Generac Enters Amazon’s Compute Stack

By September 2026, Amazon had agreed to buy up to $8 billion in Generac data-center generators—and secured a warrant for up to 2.6% of the supplier. The backup-equipment order came with an ownership bet, a response to a market in which a seven-year grid queue can leave the most expensive GPU dark.

Grid delay has turned AI infrastructure into a power-continuity business, pulling generation, storage, switchgear and load control into the schedule and economics of AI deployment. Developers are buying generation, hyperscalers are signing curtailment deals, and industrial suppliers are chasing data-center orders—parallel responses to the same delay.

A grid queue now sets the compute opening date

US data-center developers facing grid-access waits of up to seven years have turned to aeroderivative turbines and diesel generators to energize sites while the grid catches up, making backup hardware a bridge to operation.

A data center cannot monetize accelerators that it cannot power. Every month between server delivery and energization leaves installed compute unavailable, so operators must make generation and electrical-balance-of-plant decisions earlier in the build. GPU procurement and site power can no longer proceed as independent workstreams.

PJM Interconnection made the operating consequence explicit when it proposed requiring large data centers either to provide their own generation or curtail consumption to prevent a large-scale outage. Under PJM’s proposal, a campus would have to show how it will operate when the grid cannot serve it.

Operators must also solve for location and latency. A remote site with cheap electricity can work for workloads tolerant of distance, but inference often puts a premium on proximity to users and dependable supply. A low-priced megawatt loses its advantage when latency degrades the service or the interconnection arrives years after the hardware.

A ten-second gap requires an entire power stack

A generator does not cover its own startup interval. Data-center designers use an uninterruptible power supply to bridge the 10 to 15 seconds a generator typically needs to start and take full load. The UPS also selects and conditions incoming power, while transfer switching, storage, generation, fuel systems and power-management controls coordinate the transition.

Energy storage adds an economic function beyond emergency continuity. Operators can use backup-system storage to flatten load profiles, reduce expensive utility consumption and manage variability in supply. Peak capacity shapes the facility even when average consumption remains lower; the campus must either provision for the peak or gain enough control to avoid it.

Google signed demand-response agreements with five US utilities covering up to 1 GW. The agreements let Google reduce data-center consumption during peak periods.

Data centers can destabilize the wider system when controls and grid coordination fail. When campuses switch to local generators without advance notice, their “unannounced disconnections” can contribute to cascading outages. Local generation therefore needs controls that make the campus legible and responsive to the grid.

Amazon is procuring supplier continuity, not loose equipment

The long-term agreement includes a warrant for up to 2.6% of Generac.

Amazon chose a long horizon, while Generac gained a large named customer for equipment whose availability can govern opening dates. The warrant extends their shared stake beyond each shipment.

Data-center developers can wait until installation to solicit bids when equipment is interchangeable and abundant. When opening dates depend on manufacturing slots, installation crews and compatible electrical systems, they secure suppliers before the facility is ready.

Industrial suppliers capture value at the physical bottleneck

Caterpillar’s power-and-energy generator sales grew 22% to $7.8 billion in 2024 and rose another 28% year over year through September 2025 as AI data-center demand increased. Cummins, Eaton and Ford also moved into data-center power equipment.

Rising AI-campus spending drew companies that could supply generation, switchgear and related electrical machinery. The model providers got the headlines; the order books identified the scarce complements.

US data-center builders also rely on Chinese imports because domestic suppliers cannot provide enough transformers and switchgear, and shortages have delayed construction. An accelerator vendor can deliver on schedule while missing electrical equipment keeps the campus dark. Production capacity for ordinary components can therefore govern the use of far more expensive hardware, giving the breaker box excellent negotiating leverage.

In 2007, Caterpillar was framed as an unexpected beneficiary of Web 2.0; by 2026, it had reached a $300 billion market capitalization amid demand for data-center power equipment. Caterpillar benefited twice because computing booms reward suppliers that control physical complements builders cannot proceed without.

Grid behavior now enters the inference ledger

For one month in 2025, wholesale electricity prices in areas with significant data-center activity ran as much as 267% above 2020 levels. Big Tech’s planned AI data centers were estimated to require another 44 GW by 2028. Operators must price annual consumption alongside local electricity rates, peak demand, contingency capacity and curtailment.

Operators must carry AI data-center power costs into inference economics. A campus that can shift consumption away from expensive peaks, bridge interruptions and coordinate its load with utilities has a different cost structure from one drawing the same annual energy without those controls. Expensive or unavailable electricity erases the advantage of cheap accelerators.

Ireland turned local-grid impact into a license-to-operate rule. Its Bring Your Own Power policy requires new data centers to provide on-site plants or dedicated nearby energy sources, explicitly balancing AI investment against outages and higher bills for residents. The policy makes local-grid impact part of the product an operator must deliver to secure a site.

Data-center power-plant projects still face permitting and supply-chain constraints. The US has been adding less than 65 GW of generation annually against the roughly 80 GW pace cited as necessary to meet digital demand. Building behind the meter can bypass part of the interconnection queue while leaving equipment shortages, approvals and grid coordination in place.

With a commitment of up to $8 billion, Amazon is buying a measure of control over the date and cost of usable compute when the grid cannot promise either. A deployable campus must receive, condition, store, generate, shed and restore power without destabilizing the grid. The GPU now sits downstream of the breaker.