Musk says SpaceXAI may more than double Colossus 2's Nvidia chips by year-end

Why it matters

Musk is tying his AI strategy to another rapid buildout of Nvidia-powered infrastructure. The missing baseline makes the year-end target hard to translate into actual operational capacity.

Elon Musk said SpaceXAI may more than double the Nvidia chip count at its Colossus 2 computing cluster by December 31st. Bloomberg reported the target for the Memphis-area cluster as Musk pushes to expand AI computing capacity; its report gave no current or planned chip count.

Musk founded xAI in 2023 to build AI systems, including Grok. Following SpaceX's acquisition of xAI, the AI operation is now branded SpaceXAI on its company site. The expansion puts a familiar Musk bet back in focus: build computing capacity at a pace that gives his models room to grow, then connect those models to products and distribution he controls. The target is a plan, not a report of completed installations.

The number behind the promise

Bloomberg did not give Colossus 2's current Nvidia chip count, Musk's planned total, or whether his target refers to chips already installed, operational, or contracted for delivery. Those distinctions determine what "more than double" means in practice. A facility can have equipment on order or in transit without that capacity being online and available for training or inference.

Independent tracking offers a useful, but not interchangeable, reference point. Epoch AI's estimate, updated September 24th, 2026, put Colossus 2 at 440,000 Nvidia GPUs: 110,000 B200s and 330,000 B300s. Epoch estimated the site at 946 megawatts of IT power and $35.8 billion in capital cost. Those are outside estimates based on satellite imagery, company disclosures and other inputs, not a chip count supplied by Musk or SpaceXAI. Bloomberg's report does not establish that Epoch's estimate is the baseline Musk used.

SpaceXAI's May announcement of a compute agreement with Anthropic described Colossus 1 as having more than 220,000 Nvidia GPUs, including H100, H200 and GB200 accelerators. SpaceXAI's Colossus page describes its earlier cluster's growth to 200,000 GPUs. Neither figure supplies Colossus 2's official current count.

Musk's compute-first strategy

Colossus supports Grok models and related AI products; SpaceXAI's site describes the cluster as part of the infrastructure behind its models. Musk's distribution advantage includes X, where Grok is available, alongside SpaceXAI's direct products and developer services. The cluster's economics still depend on how quickly that capacity can be built and put to work.

That product push has been visible in recent releases. RuntimeWire reported on September 21st that SpaceXAI shipped Grok 4.7 with a focus on longer-running agent work. More compute can support model development and serving, but chip count alone says little about whether customers are adopting those products, whether the capacity is being used efficiently, or whether the operation can earn a return on the infrastructure.

SpaceXAI has framed rapid construction as a core strength. Its Colossus page says the first cluster reached its initial scale in 122 days and then doubled to 200,000 GPUs in 92 days. That is a company account of the buildout, and it describes Colossus 1. The new Colossus 2 target extends the same speed-first approach to a larger site, where power, cooling, networking and equipment deliveries have to keep pace with each other.

Musk's target also lands in a business that has already drawn outside compute customers. In May, SpaceXAI said Anthropic had agreed to use Colossus 1 to add capacity for Claude Pro and Claude Max subscribers. That deal shows the infrastructure can serve another model developer, while leaving Colossus 2's customer mix and utilization unclear. SpaceXAI has not supplied a public operating metric in the materials reviewed here that would show how much of the newer cluster is serving Grok, external customers, or still being brought online.

The doubling target describes construction ambition. It says nothing by itself about model performance or commercial demand. The useful test will be the installed and operational count at the deadline, alongside the power and workload available to use it. Until those figures are clear, the headline percentage describes Musk's intended pace more precisely than it describes the capacity the AI business can actually deploy.