xAI Launches Grok 3, Trained on 200,000 GPUs in Memphis

Elon Musk's AI company, xAI, released its latest flagship model, Grok 3, late Monday night, alongside new features for its iOS and web apps. It's Musk's bid to go head-to-head with OpenAI's GPT-4o and Google's Gemini, arriving after months of development and a delay from the 2024 release date the company had originally set for itself.

Grok is the assistant that powers several features on X, Musk's social network. It can analyze images and answer questions. Grok 3 expands on those capabilities and, above all, boasts an unusual scale of training.

A Data Center with 200,000 GPUs

The figure xAI has put at the center of its pitch is infrastructure. The company trained Grok 3 in a massive Memphis data center housing roughly 200,000 GPUs, the graphics processing units used for the huge calculations required to train a language model.

In a post on X, Musk claimed Grok 3 was developed with "10x" (or so) more computing power than its predecessor, Grok 2, using an expanded training set that includes filings from court cases — and more.

"Grok 3 is an order of magnitude more capable than Grok 2," Musk said during a livestreamed presentation on Monday. He added a description of his product that sums up its whole approach: "[It's a] maximally truth-seeking AI, even if that truth is sometimes at odds with what is politically correct."

That GPU count matters because it points to a broader industry trend: the AI frontier is increasingly reached by stacking capital and silicon. xAI hasn't invented some revolutionary technique — it has simply amassed enormous compute power to close the gap with labs that had a head start.

A Family of Models, Not Just One

Grok 3 isn't a single model — it's a family. There's a smaller version, Grok 3 mini, which answers faster at the cost of some accuracy. Not every variant and feature was available on day one — some remain in beta — but the rollout began Monday.

According to xAI's own figures, Grok 3 beats GPT-4o on benchmarks including AIME, which evaluates a model's performance on a sampling of math problems, and GPQA, which assesses models on PhD-level physics, biology and chemistry problems. An early version of Grok 3 also scored competitively on Chatbot Arena, a crowdsourced test where users vote on which model's response they prefer, according to xAI.

These numbers deserve the usual grain of salt: they come from the maker itself, and benchmark comparisons between models are notoriously slippery, sensitive to exactly how each test is run.

Models That "Think"

Two of the variants, Grok 3 Reasoning and Grok 3 mini Reasoning, can carefully "think through" problems, similar to so-called reasoning models like OpenAI's o3-mini or Chinese AI company DeepSeek's R1. These models try to fact-check their own steps before delivering an answer, which helps them avoid some of the pitfalls that typically trip up conventional systems.

xAI claims Grok 3 Reasoning surpasses the best version of o3-mini — o3-mini-high — on several popular benchmarks, including a newer one called AIME 2025.

In the app, users can ask Grok 3 to "Think," or, for tougher queries, switch on a mode called "Big Brain" that dedicates extra compute to reasoning. xAI describes these models as best suited for math, science and coding questions.

There's a telling detail about how the lab protects itself here: Musk explained that some of the reasoning models' "thoughts" are hidden in the app to prevent distillation, a technique developers use to extract knowledge from other companies' models. It's not an abstract worry — DeepSeek was recently accused of distilling OpenAI's models to build its own.

DeepSearch and the Race for AI-Powered Search

These reasoning models underpin a new app feature called DeepSearch, xAI's answer to AI research tools like OpenAI's deep research. DeepSearch scans the internet and X to analyze information and return a summary in response to a question.

The fact that the tool draws on X among its sources is no small detail: it ties the model to Musk's social network, an asset none of its rivals have.

Pricing and Rollout

Subscribers to X's Premium+ tier, which costs $50 a month, will get access to Grok 3 first. Other features will sit behind a new plan xAI is calling SuperGrok. Priced, according to leaks, at $30 a month or $300 a year, it would unlock additional reasoning and DeepSearch queries plus unlimited image generation.

Musk previewed several upcoming additions with rough timelines: within about a week, the Grok app would gain a "voice mode," giving the models a synthesized voice. A few weeks after that, Grok 3 models would become available through xAI's enterprise API, alongside the DeepSearch feature.

Grok 2 Heads Toward Open Source

Musk announced that xAI plans to open-source Grok 2's code in the coming months. "Our general approach is that we will open source the last version [of Grok] when the next version is fully out," he said. "When Grok 3 is mature and stable, which is probably within a few months, then we'll open source Grok 2."

It's a middle-ground policy: not an open model from launch, but not a permanent lockdown either. The previous version gets released once the new one is already on the market.

The Promise of a "Non-Woke" AI

When Musk unveiled Grok roughly two years ago, he pitched it as an edgy, unfiltered, anti-"woke" AI, willing to answer controversial questions other systems dodge. It delivered on part of that promise: told to be vulgar, Grok and Grok 2 would happily oblige, spewing language you'd be unlikely to hear from ChatGPT.

But the models prior to Grok 3 hedged on political topics and wouldn't cross certain lines. In fact, one study found that Grok leaned to the political left on issues like transgender rights, diversity programs and inequality.

Musk has blamed that behavior on Grok's training data — public web pages — and pledged to "shift Grok closer to politically neutral." It's not yet clear whether xAI has achieved that goal, or what the consequences of trying might be. That adjustment is the project's most delicate point: deliberately nudging a model's bias toward a "neutral" position raises the question of who gets to define neutrality, and by what standard.

The launch puts xAI firmly in the top tier of major AI labs, with a roadmap stretching just weeks ahead — voice mode, enterprise API — that will need to be verified once it actually arrives.