The AI startup gold rush of 2023-2024 produced over 4,000 new AI-focused companies, according to PitchBook data. Most of them are already dead or dying. Roughly 1,100 AI startups shut down or were acqui-hired in 2025 alone, per CB Insights tracking. The ones that survived did so because they had something beyond a clever prompt template and a landing page.

What remains in May 2026 is a smaller, sharper, and more stratified ecosystem. Total venture funding into AI startups reached an estimated $97 billion globally in 2025 — up 40% from 2024 — but the distribution is brutally uneven. The top 20 deals accounted for 62% of that total. If you were not OpenAI ($6.6B), Anthropic ($4B), xAI ($6B), or one of a handful of infrastructure plays, you were fighting over the scraps.

Here is where the money is going, who is winning, and what the market structure actually looks like.

Follow the Money

The most striking feature of this chart is the concentration at the foundation layer. Nearly 40% of all AI venture capital goes to a handful of companies training frontier models — a market that is effectively closed to new entrants. Training a competitive frontier model now costs $500 million to $2 billion in compute alone, before salaries, data licensing, and infrastructure. The last new entrant to raise a foundation-model-scale round was Elon Musk’s xAI in late 2024; no new company has entered this tier since.

The infrastructure layer is the second-largest category and arguably the healthiest in terms of company formation. The picks-and-shovels thesis — sell tools to the miners — is playing out as expected.

Vertical AI, at $19 billion, is where the most interesting company-building is happening. These startups combine AI capability with deep domain expertise in a way that horizontal platforms cannot easily replicate.

Developer tools and consumer AI apps split the remainder, with consumer apps at the bottom — reflecting investor skepticism about building durable businesses on top of fast-moving foundation models.

The AI Startup Stack

The ecosystem has organized itself into a clear hierarchy, and a startup’s position in this stack largely determines its economics.

The crucial economic insight: the middle layers are getting squeezed. Infrastructure and horizontal platforms face a two-front war. Foundation model providers are expanding downward (OpenAI’s function calling, Anthropic’s tool use, and Google’s Vertex AI all absorb capabilities that used to be third-party). And vertical applications are building their own middleware, reducing dependence on horizontal platforms.

The layers with the strongest moats are the top and the bottom. Foundation models benefit from massive capital requirements and data advantages. Vertical applications benefit from domain expertise, proprietary workflows, and customer lock-in through integration depth.

Winners and Losers: Specific Names

The Thin Wrapper Problem

The term “thin wrapper” has become the industry’s sharpest insult, and the economics behind it are straightforward. A thin wrapper is a startup whose core product is an API call to a foundation model with a custom system prompt and a user interface on top. The characteristics:

No proprietary data. The startup does not fine-tune on unique datasets. It uses the same foundation model, with the same weights, that any competitor can access.

No workflow integration. The product does not deeply integrate into the customer’s existing systems. It is a standalone tool that the user alt-tabs into.

No compounding value. The product does not get better as more customers use it. There is no network effect, no data flywheel, no learning loop.

Concrete examples of the thin wrapper death spiral:

• Copy.ai raised $50M at a $290M valuation in 2023 on the strength of AI copywriting. By 2025, ChatGPT and Claude could do the same thing for free. The company pivoted to “GTM AI workflows” but struggled to retain customers at its previous price points.
• Numerous AI writing tools (Writesonic, Rytr, Peppertype) saw traffic declines of 40-70% between 2024 and 2025 as free ChatGPT capabilities expanded.
• In the image space, multiple AI avatar and headshot startups (ProfilePicture.ai, HeadshotPro) experienced 90%+ revenue drops after major platforms added similar features natively.

The antidote to the thin wrapper problem is now well understood: proprietary data, deep integrations, domain-specific fine-tuning, and compound AI systems that combine multiple models with retrieval and business logic. But understanding the antidote and executing it are very different things.

Valuation Trends

The AI valuation environment in 2026 has normalized significantly from the mania of 2023-2024, with sharp divergence between tiers.

Foundation models remain in a class apart. OpenAI’s reported $300B+ valuation (at its last secondary market transaction), Anthropic’s $60B, and xAI’s $50B are justified only by the expectation that these companies will capture a significant share of global compute infrastructure spend, not just software margins.

Series A and B: AI startups at this stage are now valued at 20-30x ARR, down from 40-80x in 2023. Investors require $1M+ in ARR, net revenue retention above 120%, and a demonstrable moat beyond “we use AI.” The bar for Series A has effectively risen to what used to be Series B metrics.

What investors look for in 2026:

1. Gross margins above 60%. Many AI startups have gross margins of 30-50% because of inference costs. Investors want to see a path to software-like margins.
2. Customer retention. Monthly churn below 3% for SMB, annual net retention above 130% for enterprise.
3. Defensibility beyond the model. Proprietary data, workflow integration, regulatory compliance, and distribution advantages.
4. A clear answer to “what happens when GPT-N+1 launches?” The best startups have an answer that does not involve the model quality itself — because that is a moving target they do not control.

What Comes Next

The AI startup landscape is consolidating around a few structural trends:

Vertical depth over horizontal breadth. The era of “we can do everything with AI” startups is over. The winners are picking a specific industry, understanding its workflows at a granular level, and building AI that slots into those workflows. Harvey does not try to be a general-purpose AI; it is a legal AI company that understands attorney workflows, court filing requirements, and the specific ways law firms bill for work.

Agentic systems as the next battleground. The most funded new category in 2025-2026 is AI agents — systems that can take multi-step actions, not just answer questions. Companies like Cognition (Devin), Factory AI, and Induced AI are building autonomous agents for software engineering, business operations, and workflow automation. Whether these systems can be made reliable enough for production use at scale is the $50 billion question.

The infrastructure shakeout. The vector database market alone has over a dozen well-funded companies (Pinecone, Weaviate, Qdrant, Chroma, Milvus, etc.) serving a market that may not be large enough for all of them. Consolidation through M&A is inevitable. The same is true for LLM observability (Langfuse, Helicone, Braintrust, Patronus, Arize) and fine-tuning platforms (Predibase, Anyscale, Together AI).

Open-weight models as an equalizer. Meta’s Llama 3.1 and its successors have made high-quality models freely available, reducing the advantage of paying for API access to closed models. Startups building on open-weight models have better margin structures (inference on self-hosted or spot GPU instances costs 50-80% less than API pricing) but take on more operational complexity. This tradeoff is increasingly attractive for startups with strong engineering teams.

The Bottom Line

The AI startup ecosystem in 2026 is healthier, more honest, and more brutal than the hype cycle of 2023-2024. The easy money is gone. The thin wrappers are dying. The companies that survive are the ones that built something a better model cannot simply obsolete — through proprietary data, deep integration, domain expertise, or sheer distribution power. That is not a new lesson in technology startups. But it is one that a lot of people had to relearn the expensive way.