Elon Owns the Whole Stack: Grok 4.5 and the $60 Billion Cursor Buy

On June 28, Elon Musk announced a frontier AI model the way you’d announce a lunch order. One tweet, 7.8 million views, no model card, no API, no benchmark anyone outside his companies can run.

Grok 4.5, based on our 1.5T V9 foundation model, with Cursor data added in supplemental training, is now in private beta at SpaceX and Tesla. Early evals show performance close to, perhaps exceeding Opus.

— Elon Musk, June 28 2026

Read that as a product claim and it’s thin to the point of vapor. “Early evals” run by the vendor, against an unnamed version of a competitor’s model, with nothing published. We’ve seen this movie. But the interesting part isn’t the claim. It’s the return address. That tweet went out, and the model releases were promised, under the @SpaceX handle. Not @xAI. A rocket company is now shipping frontier language models, and it can afford to skip the receipts because of what it spent the last five months buying.

The Five-Month Land Grab

Line the announcements up and the June 28 tweet stops looking like a launch and starts looking like a victory lap:

• February 2: SpaceX and xAI merge in an all-stock deal valuing the combined entity at $1.25 trillion. The largest private merger ever. xAI, burning roughly a billion a month, gets the rocket company’s balance sheet.
• June 12: the combined entity IPOs at a market cap crossing $2 trillion. Largest IPO in history.
• June 16: four days after going public, it exercises an option to buy Cursor’s parent Anysphere for $60 billion all-stock, expected to close in Q3.
• June 28: Grok 4.5, “with Cursor data added in supplemental training,” enters private beta inside the two companies Musk controls.

Each of these was public. None of it leaked. This is not a conspiracy you have to assemble from fragments. It’s a strategy announced out loud, four times, and mostly reported as four separate business-section stories.

The Stack, Mapped

Put the pieces on one table and the shape is unmistakable. One entity, roughly $2 trillion, Musk as controlling shareholder, now owns every layer that used to belong to four different companies:

• The compute: Colossus, the Memphis supercluster reportedly targeting 555,000 GPUs.
• The model: xAI and Grok, folded in at the merger.
• The distribution: X, absorbed into xAI before that.
• The tool and its data: Cursor, the dominant AI coding IDE, with 50,000-plus enterprise customers and, pending the close, the developer-session pipeline that Musk says now trains the model.

This is the cheap-is-a-hardware-strategy argument taken to its terminal conclusion. Google owns the TPU-to-search-box stack and can give intelligence away. Musk just assembled the coding-specific version of the same thing: silicon, model, reach, and the closed loop of usage data feeding back into training. The difference is that Google shows its benchmarks.

The Eval Standard Is One Tweet

Here’s what a “perhaps exceeding Opus” frontier model looks like from the outside. It isn’t on Artificial Analysis. It isn’t on LMSYS. There’s no model card, no architecture note, no API for a third party to poke. The Hacker News thread about the announcement got 11 points and 6 comments, the top one being “Typical Elon marketing speech.” For a claim that would be the story of the year if true, the industry’s reaction was a shrug.

That shrug is earned. This is the company that in February 2025 compared Grok 3’s consensus@64 score against a competitor’s pass@1 and got publicly called on the apples-to-oranges. The same lineage that shipped the “MechaHitler” incident. When the track record is benchmark theater, “trust me, early evals look great” is not a data point.

xAI is, as Elon himself has described, a complete train wreck for its kind of building of foundational models.

— Reid Hoffman, June 24 2026

Hoffman said that four days before the Grok 4.5 tweet, adding that SpaceX is “buying its way into relevance” like a premium-priced CoreWeave. You don’t have to take his word for it either. But when the model can only be evaluated by the company that owns the model, the compute, and the customers, the outside world’s opinion becomes structurally irrelevant. That’s the point of owning the whole stack: the model never has to sit for an exam it didn’t write.

Whose Data Was It, Anyway?

The one line that should stop you is “Cursor data added in supplemental training.” Because Cursor’s own privacy policy says the opposite: it does not use user inputs to train models, or let third parties do so, absent explicit consent or a security flag.

So which is it? Nobody has said. Neither Cursor nor xAI has published a word reconciling “we trained Grok on Cursor data” with “we don’t train on your code.” And neither has denied Musk’s characterization.

There’s a quiet irony buried in it, too. Cursor is a multi-model router. For most of its life the assistant on the other end of those sessions was Claude or GPT, not Grok. If developer sessions genuinely fed Grok’s training, xAI’s coding gains are partly downstream of millions of interactions with its competitors’ models. The moat was dug with someone else’s shovel.

What This Isn’t

The vertical-integration framing is clean and I want to bound it honestly:

• It’s disclosed, not conspiratorial. Every step was announced. The critique isn’t that Musk hid a land grab. It’s that assembling this much of the stack lets a company opt out of external accountability entirely, in plain sight.
• A private beta is unfalsifiable by design. “Perhaps exceeding Opus” can’t be disproven when nobody outside SpaceX and Tesla can touch it. That cuts both ways: I can’t call it a lie, only unverified. But unverified is not a synonym for impressive.
• The Cursor data claim may be more boring than it sounds. Curated corpus, not raw code, is a real possibility, and it would defuse most of the consent alarm. The problem is that the people who know aren’t saying.
• Owning the stack is a real strategy, not just hubris. It might work. Google proves the model. The concern is what a working version does to the rest of the field’s ability to check its claims.

The Exam You Grade Yourself

Strip away the trillion-dollar numbers and this is a simple story about who holds the red pen. A frontier model is a claim about capability, and claims are supposed to be checkable. Benchmarks, APIs, model cards, third-party arenas - the whole apparatus exists so that “we built the best coding model” is a testable statement rather than a press release.

Musk just bought his way out of that apparatus. Not by breaking a rule, but by owning every part of it. The compute that trains the model, the tool whose data feeds it, the platform that distributes it, and now the customer base that “evals” it. When the same entity owns the question, the answer, and the grader, the grade is not information. It’s marketing with a market cap attached.

The call didn’t come from an outside lab this time. It came from a company grading its own homework, and reporting an A.