Musk had 30 days to decide whether to purchase Cursor. He used only a weekend.
Rewind.
If the only tech news that mattered last week was SpaceX’s largest public offering in history, then the one piece of tech news to know this week is SpaceX’s acquisition of Cursor.
This is pretty much a done deal, something they were already planning since April. Yes, so what’s left to talk about?
A lot.
To start with, this is the one thing that will impact the landscape of the enterprise coding platforms over the next 3 years.
So, I’ll answer the following questions in this one:
Why even buy Cursor to start with?
Why was Musk in such a rush to make this decision?
What Musk actually bought?
How would this impact Anthropic? (yes, Anthropic)
Two possible war-room scenarios for all parties over the next three years
Shall we?
Why Cursor?
Back in 2022, no one could have predicted that a coding platform would be the next frontier to plant a flag on.
This ‘once upon a time,‘ should start with Bubble.io, or any legacy no-code platform that was pitched to replace engineers since early 2010. They ended up as a niche, a playground for whoever believed in low-code/ no-code. In Bubble.io’s case, their market share was around $500M in 2022, still a mid-size startup after 10 years.
In comparison, within just a single year, Cursor launched in mid-2023, and by 2024, it reached about $2.6B in valuation. In the post-ChatGPT boom, devs didn’t work directly with an AI model on their terminal, but via Cursor’s interface.
For about two years, the AI coding market had a clean split. On one side, the labs that made the models (Anthropic, OpenAI). On the other hand, the apps (Cursor) that wrapped those models into something most developers would actually interact with.
While Anthropic didn’t have its own harness platform, its models consistently ranked as the primary choice for developers. So, one day, Amodei realized the potential of the enterprise AI coding platform.
Soon, the marketing lines and sales pitches of any new Anthropic models shifted from a general-purpose assistant in early 2024 to ‘developer productivity and code understanding’ entering 2025.
Unlike OpenAI’s lack of direction, I admire Anthropic for its relentless product focus on enterprise customers.
As soon as Amodei realized that coding is the future (or the only use case) for LLMs, the only sensible next move was to build their own coding assistant.
Hence, Claude Code launched in Spring 2025.
That was pretty much the beginning of the end for Cursor. Think about it — as a wrapper, you’re carrying all the usual operating costs, but your biggest cost, the tokens, isn’t even in your control. There’s no way Cursor can win that fight once Anthropic decides to come for it.
Under this condition, Cursor was forced to raise its prices, and developers pushed back hard. Active users dropped steeply, and the company had little choice but to reverse course and make a public apology for it.
While the data can’t be found anywhere in the official announcement, either in the wrapper or the AI labs, a well-known engineering newsletter did a small-scale survey and found that Claude Code is now the #1 AI coding tool, whereas Cursor ranked #4. Not to mention how Sam Altman finally, finally, came around to the idea that enterprise is the only way for OpenAI to profit.
Hence, OpenAI launched Codex three months after Claude Code.
It’s only a matter of time until Cursor is replaced by AI labs. While the clock is ticking, they still have options. Let’s count the cards in their hands:
over 1 million daily active users
high-quality structured data from developers
Only these two, yes, but each is mission-critical for any AI labs.
They could become a model provider themselves, which they did; they have their own built on open-weight models. But they also ran into an issue that all AI labs have: money.
So, Cursor, a popular AI wrapper used by millions of devs in enterprise, is now looking for a way out.
Let’s turn to SpaceX and Musk for a bit.
Why Now? Just a few days after IPO?
Priced at $135, IPO opened at $150, and it spiked up to $216 by the following Tuesday. On the same day, Musk agreed to buy Cursor for $60 billion.
All unfolded in less than a week. Why the rush?
To understand this, you need two things: first, there is a clause about this deal in SpaceX's S-1; second, you need to understand how timing and share price affect the acquisition cost.
The Cursor Acquisition Clause
On page 12 or F-96 in the S1 (the key doc. must file with the SEC when going public).
The option to acquire Cursor was already signed in April, and SpaceX has the right to buy Cursor in one of the two following timelines, whichever happens sooner:
30 days + 7 trading days post IPO
before the end of September 2026
Given that the IPO happened in June, the 2nd option automatically became invalid. So the ~40 days countdown started on 12th June. To understand why not wait for a few more days, here’s the cost structure of this acquisition that Musk plays with.
$60B paid with SpaceX’s Share
The $60 billion is fixed.
However, the number of shares that fills that figure was not.
What do I mean by this?
Because in S1, they’ve agreed the acquisition will be paid with SpaceX’s shares rather than cash.
So the higher the share price, the cheaper it is for SpaceX to buy Cursor. Also, to keep in mind, the more shares SpaceX issues, the more diluted the value is per share (same ideas as gold, and everything, the volume decreases the value).
As the stock price climbs, your money simply gets you fewer shares, and the stock, four days into public life, surged from $135 to north of $200.
If the acquisition happened at the IPO valuation, handing over $60 billion of stock dilutes existing owners by ~3.4%. But when the price hit north of $200 than, the cost of getting Cursor is closer to ~2.2% of SpaceX’s share.
Knowing that $200/share is about as good as it gets for the next 30 days, and that most analysts believe SpaceX is trading on aspiration rather than anything fundamental, it’s only rational (rare for Musk) to lock in the acquisition.
What did Musk actually buy?
Certainly not $60B just for a code editor.
Remember that two cards remain in Cursor’s hands? They have over 1 million active users per day and high-quality, structured data from developers.
The S-1 just flat-out says it.
Every time a developer works in Cursor, they generate high-quality training data: the prompt to generate code, the code that comes back, the architecture decision, and the feedback loop. All the highly valuable training data that AI labs scramble for.
Then, on page 145, it states what this clean, high-quality data is for.
accelerate the development of our existing AI models, including Grok
The developers on the AI code editor are the customers, yes, but they are also the product to help improve xAI.
Which is why the marriage makes sense, once you see what each side was missing.
Cursor needed two things that none of its rivals could offer.
First, a model owner without a coding tool would be the perfect fit, except that the problem is, every major lab that comes to mind (Google, Anthropic, OpenAI) already has one. Second, the computing power.
All in all, lack of funding is the main reason (if w/o this acquisition), Cursor could only post-train someone else’s open-weight model and has little flexibility to move this business forward.
xAI, on the other hand, was short of the opposite two things.
It has compute (Colossus I, and building two more), but lacked direct access to the developer community, not to mention their coding activities. Without this, they’d be further behind than those advanced AI labs that have their own code-harness platforms.
So each side plugs the other’s gap. Cursor gets a super sugar backer with cheap compute; xAI gets the audience it could never win on merit.
It’s a no-brainer that Cursor’s default model will be Grok in no time. But the war-room play-through is so much more than just which model is the default for millions of developers.
We have one more puzzle to reveal before we put them together: the Anthropic and SpaceX compute deal.
Anthropic Is Urgently Behind In Compute Power
In May, Anthropic signed a deal to rent all the compute capacity at SpaceX’s Colossus I data center for $1.25B/month until 2029.
Anthropic expects to draw 300 megawatts of power and run over 220k Nvidia GPUs through SpaceX, as stated in their announcement.
If you’re a heavy Claude model user, you’d likely remember the downtime experience in the past few months. So this deal is part of their capacity boost solution: doubling Claude Code’s rate limits, removing the peak-hours throttle on Pro and Max tier users.
Hold onto that, because the rest of the section is about how much this matters to Anthropic.
To explain this, we need to go back to this screen:
If you didn’t read carefully between the lines, you’d think they already have north of 10GW online.
Because their compute deal includes:
up to 5 GW with Amazon,
another 5 GW with Google and Broadcom,
$30bn of Azure with Microsoft (capacity didn’t specify),
$50bn with Fluidstack (unknown capacity)
Next to a total of 10 gigawatts and another $80bn in deals, SpaceX’s Colossus 0.3 GW looks like a rounding error, might be less than 3% of the total capacity Anthropic has announced.
Except that if you read the Anthropic announcement again, all these capacity claims are at some point in the future.
If we take a look at when each one actually arrives.
The Amazon deal is “nearly 1 GW of new capacity by the end of 2026.” The Google and Broadcom one begins coming online in 2027. Colossus is the only one that’s a firm current.
So if you look at what's actually live and available for Anthropic this summer, Colossus isn't 3% of anything. It's carrying a significant chunk of their compute right now.
And this doesn’t come cheap.
As I’ve pointed out, Anthropic pays $1.25bn a month for it through 2029. If no party withdraws early, it’s a total ~$45bn deal.
Now layer this in, Anthropic is xAI’s competitor.
Especially at a time when operating AI models and selling tokens are money-losing operations and can’t scale their way out, whereas, on the other hand, selling the land (like the data center, the compute) is the only money-making part of the whole AI supply chain.
Anthropic must be desperate.
What’s worse is that this whole deal can fall apart with just 90 days’ notice, and this one adds to that:
still permitting reallocation of the capacity for our own internal initiatives if needed in the future.
In plain English, SpaceX can pull back capacity with as little as 90 days ’ notice if they ever need to.
Now that we have all the players on the board and know what cards they each hold, we are ready for the war room.
Unlike most business strategy cases, where you can rely on decision-makers following the standard MBA playbook, these two scenarios are different — because one party is Elon Musk.
Two PlayOuts
The landlord is a direct competitor, and the lease is short.
And the capacity propping up Claude’s consumer experience belongs to the one company now assembling the full stack against it: model (Grok), tool (Cursor), and data center (Colossus).
When Cursor makes Grok its default model, because of course it will; you don’t pay $60bn for a distribution channel and aim it at a competitor model. This is the only path to complete the vertical integration that the S1 brought up 117 times.
Musk, who spent the winter calling Anthropic “evil” and a play on its own name, is currently its most important landlord, while the Pentagon blacklisted Anthropic and separately struck a deal to use Grok.
Scenario one: happy day, no one ends the deal early
This is the most likely case in the next ~12–18 months.
You see, not 36 months until the full contractual date of May 2029, but half the term.
With this, a few assumptions are needed:
First, assuming Grok's demand stays under the spare capacity. Which is unlikely, at least, having paid $60bn for Cursor, I’d believe the investment is worth something. Seeing the immediate impact is a must.
To start with, the least they can hope for is a ramp-up of Grok’s usage. Even though more people are using AI these days means losing more money, it’s obvious that since the start of this deal, Musk wants to compete with OpenAI and Anthropic on the enterprise coding market share.
Second, and more importantly, this $45bn in three years deal is not insignificant. Provided Colossus 1 was at ~11% utilization (when w/o Anthropic) and only the connectivity (Starlink) department makes a profit, $1.188bn (quarterly), of all SpaceX. So yes, $45bn is some duct-tape for the $16bn annual losses hole.
Given Musk’s confidence in constructing a data center at a much faster pace compared with competitors (and that no one’s using Grok), he shouldn’t be worried about not having enough for a while.
So rationally speaking, this is the best option for SpaceX’s business and shareholders alike.
But the word "rational" usually doesn’t apply to Musk.
Scenario two: Musk pulls the capacity
The chance is low in the next 12–18 months; however, it is high over the full term to 2029.
I'd put it above 50% that he terminates or materially changes the deal before it runs its course.
Two reasons for the pulling case: 1) Grok’s usage ramped up as they’d hoped; 2) what’s more likely: who Musk is.
He never rents his crown jewels, and Musk is always right. This should be obvious enough over the years, but if not, just look at how SpaceX/Tesla is building and hoarding critical inputs (engines, satellites, chips), and the whole S-1 itself brags about the ‘value chain’ and ‘verticals.’
As long as he can keep funding it (and, given where the stock price is, he has access to relatively cheap capital through further share issues), especially given how ownership is currently structured.
Renting Colossus to a rival is the anomaly in his record.
He’s also renting to someone he publicly calls evil, and he framed the deal as a reluctant favor, which is not how you describe a relationship you intend to keep. And this isn’t a one‑off. Remember the time he signed a binding deal to buy Twitter? He tried for months to back out, and only closed the sale when a judge in Delaware was about to drag him into court.
Contracts with Musk aren’t a done deal, but something renegotiable the moment they stop serving him. On top of that, nothing institutional meaningfully stops him here: SpaceX is a controlled company, Musk retains full operational control of the business — he holds 85% of the voting power — no one gets to out-vote and overrule him. See S1 page 253.
The only thing holding the deal together is cash, as we covered in scenario one.
So, near-term, he keeps cashing the cheques.
The pull comes when his own demand catches up to the capacity he’s renting out, and on the S-1’s own wording, that collision is built in: Grok 5 trains on Colossus II, as Grok 5/6 and Cursor’s sessions scale, his demand and Anthropic’s contracted capacity fight over the same racks. That’s not an “if.” It’s a “when.”
For Anthropic, this means that even if it's not a total disaster, they’ll still lose customers due to this sudden loss of capacity.
It’d be interesting to see how it goes.
One thing is for sure: when it comes to LLMs, software development is the arena.
