In February 2026, Cursor hired Boris Cherny and Cat Wu away from Anthropic. Cherny was the lead engineer on Claude Code. Wu was the product manager. Together, they represented an unusual concentration of institutional knowledge about the most threatening product on Cursor's horizon.
They came back two weeks later.
No official explanation was given. The internet compared it to LeBron returning to the Cavaliers. But the metaphor undersells the damage. Cursor didn't just lose two executives. It demonstrated, publicly, that the company it depends on for its core technology also controls the talent market that could save it.
This is a story about what happens when the fastest-growing SaaS company in history discovers that its supplier is its competitor, its moat is leased, and every escape route it tries leads somewhere worse.
The Numbers Behind the Narrative
Cursor's growth metrics are real and remarkable. By any conventional SaaS measure, the company is performing at historically unprecedented levels.
But growth metrics tell you where a company has been. Margin structure tells you where it's going. And Cursor's margin structure is a warning sign that no amount of ARR can obscure.
In August 2025, Cursor was spending approximately $650 million annualized on Anthropic API calls alone — against roughly $500 million in total revenue. The company's $200/month Ultra tier can cost ~$5,000 in actual compute per subscriber. Consumer subscriptions are explicitly negative margin. When Cursor introduced a credit pool overhaul in June 2025, effective Pro-tier requests dropped from roughly 500 to 225 per month — not because usage declined, but because the company needed to slow the bleeding.
This is the core structural problem: Cursor's primary cost is controlled by its primary competitor.
The Supplier Problem
Anthropic isn't merely a vendor in the way that AWS is a vendor. AWS doesn't ship a product that directly competes with the applications running on it. Anthropic does.
Claude Code has grown to a $2.5 billion+ run-rate, more than doubling since January 1. It has 300,000+ business customers. In the Pragmatic Engineer survey, 46% of developers called it their "most loved" tool versus 19% for Cursor. It now accounts for 4% of all public GitHub commits. And Anthropic has been systematically tightening the screws:
- 66.7% API price cut — Opus 4.6 at $5/$25 vs. Opus 4's $15/$75. Cheaper for Cursor to use, but also cheaper for everyone to bypass Cursor entirely.
- 1M token context at standard pricing — Eliminates one of the IDE wrapper's core value propositions: managing context windows.
- Agent Teams, self-serve enterprise, enterprise analytics — Direct enterprise pitch. No middleman needed.
- Harness crackdown on third-party tools — January 9: blocked tools spoofing Claude Code subscriptions. DHH called it "very customer hostile."
Each move individually is defensible. Together, they constitute a systematic effort to collapse the gap between the model and the user — removing the intermediary.
Five Escape Routes, Five Dead Ends
Cursor hasn't been passive. The company has tried, visibly and aggressively, to break free from platform dependence. But every escape route so far has revealed a new problem.
1. Build Your Own Model
On March 19, Cursor launched Composer 2, marketed as a proprietary advancement. Within 24 hours, a developer named Fynn found the model ID kimi-k2p5-rl-0317-s515-fast in the API response. The post went viral — 444,000 views. Elon Musk amplified it.
Composer 2 was a fine-tuned version of Kimi K2.5, an open-weight model from the Chinese AI company Moonshot. Cursor co-founder Aman Sanger acknowledged the base model, saying it was "a miss to not mention" it and that roughly 25% of the compute came from the base.
The partnership was technically authorized — a commercial deal through Fireworks AI. But Kimi K2.5's modified MIT license requires prominent "Kimi K2.5" branding once a deployer exceeds $20 million per month in revenue. Cursor was at approximately $167 million per month — more than 8x the threshold. No lawsuit has been filed, but license experts have cited Jacobsen v. Katzer (2008), the federal case establishing that open-source license violations can constitute copyright infringement, not just breach of contract.
The irony is thick. Cursor's attempt to build its own model turned into a rebranding scandal involving a Chinese open-source model with unclear license compliance. The escape route didn't just fail — it generated a trust crisis at exactly the moment Cursor needed to project independence.
"It was a miss to not mention the Kimi base."
— Aman Sanger, Cursor co-founder, after model ID was discovered in API response
2. Hire the Competition's Engineers
This is the boomerang. Boris Cherny, Claude Code's lead engineer, and Cat Wu, its product manager, were recruited as Cursor's Chief Architect/Head of Engineering and Head of Product respectively. Two weeks later, they returned to Anthropic.
The move was audacious and logical — if you can't build what your competitor has, hire the people who built it. But the return exposed something deeper than a failed hire. It suggested that whatever Anthropic is building is compelling enough that even people who've seen the other side choose to come back. The institutional knowledge didn't just refuse to transfer — it actively rejected the new host.
Cursor has also lost its heads of engineering — Andrew Milich and Jason Ginsberg both departed for xAI on March 12.
3. Raise More Capital
Cursor's November 2025 Series D valued the company at $29.3 billion. Bloomberg reported in March that a $50 billion round was in early-stage talks. But capital doesn't solve the margin problem — it subsidizes it. At negative consumer margins and a cost structure where API spend exceeds total revenue, more money buys more time but not structural independence. At some point, investors will want to know when the economics work without subsidies.
4. Pivot to Enterprise
Enterprise already represents 60% of Cursor's revenue — a genuine achievement. But Claude Code's enterprise growth is accelerating faster. Anthropic claims 70% of the Fortune 100 as customers, with business subscriptions quadrupling since January 1. Anthropic is selling directly to the same CTOs who approved Cursor licenses, with the added pitch: "Why pay for a wrapper when you can use the model directly?"
5. Go Multi-Model
Cursor supports multiple model providers. In theory, this reduces dependence on any single supplier. In practice, it transforms Cursor into a UI layer over commodity APIs — exactly the position with the worst margin structure in software. If the model is interchangeable, the value lives in the user experience. But user experience in coding tools has a ceiling: eventually, the terminal is good enough, and the agent does the rest.
The Valon Canary
In February 2026, real estate startup Valon canceled its entire engineering team, attributing the decision to AI coding tools. The tweet from Aakash Gupta went viral and catalyzed a "Cursor is dead" narrative — not because Valon specifically used Cursor, but because it demonstrated the ceiling: if AI coding agents can replace engineering teams, the IDE that hosts those agents is a feature, not a product.
Cursor co-founder Michael Truell — age 25, sometimes called "Gen Z's Patrick Collison" — told Fortune in March: "We're perplexed by this continuous flow of 'You're going to be dead tomorrow.'"
The perplexity is understandable. $2 billion in ARR does not suggest imminent death. But the critique was never about today's revenue. It was about the structural question that the Boris Cherny boomerang, the Kimi K2.5 revelation, and the margin math all point toward:
What happens when the IDE is a feature of the model, not the other way around?
The Structural Question
Cursor's situation is not unique in tech history. It echoes every platform risk story: app stores that competed with their own developers, social platforms that replicated their most popular third-party features, cloud providers that launched managed services over their customers' open-source projects. The playbook is well-documented.
What makes Cursor's version unusual is the speed and completeness of the squeeze. In most platform risk stories, the incumbent at least had time — years to diversify, build switching costs, or establish a brand moat. Cursor went from founding to $2B ARR faster than almost any company in history. But its supplier went from API provider to direct competitor in the same window.
Anthropic's revenue grew from $1 billion in December 2024 to $14 billion annualized by February 2026. Claude Code alone is now on a $2.5 billion run-rate. The supplier isn't slowly encroaching — it's growing faster than the company it supplies.
The Composer 2 episode reveals the depth of the trap. Even the escape route of building your own model leads back to dependency — on open-source models with license strings, on fine-tuning that requires base model access, on a talent pool that your competitors also recruit from (and apparently retain better). The five escape routes aren't independent options. They're all manifestations of the same underlying constraint: you cannot build a durable business on top of a supplier who wants your customers.
Cursor may yet find a path. The company has real revenue, real enterprise adoption, and a product that millions of developers use daily. The 25-year-old CEO running a $30 billion company at the center of AI's most important market deserves more than dismissal. Platform risk stories don't always end in platform death. Sometimes the app becomes indispensable enough that the platform can't afford to kill it.
But the Cherny-Wu boomerang keeps echoing. The most vivid illustration of Cursor's predicament isn't in the financials or the benchmark scores. It's in the fact that two people saw the inside of both companies and chose the supplier over the customer. Twice.
"I don't believe the 'Cursor is dead' memes, but 'The IDE is dead' is real."
— Zach Lloyd, CEO of Warp
The distinction matters. Cursor might survive. But what it sells — an intelligent IDE — might not be a category that outlasts this transition. When the agent is the interface, the editor around it is chrome.