On June 25, Caterpillar rose more than 6 percent. Apple fell 6 percent. The company that makes yellow excavators outperformed the company that makes the devices AI runs on, and it has been doing so for a year. Caterpillar stock is up more than 160 percent in twelve months. It is the best-performing industrial name in the Dow and one of the best-performing AI trades in the market, despite making no chips, writing no software, and building no servers.
It makes generators.
The bottleneck in AI is not compute. Nvidia solved compute. The bottleneck is electricity. A hyperscale inference campus can consume a gigawatt, the output of a nuclear reactor. The demand is arriving faster than the grid can absorb it.
The U.S. grid interconnection queue now contains approximately 2,600 gigawatts of pending projects. That is more than double the country's entire installed generation capacity. The median time from application to commercial operation is approaching five years. In Northern Virginia, the largest data center market in the world, the wait extends to seven years.
Seven years is forever in AI. The companies building infrastructure for models that do not yet exist as concepts cannot wait for a utility to build a substation and clear the queue. So they are doing something else.
They are building their own power plants.
Behind the Meter
Behind-the-meter generation means producing electricity on site, inside the property boundary, without connecting to the grid. No utility interconnection. No transmission fees. No queue. The data center and the power plant share a fence line and a wire.
The Monarch Compute Campus, a 2,380-acre site near Point Pleasant, West Virginia, is the most visible example. American Intelligence & Power Corporation signed a strategic alliance with Caterpillar in March to deploy 2 gigawatts of dedicated natural gas generation, one of six announced gigawatt-scale power agreements. The equipment is Cat G3516 Fast Response natural gas generator sets that ramp from zero to full load in seven seconds, built for the volatile power profiles of AI workloads. Deliveries begin September 2026. The first phase powers up in 2027. Total planned capacity is 8 gigawatts.
Eight gigawatts. The output of eight nuclear reactors. On a single site in West Virginia. Running on natural gas. Fully self-supplied. No grid connection required.
The Numbers
Caterpillar's first quarter told the story in revenue. Total sales reached $17.4 billion, up 22 percent year over year. Power generation sales to users surged 48 percent, driven almost entirely by large data center applications. The order backlog hit a record $63 billion, up 79 percent from a year earlier. The backlog for large reciprocating engines, the product line serving data centers specifically, has grown more than 3.5 times since January 2024.
The company responded by tripling down. Large reciprocating engine capacity targets were raised to nearly three times 2024 levels by 2030, up from the prior two-times goal announced eighteen months earlier. Caterpillar committed $725 million to expand its Lafayette, Indiana manufacturing plant. Even so, lead times for large lean-burn generators run 24 to 30 months. Demand exceeds supply for the foreseeable future.
The Parallel Grid
What Caterpillar's backlog represents is not a product cycle. It is an infrastructure buildout.
The $63 billion in orders is a map of where a parallel electrical grid is being constructed, site by site, generator by generator, outside the jurisdiction of utility regulation and grid planning. Each behind-the-meter campus is a node that will never appear on the grid operator's dispatch stack. It generates power, consumes power, and never touches the transmission system.
The popular narrative frames AI's power demand as bullish for utilities. More load, more revenue, more rate base. But behind-the-meter generation removes the load from the utility entirely. The data center that bypasses the grid is a customer the utility will never have. The generation it builds is capacity the grid will never see.
Caterpillar estimates $2.2 to $2.4 billion in tariff headwinds for full-year 2026. It does not matter. When your backlog is $63 billion and your lead times are two years, you pass through costs. The customers building behind-the-meter campuses are not price-sensitive. They are time-sensitive. A 24-month generator delivery is still five years faster than a grid interconnection.
The AI industry is rebuilding the electrical system from scratch, one gas turbine at a time. The company selling the turbines just reported its best quarter in 99 years. The bypass is the trade.