AI Data Centers Cost Billions to Build. This Is Why

Meta, Amazon, Microsoft and Alphabet are expected to spend about $750 billion on capital expenditures this year, primarily on data centers; a theoretical 1-gigawatt AI facility in Abilene, Texas, could cost roughly $30 billion. More than 60 gigawatts of US capacity is expected to be added over the next four years, while data-center memory prices have risen more than fivefold in a year and power-equipment suppliers face backlogs extending beyond 2030. The build-out is boosting activity across energy, construction, real estate and financing markets, but its economics remain uncertain amid rising costs and warnings of an AI bubble.

Analysis

The key risk is a timing mismatch: infrastructure costs are paid upfront, while utilization and AI revenue arrive later. Higher cost per usable megawatt and power delays can therefore weaken returns even if AI demand ultimately grows; if costs cannot be passed through, cloud margins take the hit, and if they can, customer adoption may slow. The likely near-term winners are constrained equipment and power suppliers, but long backlogs are not proof of higher earnings: delivery capacity, contract pricing and cost pass-through determine who captures value. Cushman & Wakefield’s advisory exposure may benefit from activity, but this does not establish material earnings leverage. Oracle’s project delay is a project-specific warning, not evidence that all builds are stalled.

Over the next 1–3 months, watch hyperscaler capex guidance alongside cloud growth, AI workload utilization and evidence of customer monetization; capex alone is a poor signal of returns. Over 6–18 months, sustained build costs without improving utilization would raise depreciation and free-cash-flow pressure and could trigger project deferrals. The contrarian point: a build slowdown would not necessarily mean AI demand has failed—it could reflect power and equipment bottlenecks or developers rationing uneconomic capacity. Conversely, supplier backlogs may already be capitalized before delivered capacity earns a return.

Key Decisions for Investors

• Avoid treating the capex cycle as a blanket long for AI infrastructure. For existing AI exposure, consider a relative-value hedge: short NVIDIA against a beta-sized basket of Amazon.com, Inc. (AMZN), Alphabet Inc. (GOOG), Meta Platforms, Inc. (META) and Microsoft Corporation (MSFT), conditional on valuation and borrow being acceptable. The thesis is that a capex pause can hit semiconductor orders sooner, while hyperscalers retain discretion to phase projects; this is a hedge, not a high-conviction standalone short.
• Do not initiate that pair solely on construction-cost inflation. Reassess after the next company updates: rising capex with slowing cloud growth, weak utilization or lower return commentary supports the hedge; improving AI monetization and sustained NVIDIA order visibility falsify it.
• Track transformer, turbine and power-delivery supplier lead times, contract terms and delivery schedules as a 1–3 month alert. Backlogs can support future revenue, but do not assume supplier margin expansion without evidence of pricing power and successful delivery.
• Treat Oracle Corporation’s disclosed delay as an execution-risk indicator for projects dependent on constrained energy infrastructure, not a broad short signal. A pattern of additional permitting or power delays, or hyperscalers deferring capacity, would strengthen the downside case; isolated project resolution would weaken it.