NVIDIA Confirms xAI’s Grok 4.6 AI Model Trained on Its Most Powerful Blackwell Ultra Hardware

NVIDIA posted a congratulatory message on X to the “@SpaceXAI team” after the release of Grok 4.6, saying the model brings “frontier intelligence” and is “running and trained on NVIDIA GB300 NVL72 with NVLink.” The post is one of a series NVIDIA has made about xAI’s Grok releases, following a similar announcement when Grok 4.5 launched on the same hardware.

The move puts one of the most powerful AI systems currently available squarely on NVIDIA’s newest and most capable platform — and gives the chip giant another high-profile name to attach to its Blackwell Ultra architecture.

What Is the GB300 NVL72?

The GB300 NVL72 is not a single chip. It’s a full rack-scale system — liquid-cooled, built for industrial-scale AI work, and packing 72 NVIDIA Blackwell Ultra GPUs alongside 36 Arm-based NVIDIA Grace CPUs. NVIDIA’s own documentation says the platform can deliver up to a 50x increase in AI factory output performance compared with its older Hopper-based systems.

That’s a staggering leap on paper. NVLink, the high-speed chip-to-chip interconnect NVIDIA uses instead of standard network cables between GPUs, allows all 72 GPUs to share memory and communicate at speeds that would be impossible over conventional links. The result is a system designed to train and run the kind of large, demanding models that would otherwise require dozens of separate servers stitched together.

Both Grok 4.5 and now Grok 4.6 have been trained on this platform, according to NVIDIA’s posts.

What NVIDIA Is Claiming About Grok 4.6

NVIDIA’s LinkedIn post goes further than the X announcement. It describes Grok 4.6 as faster than comparable models and says it can handle more challenging tasks than Grok 4.5. The company frames it as a meaningful improvement — not just a minor update — delivered at the same price point.

That price point, as stated by NVIDIA, is around £1.55 per million input tokens and around £4.65 per million output tokens — converted from the posted figures of $2 and $6 respectively.

For context, token pricing is how AI companies charge developers and businesses for using a model via an application programming interface, or API. The more text a model reads and generates, the more tokens it uses, and the higher the bill. Keeping that price flat while improving capability is the kind of thing that gets developers’ attention.

Where Can You Use It?

According to NVIDIA’s LinkedIn post, Grok 4.6 is available through several channels. Developers can access it via the API, and it’s also integrated into Grok Build, Cursor — the AI-assisted coding tool — and Grok Bot. That spread of availability suggests xAI is pushing the model into both professional and consumer-facing products at the same time.

The exact release date of Grok 4.6 has not been confirmed by an official xAI statement in the materials available, though community references point to August 2026. The model’s existence and hardware details are confirmed by NVIDIA’s own posts.

NVIDIA and xAI: A Growing Partnership

This isn’t a one-off endorsement. NVIDIA has now publicly backed multiple Grok releases, each time naming the specific hardware underneath. That pattern matters. When NVIDIA calls out GB300 NVL72 by name in a congratulatory post, it’s partly a product announcement of its own — a demonstration that the world’s most talked-about AI labs are choosing Blackwell Ultra over rival platforms.

Jensen Huang, NVIDIA’s chief executive, has spoken repeatedly about the company’s ambition to be the backbone of the AI industry rather than a component supplier. These posts fit that strategy neatly. XAI, Elon Musk’s AI company, has not issued a separate public statement about Grok 4.6 that was available at the time of writing. The details circulating publicly come primarily from NVIDIA’s own social media activity.

But the hardware story is clear enough. A model trained on 72 Blackwell Ultra GPUs, connected by NVLink at rack scale, represents the current ceiling of commercially deployed AI infrastructure. Whether Grok 4.6’s actual benchmark performance matches NVIDIA’s billing is something independent researchers will test in the coming weeks.

What This Means for Kent Residents

There’s no direct Kent angle here — no local employer, university, or public body is named in connection with Grok 4.6. But for developers, small businesses, and public-sector teams across the county who use AI tools in their day-to-day work, the arrival of a more capable model at unchanged API pricing is straightforwardly good news. Lower cost per token, better performance, and wider availability across tools like Cursor means more capable AI assistance without a higher bill.