Musk says Grok 3 is done pretraining with 10x more compute

Elon Musk just confirmed that xAI has finished pretraining Grok 3, and the next-generation model is coming soon.

According to Musk, training Grok 3 took 10 times more compute power than Grok 2. He dropped the news on X while replying to a user who claimed they canceled their ChatGPT subscription after four weeks of using Grok, praising its speed and reliability.

Powered by 100,000 H100s

While Musk did not share detailed specs, the timing indicates that Grok 3 was trained on xAI's Colossus supercluster in Memphis, which houses roughly 100,000 Nvidia H100 GPUs.

Training modern frontier models with hundreds of billions of parameters requires trillions of floating-point operations. The tenfold jump in compute directly reflects the massive hardware horsepower of Colossus.

Naturally, Grok 3 also fed on real-time data generated by users across X, giving it a unique pipeline of conversational information.

And xAI is not slowing down. Reports indicate Colossus will soon expand to 200,000 GPUs, combining H100 and newer H200 chips for the next model iteration, with long-term ambitions targeting a cluster powered by more than one million GPUs.

Why it matters

For users, Grok 3 represents the first real test of whether xAI can deliver a top-tier frontier model that genuinely rivals GPT-4o. If the model translates its massive compute budget into sharper reasoning and fewer hallucinations, it could become a serious daily driver for professionals.

For the AI race, xAI's blistering speed should put serious pressure on OpenAI, Google DeepMind, and Anthropic. Building a 100,000-GPU cluster and pushing a major model through pretraining in record time proves their infrastructure execution is world-class.

We think Musk's brute-force approach to scaling compute is both awe-inspiring and risky. Ramping up hardware at this scale guarantees fast iterations, but whether throwing 10X the silicon at the problem produces proportional gains in deep reasoning remains the real question.