Your Favourite AI Startup Will Be Dead in 18 Months

90% of AI startups are building on sand. When the foundation models consolidate, most will vanish.

9 min read

The 15,000 Problem

There are 15,000+ AI startups funded in the last 18 months. In 18 months from now, 90% of them will be gone. Not because they're bad companies. Because they're building on foundations that are shifting beneath their feet.

This isn't doomerism. It's math. The historical survival rate for venture-backed startups is around 25% over five years. AI startups face all the normal startup risks plus one unique one: the platform they're built on improves faster than they can differentiate from it.

The Great AI Startup Deception

Most "AI companies" aren't AI companies. They're API companies with nice branding. Take your typical AI writing tool: OpenAI's GPT + custom prompts + some UI polish = $50M valuation. The problem? OpenAI can build your entire product in a weekend if they decide they want to. And they will.

This has already happened. Remember when dozens of startups built AI image editing tools? Then Adobe added Generative Fill to Photoshop. Remember the wave of AI meeting note-takers? Then Zoom, Google Meet, and Microsoft Teams all added native transcription and summaries. The pattern repeats every quarter.

The venture-backed AI startup graveyard is already filling up. Jasper, once valued at $1.5 billion, has gone through multiple rounds of layoffs. Copy.ai pivoted twice. Dozens of smaller writing tools have simply shut down. And this is just the first wave.

The Infrastructure Layer Is Consolidating

Here's what's happening at the foundation level: OpenAI dominates enterprise. Google owns search integration. Meta controls open source. Anthropic leads safety-conscious buyers. Microsoft has the enterprise sales machine.

That's it. Those are the players that matter. Everyone else is fighting for scraps. And the scraps are getting smaller as each foundation model release absorbs capabilities that used to require specialised companies.

The consolidation is accelerating. Every major model update eliminates an entire category of startups. When Claude got better at code, dozens of code-generation startups became redundant. When GPT-4 added vision, image-analysis startups lost their reason to exist. When Gemini integrated with Google Workspace, countless productivity AI wrappers became features.

The Wrapper Apocalypse

Most AI startups fall into predictable categories:

Content wrappers: ChatGPT for lawyers/marketers/doctors/whatever. Differentiation: custom prompts and compliance theatre.

Workflow wrappers: AI + existing tool integration. Differentiation: which SaaS tools they've integrated with.

Data wrappers: Fine-tuned models on industry-specific data. Differentiation: their training dataset (until foundation models get better at few-shot learning).

The harsh reality: if your competitive advantage is your prompts, you don't have a competitive advantage. Prompts are text files. They can be reverse-engineered in an afternoon. And the foundation models are getting so good at zero-shot tasks that fine-tuning advantages shrink with every release.

Who Survives?

Three types of AI companies will make it:

1. The Infrastructure Players. The ones training foundation models. That's maybe 10 companies globally with the capital and talent to compete.

2. The Data Monopolists. Companies with unique, valuable datasets that can't be replicated. Think medical imaging companies with 20 years of annotated scans, or financial data providers with proprietary market signals. The key word is "can't be replicated" — not "haven't been replicated yet."

3. The Integration Winners. Companies that become so embedded in critical workflows that switching costs are too high. They're not just AI tools — they're AI-powered business systems. They touch money, touch customers, and touch outcomes. Ripping them out would cost more than keeping them.

The Great Shakeout

We're about to see the fastest startup die-off in tech history. Here's how it plays out:

Phase 1 (Next 6 months): Foundation models get dramatically better. Most fine-tuned models become obsolete overnight.

Phase 2 (6-12 months): OpenAI, Google, and Anthropic start building vertical applications. The wrapper companies get squeezed from above.

Phase 3 (12-18 months): Funding dries up for anything that looks like a wrapper. VCs realise they've been funding features, not companies.

If you're building an AI startup today, ask yourself: what happens when OpenAI adds your core feature to ChatGPT? If the answer is "we're screwed," you're building on sand. The winners will be the companies building moats that AI can't easily cross: proprietary data, network effects, or deep integration into mission-critical workflows. Everyone else is just building very expensive demos.