On September 2, Elon Musk said Grok 4.7 would arrive in about ten days — a target of roughly September 12 — and by the time the Hacker News thread on it cleared 513 points and 428 comments, the argument had already assembled itself. Two-point-one trillion parameters, up from Grok 4.6’s 1.5 trillion at its August 12 release. Forty percent bigger, therefore better, or at least more. The benchmark chart will be produced by xAI, disputed by rivals, and forgotten by Halloween.

The number that matters is the one nobody has put a price on.

The Parameter Count Is the Decoy

Frontier model releases used to be events. Now they’re a metronome, and the metronome is the point. Grok 4.5, 4.6 and 4.7 have landed inside roughly two months. Each drop resets the conversation before the previous one can be examined. Notice what was missing from the September announcement: pricing. Grok 4.6 carried public rates — $2.00 per million input tokens, $6.00 per million output. Grok 4.7’s listing has no equivalent. You cannot comparison-shop a model whose price is still TBD, which is a tidy arrangement if your business depends on nobody comparison-shopping.

So the scoreboard becomes a parameter count: a metric that flatters the seller and cannot be verified from outside the building. Everyone knows this. What’s stranger is that the same crowd that has spent three years demanding transparency in AI will spend release week arguing about a chart.

SpaceX Data Is Not a Market Good

In mid-August, Musk described Grok 4.7’s supplemental training as drawing on a large volume of SpaceX company data. Read that slowly, because it describes something structurally different from every training-data fight we have had so far.

When OpenAI scrapes the open web, or licenses an archive, there is at least a counterparty. When a model ingests public code, there is a public. Those fights are ugly, but they are market fights — somebody owns the thing, somebody else wants it, and a price eventually emerges or a court imposes one.

SpaceX’s engineering records are not a market good. They cannot be bought at any price by Anthropic, by OpenAI, by Google, by you. They moved from one entity under Musk’s control to another entity under Musk’s control, in a transaction with no disclosed price, no disclosed scope, and no disclosed terms. That is the moat. Not the parameters. Not the context window. A captive corpus is the one training asset a competitor cannot assemble with capital.

Which makes Grok 4.7 less a product launch than a related-party transaction wearing a product launch’s clothes.

The Right Has the Right Tools and Won’t Pick Them Up

Here is where I part company with my own side.

If you believe in shareholder primacy, disclosure, arm’s-length dealing, and a healthy suspicion of insiders shuttling assets between companies they happen to control at both ends, this is your fight. Tesla ships Grok into its vehicles. SpaceX data trains the model. xAI sells it. One man sits at every vertex. Ask the obvious questions. What did SpaceX receive for its engineering records? What did Tesla shareholders get for turning their cars into a distribution channel for a private company’s product? What was the transfer price, and who approved it?

A second-year associate on a public-equity desk put it better than any think tank will. “We don’t have a line item for it,” she wrote, mid-session, in a Slack thread that was forwarded to me. “That’s the problem. The moat is an internal transfer, so it shows up nowhere and it can’t be valued.”

Nothing xAI has published answers her. And the instinct on the right — to wave this off because the man is on our team, or because the alternative is a new federal agency with a new letterhead — is how you end up defending vertical integration you would denounce in any other industry.

Ninety Minutes of Shared Infrastructure

One more item from the same week, logged by an industry newsletter that tracks xAI’s release cadence: ChatGPT, Claude and Grok all went down within about ninety minutes of one another. Coincidence, reportedly. A useful one. The industry’s entire public argument is about which model is smartest, while the more consequential fact is that the three leading systems share enough underlying plumbing to fail in formation.

You don’t need a new regulator to care about that. You need procurement officers who ask what happens on the day the shared thing breaks.

Grok 4.7 will be benchmarked, praised, dunked on, and superseded by 4.8 before the leaves turn. The version number is marketing. The question worth carrying into October is narrower and duller: what exactly came out of SpaceX, what was it worth, and who signed off?

If nobody asks, the answer will be whatever number is most convenient — which is how these things always get priced.

Sources