The AI Capex Cycle: When Does Infrastructure Spending Peak?
Hyperscaler capital expenditure has tripled since 2022. The familiar question — when will it stop growing — is the wrong one. The more useful question is which line items peak first, and what replaces them in the budget.
The setup
Hyperscaler capital expenditure — the combined spend of the largest cloud providers on servers, networking, land, and power — has moved from the low hundreds of billions of dollars per year to the high hundreds in less than four years. The proximate driver is AI model training and inference, but the shape of the spend is driven by constraints that are not, strictly speaking, about models at all.
The question most investors ask is when does the capex cycle peak. We think that framing is too blunt. The cycle is not a single line. It is a stack of line items with very different half-lives: GPUs, power contracts, land, networking, and the long-lived civil engineering of substations and fiber. Each peaks on its own clock.
Why the headline number is misleading
Headline capex captures the dollar total but hides the mix. Roughly three buckets matter:
- Compute silicon. Accelerators and high-bandwidth memory. Replacement cycles are short. Pricing power depends on the supplier and the generation. Demand is highly sensitive to what the previous generation can still do.
- Power and land. Substations, transformers, water rights, and the multi-year permitting that surrounds them. Replacement cycles are measured in decades.
- Networking and interconnect. Switches, transceivers, dark fiber. Roughly between the two in durability.
The first bucket can halt in a single quarter if the next generation's training-efficiency gains disappoint. The second almost cannot — the contracts were signed years ago, and the substations will be built either way.
The supply-side floor
Even if model demand plateaued tomorrow, three supply-side commitments keep a floor under the capex number:
- Power interconnection queues. Utility interconnection queues in the U.S. and Ireland were already years long before AI. The hyperscalers' power purchase agreements typically pre-commit spend that cannot be cancelled without penalty.
- Long-lead civil work. Dirt, steel, and cooling take time no software cycle can compress.
- Silicon pre-orders. Leading-edge accelerator allocations are booked through multiple quarters in advance.
These commitments mean the trough of any capex reset will arrive with a lag — and will be shallower than demand-side analyses suggest.
The mix shift that's already happening
The real move worth tracking is not total spend but the mix. Over the past eighteen months, the share of capex going to power-adjacent assets (generation, transmission, grid-scale storage) has grown faster than the share going to compute. If this continues, the constraint on AI build-out shifts from chips to electrons, and the identity of the marginal winner shifts with it — toward grid operators, turbine manufacturers, and interconnection-queue specialists rather than accelerator designers.
What we're watching
- Hyperscaler disclosures of power contracts vs. equipment orders.
- The pace at which second-tier cloud providers sign their own PPAs.
- Whether any hyperscaler takes a write-down on pre-ordered silicon — the first sign that the mix is tipping from compute-bound to power-bound.
Takeaways
- The aggregate number is a distraction. Watch the mix.
- Compute silicon can unwind quickly; power and land cannot.
- The next leg of the AI build-out will likely be dominated by electrons, not chips.
- Expect the cycle's trough to be shallower and its peak less dramatic than demand-side forecasts suggest.
- The interesting beneficiaries over the next 36 months are further up the supply chain than most investors have positioned for.