Elon Musk has moved into an Airstream trailer parked beside a Memphis warehouse to oversee Colossus, SpaceXAI's massive compute campus that spans two states and generates enough computing power to train some of the world's most advanced AI models. The facility, which carries approximately 1.29 gigawatts of computing load and houses roughly 670,000 physical accelerators, represents one of the largest infrastructure bets in artificial intelligence history.

The decision to live on-site reflects Musk's pattern of embedding himself at production bottlenecks. SpaceX president Gwynne Shotwell described the Airstream as his "palace" during a recent public appearance, acknowledging the unconventional setup. This hands-on approach mirrors his previous moves: sleeping at Twitter headquarters after the 2022 acquisition and living near Starbase during early Starship development. For Musk, the constraint becomes the destination.

What Makes Colossus Different From Other AI Data Centers?

Colossus is not a single building but rather a sprawling cluster of halls and power generation assets straddling the Tennessee-Mississippi border. The facility transformed from a private training operation into a commercial compute rental business after SpaceX acquired xAI in early 2026 and folded it into an operating division branded SpaceXAI. This shift changed everything about how the facility operates and who benefits from its capacity.

The campus consists of three major sections. Colossus 1 occupies the former Electrolux appliance plant at 3231 Paul R. Lowry Road, spanning approximately 785,000 square feet across 81 acres. Colossus 2, built at 5420 Tulane Road, covers nearly 1.0 million square feet on roughly 79 acres, with an adjacent denser hall called Minihard rising nearby. A third site in Southaven, Mississippi, called Macrohardrr or Colossus 3, adds another layer to the complex. Taken together, the campus encompasses approximately 2.6 million square feet of building and yard space.

How Does the Computing Power Scale Across the Three Facilities?

Colossus 1 operates on 340 megawatts of computing power and houses approximately 276,000 graphics processing units (GPUs) in H100-equivalent capacity, according to Epoch AI's analysis. The facility is a fossil record of its own construction, with different GPU generations arriving at different times. The original 150,000 H100 GPUs, installed in 2024, made Grok competitive as a large language model. An additional 50,000 H200 units arrived by February 2025, nearly doubling memory capacity to 141 gigabytes per card compared to the H100's 80 gigabytes. Finally, 30,000 B200 units arrived in July 2025, representing only 13 percent of the card count but delivering 28 percent of the total computing power through advanced processor technology.

Colossus 2 represents a dramatic scaling up of the operation. Epoch AI models the facility at approximately 1,112,000 H100-equivalent GPUs running on 946 megawatts of computing power, with a projected capital cost of $35.8 billion. By the first quarter of 2027, that capacity is expected to grow to 1,824,000 H100-equivalents consuming 1,531 megawatts and representing $58.0 billion in modeled capital investment. Unlike Colossus 1, which uses multiple GPU generations, Colossus 2 standardizes entirely on Blackwell architecture, the latest generation of NVIDIA accelerators.

The silicon composition of Colossus 2 currently includes 110,000 B200 units and 330,000 B300 units, totaling 440,000 physical accelerators. Projections for early 2027 hold the B200 count flat while increasing B300 capacity to 612,200 units. This focus on the newest Blackwell generation reflects the facility's design as a cutting-edge training platform rather than a legacy system.

Why Is Power Infrastructure the Real Bottleneck?

Building Colossus required solving an enormous power puzzle. The inherited electrical interconnection at the Electrolux site provided only about 8 megawatts against a 340-megawatt load, forcing SpaceXAI to invest heavily in grid infrastructure. The company spent $35 million on the first substation and committed another $20 million to a second, while also parking 69 temporary mobile turbines in the yard to meet immediate demand. However, an agreement signed with the Mississippi Department of Environmental Quality on July 30, 2026, requires removal of all temporary turbines by July 2027, with removals beginning as early as August 2026.

To replace that temporary power, SpaceXAI is constructing a permanent 1.2-gigawatt gas plant in Southaven, Mississippi. This represents a race against the clock: the commercial clock has already started. Google signed an agreement disclosed on June 5, 2026, to rent approximately 110,000 GPUs from SpaceX from October 2026 through June 2029 at $920 million per month. These are not research-budget numbers but factory-utilization numbers, meaning any delay in energization directly impacts revenue.

How Does Cooling Technology Enable This Scale?

Cooling infrastructure tells the story of a facility that outgrew its initial design. Colossus 1 features 33 evaporative cooling towers with 310.4 megawatts of thermal capacity. By February 2026, the air-cooled chillers that carried early phases had been disconnected and removed, replaced by more efficient systems. Colossus 2 employs a different cooling architecture entirely, with 217 air-cooled chillers providing 412.3 megawatts of capacity and 144 condensers. By August 22, 2025, approximately 119 chillers were operational, representing roughly 200 megawatts of cooling capacity.

The cooling challenge stems from NVIDIA's GB200 NVL72 architecture, which changed data-center engineering fundamentally. Each cabinet contains 72 Blackwell GPUs and 36 Grace CPUs connected by a fifth-generation NVLink fabric rated at 130 terabytes per second within the rack. This creates a memory-coherent domain where computing elements behave as a single system rather than independent servers sharing a network switch. A single GB200 NVL72 rack typically consumes about 120 kilowatts, meaning 110,000 GPUs require roughly 1,530 racks consuming approximately 210 megawatts including electrical overhead.

How to Understand the Commercial Timeline Driving Musk's Presence

• Immediate Constraint: The next 90 days couple civil works, high-voltage interconnection, liquid-cooling commissioning, and rack bring-up into one compressed schedule that determines whether SpaceXAI meets its October 2026 contract start date with Google.
• Revenue Dependency: Google's $920 million monthly rental agreement creates a factory-utilization clock where missing an energization date directly translates to missed invoices, making on-time delivery non-negotiable.
• Regulatory Pressure: The July 30, 2026 agreement with Mississippi environmental regulators requires removal of all 69 temporary turbines by July 2027, forcing completion of the permanent 1.2-gigawatt gas plant before that deadline.
• Multi-Tenant Operations: Beyond Google, the campus hosts training runs for Anthropic, Cursor, and SpaceXAI's own Grok model, requiring careful orchestration of compute allocation and power distribution.

Musk's decision to live in an Airstream beside the Memphis facility reflects the intensity of this timeline. When production systems become constraints on companies he controls, his pattern is consistent: he moves to the constraint. The warehouse roof is even painted "MACROHARD" in letters visible from light aircraft, a playful reference to Microsoft that signals the scale and ambition of the operation.

The Colossus campus represents a fundamental shift in how AI infrastructure is built and financed. Rather than serving as a private research tool, it has become a commercial factory generating nearly $1 billion per month in revenue while simultaneously training some of the world's most advanced AI models. The facility's success will likely shape how future AI companies approach compute infrastructure, balancing the need for cutting-edge technology with the financial discipline required to operate at industrial scale.