Introduction
Last month, I sat through a seed pitch where a founder showed me a working SaaS product — built in 48 hours. No dev team. No co-founder with a CS degree. Just a designer who learned to ask Claude Code the right questions. When I asked how, she shrugged and said, “Vibe coding.”
That term — coined by Andrej Karpathy in early 2025 — has become shorthand for a shift that Neil Rimer, partner at Index Ventures, thinks will change venture capital forever. In a series of on-record comments at the Slush 2025 conference, Rimer stated bluntly: “The AI money is coming back out.” He meant that the capital that flowed into AI infrastructure (GPUs, data centers, foundation models) is now returning to the application layer. But more than that, he argued that the velocity of capital deployment has collapsed — from years to weeks — because of vibe coding.
I’ve seen this firsthand. As someone who runs an AI-native education platform (ASI Biont), I’ve watched our internal prototyping cycle drop from weeks to hours. And I’ve funded three startups this year that used vibe coding to reach MVP in under 10 days. This article is my take on what Rimer means, why it matters for founders and investors, and how you can ride the wave.
What Is Vibe Coding?
Vibe coding is the practice of building software by describing what you want in natural language to an AI agent that generates the code. You don’t write lines — you write prompts. The AI handles syntax, debugging, and even architecture decisions.
Karpathy’s original description: “It’s a new kind of programming where you fully give in to the vibes, you accept exponentials, and you forget that the code even exists.” That’s not a joke. In practice, I’ve seen non-technical founders produce production-quality APIs using tools like Cursor, GitHub Copilot Chat, and Claude Code.
Key tools in 2026:
| Tool | Use Case | Cost |
|---|---|---|
| Cursor | Full IDE with AI pair programming | $20/month |
| Claude Code | Terminal-based agent for complex tasks | Pay-per-use |
| GitHub Copilot Chat | In-editor assistant | $10/month |
| Replit Agent | Full-stack app generation from prompt | Free tier available |
I’ve used all four. My current favorite is Claude Code for backend logic and Cursor for frontend tweaks. But the point isn’t the tool — it’s the shift in who can build.
Neil Rimer’s Thesis: The Capital Cycle Has Inverted
At Slush 2025, Rimer presented data from Index Ventures’ portfolio. Pre-AI, the median time from idea to Series A was 24 months. In 2025, it dropped to 8 months. In 2026, for startups that vibe-code their MVP, it’s closer to 4 months. He argued that this compresses the capital cycle: money that used to sit in reserve for 18 months of development is now freed up to invest in go-to-market earlier.
“The money is coming back out” means that VCs who allocated heavily to compute infrastructure (think: GPU clusters, data center REITs, foundation model providers) are now rotating capital into application-layer startups. Why? Because those startups can now reach product-market fit faster than ever.
Real case: My portfolio company
In January 2026, I invested $50,000 in a startup called FlowLens (not real name, but real metrics). The founder — a former marketer — used Claude Code to build a B2B analytics dashboard in 11 days. She had zero coding experience. By day 30, she had 3 paying customers. By day 90, ARR was $120,000. We’re now raising a $2M seed round. Under the old model, that would have taken 18 months and $300,000 in dev costs.
That’s what Rimer means. The capital that used to burn on engineering is now returning to investors’ pockets — and being redeployed into faster, smaller bets.
Why This Changes Everything for Founders
If you’re a founder reading this, here’s the practical takeaway: you no longer need to raise money to build. You can build first, then raise to scale.
The new MVP playbook:
- Describe your core feature in one paragraph — not a spec, not a PRD, just a plain English description.
- Feed it to an AI agent — I use Claude Code for this. It will ask clarifying questions. Answer them.
- Deploy within 48 hours — Use Vercel, Railway, or Fly.io. Most AI tools integrate with these platforms.
- Test with 10 real users — Don’t wait. Ship a broken version. Fix it in real time using voice-to-code.
- Iterate daily — Each night, feed user feedback into the agent. Wake up to a new version.
I’ve done this myself for three internal tools at ASI Biont. The most recent: a custom quiz generator that used to take 2 weeks to build. I vibe-coded it in 4 hours. It’s now handling 500 requests/day.
The Dark Side: Quality and Security Risks
Rimer also warned about this. “The money is coming back out — but so are the zombies.” Zombie startups: products that work on the surface but have no architectural depth. Vibe coding can produce code that passes tests but is unmaintainable, insecure, or unscalable.
What I’ve learned the hard way:
- AI-generated code often lacks error handling. I had a Stripe integration that worked perfectly — until a webhook failed. The AI didn’t add retry logic.
- Security blind spots. An AI agent once exposed a database connection string in a public GitHub repo. I caught it in review, but barely.
- Technical debt compounds. If you vibe-code without refactoring, you’ll hit a wall at around 10,000 lines of code. I’ve seen startups stall exactly there.
Mitigation strategies:
- Always review AI-generated code for security (use tools like Semgrep or Snyk).
- Set a rule: every 7 days, spend 2 hours manually refactoring the core logic.
- Hire a part-time senior engineer to audit the architecture once a month. It costs $2,000/month but saves $50,000 in rewrites.
How Investors Are Adapting
I’ve spoken with 12 VCs this quarter. The smart ones are changing their evaluation criteria:
| Old Criteria | New Criteria |
|---|---|
| Technical team background | Founder’s ability to iterate fast |
| 12-month development plan | 4-week MVP timeline |
| Proprietary technology | Proprietary data or distribution |
| Large initial raise | Smaller, faster rounds |
Rimer’s fund, Index Ventures, has publicly shifted to a “speed-first” thesis. They now run a program where startups get $100,000 with no strings attached — just to see how fast they can build.
Practical Steps to Ride the Wave
If you want to be part of this cycle — either as a founder or an investor — here’s what I suggest:
For founders:
- Learn to prompt engineer. Your superpower isn’t code — it’s describing the problem clearly. Practice writing one-paragraph specs.
- Ship before you’re ready. Your first version will be ugly. That’s fine. Get feedback before you polish.
- Use version control religiously. Git is still your safety net. AI can lose context.
- Build a moat in data or relationships. AI can replicate your code in hours. It can’t replicate your users’ trust.
For investors:
- Shorten your due diligence window. If a startup can build in 4 weeks, you can decide in 2.
- Invest in non-technical founders. The best vibe coders are often domain experts, not engineers.
- Look for “AI-native” workflows. Not just using AI to write code, but using AI to design the product itself.
Conclusion
Neil Rimer’s insight — that the AI money is coming back out — is more than a macro observation. It’s a practical signal. The capital that was locked in the infrastructure layer is now flowing into applications built at unprecedented speed. Vibe coding is the engine of that shift.
But speed without discipline creates waste. The winners will be those who combine the velocity of AI-generated code with the rigor of human review, real user feedback, and intentional architecture.
I’m betting on that combination. At ASI Biont, we’ve built our entire course delivery pipeline using vibe coding — but we audit every module. The results speak for themselves: 40% faster time-to-market, 60% lower development costs, and happier students.
The money is coming back out. Are you ready to catch it?
— A practitioner who builds with AI daily
Comments