. 2026 capex |
|---|---|
| Amazon | $220B |
| Alphabet | $200B |
| Microsoft | $175B |
| Meta | $138B |
| Total | $733B |
This is a useful measure of strategic intent. It is a poor proxy for any single supplier's demand.
A dollar reserved for land is not accelerator revenue. A signed power agreement is not an energized data hall. An installed GPU is not automatically monetized cloud capacity.
The market keeps collapsing those distinctions into one number.
One number, four different baskets
Alphabet has indicated a mix of roughly 60% machines and servers and 40% data centers and networking.
Microsoft's recent quarterly mix was roughly two-thirds shorter-lived assets, primarily CPUs and GPUs, with the balance in longer-lived infrastructure. That was a quarterly snapshot, not a full-year allocation.
Meta identifies servers, data centers and networking, but does not publish a precise numerical mix.
Amazon's cash capex includes AWS land, power, buildings, chips, servers and networking. It also includes robotics and satellites.
Some related investment may never appear in hyperscaler capex at all. Utilities can fund generation and grid upgrades. Developers can finance shells. Leases and joint ventures can move physical infrastructure outside the cleanest version of the headline.
The conversion chain
The physical system has several gates in between.
Capex guidance sits at the top of this funnel. Revenue-producing compute sits at the bottom.
The distance between them is where timing risk lives.
The hidden lag
A project can look advanced from one angle and remain years from full monetization.
A utility can sign a large load agreement before power flows.
A data center developer can book capacity before the shell is complete.
An equipment supplier can report backlog before the project is ready to accept delivery.
A chip vendor can ship into a cluster that is still waiting on networking, power or final commissioning.
A cloud provider can place an asset into service before utilization reaches the level needed to earn an attractive return.
The same project can therefore appear in different parts of the market at different times. It may be a power contract today, construction revenue next year, hardware revenue after that and cloud depreciation for years to come.
Why record spending can coexist with shortages
Record capex does not prove supply is abundant. It can prove the shortage is expensive.
If demand grows faster than the conversion chain clears, every stage can be busy while usable capacity remains scarce.
Power can be contracted but not energized.
A building can be standing but not fitted out.
Accelerators can be ordered but not networked.
A cluster can be energized but already spoken for.
This is how cloud providers can raise spending and still tell customers that capacity is tight.
Who gets paid, and when
| Stage | Primary activity | What confirms conversion |
|---|---|---|
| Site and power | Land, generation, utility service, electrical equipment | Signed contracts, interconnection milestones, firm backlog |
| Construction | Data halls, substations, cooling, fiber, power distribution | Project starts, equipment delivery, backlog conversion |
| Equipment | Accelerators, CPUs, memory, storage, networking | Purchase orders, shipments, installed systems |
| Operation | Cloud capacity and managed services | Available capacity, utilization, revenue, margin |
The same $1 of hyperscaler intent can support one supplier's backlog years before it supports another supplier's revenue. It can also support a cloud platform's depreciation before it supports an acceptable return on capital.
The supplier order book error
First, the full $733 billion gets handed to every supplier category. The same budget cannot be counted in full by power, cooling, networking, memory and accelerators. Each category receives a slice, and those slices differ by company, project and year.
Second, all 2026 capex gets treated as demand created in 2026. Some of it has already been contracted. Some supplier revenue in 2026 will come from budgets approved earlier. Some 2026 budgets will not become supplier revenue until later.
Third, capital outside the hyperscaler line item gets ignored. Utility-owned generation, developer-financed shells, operating leases and joint ventures can expand the physical build without appearing cleanly inside the headline number.
The conclusion is straightforward. $733 billion is the top of a multi-year conversion funnel.
The scorecard that matters
- Energized megawatts, not announced megawatts
- Interconnection dates and any slippage
- Construction backlog conversion, not backlog alone
- Power, cooling and networking lead times
- Accelerator, memory and switch shipments
- Available cloud capacity by region
- Utilization once capacity is live
- Revenue and operating income attached to the new capacity
- Depreciation and return on invested capital
What could break the thesis
The risk rises if demand slows before projects energize, if power delays strand completed infrastructure, or if cloud pricing weakens as capacity arrives.
Efficiency is another variable. Better models, caching and specialized silicon can reduce compute needed per task. The bull case requires usage to expand faster than efficiency lowers unit demand.
The final test is economic. If depreciation, financing and operating costs grow faster than cloud revenue, the industry can build an extraordinary amount of infrastructure and still disappoint shareholders.
What we are watching
- Which sites receive firm power?
- Which projects slip?
- Which suppliers turn backlog into revenue?
- Which cloud providers remain capacity constrained after the new clusters arrive?
- Which platforms turn each new megawatt into durable revenue and cash flow?
Those answers will matter more than the next round number.
Capex is the top of the funnel. Energized, utilized compute is the output.
The gap is where the investment work lives.
That is what we read this week.
Otto Analytics
Download this issue (PDF)The Weekly Read returns next week.
Past issues
No. 07 · July 31, 2026 · When Does Space Become the Cheapest Place to Compute? Read → No. 06 · July 24, 2026 · Open Models, Closed Economics: Who Actually Pays for “Free” AI? Read → No. 05 · July 19, 2026 · The Cheaper Intelligence Gets, the More Infrastructure It May Need Read → No. 04 · July 16, 2026 · The Machine Is Only the First Sale Read → No. 03 · July 10, 2026 · The Second Door Read → No. 02 · July 3, 2026 · The Purchase Order Read → No. 01 · July 2, 2026 · The Robot Is Not the Product Read →
This edition uses Otto Analytics' current synthesis of 2026 capital spending guidance from Amazon, Alphabet, Microsoft and Meta. The figures are approximate because company definitions and accounting treatment are not homogeneous.
Alphabet's 60/40 split reflects company commentary on machines and servers versus data centers and networking. Microsoft's two-thirds mix refers to a recent quarterly disclosure and should not be treated as a full-year allocation. Amazon's capex includes non-AWS investments, including robotics and satellites. Meta does not provide a precise numerical mix.
Amazon Investor Relations
Alphabet Investor Relations
Microsoft Investor Relations
Meta Investor Relations