Cursor admits defeat, Elon Musk doesn't win

In 2025, Cursor's annualized revenue exceeded $2 billion. In just three years since its establishment, it went from zero to $2 billion, the fastest in B2B history. Its per - capita revenue generation outperformed any Silicon Valley software company.

Everyone thought it would become the next Adobe.

One year later, it was sold. The acquisition was a $60 billion all - stock, share - exchange merger, and the acquirer was Elon Musk's SpaceX, which had just gone public four days prior.

It sounds like a victory, but in fact, it's more like a retreat. There are two figures hidden in the transaction contract: if the deal falls through, SpaceX has to pay $10 billion in compensation. If it gets stuck due to antitrust issues, an additional $4 billion in compensation is required.

When the contract is written like this, it's not an acquisition; it's a rescue.

Cursor wasn't short of money. It was cornered by its own suppliers. And Cursor's experience is making the entire AI application industry collectively lose its sense of security.

1. The Closest Supplier Turns Against First

Cursor's business model is brutally simple. Users pay a monthly fee, and Cursor then turns around and buys the Claude API from Anthropic. At its peak, Cursor accounted for 40% to 50% of Anthropic's revenue. It wasn't really developing products; it was acting as a sales channel for Anthropic.

The two companies were so interdependent that they reached a symbiotic relationship. One needed the model to drive the product, and the other needed the revenue to sustain the company.

Whoever turns against the other first is committing murder.

Anthropic was the first to turn against.

In May 2025, Claude Code was released. Before the release, Anthropic's management gave an oral promise: "Don't worry; it's more of a research project." It meant that they were going to do it themselves, and they asked Cursor not to leave.

Six months later, Claude Code's annualized revenue exceeded $1 billion, even faster than Cursor's at that time. By the end of the year, its user base officially surpassed Cursor's.

While Cursor was still digesting this news, the second blow landed.

On January 9, 2026, Anthropic blocked all third - party access to the Claude model from the server side. There was no warning and no transition. Cursor, OpenCode, and Cline all became invalid overnight. The supplier turned into an executioner. The other party was not only a competitor but also the only supply chain. If you compete with it in products, it will cut off your lifeline.

2. Two "Self - developed" Attempts, Both Exposed

After being blocked, Cursor had only one way out: develop its own model.

Cursor's founder, Truell, held an emergency all - hands meeting on January 5, canceling all non - essential meetings and fully shifting to the Composer programming project. The goal was straightforward: to create a programming model that was no worse than Claude Opus but with lower costs.

In March 2026, Composer 2 was released. The official claimed that it outperformed Claude Opus in multiple programming benchmarks. The developer community was in an uproar. An application - layer company had outperformed its own supplier in model capabilities.

The excitement lasted less than 24 hours.

A developer caught the model ID "kimi - k2p5 - rl - 0317 - s515 - fast" while debugging the API. It turned out to be a re - skinned version of Kimi K2.5.

Elon Musk personally replied: "Yeah, it’s Kimi 2.5."

What hurts more than being exposed is that the one who exposes you later becomes your buyer.

The same thing happened two months later. Cursor 3.0 was released, claiming a full reconstruction of the Agent mode. When developers disassembled it, they found that the underlying engine was the Claude Code SDK. The brand logo was replaced, the entire engine was modified. Claude Code was changed to Cursor Agent, CLAUDE.md became AGENTS.md, and even the Git commit signatures were changed.

The first level is borrowing someone else's model and pretending it's self - developed. The second level is directly packaging someone else's engine into the installation package and selling it under a different name.

Cursor proved the cruelest reality in the AI era: there is no moat for pure application - layer companies. With $2 billion in ARR, millions of developers, and 70% of Fortune 500 companies as customers, nothing could stop the company when the supplier turned against it.

3. The Buyer's Wounds

Seeing this, many people may think that Elon Musk got a bargain.

But the buyer's wounds are not small either.

In February this year, SpaceX completed an all - stock acquisition of xAI. The combined entity was valued at $1.25 trillion. The public statement was nice: the integration of rockets, AI, and social data to build an empire. But the dark side was that xAI was falling apart.

All 11 co - founders who started xAI with Elon Musk left within two months. Only he was left in the 12 - person founding team. The heads of reasoning, pre - training, and core product lines all left. There is no second such case of a complete exodus in Silicon Valley.

Elon Musk's Grok model has good reviews but poor sales in the programming track. Even he himself admitted in March that the programming tools were behind the competitors. In the first quarter of 2026, xAI's AI business had $818 million in revenue but a loss of $2.5 billion.

What's more troublesome is that Colossus, SpaceX's most powerful asset, couldn't save Grok either. Colossus 1 has more than 220,000 NVIDIA GPUs. With mixed hardware and serious network latency, it can't run cutting - edge model training at all. Its utilization rate is only 11%.

Elon Musk made a smart shift and turned the data center into a rental business. Anthropic rents Colossus 1 for $1.25 billion per month, and Google rents 110,000 GPUs for $920 million per month. The annualized revenue is $26 billion, exceeding SpaceX's total revenue of $18.7 billion in 2025.

But an embarrassing fact lies behind this: SpaceX rents its largest computing power asset to two of its competitors in the AI track.

It's not that they don't want to use it themselves. They just can't.

So, looking at this acquisition again, Elon Musk isn't buying Cursor; he's rescuing SpaceX AI. The $60 billion isn't buying a product or users; it's buying a possibility that someone will continue to develop products based on Grok.

4. For the First Time, Silicon Valley Trusts a Chinese Model with Its Future

Let's go back to the incident of Cursor Composer 2 being re - skinned. There is a truly important detail in that storm.

After it was exposed that Cursor used Kimi K2.5, Cursor's vice - president publicly apologized. Cursor actively evaluated all open - source bases on the market and chose Kimi. Co - founder Aman Sanger's exact words were: "K2.5 has been proven to be the strongest."

There was no geopolitics and no administrative orders. It was a purely technical decision.

In the past few years, the narrative of the Chinese AI industry has been just one word: catch - up.

When others create something, we follow. In Silicon Valley's perception, Chinese models have always been labeled as cheap but slightly inferior. But at the moment when Kimi K2.5 was actively selected by Cursor, this label was torn off.

A Silicon Valley company valued at nearly $30 billion actively chose a Chinese base at its most core model layer. It wasn't because of lack of choices or cost considerations; it was because it was the best to use.

In the past, the highest praise for Chinese technology companies was to be as good as those in Silicon Valley. Today, for the first time, there is a new narrative: a Silicon Valley star company is entrusting its future to a Chinese model.

There are three things worth remembering for Chinese AI players.

First, the Achilles' heel of the application layer lies not in the product but in the supply chain. Relying on others' infrastructure for core competitiveness means that the higher the building, the greater the risk. This is a repeat of the lessons from the chip industry. If you don't develop the underlying technology, you're always at risk of supply disruption and being sold.

Second, the walls of Silicon Valley's open - source ecosystem are getting higher. OpenAI is narrowing API access, Anthropic is cleaning up third - party integrations, and Google is locking down top - tier models. Every big company is becoming a competitor to its customers. The application layer has only two ways out: either have your own base or have something that the supplier can't take away.

Third, Chinese open - source models are becoming the only neutral base in the eyes of global developers. A supplier with strong technology, open access, and no competition with any front - end product lines is creating a rare structural opportunity in today's geopolitical and business ecosystem.

5. No More Independent Companies in AI Coding

As Chinese models become a safe haven, a generational purge is taking place.

In the past two decades, there has been a default rule in the Internet industry: platforms build the roads, and startups open the shops. Google does search, Adobe does design software, and Salesforce does CRM. Each makes its own money without interfering with each other.

This rule no longer works in the AI era.

Today's model companies not only build the roads but also open the shops. They even open the shops right in the middle of the road.

OpenAI has Codex. Anthropic has Claude Code. Google has Gemini Code Assist. Every model company is developing applications, and they're all developing the same application: helping you write code.

This is not a coincidence.

Coding is the most profitable track in AI today. A model that can write good code is like having a ticket to the entire software development industry. No one wants to hand this market over to others.

The deeper logic is here: the model is the product.

When a model is strong enough, there's no need for an intermediate layer. You can directly talk to it, and it will write the code for you. At this time, the editor itself becomes almost non - existent. What Cursor fears most is not that Claude Code will steal its users, but that users will no longer need an editor to write code one day.

So, the platform starts to swallow the applications.

Just like Apple swallowed the flashlight app, weather app, and recording app in the iOS system back then. You thought it was just a feature integration at that time, but later you realized that it was a systematic clearance of the application layer by the platform. Independent developers who were doing well one day found that the system - built - in functions were better and free the next day.

Cursor is the first giant to die in this clearance. It didn't lose to Claude's product strength; it lost to Claude's owner.

AI application - layer companies are facing an unprecedented dilemma. Your supplier can turn into your competitor at any time, and you have no countermeasures.

Anthropic didn't buy Cursor not because it couldn't afford it. It's because keeping Cursor alive is more valuable than killing it. A living Cursor is a sword hanging over the entire AI programming ecosystem. Every independent developer tool company can see that Cursor's fate could be their own. The only way out may be to find a bigger buyer.

In the AI Coding track, the term "independent company" no longer exists. Cursor has found a new owner. The next one is already in line.

Maybe it won't be the next programming tool. But it could be the next AI writing tool, design tool, or customer service tool.

The principle is the same. When a model is strong enough to replace your product logic, your company is no longer a product company but a distribution channel. And the valuation of a channel is never high.

Back to the $60 billion. It's expensive, but it's also not that expensive. SpaceX isn't buying revenue, technology, or a team. It's buying a dream, a dream that a company with the world's most powerful computing power cluster won't be empty - handed in the hottest track.

But whether the dream can come true doesn't depend on how many contracts Elon Musk signs.

The computing power of Colossus 1 can be rented to Anthropic. The GPUs of Colossus 2 can be rented to Google. No one can pay for developers' habits and choices in the editor.

The $60 billion only buys an entry ticket, that's all.

[Beyond the Page] Words:

The truly terrifying thing about Cursor being sold to Elon Musk is that it makes all AI application companies realize for the first time that they may have no future.

In the past three years, the hottest track in AI startups has been the application layer. There's Cursor for writing code, Lovable for design, and various startups for Agents.

Everyone believed that as long as the product was good enough, the company could grow.

Until Anthropic got involved and Claude Code started to swallow Cursor. Until a star company with $2 billion in ARR chose to be sold.

At that moment, the entire industry suddenly realized that the biggest risk has never been the competitor but the supplier.

This article is from the WeChat official account "Beyond the Page", written by Huahua, and is published by 36Kr with authorization.