I’ve sat through 73 AI startup pitches in the last six months. I funded two. The other 71? Most were some variation of “we put a nice interface on GPT and charge $29/month.”

That’s not a startup. That’s a weekend project with a Stripe account.

The Graveyard Is Getting Crowded

Here’s what’s actually happening on the ground: AI wrapper startups that raised $2M-$5M seed rounds in 2024 are hitting a wall. Their 18-month runway is burning down, and Series A investors aren’t biting.

Why? Because the math doesn’t work.

A typical AI wrapper startup I see has gross margins around 40-55% after API costs. Compare that to traditional SaaS at 75-85%. You’re starting with worse economics and no structural advantage. When OpenAI raises prices (which they’ve done) or changes their terms (which they have), your entire business model shifts overnight and you had zero say in the matter.

I talked to a founder last month who built an AI writing tool for real estate agents. Good product, honest founder, real customers. But his $4.50 per-user API cost on a $19/month subscription left him with razor-thin margins after infrastructure and support. He’s got 2,200 paying users and still can’t make the numbers work for a Series A.

That’s not a founder problem. That’s a structural problem.

What Investors Actually See

I’ll be blunt because I think founders deserve honesty instead of polite rejections.

When an investor looks at your AI wrapper, here’s the mental checklist:

Can OpenAI or Anthropic ship this feature natively? If yes — and it usually is — your entire company is one product update away from irrelevance. We’ve watched this happen in real time. ChatGPT added custom instructions, memory, file analysis, image generation. Each feature killed a crop of startups.

What happens when a well-funded competitor copies you? Your prompt engineering and UI can be replicated in 6-8 weeks by a decent team. That’s not a moat. That’s a head start measured in days.

Are your customers loyal or just curious? Early AI adoption is driven by novelty. Month-3 retention tells the real story, and for most wrappers I’ve seen, it’s ugly — 30-40% monthly churn in some cases.

TechCrunch has been covering the AI funding slowdown, and the data backs up what I’m seeing in pitch meetings. The easy money for “AI-powered [noun]” is gone.

What Actually Makes an AI Startup Defensible

Not all AI startups are wrappers. Some are building real companies. Here’s what separates them.

Proprietary data loops

The best AI startups I’ve funded generate data through usage that makes the product better over time. One company in my portfolio processes 50,000+ legal documents monthly. That corpus, annotated by actual lawyers using the tool, is worth more than the underlying model.

You can’t download that from Hugging Face.

Vertical depth that hurts to build

If your product requires six months of domain expertise just to understand the problem, that’s a moat. Not because the code is complex, but because the knowledge embedded in the product is hard-won.

I backed a startup doing AI-assisted compliance for Australian financial services. They spent 14 months just mapping regulatory requirements before writing product code. No one’s copying that on a weekend hackathon.

Owning the workflow, not just the intelligence

The AI is one piece of a larger system. If you’re embedded in how a company operates — connected to their data, woven into daily processes, trained on their specific context — switching costs are real even if the AI layer is commoditised.

Salesforce isn’t defensible because of any single feature. It’s defensible because ripping it out costs hundreds of thousands of dollars. Build that kind of stickiness around your AI product.

Infrastructure, not applications

The companies selling shovels during a gold rush tend to do better than the miners. AI infrastructure — monitoring, evaluation, security, deployment tooling — has better margins and stickier customers than end-user applications.

The Uncomfortable Question

If you’re running an AI wrapper startup right now, here’s what I’d ask over a beer:

Could a smart engineer recreate your core product in two weeks using the same APIs? If the honest answer is yes, you don’t have a company. You have a feature.

That doesn’t mean shut down. It means pivot toward something defensible while you still have runway. Some of the best companies I know started as wrappers and evolved into something more interesting once they found a real problem hiding underneath.

What I’m Actually Funding

I’m writing checks for AI startups that have at least two of these:

• Proprietary data that improves with usage
• Deep domain expertise that takes years to acquire
• Workflow integration that creates genuine switching costs
• A business model where AI costs decrease as a percentage of revenue over time

If your pitch is “we’re like ChatGPT but for [industry]” — save us both the meeting. But if you’ve found a problem where AI is necessary but not sufficient, and you’re building the hard stuff around it? Let’s talk.

The AI startup wave isn’t over. But the easy part is. According to Startup Daily, Australian AI startups raised over $800M in 2025. The question for 2026 isn’t whether there’s money — it’s whether founders are building things worth putting money into.

Most aren’t. Some are. The difference is defensibility, and that was always the game.