Inside Colossus 2: The Race Is Time to Useful Work

Picture a building that, according to Elon, used to house Hormel product. Loading docks, pallets of food. Now it's called Spam, and it's scheduled to hold around 200,000 Nvidia chips by December 2026.

Now a hypothetical, and I'm labeling it as one. The chips arrive. They're paid for. A customer is ready. The power hookup isn't done. Every machine in the room waits on the same thing, and buying another machine fixes nothing.

A chip sitting in a warehouse can't answer your question or train a model. It needs electricity, a way to get rid of heat, and a link to every other computer on the job. I think Colossus 2 is one of the most important projects in the global economy, and Elon's latest numbers raise the stakes. Buying a million chips does not give you a million chips doing useful work. If you use AI at work, that gap shows up in the service you pay for. So what lets SpaceX close it faster, and how does that become earnings?

I've written before about how much power the AI build needs. This piece stays inside the buildings.

Five halls, two jobs

On February 2nd, 2026, SpaceX acquired xAI, the company behind Grok. The computers running that AI business became SpaceX computers. Colossus 2 is a campus stretching across Memphis, Tennessee and Southaven, Mississippi, with different buildings doing different jobs.

I picture it as a factory. The GPUs are the machines doing the calculations. Power and cooling keep them running. The network lets them pass work to one another. Blackwell is Nvidia's chip family here, and GB200 and GB300 are the two versions that show up in Elon's building counts.

The 1.2 million comes from Elon. On October 31st, 2026, he set a target of around 1.2 million Blackwells at Colossus 2 by the end of 2026, and his building update shows how he's stacking it. Macrohard, the building with its name written across the roof, had 110,000 GB200s and 220,000 GB300s. Macroharder and Minihard each had 220,000 GB300s, and Elon described both as configured for training. His schedule put Gun Range on track to operate by the end of November 2026 with the same training setup. Spam was scheduled for December 2026 with around 200,000 GB300s configured for inference.

The names are literal. Gun Range was a local police department's gun range. My read is that the names tell you the method: existing places get turned into pieces of an AI campus.

Training is teaching a model by having it learn from data. Inference is a trained model working for you: you ask, the computers run the model, you get an answer. Each job puts different demands on everything around the chip. So adding a building starts with deciding what work it will do.

Power before the grid shows up

The expensive part starts when chips arrive and can't run. In its 2026 outlook, JLL put the average grid connection wait in major data center markets at more than four years.

So the Colossus method gets useful capacity running before every permanent piece is finished. In 2024, the original Colossus reused the shell of an Electrolux factory in Memphis. Colossus 2 began with a warehouse purchase in March 2025. A shell that's already standing lets you start fitting it out for computers right away.

At the first Colossus, temporary gas turbines supplied electricity while the grid connection was built. By May 1st, 2025, its first phase was running on 150 megawatts of grid power. The same update described 150 megawatts of Tesla Megapacks integrated into that phase. Megapacks are large batteries that store electricity and release it when needed. The stated job was helping the system get through outages or events when it had to cut its demand on the grid. Turbines make power. Batteries give the site a reserve. You plan for supply and for the moment supply changes.

Colossus 2 draws on the Southaven plant on Stanton Road. In its July 30th, 2026 update, SpaceX described 69 temporary mobile turbines powering the Southaven facility, plus a permanent plant under construction with 41 turbines rated at 1.2 gigawatts, or 1,200 megawatts. SpaceX's schedule is to remove all the temporary turbines by July 2027.

This is physical equipment next to real communities. On October 30th, 2026, a federal judge allowed the Clean Air Act lawsuit over the Southaven turbines to proceed. Power delivery runs through permits and courtrooms too, and I'd count that as part of the schedule.

Where the heat goes

Every chip doing math throws off heat, and the heat has to leave so the chip can keep running. The October 2024 ServeTheHome tour of Colossus 1 shows it clearly: the camera follows rows of computer racks and the liquid cooling equipment built in alongside them. Cooling is part of how these machines operate.

At Colossus 2, construction imagery analyzed in September 2025 showed large banks of chillers going in outside the building. Those chillers belong to the system that pulls heat out of the computing equipment.

The switch sets the block size

During training, the work is spread across many chips. Each does a piece, they swap results, and they use those results to take the next step. Picture the factory developing a new design. Different teams test different pieces, and every result has to come back together before the next version. A team that finishes early still waits. At the moments where chips sync up, one slow transfer holds up the whole group.

Add chips without enough capacity to trade their work and you just buy more waiting: paid-for computing power sitting idle until data arrives. That's why the network decides so much of what a company gets from its hardware.

In his October 31st update, Elon described Macroharder and Minihard with network cards rated at 800 gigabits per second. A network card sends and receives data. It connects through cables to switches, which route data toward the computers that need it. As the group grows, that whole path has to keep up.

On September 25th, 2026, Elon explained that the unusual 110,000-chip increments came from how many fiber optic cables fit into a central switch. The size of a chip group is set by a physical limit in the gear connecting the machines. That helps explain why these projects grow in blocks, and why 110,000 and 220,000 keep showing up in the building counts.

Nvidia's published Blackwell rack designs use copper cables to link chips inside a rack and fiber for the larger network tying groups together. Short copper links let neighbors trade work fast. The fiber network lets those groups cooperate on one big task.

Spam gets less fiber

Back to Spam. In the October 31st update, Elon said its inference configuration would have much simpler networking and less than half the fiber.

Why would the job change the wiring that much? Picture the factory once the design is finished. Separate production lines fill separate customer orders, and no line waits on the others before shipping. Independent copies of an AI model serve different people the same way. Your request goes to one group of chips. Somebody else's goes to another. Those groups still need connections to receive requests, pull data and send answers back. They don't need the whole building to share a training update before moving forward.

Nvidia's reference designs describe the case where the whole model fits inside one server. For that job, the separate network that coordinates training across servers can be left out. Larger models can still need several groups of chips talking during inference. The design follows the job.

Two clocks from the first Colossus

How fast is fast? On September 2nd, 2024, Elon announced that the original Colossus had its 100,000 H100 training chips online. The build took 122 days from start to finish, about four months. Elon's end of 2026 target for Colossus 2 is about 12 times that chip count.

Nvidia supplied a second clock in October 2024: 19 days from the first rack arriving on the floor to training beginning. That's the interval I care about. Taking delivery gives you equipment. Starting a training run means the equipment is doing the job you bought it for.

The next project got faster. In its June 2026 prospectus, SpaceX put the first Colossus 2 cluster at 91 days, about three months.

The competition's best case

Give the other side its strongest argument. Other companies can buy Nvidia chips, hire great engineers and spend serious money on buildings. So where does a lasting edge come from?

I think it comes from how fast the whole system becomes usable and what it costs to get there. Two buyers can own the identical chip and face very different amounts of work before that chip earns anything.

Put some real projects on the calendar. Construction at Stargate's Abilene campus began in summer 2024, and its first phase was live by September 2025, about 15 to 16 months. Microsoft's first operational Fairwater building in Wisconsin was completed in June 2026, a little over two years after the plans were announced. Those clocks don't start at the same event, so treat the comparison as rough. Still, starting from an existing shell and bringing in power to begin the work goes a long way toward explaining the Colossus schedule.

SpaceX draws this link itself. Its June 2026 prospectus describes faster deployments as a way to put high-performing hardware to work sooner and keep the cost of producing AI output down. It says Colossus 2 construction costs per megawatt were considerably lower than industry benchmarks, and it ties the speed to keeping more of the work inside the company plus a shared culture. In his October 31st post, Elon credited the SpaceX team and contractors with the progress.

Buying the equipment commits the money. Getting it working gives that money something to do. Call it time to useful work: how long between committing the resources and having a machine that can do the job. That's the yardstick I'd use on any AI build.

Anthropic rents the factory

SpaceX can put this infrastructure to work two ways: develop its own AI, or supply computing power to somebody else's AI business.

In May 2026, SpaceX entered cloud agreements with Anthropic covering approximately 325,000 Nvidia GPUs across Colossus 1 and Colossus 2. In SpaceX's filing, Anthropic agreed to pay $1.25 billion per month through May 2029, following reduced fees during the initial ramp, and either side could end it with 90 days' notice after the first three months. Anthropic's own October filing shows amended deals: initial periods ending in January and March 2027, then 90 days' notice, with monthly payments rising to about $3.1 billion once all capacity is delivered.

Anthropic is buying access to working capacity. SpaceX has to keep the service available, and the customer pays monthly for that availability. At the stated rate, 12 months is $15 billion. The gap between equipment arriving and equipment becoming usable decides when there's something to sell.

Doesn't that help a competitor? Yes. Anthropic gets compute for its own business, and SpaceX gets a paying customer. Both sides have a reason to sign. My read is that this makes the project more interesting than judging it on one Grok release. A customer can pick another AI provider and still be running on SpaceX capacity.

Big revenue, thin margin

For the quarter ending September 30th, 2026, SpaceX's Super Intelligence segment generated $7.56 billion in revenue, the company's largest segment that quarter. Solutions and infrastructure, which covers cloud services and AI products, produced $7.022 billion. Advertising supplied $538 million. Computing is driving the mix.

Operating profit is what's left after operating costs. Super Intelligence produced $431 million. The Connectivity segment, which includes Starlink, produced $2.367 billion. AI led on revenue, Connectivity led on profit, and less than $6 of every $100 in AI segment revenue stayed as operating profit.

That margin is the right lens for the next building. Spend less to build useful capacity and there's less construction cost to recover. Keep chips working instead of waiting on the network and more hardware time goes to the job. Tailor an inference hall to its actual work and you skip gear it doesn't need. That's the business meaning of the training and inference split.

The next chips bring the same problem

On October 24th, 2026, SpaceX signed a binding commitment letter for a finance lease of about $40 billion in principal, with an Apollo managed lessor, covering Nvidia Vera Rubin GPUs, the next chip generation. The deal was expected to close before the end of 2026, with deliveries expected to begin in the first half of 2027.

The same problems follow SpaceX into orbit. On October 15th, 2026, Elon said the Star Mind computer, the orbital computer SpaceX and Nvidia are designing together, would reach volume production in 2027. In orbit, a computer still needs power, still has to shed heat, and still needs connections to whoever uses it. Elon was discussing orbital heat rejection as a design problem in August 2026. The ground build is where the earnings link is already concrete, and with a 90-day exit on the table, keeping Anthropic means the service has to keep working after the construction crews go home.

Back to the warehouse

A warehouse that held food. Land that used to be a gun range. Rows of computers that need electricity and cooling before they can do anything for you.

The AI answer on your screen depends on somebody getting all of that to work, and SpaceX is building a business around doing it fast. The useful part starts when those machines can handle the work you're asking for. That's when the factory starts delivering.