Elon Musk’s AI ambitions just took another giant, silicon‑packed step forward, with SpaceXAI outlining a roadmap to bring roughly 1.44 million Nvidia GPUs online across its Colossus supercomputer complex by the end of 2026. The latest expansion centers on a surge of Blackwell‑generation GB300 chips, positioning Colossus as one of the most aggressively scaled AI compute clusters on Earth despite SpaceXAI being only about three years old.
In a series of posts, Musk broke down the current and planned configuration: Colossus 1 runs a mix of 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs, while Colossus 2 already hosts 110,000 GB200s and 440,000 GB300s. SpaceXAI expects another 220,000 GB300 units to be fully operational within days, followed by 220,000 more in November, and a final 220,000 “if we get lucky” by late December. If the schedule holds, Colossus 2 alone would field around 1.21 million GB200 and GB300 processors, bringing SpaceXAI’s GB300 count to roughly 1.1 million and its total GPU fleet to about 1,440,000 chips. That pace puts Musk’s operation in direct contention with hyperscale AI rivals such as OpenAI and Anthropic, even though those labs are six and ten years old respectively.
Under the hood, the design shows how SpaceXAI is trying to squeeze every drop of performance from Nvidia’s high‑end stack. Colossus 1, a Hopper‑era build, now functions as a sort of “legacy” cluster, mixing H100, H200, and early Blackwell GB200 cards—and, according to reporting, is relatively inefficient at training Musk’s Grok models. Rather than let that compute go idle, SpaceXAI has reportedly leased Colossus 1 capacity to Anthropic, using it primarily for inference workloads to serve Claude and related systems, while reserving Colossus 2’s pure‑Blackwell configuration for next‑generation training runs. Colossus 2, built around GB200 and GB300 GPUs with high‑bandwidth networking, is designed explicitly to avoid the bottlenecks that can plague mixed‑architecture clusters.
The scale‑up is not just about chips—it’s about electricity, and that’s where Colossus has already collided with the real world. To power its Memphis–Southaven complex, SpaceXAI initially deployed a fleet of 69 mobile gas turbines without traditional long‑term power plant infrastructure, triggering a Clean Air Act lawsuit and community backlash over pollution and noise in nearby neighborhoods. Environmental groups including the NAACP and the Southern Environmental Law Center alleged SpaceXAI and affiliated contractors were operating dozens of turbines without proper permits, prompting regulators and courts to scrutinize the project’s fast‑and‑loose build‑out. SpaceXAI has since agreed to phase out the temporary turbines and replace them with a 1.2‑gigawatt natural‑gas power plant built under a formal permit, with 41 permanent turbines authorized and a removal timeline for the mobile units stretching into mid‑2027.
Even as parts of the turbine fleet are being powered down—local coverage has documented 11 of the 69 units going offline at the Southaven site—SpaceXAI is racing to complete permanent power so the GPU growth curve does not stall. The company’s own materials describe a roadmap from an initial few hundred megawatts of capacity up to a total in the 1–2 GW range as Colossus 1, Colossus 2, and additional buildings come online. That energy footprint puts the complex in the same ballpark as a large nuclear plant or multiple big‑city data centers combined, underscoring why surrounding residents and environmental advocates have treated the project as a major test case for AI infrastructure’s real‑world costs.
Musk, unsurprisingly, is already thinking much bigger than a single terrestrial campus. He has floated plans for SpaceXAI to grow its overall data center capacity sevenfold by 2027 and ultimately deploy the equivalent of 50 million H100‑class GPUs by 2030, a scale that would dwarf even the most aggressive current hyperscaler roadmaps. Broadcom has said it is working with three unnamed hyperscale customers that each aim to run a million‑GPU cluster by around 2027, suggesting SpaceXAI is not alone in chasing that milestone—but Musk’s numbers blow past it. He has also teased an Orbital Data Center System with up to a million satellites handling compute in space, a vision Nvidia CEO Jensen Huang has politely described as a “dream” for now, but one that fits neatly into Musk’s pattern of fusing space hardware with compute infrastructure.
For the broader AI and tech ecosystem, Colossus hints at where the arms race is headed: fewer modest clusters, more mega‑scale “compute reactors” whose GPU counts rival global supply and whose power needs reshape local grids. Between the Anthropic leasing deal, the regulatory fights over turbines, and Musk’s orbital aspirations, SpaceXAI’s path is becoming a template—and a warning—for anyone trying to compete at the frontier of model training. If Colossus hits its 1.44‑million‑GPU target this year and continues toward the tens of millions Musk has sketched out, the story will be less about whether the hardware can be built and more about whether the world around it is willing to live next door to the engines of AI.
