The software industry faces a structural threat that goes beyond competition within existing categories. While Andreessen Horowitz (a16z) recently published a defense of software moats in an AI world, arguing that brand, process power, and switching costs will hold firm, the analysis misses a more fundamental shift: AI is collapsing the cost of building custom software, making the entire "buy versus build" equation favor building for the first time in decades .

In February 2026, $285 billion in software market capitalization evaporated in 48 hours as investors panicked over what became known as the "SaaSpocalypse." The trigger was immediate: Anthropic's launch of Claude Cowork, an AI agent platform with plugins spanning 11 business categories, combined with disappointing earnings reports showing customers actively reducing software seat counts. But the underlying fear was structural. If AI agents can do the work that software helps humans do, the subscription model that built the industry starts to crack .

What Makes a16z's Moat Defense Incomplete?

a16z applied business strategist Hamilton Helmer's Seven Powers framework to argue that most software advantages hold or strengthen in an AI world. The firm's strongest claim centers on "process power," the idea that SaaS platforms encode years of institutional knowledge about how organizations actually operate. Their example is Harvey, a legal AI platform: if Harvey deeply understands how a specific law firm structures its work, no new entrant can replicate that overnight .

But this argument contains a critical blind spot. It assumes process knowledge must live inside the vendor's platform. Tools like Claude Skills already let users encode their own workflows and institutional knowledge in their own environment, without relying on a SaaS vendor at all. This erosion happens in two ways: first, AI tools lower the barrier for users to replicate operational logic that vendors previously held as proprietary; second, they let users encode operational context that no standardized product could ever accommodate .

The real issue is that a16z's analysis assumes competition happens within existing software categories. The version they do not fully engage with is the one where the SaaS category itself becomes optional. This is what business strategist Clayton Christensen called "New-Market Disruption," where agent workflows serve users who never justified a dedicated SaaS product in the first place .

How Is AI Changing What Gets Built?

The shift has a name: "vibe coding," a term coined by AI researcher Andrej Karpathy in early 2025 to describe a workflow where developers describe intent in natural language and AI generates functional code. What began as a novelty has become a default. By early 2026, over 80% of developers report using or planning to use AI coding tools, and an estimated 41% of newly written code is AI-generated .

But the more consequential shift is not that existing developers are faster. It is that the threshold for who can build software at all has dropped dramatically. Non-technical user adoption of AI coding tools surged over 500% year-over-year through early 2026. A founder who previously needed to hire a development team to test an idea can now prototype a working application in hours. An operations lead who managed a process through email and spreadsheets can now spin up a dedicated internal tool .

The economic logic is straightforward: when the cost to build a custom application drops by an order of magnitude, the number of applications worth building rises by a corresponding amount. Problems that were too niche, too internal, or too low-value to warrant a development project are suddenly viable. The long tail of enterprise software demand, suppressed for decades by the cost of supply, gets unlocked .

Steps to Understanding the New Software Economics

• Recognize the boundary shift: Organizations that previously accepted the compromises of general-purpose software, the unused features, the workaround processes, and the imperfect data models, now have a credible alternative in custom-built tools.
• Understand the timeline problem: a16z is mostly right in the short term; moats hold. But in the medium term, category boundaries blur and the orchestration layer fight intensifies. In the long term, structural dissolution occurs for codifiable workflow categories.
• Account for margin compression: The SaaSpocalypse is not just an earnings collapse; it is the market repricing what a dollar of software revenue is worth when gross margins may be structurally lower due to AI competition.

Enterprise adoption of AI is accelerating faster than historical technology transitions. a16z data shows that 29% of the Fortune 500 and approximately 19% of the Global 2000 are already live, paying customers of at least one leading AI startup. This is not experimentation; it is production deployment. The market has moved through three stages: Wave 1 (2023) was exploration, Wave 2 (2024) was experimentation with spending increasing by 130%, and we have now entered Wave 3, which a16z calls "Accountable Acceleration" .

In this current phase, 82% of enterprise leaders use generative AI weekly, and 46% use it daily. The a16z survey of Global 2000 chief information officers suggests the market is consolidating around an emerging oligopoly of closed-source providers. OpenAI remains the clear leader, with a strong majority of enterprises using its models in production and capturing the largest share of enterprise spend. Anthropic has emerged as the fastest-growing challenger, with rapidly increasing enterprise penetration, particularly in coding and data-intensive workflows .

The data contradicts early predictions that the "app apocalypse" would lead to purely in-house builds. Instead, enterprises are increasingly migrating from DIY large language model implementations to packaged third-party applications. This spend is being directed toward a multi-model strategy; 81% of respondents now utilize three or more model families in production, up from 68% a year ago .

Where Does the Real Value Actually Sit?

The a16z report highlights a "barbell" distribution of value, where adoption is concentrated in three horizontal categories: coding, support, and search. Coding is the dominant use case for AI by nearly an order of magnitude. Code is the "ideal" substrate for AI because it is data-dense, text-based, precise, and verifiable; anyone can run it and know if it works, creating tight feedback loops for models to learn. Best-in-class engineers can increase productivity by 10 to 20 times using AI coding tools. Because code is the core building block of all software, AI's impact here accelerates every other domain .

On the other side of the barbell is customer support. Support interactions are usually time-bound with constrained intent, making them perfect for agentic automation. This is where the real economic dividend of AI will be realized, through its diffusion into the services economy, which accounts for over 70% of GDP in developed nations. Intelligence is becoming an industrial input, something you can call on demand, price per unit, and measure in rigorous return on investment terms .

The tooling ecosystem supporting AI-assisted development, including platforms like Cursor, Claude Code, Bolt.new, and Lovable, attracted $9.4 billion in equity funding between 2022 and 2025, a signal that investors see this as structural, not cyclical . This massive investment reflects confidence that the explosion of custom software does not eliminate the need for professional services; it reshapes and multiplies it. Someone still needs to architect systems that scale, integrate custom tools with existing infrastructure, manage data governance, and maintain what gets built. The gap between a working prototype and a production-grade application remains wide, and that gap is where services firms operate .

The parallel to an earlier transition is instructive. Cloud computing did not reduce demand for IT services; it shifted what those services looked like and expanded the population of buyers who could afford them. AI-assisted development is following a similar pattern: lowering the floor for what can be built, while raising the ceiling for what needs professional support .

The fundamental question is not whether software moats hold. It is what happens when the boundaries between a SaaS product, an internal automation, and an AI agent start to dissolve, and critically, over what time horizon. For large organizations, SaaS products do more than encode process; they bind everyone into a shared process, and that standardization is often the point. A 500-person sales team needs everyone working the same pipeline, not 500 bespoke agentic workflows. But for smaller teams, specialist functions, and workflows where contextual fit matters more than standardization, the constraint that once justified SaaS adoption is disappearing .