Musk says Grok 2 is coming in August, Grok 3 by year-end

Training an AI on the modern internet is turning into a nightmare because the web is full of content written by other AI models.

Elon Musk just revealed that xAI is working hard to purge LLM-generated junk from its datasets, and the upcoming Grok 2 will deliver a major leap forward when it arrives in August.

Musk also confirmed that xAI is not stopping there, with an even more capable Grok 3 scheduled to drop before the end of the year.

Cleaning dirty data and channeling JARVIS

When xAI first rolled out Grok last November to challenge OpenAI, it leaned into wit and real-time updates. With Grok 2, Musk claims the assistant draws inspiration from iconic sci-fi entities like Iron Man's JARVIS and The Hitchhiker's Guide to the Galaxy.

The real technical battleground is data hygiene. Musk admitted that filtering synthetic model output from raw web data is brutal, but promised Grok 2 represents a giant improvement on this exact front.

Alongside real-time web access and image generation features, the new model is said to show noticeable gains on core benchmarks like MATH and HumanEval, pointing toward sharper reasoning and fewer factual hallucinations.

A 100,000-GPU monster waiting in the wings

While Grok 2 handles the summer update, the heavier artillery arrives late this year with Grok 3.

xAI is reportedly preparing a cluster powered by 100,000 Nvidia H100 GPUs to train Grok 3, aiming to brute-force a new frontier in raw performance.

Hardware vendors are lining up behind the effort. Dell Technologies CEO Michael Dell confirmed that his company is teaming up with Nvidia to build a specialized AI factory explicitly backing Musk and xAI.

Why it matters

For regular users and X subscribers, Grok 2 should immediately show whether xAI can deliver a serious daily driver rather than just a novelty bot. If the reasoning gains hold up, users will get a much sharper research tool inside their social feed.

On the industry side, Musk is attacking the two biggest bottlenecks at once: data quality and compute scale. Purging recycled synthetic text from datasets is now a survival requirement for frontier labs to avoid model degradation.

Backing that clean data with 100,000 H100 chips is an aggressive flex that directly challenges OpenAI and Google. Still, Musk's ambitious timelines are notorious for slipping, and successfully wrangling a cluster of that scale by year-end is an immense engineering gamble.