Swaleh Asif

Co-founder of Cursor, an AI-native code editor.

Claims

As speaker

Mentioned in

Key positions and views

Running custom, scaled inference infrastructure is critical for delivering a cost-effective and high-performance AI coding product.

DeepSeek models serve as exceptionally effective, pre-trained base models for building specialized, production-grade AI applications.

The ultimate test for a developer tool feature is whether its own creators use it daily; features that don't meet this 'dogfooding' standard should not be shipped.

For interactive AI tasks like coding agents, model latency and coherence are often more important than raw intelligence, making models like Anthropic's Sonnet preferable to more powerful but slower alternatives.

The future of AI coding tools is not just about assisting developers within existing paradigms but about creating a fundamentally new, more automated development environment.

Podcast consensus on Asif

Points of consensus

▶Swaleh Asif consistently emphasizes the importance of running custom AI infrastructure at a massive scale, citing Cursor's ability to handle hundreds of millions of daily calls on its own DeepSeek V2 inference and index billions of files per day.Apr 2026

▶Asif views DeepSeek models as foundational to Cursor's technical strategy, highlighting them as excellent pre-trained base models and noting Cursor's long-term use, from the 1.5 series to running the V2 series for over 8 months.Apr 2026

▶A core tenet of Asif's product philosophy is rigorous internal validation, stating that the primary metric for shipping a feature is whether Cursor's own developers use it daily, a process that led to rejecting three internal agent prototypes before release.Apr 2026

▶Asif's perspective on the AI model market is pragmatic, prioritizing specific model characteristics for different tasks, such as preferring Anthropic's lower-latency Sonnet for agents over the more powerful Opus model.Apr 2026

Points of debate

▶Asif articulates a tension in Cursor's privacy policy, stating the current policy is not to store user code, while simultaneously revealing future plans to privately store code to enable faster and more reliable features.Apr 2026

▶There is a potential conflict between Asif's stated mission to keep the developer 'in control' and the simultaneous goal to 'automate as much of the coding process as possible,' which could reduce developer agency over time.Apr 2026

▶Asif criticizes GitHub Copilot for a lack of significant improvement post-launch, yet also credits its initial release as the key catalyst that convinced some of his co-founders to switch from Vim to VS Code, highlighting its foundational impact.Apr 2026

▶Asif's vision for Cursor is to become 'very different from VS Code' within a year, a position that contrasts with its current identity as a fork of VS Code, creating a strategic challenge of evolving the product without alienating its existing user base.Apr 2026

Key themes

▶Pragmatic AI Infrastructure at ScaleApr 2026

Asif details Cursor's commitment to building and operating its own high-throughput AI infrastructure. This includes running a custom DeepSeek V2 model for hundreds of millions of daily calls, a code completion model handling 100 million daily requests, and an indexing system designed for billions of files, all while managing costs by using tools like Turbopuffer on S3.

Cursor's investment in bespoke, scaled infrastructure suggests a strategy to compete on performance, cost, and customization, creating a defensible moat against competitors who rely solely on third-party model APIs.

▶Developer-Centric Product PhilosophyApr 2026

Asif's approach to product development is rooted in solving his own team's problems. The primary metric for shipping a feature is its daily adoption by Cursor's own developers, a practice that led to multiple internal rejections of an AI agent before finding a version they deemed truly useful.

This 'dogfooding'-centric approach ensures high-quality, practical features but may risk creating a product that is over-optimized for a small group of power users, potentially overlooking the needs of a broader developer market.

▶Strategic Model Selection and ApplicationApr 2026

Asif demonstrates a nuanced view of the AI model landscape, selecting models based on specific use cases rather than raw power. He praises DeepSeek as a base for custom applications, prefers Anthropic's Sonnet for agentic tasks due to its low latency and coherence, and notes that GPT-4 was a key inflection point that expanded the team's ambitions.

This indicates a mature AI strategy focused on a portfolio of specialized models, suggesting that the future of AI development tools lies not in a single 'best' model but in orchestrating multiple models optimized for specific sub-tasks like completion, editing, and agentic workflows.

▶Evolving Vision for AI-Native CodingApr 2026

Asif articulates a long-term vision for Cursor that moves far beyond its current form as a VS Code fork. He envisions a future product that is fundamentally different, enabled by new infrastructure that can privately store user code and new interaction paradigms like `readme.ai.md` files to provide architectural context to LLMs.

Asif is signaling a move from building an 'AI-assisted' tool to a truly 'AI-native' development environment, a strategic pivot that will require overcoming significant technical and user-trust hurdles related to code privacy and storage.

Source episodes

Timeline

Pre-Cursor

Credits the release of GitHub Copilot as the key event that prompted co-founders Amon and Arvid to switch from Vim to VS Code, setting the stage for building within that ecosystem.

Post-GPT-4 Release

States that the release of GPT-4 significantly raised the Cursor team's perception of what was possible, acting as a catalyst for building more advanced AI coding products.

Last 8-12 months

Asif reports that Cursor has been running the DeepSeek V2 model on its own inference infrastructure for this period, indicating a long-term strategic bet on the model series.

Present

Describes Cursor's current operational capacity at a massive scale, handling 100 million+ daily requests for its code completion model and running indexing infrastructure for billions of files per day.

Future

Outlines plans to evolve Cursor's infrastructure to privately store user code, a significant shift from the current policy of not storing code beyond a single request.

Suggested prompts

How does Cursor's strategy of running its own inference on models like DeepSeek V2 provide a competitive advantage against competitors relying on third-party APIs? &nearr;What are the primary technical and user-trust challenges Cursor must overcome to transition from its current no-storage privacy policy to one that privately stores user code? &nearr;How does the internal metric of 'daily use by our own developers' shape Cursor's product roadmap and potentially create blind spots to the needs of the broader market? &nearr;Analyze the potential conflict between Cursor's mission to keep the 'developer in control' and its goal to 'automate as much of the coding process as possible'. &nearr;

Key concepts

AI Model Strategy 7 ep Cursor's Infrastructure & Scale 5 ep Product Development & Vision 5 ep DeepSeek Models 3 ep Business & Privacy 2 ep Anthropic Models 2 ep GitHub Copilot 2 ep Code Completion 2 ep

Notable quotes

“So we've been running DeepSeq V2 model for eight or 10 months now, probably 12 months, something like that on our, you know, own inference. And that's the model we've scaled up to hundreds of millions of calls.”

Swaleh Asif · Inside Cursor: The future of AI coding with Co-founder Sualeh Asif

“We spent a lot of time optimizing and running at quite enormous scales, like billions of files per day kind of infrastructure.”

Swaleh Asif · Inside Cursor: The future of AI coding with Co-founder Sualeh Asif

“Cursor's mission is to automate as much of the coding process as possible while ensuring the developer remains in control.”

Swaleh Asif · Inside Cursor: The future of AI coding with Co-founder Sualeh Asif

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report