xAI's Colossus supercomputer started with 100,000 GPUs. Three months later they doubled it to 200,000. Now they're planning to hit one million GPUs, making it the largest AI training system in the world by a ridiculous margin. It's impressive engineering. It's also burning through Memphis's power grid, polluting the air with gas turbines, and sidestepping environmental permits because the equipment is technically "portable."
The Scale Is Absurd
Colossus currently runs 150,000 H100 GPUs, 50,000 H200 GPUs, and 30,000 GB200 GPUs. That's 230,000 total GPUs in active operation. Another 110,000 GB200 GPUs are being added at a second Memphis data center. The expansion to one million GPUs would make Colossus roughly 25 times larger than most national supercomputers.
For perspective: El Capitan, currently the world's fastest supercomputer, has about 44,000 GPUs. Frontier has around 38,000. Colossus is already five times larger than those, and Musk wants to make it 20+ times bigger still.
xAI built the first phase—100,000 GPUs—in 122 days. That's insane. Most data centers take years to plan and construct. They went from concept to operational in four months by repurposing an abandoned Electrolux manufacturing facility in South Memphis.
The Environmental Disaster
Here's where it gets ugly. The site only had an 8MW grid connection. Running 200,000+ GPUs requires hundreds of megawatts. So xAI set up more than 35 gas turbines—portable units that burn methane to generate about 422MW of power.
The Southern Environmental Law Center says xAI is operating those turbines without proper air permits, which violates federal law. The emissions make Colossus likely the largest industrial emitter of nitrogen oxides (NOx) in Memphis—between 1,200 and 2,000 tons annually.
NOx creates smog and aggravates respiratory issues. Memphis residents living near the facility have voiced complaints about air quality at public hearings. One woman said she smells "everything but the right thing and the right thing is the clean air."
xAI argues the turbines are "portable" so they don't need major source permits. That's technically legal but ethically questionable. You can't permanently run a data center on dozens of gas generators and call it a temporary setup.
The Power Grid Problem
The Tennessee Valley Authority approved a deal to supply 150MW of grid power to the site. But even that won't be enough long-term. xAI is installing 168 Tesla Megapacks for battery backup and plans to add more to reduce dependence on Memphis's gas and water grid.
The problem is that Memphis's grid was already strained. TVA proposed building additional gas-powered turbines last year because they couldn't meet demand. Now xAI is pulling massive amounts of power for AI training while local residents deal with the pollution.
KeShaun Pearson, president of Memphis Community Against Pollution, called it "policy violence" and said the community deserves "clean air, not silent strangulation." Environmental groups argue TVA should have studied the impact before approving the power deal.
Why Memphis?
xAI chose Memphis because the abandoned Electrolux building could be repurposed quickly and the site was near existing power infrastructure. The Greater Memphis Chamber celebrated it as the largest capital investment in the region's history, bigger than Ford's $5.6 billion Blue Oval City plant.
The expansion is expected to create 320 permanent jobs. Nvidia, Dell, and Supermicro are all establishing operations in Memphis to support Colossus. From an economic development perspective, this is a massive win for the city.
But economic benefits don't offset environmental harms if you're the person living downwind from gas turbines burning methane 24/7. The jobs and tax revenue go to different people than those dealing with the air quality issues.
The Technical Achievement
Setting aside the environmental concerns for a moment: what xAI built is genuinely impressive. Constructing and deploying 100,000 GPUs in 19 days is unprecedented. Doubling it to 200,000 in three months is absurd. Scaling to one million is technically feasible but logistically insane.
Colossus uses Supermicro 4U liquid-cooled servers, each with eight H100 GPUs. The liquid cooling is necessary because these chips generate enormous heat. The entire facility is essentially a massive heat management problem combined with a power generation challenge.
NVIDIA CEO Jensen Huang said "there's only one person in the world who could do that; Elon is singular in his understanding of engineering and construction and large systems." That's not just hype—the speed and scale are legitimately remarkable.
The Strategic Play
Colossus trains Grok, xAI's ChatGPT competitor. The latest version, Grok 4.1, scored competitively with GPT-5 and Claude on benchmarks. Musk claims Grok 3 (currently training on Colossus) will be "the smartest AI on Earth."
Having the largest training cluster gives xAI a major advantage. They can iterate faster, train larger models, and run more experiments than competitors. OpenAI and Anthropic rely on cloud providers. xAI owns their infrastructure outright.
The compute capacity also supports Musk's other companies. Colossus can train models for Tesla's autonomous driving, improve X's recommendation algorithms, and support SpaceX's operations. It's shared infrastructure across the entire Musk empire.
The Funding Question
xAI is reportedly raising $15 billion at a $230 billion valuation. The majority of that money will fund infrastructure expansion, including scaling Colossus to one million GPUs.
The math is staggering. NVIDIA H100 and H200 GPUs cost roughly $25,000-40,000 each. A million GPUs at $30,000 average is $30 billion just for the chips, not including servers, cooling, power infrastructure, or the data center itself.
xAI needs to raise that capital because nobody else—not even OpenAI or Google—is committing this level of resources to a single training facility. It's a massive bet that having the largest compute cluster will translate to the best AI models.
What Could Go Wrong
Besides the environmental issues, there are technical risks. Power outages could crash the entire system. Hardware failures across a million GPUs would be constant. Cooling a facility this large in Memphis summers is non-trivial.
There's also the question of whether this much compute is necessary. DeepSeek trained competitive models in China using far fewer GPUs through algorithmic efficiency. If the bottleneck shifts from compute to algorithms, xAI's infrastructure advantage evaporates.
And politically, building the world's largest AI supercomputer in a neighborhood already dealing with pollution issues creates backlash that could lead to regulatory intervention. Tennessee officials have been hands-off so far, but that could change if pressure builds.
My Take
I'm torn on this one. The engineering is genuinely impressive and I respect the ambition. Musk is doing what he always does—pushing scale beyond what anyone thought possible and moving faster than bureaucracy can respond.
But the environmental impact is real and the community concerns are legitimate. You can't just roll into South Memphis, set up 35 gas turbines, and pollute the air because you're racing to beat OpenAI. That's not acceptable regardless of how important AI development is.
The right move would have been to delay construction until proper permits were secured and cleaner power sources were available. But Musk doesn't do delays. He builds fast, deals with consequences later, and often gets away with it because the final result is impressive enough to retroactively justify the mess.
Whether that works this time depends on how bad the air quality gets and how much political will exists to force xAI to clean up their operation. Based on past patterns, they'll probably get a slap on the wrist, pay some fines, and keep building.
But watching the largest AI infrastructure project in history being built on the backs of a low-income community dealing with the pollution? That's the part nobody's talking about when they celebrate xAI's technical achievements.