The Economics of Vibe Coding: Productivity Without Defensibility?
As developers shift from writing code to supervising AI systems, the economics of software are changing and so is the distribution of value.
News
[Image source: Chetan Jha/MITSMR India]
AI systems are increasingly writing the code that once defined software engineering. The shift, described as “vibe coding,” is accelerating development cycles and attracting capital, while raising deeper questions about defensibility and long-term value creation.
In early 2025, Andrej Karpathy, a Co-Founder of OpenAI and former head of AI at Tesla, gave the trend its name. “There’s a new kind of coding where you fully give in to the vibes, embrace exponentials, and forget that the code even exists,” he wrote. “I call it vibe coding.”
A couple of years earlier, he had suggested that the “hottest new programming language is English.”
Vibe coding involves describing software requirements in plain language and allowing AI systems to generate the underlying code. Rather than writing functions line by line, developers specify features and architecture, then review, test, and refine what the system produces.
Tools such as the AI-powered code editor Cursor, the coding platform Replit, and Anthropic’s Claude Code now translate natural language into working applications.
NVIDIA Founder and Chief Executive Officer Jensen Huang captured the mood at a recent Cisco AI summit when he quipped, “Programming, as it turns out, is just typing.”
The productivity gains are measurable. Anthropic executives say that for many of the firm’s products, AI effectively writes all the code, with developers supervising agents that implement features, run tests, and fix bugs in parallel.
A 2023 randomized controlled trial by researchers from MIT, Princeton and other institutions found that developers using GitHub Copilot completed a coding task 55.8% faster than those working without AI assistance, reinforcing claims that AI copilots can materially reduce development time and costs.
That helps explain the surge of capital into the sector.
Emergent, an Indian startup building a vibe coding platform, has raised $100 million within seven months of launch and claims $50 million in annual recurring revenue with five million users across more than 190 countries.
Investors are wagering that the addressable market will grow alongside productivity gains.
Nandagopal P., CEO of product engineering firm Asymmetri, Chief Technology Officer of investment platform Gacsym Ventures, and limited partner at India-focused venture firm Arya Ventures, argues that the market could grow from 30 million professional developers to perhaps 300 million product builders. “Investors are underwriting not just incremental productivity gains, but a potential redefinition of who can build software,” he says.
Yet the optimism is mixed with unease. Even the builders feel it.
OpenAI Chief Executive Officer Sam Altman recently admitted that when an AI coding agent suggested ideas better than his own, “I felt a little useless, and it was sad.”
Aditya Agarwal, Founder and Chief Executive Officer of the health-tracking app Bevel, described feeling “Happy but also sad and confused,” along with wonder, at the prospect of never writing code by hand again.
Beneath the emotion lies a structural concern. Faster output does not eliminate long-term maintenance, which still depends on human domain knowledge. Several startups report rising “AI debt”, the hidden cost of untangling and maintaining code generated at speed.
Investors are wary.
“The biggest concern is the absence of a defensible moat,” says Nandagopal. Many startups, he argues, are wrapping a foundation model with prompts and a user interface. “Without proprietary data, deep workflow integration, or meaningful differentiation, the defensibility is extremely weak.”
He also warns of “weak evaluation frameworks” and shallow enterprise readiness, especially in security and compliance.
Ashish Bhatia, Founder and CEO of early-stage venture firm India Accelerator, shares concerns about superficial differentiation.
“Many products look impressive in demos but lack proprietary IP, deep workflows, or defensibility,” he says, adding that early-stage firms remain “heavily dependent on large foundational model providers, which creates margin and platform risk.”
The winners, he says, will be those who “build on top of the model via fine-tuning, orchestration layers, or domain-specific intelligence.”
Platform risk recurs in investor conversations.
Ranjeet Shetye, a mentor at deep-tech venture fund YourNest Venture Capital and the Chief Product and Technology Officer at agri-tech platform MapMyCrop, notes that most AI coding startups rely on a handful of providers, such as Google, Anthropic, and OpenAI.
“Any modifications to model access, pricing, or licensing may have a significant effect on the company,” he says. For him, reducing that reliance has become a crucial criterion for evaluation.
Security and reliability also loom large. Shetye warns that AI tools generating unsafe or non-compliant code could become “an invisible zero-day attack vector at a massive scale.” He stresses that embedding within real engineering processes matters more than dazzling surface-level features.
Even enthusiasts draw boundaries. Sundar Pichai, Chief Executive Officer at Google and Alphabet, has said he would not rely on vibe coding for large codebases where security must be watertight.
Boris Cherny, an engineer at Anthropic and contributor to Claude Code, has argued that while the approach is powerful for prototypes, “You want maintainable code sometimes. You want to be very thoughtful about every line sometimes.”
Still, few doubt the durability of the shift. “It’s obvious that this is a platform change rather than a passing trend,” says Shetye.
AI is not merely speeding up individual lines of code, he argues, but transforming how software is created, evaluated, and deployed.
Bhatia calls it “a platform opportunity, but only for a small subset of players.” Many tools will fade. Those that integrate deeply into development lifecycles could become core infrastructure.
The returns may be concentrated. “Strong returns are likely to accrue only to the top 0.1% of companies in this space,” says Nandagopal. The rest may struggle with commoditisation and platform encroachment.
If foundation model providers capture the entire workflow from generation to deployment, he adds, “independent startups will struggle to defend meaningful value.”
For India, the opportunity may lie at the application layer. With a vast developer base and a strong SaaS (software-as-a-service) talent pool, the country may build globally competitive application-layer agents rather than foundational models.
Shetye points to sectors such as fintech, banking and industrial systems where specialized AI tools can deliver measurable commercial outcomes.
Vibe coding compresses development cycles and lowers barriers to entry. But taste, architecture, and judgement become more valuable. As Nandagopal puts it, differentiation must come from depth and defensibility, “not just interface layering on top of a general purpose model.”
