Elon Musk's SpaceXAI to add another 660,000 AI GPUs this year, nearing a total of 1.44 million in operation — firm is building 1.2-gigawatt power plant to bring systems fully online
That's a lot of GPUs.
The end goal is finally in sight for Elon Musk, nearly two years after he announced plans to expand the Colossus supercomputer to over a million GPUs. The billionaire said on X that 220,000 Nvidia GB300 GPUs will be operational by next week, with another 220,000 coming online in November. He also added that another 220,000 units will come online by late December “if we get lucky.”
Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300. Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December.September 25, 2026
These numbers would add to the 110,000 GB200 and 440,000 GB300 GPUs already operating at Colossus 2, plus the 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs at Colossus 1. This would bring SpaceXAI’s GB300 GPUs to 1.1 million units, with its total GPUs in operation to 1,440,000 units. This is quite an achievement, especially given that the company is behind its rivals by several years — SpaceXAI is only three years old, while Anthropic and OpenAI are six and ten years old, respectively. Interestingly, the Colossus 1 site, which features a combination of Hopper and Blackwell GPUs, is inefficient for training Grok, so Musk rented it out to Anthropic for inference instead. On the other hand, Colossus 2 solely uses Blackwell GPUs, ensuring that there won’t be any bottlenecks.
While Elon Musk essentially begged for Jensen Huang to give him these GPUs, getting his hands on them isn’t currently the hardest part. One of the biggest issues facing data center build-outs right now is power, and SpaceXAI solved this by bringing its own. This has resulted in some controversies with the surrounding neighborhood, ending in a lawsuit against the company for running unpermitted gas turbines. It has since promised to remove them, but only over a span of one year as its own 1.2-GW power plant comes online.
Running a million GPUs or more on a single site is an impressive achievement, but SpaceXAI isn’t the only one trying to reach this goal. Broadcom said in 2024 that it has three hyperscale customers gunning for this target by 2027 but didn’t mention who these customers were. As for Elon Musk, the 1-million-GPU target is just the beginning. He said that SpaceXAI will grow its data center capacity sevenfold by 2027, and he’s even aiming for 50 million H100-equivalent GPUs by 2030. These data centers won’t be limited to the ground as well, with SpaceX planning to launch an Orbital Data Center System with a million satellites, despite Jensen Huang saying that it’s a “dream” for now.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.
-
DS426 Quit saying "online" -- what does that mean exactly? They are installed and now powered on for the first time, actually starting utilization for AI training or inference, or what? The market demands so much AI compute allegedly, so what are those compute numbers?Reply -
alan.campbell99 Reply
Indeed, somewhat like Micro$oft claiming they bring a GW of data centre capacity online every quarter, trying to imply this is AI capacity when there's no proof of such. Or claims about Stargate Abiline coming online when only a couple of the 8 buildings planned are actually energised, said project is limping along and it's not the only one.DS426 said:Quit saying "online" -- what does that mean exactly? They are installed and now powered on for the first time, actually starting utilization for AI training or inference, or what? The market demands so much AI compute allegedly, so what are those compute numbers?
I'd also like to see a bit better reporting about the so-called 'massive demand' for GPUs, that millions of these are getting bought when these data centres are taking 2-3 years to complete so in the meantime these chips are just sitting around somewhere. How many of these millions of sold chips have actually been energised, such numbers seem to be a challenge to find. I've heard estimates across the hyperscalers and neo clouds that it could be hundreds of billions worth of idle chips. This is from what can be gleaned from construction in progress details in their financial reports . -
Bigshrimp This whole thing is just one of the biggest Ponzi schemes ever. I wonder how much longer until we see something like what happened to Enron...Reply -
blppt Reply
Well, its not likely to happen until 2029 at the earliest given that most of the Federal government is under the thumb of one man who happens to love the sugar rush of the AI industry.Bigshrimp said:This whole thing is just one of the biggest Ponzi schemes ever. I wonder how much longer until we see something like what happened to Enron... -
Zaranthos They're not buying this much compute for no reason. It will get used to rapidly train new models. Grok has advanced remarkably fast and more compute coming online makes perfect sense if the goal is training larger models quickly along with parallel reinforcement learning. Full capacity compute might not be needed all the time but it definitely has value for a wide variety of situations.Reply
Anyone who has used Google AI features is probably well aware of how often it hallucinates and spits out pretty convincing AI slop. I now use Grok more often because it's usually far more accurate. I can prod Google and sometimes get much better answers when I know it's failing but that doesn't do me much good when I'm asking it about things I don't know enough about to recognize problems with accuracy. More compute was required for Grok when they baked in extra fact checking algorithms but the results were tangible improvements.
AI is not nearly good enough yet. It hasn't reached anything close to the potential I'd like to see. As it gets even better the user base will grow exponentially and the required compute will increase as well. It's very likely some very big AI companies will fail, go bankrupt, and be gobbled up by competitors, and that's also how business works. Before it's all said and done human progress should advance and improve a lot. I'm exited that it's very likely in my lifetime I'll see SMR nuclear power widely available which would make a much healthier and robust power grid along with cheaper electric rates. The Internet will continue to get faster, more robust, and more reliable as terrestrial and space based broadband compliment each other. I also want an advance AI humanoid robot that I can put to work doing all kinds of random tasks. -
justrudi Reply
It is always a wonder to me how one can outsource and trust thinking, judgment, or research to a tool made by a fascist-leaning billionaire arguing about the humanity, truth and even right for morality's existance.Zaranthos said:They're not buying this much compute for no reason. It will get used to rapidly train new models. Grok has advanced remarkably fast and more compute coming online makes perfect sense if the goal is training larger models quickly along with parallel reinforcement learning. Full capacity compute might not be needed all the time but it definitely has value for a wide variety of situations.
Anyone who has used Google AI features is probably well aware of how often it hallucinates and spits out pretty convincing AI slop. I now use Grok more often because it's usually far more accurate. I can prod Google and sometimes get much better answers when I know it's failing but that doesn't do me much good when I'm asking it about things I don't know enough about to recognize problems with accuracy. More compute was required for Grok when they baked in extra fact checking algorithms but the results were tangible improvements.
AI is not nearly good enough yet. It hasn't reached anything close to the potential I'd like to see. As it gets even better the user base will grow exponentially and the required compute will increase as well. It's very likely some very big AI companies will fail, go bankrupt, and be gobbled up by competitors, and that's also how business works. Before it's all said and done human progress should advance and improve a lot. I'm exited that it's very likely in my lifetime I'll see SMR nuclear power widely available which would make a much healthier and robust power grid along with cheaper electric rates. The Internet will continue to get faster, more robust, and more reliable as terrestrial and space based broadband compliment each other. I also want an advance AI humanoid robot that I can put to work doing all kinds of random tasks. -
Zaranthos Replyjustrudi said:It is always a wonder to me how one can outsource and trust thinking, judgment, or research to a tool made by a fascist-leaning billionaire arguing about the humanity, truth and even right for morality's existance.
I'd prefer AI slop to what you just regurgitated. Our own brains produce AI slop when we feed them garbage source material that reduces everything to OMG a fascist!
Might as well lock the topic, it only goes downhill from here.