Writer: Jordan Sprogis
Editor: Lillian Castro
Reviewer: Cristian Lopez
Key Takeaways
- xAI revealed Grok 3, a new chatbot that can perform advanced reasoning tasks, self-correct and refine, and even create games
- Compared to its competitors, Grok 3 uses more GPUs — but also outperforms on several benchmarks
- xAI built its own data center, Colossus, to power Grok 3, scaling to 100,000 GPUs in just four months
Elon Musk took to X for a livestream to unveil Grok 3, xAI’s most powerful chatbot yet. Grok 3 is said to be “an order of magnitude more capable than Grok 2.” Its predecessor was released only in August 2024, and the third iteration comes with some major upgrades, having undergone 10 times more training.
Musk introduced Grok 3 by reiterating xAI’s core mission: “In order to understand the nature of the universe, you must absolutely rigorously pursue truth, or you will not understand the universe; you will be suffering from some amount of delusion or error,” he said on February 17. “So that is our goal: to figure out what is going on.”
He alluded that Grok (likely some future version of it) may help answer some of life’s biggest existential questions: where aliens come from, the meaning of life, and how the universe might end.
Three members of xAI’s team — Igor Babuschkin, Jimmy Ba, and Tony Wu — joined Musk for the livestream event, where they led a 50-minute demonstration showcasing Grok 3 and future goals.
Inside Grok 3
The word “grok” was coined by Robert A. Heinlein in his 1961 science fiction novel, “Stranger in a Strange Land.” “Grok” is an adjective that means to understand something deeply.
With that said, Grok 3 has several thinking modes, depending on the use case:
- DeepSearch: Designed to compete with search engines like Google or Bing; ideal for tasks like market research and content creation
- Think: A problem-solving mode that breaks down complex questions, fact-checks each component, and provides logical answers; ideal for academics, strategy, and analysis
- Big Brain: A computation-heavy mode that boosts Grok 3’s power to tackle advanced and complex tasks, including scientific problem-solving and coding
Grok 3’s continuous reinforcement training makes sure it keeps improving. The goal is to apparently enhance its advanced reasoning to refine its abilities during both training and real-world use.
Babuschkin demonstrated Grok with two complex prompts: create a game based on Tetris and Bejeweled and answer a physics question. In just minutes, Grok created a full script with in-app play capabilities.
For the physics prompt — “Generate code for an animated 3D plot of a launch from Earth, landing on Mars, and then returning to Earth at the next launch window” — Grok reasoned through the solution in real time.
The chatbot even verbalized its thought process, making statements like, “Perhaps it’s easier to use the concept of phasing,” before refining its approach.
Musk connected this to SpaceX’s goals, referencing the next Mars transit window in November 2026. He said: “If all goes well, SpaceX will send Starship rockets to Mars with Optimus robots — and Grok.”
Grok 3 vs. DeepSeek R-1 vs. ChatGPT-4o
The million dollar question is how Grok 3 compares to other leading AI models.
It’s obvious Grok 3 is a response to ChatGPT. Besides the fact every major tech company is releasing its own bot nowadays, xAI was founded in early 2023, just months after ChatGPT took off.
Musk also said xAI’s founding members and engineers are driven by curiosity, ultimately describing Grok 3 as a truth-seeking AI, “even if that truth is sometimes at odds with what is politically correct.”
It’s a clever slight to OpenAI, whose model is known for its stricter content moderation policies, especially when it comes to political, violent, or sexual content.
Grok 3 has also outperformed other models, including GPT-4o and Gemini, in several benchmarks, with a #1 ranking on Chatbot Arena.
There is little else we know about Grok since xAI hasn’t released more information, like a technical report. But we do know the xAI team has big goals.
“The best AI needs to think like a human, needs to contemplate about all the possible solutions, self-critique, verify, all the solutions, backtrack, and also think from the first principle,” Ba said.
Grok 3’s power comes at a steep cost. Grok 3 required more than 100,000 NVIDIA H100 GPUs — four times more than its closest competitor. In contrast, DeepSeek R-1 runs on just 2,048 H800 GPUs, while ChatGPT-4o uses 25,000 A100 GPUs.
Powering an AI of this scale demands an immense infrastructure. It’s why xAI built its own data center.
“Elon decided that really the only way for xAI to succeed, for xAI to build the best AI out there, is to build our own data center,” Babuschkin said.
The data center, Colossus, was completed in September 2024. Housed inside a former Electrolux factory in Memphis, Colossus started with 100,000 NVIDIA H100 GPUs, which Tom’s Hardware estimates consumes around 150 megawatts (MW) of power. By December 2024, xAI decided to double its GPU count to 200,000, or around 250 MW.
“We believe it’s the biggest, fully connected H100 cluster of its kind,” Babuschkin said.
Whether it is or not, it’s clear that xAI doesn’t intend to compete with the likes of OpenAI or Google; it wants to create something with the DNA to truly “understand the universe.”
