The four-trillion-dollar bet.

The scale is large even before anyone tries to compress it into a single headline number. Amazon expects about $200 billion of capital expenditures across the company in 2026. Alphabet currently guides to $180 billion to $190 billion. Meta guides to $130 billion to $145 billion. Microsoft expects roughly $190 billion of calendar-year 2026 capex. Taken together, those four corporate plans are roughly $700 billion to $725 billion.

Source and scope note: These are corporate capital-expenditure plans, not an assertion that every dollar is AI-specific. The companies themselves describe AI, cloud, technical infrastructure, chips, data centers, and related capacity as major drivers. Oracle is discussed separately because its FY2026 reporting horizon is not the same calendar-year basis.

The useful question is not whether every dollar in those budgets can be labeled AI. It cannot. The useful question is what kind of demand must arrive for this much technical infrastructure to earn an acceptable return. That turns an engineering story into a capital-allocation story.

What they are buying.

The public disclosures point to a mix of short-lived compute assets and long-lived infrastructure. Amazon describes its 2026 investment opportunity as spanning AI, chips, robotics, and its low-earth-orbit network. Alphabet says the overwhelming majority of its capex is technical infrastructure. Meta's April outlook attributed its higher capex range to component pricing and, to a lesser extent, additional data-center costs. Microsoft reported that roughly two thirds of its fiscal third-quarter capex was for short-lived assets, primarily GPUs and CPUs, with the remainder supporting long-lived data-center capacity.

That mix matters because compute assets and the buildings, power, networking, and land around them do not share the same economic life. The return calculation therefore depends on more than model demand. It depends on utilization, product pricing, hardware refresh, financing cost, and the ability to monetize capacity before the short-lived assets age out.

The financing question is not one thing.

The financing structure differs by company. Large hyperscalers continue to generate substantial operating cash flow, even as higher infrastructure spending pressures free cash flow. Oracle illustrates a different edge of the cycle: in February 2026 it announced a plan to raise $45 billion to $50 billion of debt and equity financing to expand OCI capacity against contracted demand.

The accounting pressure is also visible without relying on a speculative depreciation forecast. Alphabet has warned that rising technical-infrastructure investment will increase depreciation and data-center operating costs. Microsoft separates short-lived GPUs and CPUs from long-lived site assets in its capex disclosures. The central issue is mechanical: large infrastructure programs create future fixed-cost and depreciation burdens that have to be matched by revenue and utilization.

The wager is not that AI exists. The wager is that demand, monetization, and utilization arrive on a timeline that earns a return on the infrastructure being built.

Three analytical scenarios.

Demand absorbs the build. AI usage and cloud demand grow fast enough to keep capacity constrained, pricing and product mix support attractive returns, and the infrastructure becomes a productive base for a larger market.

Demand arrives more slowly. The infrastructure is useful, but capacity arrives ahead of monetization. Utilization and pricing weaken, weaker operators restructure, and assets migrate toward stronger balance sheets. The technology succeeds while some capital earns poor returns.

The profit pool is smaller than the build assumes. AI remains widely useful, but the economics of serving it do not support the amount or pace of capital deployed. In that case, write-downs, slower expansion, and a broader reassessment of AI infrastructure returns follow.

These are scenarios, not probability forecasts. The evidence available today establishes the scale of the investment and the demand signals companies say they are seeing. It does not establish which return path will dominate over the next several years.

What it means for the rest of the market.

A build this large creates second-order questions for power supply, construction, semiconductor capacity, financing, corporate free cash flow, and the geography of data-center development. Those effects should be measured individually rather than bundled into one deterministic macroeconomic story.

That is the editorial task: separate what is disclosed from what is inferred, and separate what is happening now from what might follow. The capex plans are public. The eventual utilization, pricing, and return on those assets are not yet known.

Sorso View's role is narrower than predicting the winner. The Sorso View Compute Index measures the reproducible public price layer of AI compute. Essays and Reports widen the frame to the capital and market structure around it. The Index now publishes on recurring issue cycles under A1-R1, not on a guaranteed weekly schedule.

The title remains deliberately larger than the one-year table. It names the multi-year wager implied by the infrastructure cycle, not a single sourced four-trillion-dollar total. The disciplined conclusion is simpler: the build is real, the capital is large, and the return remains uncertain.