AINEWS
xAI, Musk's AI firm, has $27.8B a year in signed AI compute leases from Google, Anthropic, and Reflection, and a blogger's model projects the run rate could triple by year end on gas turbines.
xAI, Elon Musk's artificial-intelligence company, has spent the last year turning its Memphis data centers into a commercial leasing platform. The result is a $27.8 billion annualized book of signed customer contracts. A financial blogger's month-by-month model projects that figure could roughly triple by year-end, if the buildout keeps pace and one or two more major leases close.
In the United States, a grid-interconnected data center typically waits three to seven years for a utility interconnect. That wait is the bottleneck, and xAI is working around it with mobile gas turbines: truck-trailer-mounted generators made by APR Energy, a joint venture with Caterpillar. NextBigFuture's financial model pegs the planned deployment at roughly 2 gigawatts of mobile capacity at Colossus 2, xAI's second Memphis data center, by late 2026. One gigawatt is roughly the output of a large nuclear reactor, so the planned Colossus 2 build sits in the scale range of a small city's baseload supply.
Three signed deals anchor the $27.8 billion run rate. Google has agreed to pay SpaceX $920 million per month from October 2026 through June 2029 for roughly 110,000 NVIDIA GPUs and related compute, a deal anchored in an SEC free-writing prospectus filed alongside the planned SpaceX IPO. Anthropic is paying $1.25 billion per month through 2029 to rent Colossus 1 capacity near Memphis, in a deal that Wired's IPO analysis flagged as one of the largest private compute contracts on record. Reflection AI has signed $150 million per month starting July 1, 2026, with the contract potentially running through 2029. The Motley Fool's tally of the three signed deals totals roughly $27.8 billion a year.
The $82 billion figure in NextBigFuture's headline is an analyst projection, not a reported company number. It assumes Colossus 2 reaches its 2-gigawatt target on time and that one to two more major hyperscaler leases are announced before the end of 2026. Brian Wang, the blogger who built the month-by-month model, treats it as a forward scenario, not a result. Larry Goldberg reviewed the model and called it "incredible," but Goldberg's verdict comes from one person weighing one financial blog's model, not an audit or a company confirmation.
About 440,000 chips are projected to run at Colossus 1 and Colossus 2 by late 2026, per Epic AI, a third-party watcher of capacity rather than a SpaceX or xAI disclosure. The split: about 330,000 B300 chips and 110,000 B200 chips, two current NVIDIA accelerator models. The figure is directionally supported rather than company-confirmed.
APR Energy's trailer-mounted gas turbines can be deployed in weeks rather than the years a grid interconnect requires, because they need a gas pipeline and a flat pad, not a substation upgrade. That collapses xAI's path from a signed customer contract to billable compute from multi-year to single-digit months. A grid-locked competitor trying to match the pace has to solve the same power problem, because most U.S. utilities are not signing new multi-gigawatt interconnect agreements on that timeline.
Three named customers, Google, Anthropic, and Reflection, account for the entire signed book, and the SpaceX IPO prospectus flags the customer concentration as a material risk factor. A single counterparty default, a chip-supply disruption, or a turbine failure at the wrong moment would shift the math materially.
The Google contract ramps to its full $920 million monthly run rate in October 2026. If the mobile-turbine buildout at Colossus 2 holds that schedule and at least one additional major lease closes before year-end, the $82 billion projection moves from blog arithmetic to a testable hypothesis. If it slips, the gap between $27.8 billion confirmed and $82 billion projected is exactly the size of the unbuilt power capacity.