At some point Monday morning, Artificial Analysis posted a number, and by afternoon it was on the front page of Hacker News with 424 points and 345 comments: Grok 4.7 scored 46 on the Intelligence Index — two points above the previous build, enough to put the merged SpaceX-xAI entity in the world’s top four labs and past GPT-5.6 on coding agents, where Grok 4.7’s agent score jumped to 56 on a terminal-bench surge.

Two points. That is the entire event.

The two available readings are both lazy. The first is that Elon Musk said on September 2 that Grok 4.7 would ship “in ten days,” missed the September 12 target, and shipped late — again. The second is that a two-point move on a synthetic index is noise, and that AI leadership flipping every few weeks proves the leaderboard is meaningless.

Both are true. Neither is the story. The story is what a two-point gain, bought with a 40 percent parameter increase, tells you about where the moat went — and what the merger did to the one input that cannot be bought.

A Two-Point Gain That Reorders the Top Four Is a Confession

Grok 4.6 shipped August 12 at 1.5 trillion parameters, priced at $2 per million input tokens and $6 per million output. Grok 4.7 is 2.1 trillion — 40 percent more model — and buys two index points and roughly ten points on coding agents.

In any other industry, that is a marginal-return curve you would show to a board. In AI, it gets reported as a leadership change. If a top-four position can be reshuffled by an increment that small, the labs have converged on the same recipe and are now fighting over execution, not insight. Everyone can buy the compute. Everyone can hire the researchers. Everyone can read everyone else’s papers. The recipe is public; the ranking is a rounding error.

Which means the durable advantage, if one exists, has to live somewhere the recipe does not reach.

The Missed Date Was Never the Story

The coverage of the nine-day slip has been framed as a Musk-promise story, which is a category error. A standalone lab that misses its date gets punished by its next raise — the investors who put a reported $230 billion on it have a financing clock, and that clock runs on shipping. A captive division inside a private launch company has no such clock. As one post-merger analysis put it, xAI no longer has an independent governance or liquidity path; its fortunes are now tied to SpaceX’s execution on Starlink and launch. Grok 5 already slipped past its original Q1 2026 target.

Nobody’s shares get marked down when nobody’s shares trade. That is not a defense of the slip. It is an explanation of why the slip is uninformative and the merger is not.

Nobody Can Price a Corpus That Never Trades

Here is the part the leaderboard discourse cannot see, because it is not in the model card.

Grok 4.7 was trained partly on SpaceX engineering data. Think about what that is: propulsion and flight telemetry, failure histories, manufacturing tolerances, Starlink link budgets at a scale no competitor has — a frontier-grade corpus that is not for sale at any price. You cannot license it. You cannot scrape it. You cannot reproduce it with a data-center budget, because it was generated by flying hardware.

Every other frontier lab’s advantage is purchasable by someone with enough capital. This one is not. But it also has no market price, because it moves between two entities that share a board and a controlling shareholder. There is no arm’s-length transaction, no disclosure, no segment reporting that separates the value of the corpus from the value of the model.

A junior analyst on a secondary-market desk that moves pre-IPO space and defense paper, mid-session Monday, put it more bluntly than I can: “We can mark a launch cadence. We can’t mark an internal data transfer between two companies that share a board. So we don’t mark it.”

That is the actual news in Monday’s number. Not that a model got two points better, but that the only non-commodity input left in frontier AI is a property right held inside a launch contractor with federal customers — and the market has quietly declined to price it.

What to Watch Instead of the Leaderboard

Three things will tell you more than the index.

First, the API price. If Grok 4.7 lands at or above Grok 4.6’s $2/$6 per million tokens, “top four lab” is a press release, not a position. If it undercuts meaningfully, SpaceXAI is buying share with something other than capability — and given the parent’s balance sheet, it can.

Second, data provenance. Almost no frontier lab publishes where its training corpus came from. If SpaceXAI starts bragging about the engineering data in a model card, that is a tell: it means the corpus is the differentiator they intend to defend.

Third, the next tender or secondary pricing of the merged entity. If nothing in it marks xAI separately, then the $230 billion figure describes a division whose primary irreplaceable asset has no market price and no line item — a strange thing to put in a valuation, and a stranger thing to oversee, since that asset sits inside a company federal procurement already watches closely.

The people who need to update are not the Musk skeptics or the Musk faithful. Both camps filed their takes before lunch.

It is the readers who cover AI as a product market — a leaderboard, a price sheet, a release cadence. On that framing, Monday’s event was a model two points better than the last one. But a market where leadership flips every three weeks is, by definition, a market where nobody has an advantage. The standings move fast because they measure nothing durable.

And it is the readers who like vertical integration, which I do. Efficiency, aligned incentives, less leakage — all true. But this integration’s moat is not a market outcome. It is a data set that by construction cannot have a competitor, transferred internally at a price nobody sets. If competition is what disciplines firms, then the most important asset in frontier AI just moved somewhere competition cannot reach it.

Grok 4.7 may well be the best model of the week. That is the problem.

Sources