The vibecoding bet: why the next ten million developers will never write code

The first version of vibecoding was “anyone can write code.” The second version is “anyone can ship production software.” That is a different and larger transition, the infrastructure for it just shipped, and the products that will deliver it mostly do not exist yet.

Sixty-three percent of vibecoding users identify as non-developers.¹

That single number, buried halfway down most pitch decks in the category, is the most important fact in the entire $30 billion vibecoding market. It tells you the category is not a developer-tools market masquerading under a more exciting name. It is something else, and almost nobody is selling to who it actually is.

I want to walk through that mismatch carefully, because it is the foundation of a bet that I am running thirty parallel experiments against right now.

---

I. The fastest category in software history

In February 2025, Andrej Karpathy posted a short essay describing his own coding workflow. He called it “vibecoding”: instead of writing code line by line, you describe what you want in plain English, an AI generates the implementation, you accept most of what it produces without reading the diff, and you iterate by talking, not typing.

The post got modest engagement. Karpathy is not unknown (he had 800k followers at the time), but the term itself was an offhand coinage, not a movement.

Fourteen months later, in March 2026, Collins English Dictionary named “vibe coding” its Word of the Year.² What happened between those two moments is one of the most aggressive market formations in the history of developer tools.

The revenue is real, the user numbers are real, the speed of adoption is, by every measure, without precedent in developer tools history. But the part the financial analysis doesn’t capture is the orange box on the right. The category is mis-classified at the root, and the mis-classification is hiding the largest opportunity inside it.

---

II. What vibecoding actually is, beneath the hype

Let me be precise about what shifted, because the hype obscures the substance.

Before vibecoding, writing software required three things together: knowledge of how to express logic in a programming language, understanding of the surrounding ecosystem (package managers, build tools, deployment, runtime constraints), and the patience to debug. Each was a multi-year apprenticeship. The result was a profession with a high barrier to entry and a 10x productivity gradient between novice and expert.

What AI coding tools have done across three waves (autocomplete with Copilot in 2021–2023, chat-driven editors with Cursor in 2024, full-stack generators with Lovable / Bolt / v0 in 2025) is collapse the first requirement almost entirely, and substantially compress the second and third. You still need to know what you want. You still need to recognise when the output is wrong. You still need product judgment. But you do not need to remember which arguments useState takes, and you do not need to configure Vite or webpack or Next.js by hand.

The qualitative break from the 2015–2022 no-code wave matters here. Webflow and Bubble made certain categories of software easier to build, but they did so by trapping you in proprietary markup. Vibecoding tools produce actual code (JavaScript, Python, React, SQL, Tailwind) that runs anywhere and can be exported, modified, deployed independently.

That difference — real code, not lock-in artifacts — is why this wave is consequential and the previous one was not.

---

III. The cracks in the wall

If you only read the funding rounds and the ARR numbers, you would conclude that vibecoding is a one-way street that ends in software being eaten by AI. The data on what is actually happening underneath suggests something more complicated, and the complications point directly at where the opportunity is hiding.

Traffic is collapsing for prototype-focused tools. Business Insider reported in late 2025 that Lovable’s traffic was down approximately 40 percent from its summer peak. v0 by Vercel was down 64 percent since May. Bolt.new was down 27 percent.⁶ These are products that won the first wave of hype and are bleeding users who finished a one-off project and moved on.

The unit economics are fragile. Vibecoding products are compute-heavy: real-time generation, multi-step reasoning, toolchain orchestration. When startups rent best-in-class models from OpenAI or Anthropic, per-session cost can exceed average revenue per user for free-tier users. The path to sustainable economics either requires building in-house models (capital-intensive) or repricing aggressively (kills growth).

The quality is uneven. CodeRabbit’s analysis of 470 pull requests found that AI-generated code has 1.7x more major issues and 2.74x higher security vulnerability rates than human-written code.⁷ The Lovable security incident (CVE-2025-48757) revealed that 10.3 percent of audited Lovable-generated apps had critical row-level security flaws in their Supabase configurations.⁸ Not edge cases. Production reality when non-experts ship without review.

Apple banned three vibecoding apps from the App Store. Replit mobile, Vibecode, and Anything were all pulled under Guideline 2.5.2 (apps cannot download or execute code that changes functionality post-review). Apple is not anti-AI. Apple is in a structural conflict with how vibecoding tools work: continuous iteration versus one-time review. The conflict will not resolve quickly. App submissions to the App Store surged 84 percent between mid-2025 and early 2026, and Apple’s review queue, designed for professional developers, buckled.⁹

The cracks are not fatal. None of them disprove that vibecoding is real or large. But they do tell you something specific about which segment of the market is mature, which is fragile, and where the structural opportunities still lie.

---

IV. The two-tier landscape, and why nobody is selling to tier three

The market, as currently structured, has two clear tiers and a third tier that is mostly empty.

Tier 1 is mature. Cursor’s enterprise revenue is now 60 percent of total¹⁰; the willingness to pay is high; the moat is the IDE integration plus codebase context. Tier 2 is overcrowded and bleeding, willingness to pay is capped at hobbyist tiers, compute costs are high, retention is genuinely bad because users finish their one project and leave.

Tier 3 is the empty one. Think about who falls in this gap:

• A non-technical founder who has raised seed funding and needs to build an actual product, not a prototype. They do not want a Lovable MVP that breaks at 100 users. They want software they can put into market.
• A CxO at a mid-market company who needs internal tools, CRM integrations, dashboards, custom workflows, that the IT department cannot prioritise. They want a tool fifty employees use daily and IT can eventually take over.
• A product manager who wants to ship internal-facing software (analytics dashboards, ops tools, marketing experiments) without a full engineering sprint. They want something that does not embarrass them when shown to their CTO.
• A solo operator running a portfolio of small SaaS products who needs to ship at the pace of a venture portfolio, not the pace of a traditional engineering org.

These buyers exist in large numbers. They have real budgets, substantially higher than the $25/month hobbyist tier. And they care about specific things that the current Tier 2 tools do not deliver: production-grade architecture, opinionated defaults, multi-agent orchestration, autonomous iteration, and stack legibility that lets them hand the project off to a hired engineer without rewriting it.

The 60 percent of vibecoding users who are non-developers are not all hobbyists. A meaningful subset, my read is 15 to 20 percent, are professionals trying to use Tier 2 tools to do Tier 3 work, and finding the experience frustrating.

---

V. Why Cloudflare’s Agents Week 2026 changes the bet

The technical case for building Tier 3 today, versus eighteen months ago, rests on infrastructure that did not exist before April 13, 2026.

Cloudflare’s Agents Week between April 13 and 20 shipped a complete stack that turns what would have been an impossibly heavy infrastructure problem into a tractable one.¹¹ The relevant primitives:

• Sandboxes GA: persistent, isolated Linux environments for agents with shell, filesystem, and background processes. An agent can clone a repo, install packages, run long-running processes, persist state across sessions.
• Dynamic Workers: isolate-based runtime sandboxes that execute AI-generated code in milliseconds (100x faster than containers). Kills the old tradeoff between insecure, too slow, and too restrictive.
• Durable Object Facets: every app an agent generates gets its own isolated SQLite. Exactly the right primitive when a single Tier 3 tool is generating dozens of customer applications.
• Artifacts: git-compatible versioned storage for agents, supporting tens of millions of repositories with fork-from-remote. Source control as a primitive.
• Cloudflare Mesh: zero-trust private network access for agents to internal databases and APIs. Critical for the CxO use case where the tool needs to reach into customer infrastructure.
• AI Gateway across 14+ providers, Agent Memory, and VibeSDK round out the stack.

The strategic significance is that Tier 3 is no longer an infrastructure-bound problem. Eighteen months ago, building a multi-agent system that could spin up persistent environments, run code safely, manage state across sessions, and integrate with private networks required either $5–10M of custom infrastructure or compromises that broke the product. As of April 2026, the entire stack is rentable from Cloudflare at unit economics that work for a small team.

This is the kind of shift that changes which companies can be built. The 2010 cloud shift made consumer SaaS possible at solo-founder scale. The 2026 agent infrastructure shift makes Tier 3 vibecoding products possible at the same scale.

---

VI. The Tier 3 product, specifically

What does a Tier 3 product look like? It is not just “Lovable with better defaults.” The architecture is structurally different.

The user experience is goal-oriented, not feature-oriented. A Tier 2 product asks “what do you want to build?” and accepts a freeform description. A Tier 3 product asks “what business outcome are you trying to drive?” and decomposes that into specifications. The conversation starts at the business level and descends to the technical level only after intent is locked.

The system is multi-agent by default. A Product Manager agent owns scope and acceptance criteria. An Architect agent owns the technical design and stack choices. A Coder agent (or several specialised coders) writes the implementation. A QA agent runs tests, checks security patterns, validates against acceptance criteria. A DevOps agent handles deployment. Each agent has a defined role, output format, and hand-off protocol. The user interacts primarily with the PM agent; the rest run in the background.

The loop is autonomous, not interactive. The user describes the goal. The system runs an end-to-end build → test → review → fix loop until the result passes acceptance criteria or until a defined budget (time or tokens) is exhausted. The user is in the loop for goal-setting and final acceptance, not each iteration.

The stack is opinionated and legible. Not “what framework do you want?” but “we use Next.js + TypeScript + Tailwind + shadcn/ui + Cloudflare Pages + Workers + D1 + Clerk + Stripe, and here is why.” Opinionation removes choice overhead for the non-technical user. Legibility means a hired engineering team recognises everything.

The output is production-shaped. Not just code that runs. Code that includes environment configuration, secrets management, observability hooks, error boundaries, security policies, accessibility patterns, CI/CD, and deployment scripts. The Tier 3 product ships what a senior engineer would ship on day one, not what a junior would ship for a portfolio piece.

The pricing reflects business value, not seat economics. $25/month is the Tier 2 ceiling. Tier 3 is closer to $500–2,000/month per project, or outcome-based pricing.

This is what FlareCode v2, the moonshot in my own current 100-day sprint, is being built toward. I picked the architecture because, after building twenty-seven products on the previous generation of tooling, the gap between “can prototype something” and “can ship to a customer” is the bottleneck that costs me weeks per venture. If a tool exists that closes that gap end-to-end, every non-technical operator in the world has a different career within twelve months. That is the prize.

---

VII. Who will win Tier 3

The honest answer is that we do not know yet, and the category is too young to have clear winners. But the conditions for who is likely to win are unusually specific.

Distribution advantage matters less than usual. This is not a market won by the biggest sales team. The buyers are sophisticated enough to evaluate on product quality, and the product itself produces evidence, you can see what it builds. The marketing is the work.

Opinion matters more than usual. A Tier 3 product is fundamentally an opinionated bet on what production software should look like in 2026. The companies that win will have strong, specific points of view about architecture, stack, workflow, and quality. The ones that try to be flexible across all preferences will lose to the ones with strong opinions.

Founder-software fit matters more than market-product fit. The reason non-technical founders are the highest-leverage ICP is that the founder of a Tier 3 product is the proof of concept. If the founder cannot use their own tool to ship the product, the product is not real. The best operators will eat the brand-name competitors here.

Vertical specialisation will emerge faster than expected. A horizontal Tier 3 product (build any app) is harder to ship at quality than a vertical one (build CRM workflows, analytics dashboards, internal tools, e-commerce). The first commercially successful Tier 3 products will be verticalised.

Companies to watch as of mid-2026: Replit Agent is closest on the prototyping-to-production axis but DNA is consumer/educational and the enterprise pivot is incomplete. Cognition (Devin) is the most ambitious multi-agent attempt but has struggled with reliability and is closer to a developer-facing tool. Lovable’s Pro tier is pushing toward Tier 3 with enterprise features but the prototype DNA may constrain how far it can go. Vercel v0 is moving toward production-grade UI generation but stays primarily a developer tool. FlareCode v2 and similar Cloudflare-stack opinionated builds are the indie founder bet on Tier 3. A wave of vertical Tier 3 products, CRM-specific, analytics-specific, e-commerce-specific, will emerge through 2026 and 2027.

---

VIII. The contrarian thesis, made explicit

If I had to compress the bet to three sentences:

1. The current vibecoding market is dominated by tools aimed at developers (who already know how to code and are just going faster) and hobbyists (who want to ship something fast and do not care about production quality).
2. The largest underserved segment is non-technical professionals who need production-grade output (non-technical founders, mid-market CxOs, product managers, consultants), and current Tier 2 tools fail them on architecture, stack, and quality.
3. The infrastructure required to build for this segment (multi-agent orchestration, persistent sandboxes, integrated observability) did not exist eighteen months ago and exists now. The companies that recognise the gap and ship before the incumbents pivot will own a large category.

The strong version of the bet is that Tier 3 is a 10x larger market than Tier 1, because every Tier 1 user is a developer (a constrained population) but Tier 3 users include every non-technical professional with a software need (an unconstrained population). The companies that figure out how to serve this segment with production-grade output will be larger than Cursor or Lovable, by a wide margin.

I think the strong version is closer to right, with the caveat that the timeline is 18–36 months, not 6–12. The first commercial Tier 3 winners will emerge in late 2026 and 2027. By 2028, the category will be obvious in retrospect.

---

IX. What to actually do about it

If you are a non-technical operator who wants to ship software: try Lovable, Bolt, or Replit Agent for any project where the output does not need to scale beyond 100 users. They are genuinely excellent for prototypes, internal tools, and validation. For anything you want in market, treat current Tier 2 tools as production-shaped sketch tools, and invest in either a human engineer to take the sketch to production, or an emerging Tier 3 product that promises to do this end-to-end.

If you are a developer: use Cursor or Claude Code or whatever AI-native editor fits your workflow. The 2–3x productivity gain on greenfield work and 1.5–2x on existing codebases is genuinely large. Adopt aggressively. The developers who refuse to adopt AI tooling will be uncompetitive on shipping pace by end of 2026.

If you are a founder building in this space: Tier 1 is mostly won by Cursor and Anthropic. Tier 2 is overcrowded and the unit economics are getting worse. Tier 3 is where category creation is still possible and where infrastructure has just become available. Pick a specific vertical or buyer persona, not a horizontal “everyone.” Be aggressively opinionated about stack and architecture. Build for production-grade output from day one, not as an upgrade path from prototype.

If you are an investor: the Tier 2 valuations look stretched relative to the traffic decline data. Cursor’s $29.3B and Lovable’s $6.6B may or may not hold. Tier 3 does not have obvious winners yet, which is exactly the period when the highest returns are available. The pattern recognition that worked in Tier 2 (rapid revenue ramp, viral product growth, prosumer pricing) will not transfer cleanly to Tier 3, which is closer to enterprise-shaped (longer sales cycles, higher ACVs, lower velocity).

I am running this bet personally across thirty parallel ventures, Tier 3 builds and the surrounding infrastructure to support them, and the public ledger of how that is actually going is the next piece in this sequence.

---

X. The closing observation

There is a story Karpathy tells, the one that birthed the term. He describes sitting at his computer, talking to it, and accepting most of what it produces without reading the diff. The implicit framing was: this is what coding feels like when the friction is removed.

What he did not say, what nobody really said in the first year of vibecoding hype, is that the friction was not the point of coding. The friction was the price of admission. It selected for people willing to sit through the apprenticeship, and it ensured that the people who shipped software had earned the right to do so.

Removing the friction is enormously valuable. It also removes the selection mechanism. The 84 percent surge in App Store submissions is exactly what you would expect when the selection is removed. So is the 10.3 percent of Lovable apps with critical security flaws. So is the App Store ban.

The question for the next eighteen months is whether the industry builds the new selection mechanism, production-grade tools that filter for outcome quality rather than developer credentials, or whether it just keeps optimising the prototype loop until the App Store bans become structural.

Sixty-three percent of vibecoding users are non-developers. The first version of vibecoding gave them the ability to write code. The second version gives them the ability to ship production software. That is a different and larger transition. The infrastructure for it is now available. The products that will deliver it are mostly not built yet.

That is the bet.

---

Sources

1. Taskade. “State of Vibe Coding 2026.” March 2026. Lovable internal reporting, cited in multiple sources including Vestbee and Taskade.
2. Collins English Dictionary. “Word of the Year 2025.” Announced January 2026.
3. Bloomberg, March 2026, on Cursor financial metrics; Crunchbase data on Series D round ($2.3B at $29.3B valuation, November 2025).
4. Bloomberg, March 2026, on Lovable’s $400M ARR; TechCrunch on Lovable’s $330M Series funding round at $6.6B valuation, December 2025.
5. Vestbee. “Not just Lovable: who and how is driving the vibe coding revolution.” October 2025. Aggregate valuation analysis of Cognition, Lovable, Replit, Cursor, Vercel.
6. Business Insider, late 2025, on vibecoding tool traffic data, cited via Vestbee.
7. CodeRabbit. Analysis of 470 PRs comparing AI-generated to human-written code.
8. Lovable security incident, CVE-2025-48757, early 2026. 170 of 1,645 audited Lovable-generated apps (10.3 percent) showed critical row-level security flaws in Supabase configurations.
9. Apple App Store data, early 2026; Apple guideline 2.5.2 enforcement actions against Replit mobile, Vibecode, and Anything.
10. Bloomberg / CNBC on Cursor enterprise revenue share, March 2026.
11. Cloudflare. “Agents Week 2026.” April 13–20, 2026. Press release: “Cloudflare Expands its Agent Cloud to Power the Next Generation of Agents.” blog.cloudflare.com/agent-readiness.