😼 The Gigawatt That Wasn't

The claim: xAI's Colossus 2 will be the world's first gigawatt AI cluster. Elon said it on X. Multiple times. With the confidence of a man who once promised a million robotaxis by 2020.

Let's do the math nobody at xAI apparently did. 😹

Colossus 1 in Memphis launched with roughly 100,000 H100 GPUs crammed into a facility that had contracted around 150MW of grid power. When that wasn't enough, xAI parked diesel generators outside — so many that Shelby County residents filed air quality complaints and the Tennessee Department of Environment opened an investigation. The solution to not having enough clean power was, apparently, more dirty power.

Now Colossus 2. The "gigawatt" figure comes from xAI's own announcements and Musk's posts. But the actual contracted utility power for the expanded Memphis campus? Reports from local utility filings and energy analysts put it between 300-350MW, with "plans" to reach higher numbers through a combination of grid expansion, on-site generation, and what I'd charitably call optimistic future commitments. 😾

Satellite imagery tells the rest of the story. Commercial overhead captures of the Memphis site show a facility footprint consistent with a 300-400MW operation — not a gigawatt one. A true 1GW data center requires massive cooling infrastructure: think rows of cooling towers, chiller plants, and water treatment facilities on the scale of a small power station. What's visible at xAI's campus is a fraction of that. Compare it to Meta's Luleå data center in Sweden or Microsoft's Goodyear campus in Arizona — facilities in the 200-300MW range whose cooling footprints already dwarf what Colossus 2 shows from orbit. Several building pads on the Memphis site are visibly still under construction, with foundation work incomplete and no mechanical equipment installed. The "gigawatt" includes buildings that don't exist yet. 😾

Three hundred fifty megawatts is impressive. It's genuinely one of the largest AI training facilities on the planet. But it is not a gigawatt. It's a gigawatt the same way Grok 5 was a Q1 2026 release — technically a number someone said out loud.

Meanwhile, the actual output of all this power: Grok 3 benchmarks that trail Claude Sonnet on coding, a chatbot bolted onto a social network hemorrhaging advertisers, and a flagship model that missed its own deadline by at least a quarter.

The verdict: xAI built a genuinely massive cluster and then marketed it at 3x its actual capacity. In any other industry, that's called fraud. In AI, it's called a press release. 😹

The gigawatt will probably arrive eventually. So will Grok 5. So will Full Self-Driving. The pattern isn't ambition — it's the chronic inability to distinguish between "we plan to" and "we did."

🐈‍⬛ A real gigawatt cluster would power ~800,000 American homes (at average US household draw of ~1.2kW). Colossus 2 powers one man's valuation narrative.

TechCrunch · Ars Technica