VGS Logo
Artificial Intelligence⏳ 5 min readPublished on October 2, 2026

Beyond the Hype: Inside Grok's Wild New Evolution (And the Secret Engine Driving It)

Grok just got a massive upgrade, a new 500K-token memory, and persistent AI agents. But the real story isn't the chatbot. It's what Grok quietly became, and who's now renting its engine room.

V

Vendora Global Solutions

Tech Writer

Beyond the Hype: Inside Grok's Wild New Evolution (And the Secret Engine Driving It)

Here's something most coverage of Grok's latest upgrade skips entirely: the company that makes it isn't really called xAI anymore. And the supercomputer built to be its secret weapon now spends a lot of its time running other companies' AI models instead.

Both of those facts are true, both are stranger than the chatbot headlines, and both explain more about where this is all heading than another spec sheet would.

Researched October 2, 2026. AI moves fast enough that details here could shift within weeks.

What Grok 4.7 Actually Does

Grok 4.7 launched on September 21, 2026, built for coding, long research tasks, and AI agents that run for hours instead of seconds. Its headline spec is a 500,000-token context window, confirmed in the company's own developer documentation and matched by independent cloud providers like Oracle and AWS. In practice, that's enough room to hand the model an entire codebase, a stack of reports, or a long research archive in one go without it losing track of the beginning.

It also reads images now, offers four levels of reasoning effort depending on how much thinking a task needs, and was specifically trained to work inside "Grok Bot," the company's system for letting an AI agent keep working on a task across sessions instead of forgetting everything the moment you close the chat.

One number you'll see everywhere, 2.1 trillion parameters, is worth a pause. It traces back to pre-launch posts on X rather than anything the company has officially confirmed in its own spec sheets, so treat it as a widely repeated rumor, not a fact. What the company has said officially: Grok 4.7 is "twice as fast, at half the price" of comparable models. That's their own claim about their own product, which is a different thing from an independently verified benchmark.

The Part the Headlines Miss: It's Not "xAI" Anymore

Here's the twist that barely made it into most coverage. In February 2026, SpaceX acquired xAI outright, in what's been described as the largest private merger on record. By July, the companies had fully merged, and xAI stopped existing as a separate brand. It's now SpaceXAI, a division of SpaceX. Grok, the X platform, and the entire Colossus supercomputer now sit inside the same corporate structure that builds rockets.

That's not a minor branding footnote. It means Grok's "secret engine" isn't a scrappy AI startup's side project anymore. It's AI infrastructure sitting inside one of the most capital-heavy, vertically integrated companies on the planet, with aerospace engineering data, satellite data, and a social network all feeding the same machine.

The Bigger Twist: The Engine Also Runs for the Competition

This is the part that genuinely surprised us. Colossus, the Memphis-based GPU campus that trains Grok, draws gigawatt-scale power and now runs several hundred thousand GPUs, with plans to push toward a million. It's a real engineering feat, built in months rather than years.

But as of May 2026, Anthropic (the company behind Claude) agreed to rent essentially the entire compute capacity of the original Colossus facility, reportedly paying over a billion dollars a month to run its own models on it. A month later, Google struck a separate deal to rent roughly 110,000 more of those GPUs starting this October. In other words, the supercomputer built to be Grok's unfair advantage now spends a meaningful share of its time as a very expensive rental property for two of Grok's direct competitors.

That doesn't make the infrastructure less impressive. If anything, it's a sign of how much GPU capacity is worth right now, when even a company racing to build its own frontier model would rather lease out the extra capacity for billions than let it sit idle. But it does complicate the "this is Grok's secret moat" narrative. The moat, it turns out, has toll booths, and rival kingdoms are paying to use it.

Why This Still Matters

None of this undercuts what Grok 4.7 actually does well. Its live connection to X's real-time post stream remains genuinely unusual among major AI assistants, letting it react to breaking news or cultural moments other models simply can't see. For developers, a half-million-token context window paired with persistent agents is a real, practical upgrade, not hype. For creators, a model trained on a less corporate, more internet-fluent tone is a legitimately different product than its more buttoned-up competitors.

What's changed is the frame. This isn't a scrappy challenger quietly building an edge in the shadows. It's a trillion-dollar aerospace company's AI division, whose own training supercomputer now partly functions as critical infrastructure for the very companies it's trying to outpace.

So here's the actual question worth asking, more interesting than "will you let an AI agent run your projects": would you trust your AI infrastructure to a company whose core data center doubles as a rental unit for its rivals? Or does that just prove the infrastructure is as good as they say it is?

Sources: SpaceXAI developer documentation, Oracle Cloud and AWS release notes, Wikipedia's entries on Colossus and SpaceXAI, and public reporting on the SpaceX-xAI merger.