Windsurf CEO Deep Dive Interview: Speed Is the Only Moat

Windsurf CEO Varun Mohan recently sat down for an in-depth interview with 20VC host Harry Stebbings, sharing profound insights on startup pivots, product building, and the competitive landscape of AI coding tools.

The Core of Startups: Don't Fall in Love with Your Ideas

Varun candidly shared that the company has gone through multiple pivots. It started as Exafunction, focused on GPU virtualization; then pivoted to Codium for code completion; and ultimately built Windsurf, an AI-powered coding IDE.

GPU virtualization refers to abstracting and partitioning physical GPU resources through a software layer, enabling multiple applications or users to share the computing power of a single GPU. Exafunction initially targeted this direction—in deep learning inference scenarios, many models cannot fully utilize a GPU's compute capacity, resulting in massive waste of expensive hardware resources. Through virtualization, a single GPU's compute resources can be dynamically allocated across multiple inference tasks, significantly reducing infrastructure costs. This direction had some market demand during 2020-2022, but as the era of large models arrived, individual models' GPU demands expanded dramatically, weakening the value proposition of GPU virtualization—which became a key factor behind Varun's team deciding to pivot.

He argues that founders need to simultaneously hold two seemingly contradictory beliefs: irrational optimism and uncompromising realism. Every day, you must ask yourself: do we have a reason to exist? If the answer is no, you must pivot quickly.

"Startups don't win awards for persisting longer at the wrong thing. When you eventually fail, nobody cares how long you stuck with it."

Speed Is the Only Moat

Varun states plainly that the traditional "7 Powers" framework is premature for startups. The 7 Powers model comes from Stanford professor Hamilton Helmer's book 7 Powers: The Foundations of Business Strategy, a strategic analysis framework widely cited in Silicon Valley investment circles. The seven powers are: Scale Economies, Network Effects, Counter-Positioning, Switching Costs, Branding, Cornered Resource, and Process Power. Helmer argues that a company's long-term excess profits must be supported by at least one of these "powers." However, Varun's view is that for early-stage AI startups, these traditional moats haven't been established yet, and pursuing moats too early actually distracts from product iteration speed. During windows of dramatic technological paradigm shifts, the only reliable competitive advantage is learning and iterating faster than your competitors.

Even for Nvidia, the real advantage isn't CUDA—it's the speed at which they continuously ship faster hardware every year. CUDA (Compute Unified Device Architecture) is a parallel computing platform and programming model launched by Nvidia in 2006, allowing developers to use languages like C/C++ to directly leverage GPUs for general-purpose computing. After nearly two decades of accumulation, CUDA has formed a massive software ecosystem including cuDNN (deep learning acceleration library), TensorRT (inference optimization engine), and other core components—virtually all mainstream deep learning frameworks deeply depend on CUDA. The industry widely considers the CUDA ecosystem to be Nvidia's deepest moat, but Varun offers a deeper insight: CUDA itself isn't irreplaceable. What truly keeps Nvidia ahead is its organizational capability and execution speed in shipping dramatically improved next-generation hardware every year—from V100 to A100 to H100 to B200, each generation delivering multiple times the performance improvement.

He believes the core value of first-mover advantage lies in two things:

1. It demonstrates an organizational culture willing to disrupt itself
2. Learning faster from the market and discovering the next opportunity first

Windsurf ships a major update every 1-2 weeks. If you go three months without releasing something impressive, you become irrelevant—the rise and fall of GitHub Copilot and Devin are perfect examples.

GitHub Copilot was launched in 2021 as a joint effort between GitHub and OpenAI, becoming one of the earliest large-scale commercial AI code completion tools. Built on OpenAI's Codex model, it automatically generates code suggestions based on code context. Copilot was once considered the undisputed king of AI coding tools, backed by GitHub's massive developer user base and Microsoft's resources. However, as products like Cursor and Windsurf emerged with deeper IDE integration and Agentic experiences, Copilot gradually lost its cutting-edge aura due to relatively slow iteration speed and insufficient product innovation. Devin, on the other hand, was released by Cognition Labs in early 2024 as a "fully autonomous AI software engineer," generating enormous hype and being hailed as a milestone product that could potentially replace human programmers. But subsequent real-world testing revealed that Devin's performance in complex real scenarios fell far short of its marketing claims, with significant issues in completion rate and code quality, and its hype quickly faded. These two cases perfectly validate Varun's point: in the AI coding tools space, if you don't continuously innovate, you'll be forgotten quickly.

Product Building Methodology: Small Teams Validate, Fast Elimination

Varun shared his core product philosophy: new ideas are validated with small teams of just 3-4 people. The key judgment criterion is—even a rough version of a good idea will demonstrate astonishing value.

Over 50% of internal projects at the company ultimately fail, but that one successful project is enough to cover the cost of all failures. He emphasizes that the company succeeds not because it does many things, but because it does one thing exceptionally well.

The Real Challenges of Async Agents

Regarding the much-discussed async remote Agents, Varun maintains a pragmatic stance. He identifies three key dimensions: latency, quality, and correctability.

Async remote Agents refer to AI coding agents that run independently in the cloud without requiring real-time user interaction. Unlike synchronous Agents embedded in IDEs like Windsurf, async Agents receive high-level task descriptions from users, then autonomously complete code writing, testing, debugging, and other work in the background, ultimately returning results to the user. The appeal of this model is that it can process multiple tasks in parallel, theoretically dramatically improving development efficiency. However, the three dimensions Varun identifies reveal its core dilemma: when task execution time extends from seconds to hours, users lose the ability to correct the Agent's direction in real-time (reduced correctability), meaning the Agent must make numerous decisions without human feedback, errors accumulate at each decision point, and output quality becomes difficult to guarantee. This is fundamentally an "autonomy vs. reliability" tradeoff, and a common challenge facing all Agentic AI systems today.

If async Agents take hours to complete tasks, users will have extremely high expectations for output quality—requiring at least 90-99% accuracy. Therefore, in the short term, complex tasks will still be completed locally in Windsurf, while async Agents are only suitable for simple tasks with high determinism.

Measured Responses to Hot Industry Takes

On the "one-person billion-dollar company": Varun explicitly says he doesn't believe it. If one person can do it, two equally smart people can replicate it, and competition will compress margins.

On model moats: In the short term, no single model provider will dominate the market. New technologies can emerge at any time to let laggards leapfrog ahead.

On competing with Cursor: He candidly acknowledges that Cursor excels at UI/UX, while Windsurf focuses more on technical breakthroughs driving product experience. Over 50% of the company's revenue comes from enterprise customers, and JetBrains plugin support is a key differentiator.

JetBrains is a software development tools company headquartered in the Czech Republic, with a portfolio of widely popular IDE products including IntelliJ IDEA (Java), PyCharm (Python), and WebStorm (JavaScript). In enterprise development, JetBrains holds an extremely significant market share—especially in the Java/Kotlin ecosystem, where IntelliJ IDEA is effectively the standard development tool. Many large enterprise development teams have built complete workflows, plugin ecosystems, and compliance configurations around JetBrains IDEs, making migration to VS Code or other editors costly. By providing JetBrains plugin support, Windsurf enables enterprise developers to use AI coding capabilities without changing their existing work environment, dramatically lowering the decision threshold for enterprise procurement. In contrast, Cursor currently only exists as a VS Code fork and cannot serve the JetBrains user base, giving Windsurf an important differentiated competitive advantage in the enterprise market.

The Future of Engineers

Varun believes the term "engineer" will become very fluid. Just like the evolution from assembly to Python, in the future some people will work purely at the natural language level.

The history of programming languages is essentially a history of continuously raising abstraction levels. The earliest programmers wrote machine code (binary instructions) directly, then assembly language replaced binary with mnemonics, C introduced structured programming and hardware-agnostic expression, and high-level languages like Java and Python further abstracted away memory management and platform differences. Each elevation in abstraction level enabled more people to participate in software development while making low-level experts more scarce and valuable. Varun sees natural language programming as the latest stage in this evolution: future "engineers" may only need to describe requirements in natural language, and AI systems will generate, test, and deploy code. But he also emphasizes that just as kernel developers still need to write C and assembly code today, critical production systems—such as bank core trading engines, aviation control systems, and medical device firmware—will still require engineers who understand underlying computing principles to ensure system correctness, security, and performance.

But critical production systems (like bank transaction processing) will still need engineers who can go deep into the underlying layers—just fewer of them.

His company's mission: Reduce the time to build technology by 99%.

Key Takeaways

• Speed is the only moat for AI coding tools; traditional moat theory doesn't apply to startups
• Product validation uses 3-4 person teams; good ideas show astonishing value even in rough versions
• Async Agents are constrained in the short term by quality and correctability; complex tasks still require local interaction
• Doesn't believe in the one-person billion-dollar company; competition inevitably compresses margins
• Enterprise customers account for over 50% of revenue; JetBrains support is a key differentiator