Grok 4.7 in Ten Days: The Model War Is Now a $2 Trillion Stock Story

Ten days or so from now, Elon Musk says, the next version of his AI model drops. Grok 4.7 is reportedly done with its main training; what remains is a final pass over a dataset no other lab has — roughly 25 years of SpaceX's own rocket-building and flight data.

Here's what a headline like that used to mean: a teaser for people who follow AI on X. Here's what it means now: a scheduled event at a publicly traded company that the market has already begun to grade on these releases. Grok no longer belongs to xAI, the private lab Musk folded into SpaceX in February at a $250 billion valuation. It belongs to SpaceX, which went public in June at $135 a share, closed its first day up 19% and worth more than $2 trillion, and by mid-June had touched a $2.6 trillion peak before giving much of it back.

That is the part an AI-headline reader misses: every model date is now a stock date. And the dates are coming faster than they ever have.

Run the calendar. Grok 4.5 shipped in mid-July. Grok 4.6 shipped on August 12. Grok 4.7 lands in the middle of September. Three flagship models in about nine weeks. The frontier race has gone from an annual, Apple-style release rhythm to a monthly shipping war, and Grok is running the fastest schedule at the biggest valuation.

Musk's own dates have slipped all summer. On July 24 he said Grok 4.6 was "2 weeks" out, with 4.7 four weeks after that. Grok 4.6 actually shipped on August 12 — after he'd pushed the date again on an earnings call. On August 12 he reset Grok 4.7 to "3 to 4 weeks," which lands it precisely in the ten-day window you're reading about, already behind his original plan. The pace is real; the calendar is a promise, not a contract.

Why should a shipping date command your attention? Because the market has already answered the question "what is this cadence worth?" — and the answer is not the round number Musk wanted.

SpaceX's first earnings report as a public company showed revenue nearly doubled to $7.8 billion in the quarter. It also showed capital spending up sixfold, to $18.4 billion, in a single quarter — most of it AI compute. The stock fell 13% the day after and kept sliding. At the core of it, xAI was running on roughly $500 million of annualized revenue against about a billion dollars a month of spending, with guidance to push standalone sales past $2 billion this year. Put that next to the competition: OpenAI is projecting more than $30 billion in 2026 revenue, Anthropic $20 to $26 billion. The burning money is one or two orders of magnitude larger than the selling money.

That is the tension under every Grok announcement. And Grok 4.7 is the cleanest example yet of why the bet might work anyway.

Push on what's actually changing. The first thing abundance does is kill compute as a moat. Anyone can buy frontier-grade compute now — including Musk's own rivals, who rent his. In May, SpaceX leased its entire Colossus 1 cluster, the 220,000-GPU facility in Memphis, to Anthropic — reported at roughly $1.25 billion a month through 2029, a contract north of $40 billion — while moving its own training to a new site. When your biggest competitor runs its models on your chips, GPUs are not the advantage. Token prices across the industry keep falling. Compute is abundant, and growing more so.

So what becomes scarce? Proprietary data. Which is exactly where Musk is pointing Grok 4.7. The reporting is that the final training pass runs on a "massive amount" of SpaceX company data — the accumulated engineering record of nearly 25 years of launches, landings, and failures. You cannot scrape that. You cannot rent it. OpenAI can outspend SpaceX on chips; it cannot buy a quarter-century of private rocket and manufacturing data. That is the asymmetry: the scarce complement inside a commoditized input, owned by the only frontier lab that also has a consumer distribution layer in the hundreds of millions of users attached to X.

Take the counterargument seriously, because it could change the read. One: Musk's "beats everything" claims have arrived without benchmark sheets before. Two: five hundred million of revenue against $18 billion of quarterly capex is not a business yet, and a monthly release cycle does not by itself fix the monetization gap. Three: the share count is about to grow — insiders were already freed to sell roughly $101 billion of stock when the first lockup expired in August, and another 319 million shares unlock on September 9. A release that slips, underwhelms, or just fails to convert into paying customers arrives into a rising float.

But none of that erases the mechanism. The order of operations is: compute commoditizes, data does not, and this is the only frontier lab with a decades-deep proprietary engineering dataset bolted to real user distribution. The evidence so far says Grok can ship with the leaders — Grok 4.6 matched OpenAI's GPT-5.6 Sol on the independent Artificial Analysis index, within a single point of Anthropic's Claude Fable 5 — and that Musk will spend whatever it takes to stay there.

If the story matters to you, the direct way to own it is SpaceX itself: ticker SPCX, lately around $141 — above the $135 IPO price, down roughly 37% from the June peak, a market value near $1.9 trillion, with quarterly earnings and a lockup calendar now attached. The indirect way is Tesla, which converted a $2 billion investment into a small SpaceX stake when xAI was folded in and booked a $1 billion paper gain on it in July — a real but small claim on Grok buried inside a far larger company.

Say the honest part plainly: the asymmetric setup that made xAI interesting at $230 billion — breakthrough tech, small market, uncapped upside — is gone; that gap closed at the IPO. What remains is a mega-cap paying out of pocket for a monthly frontier-model war, betting that scarce data beats abundant compute. Grok 4.7, about ten days out, is the first release where that bet is fully testable in public. If it ships on time and beats the leaders on hard engineering work, the treadmill is working. If it slips or underwhelms into the September unlock, you have learned what the market already suspects — and a release date is the cheapest signal in the whole trade.