AI Startup

An AI startup is a company whose product depends on artificial intelligence or machine learning as a core differentiator. The category breaks into three distinct archetypes: foundation model labs (OpenAI, Anthropic, Google DeepMind, Meta AI training the largest models), AI infrastructure (Hugging Face, LangChain, Pinecone, Weights & Biases providing tooling), and AI application companies (Cursor, Perplexity, Harvey, Glean building products on top of foundation models). Each archetype has fundamentally different economics, capital requirements, and defensibility characteristics. Understanding which category your AI startup falls into is the first step in evaluating its moat.

The three categories:

Foundation model labs:

• Train and improve the largest models (GPT, Claude, Gemini, Llama).
• Examples: OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, xAI.
• Capital requirements: $100M-$10B+ (compute + talent + data).
• Moat: scale, talent, infrastructure, brand.
• Few companies; very high barriers.
• Most VCs cannot play here; only mega-funds.

AI infrastructure:

• Tooling, platforms, vector databases, observability, fine-tuning platforms.
• Examples: Hugging Face, LangChain, Pinecone, Weights & Biases, Modal, Replicate, Together AI.
• Capital requirements: $5M-$100M typical.
• Moat: platform effects, integrations, developer adoption, ecosystem.
• More accessible than foundation models; significant infrastructure expertise required.

AI applications:

• Products built on top of foundation models, solving specific user problems.
• Examples: Cursor (coding), Perplexity (search), Harvey (legal), Glean (enterprise search), Notion AI, Cursor, Glean.
• Capital requirements: $2M-$50M typical.
• Moat: workflow integration, data flywheel, distribution, brand.
• Most accessible category; also most crowded.

The 2024-2026 funding pattern:

Foundation model labs: mega-rounds ($1B+ at valuations from $10B all the way to approaching $1T post-OBBBA, Anthropic closed its Series H at $965B in May 2026, OpenAI and SpaceX are in the same tier). Few participants.

AI infrastructure: strong Series A/B activity. $20M-$100M rounds at $200M-$1B valuations common.

AI applications: explosion of Series A activity. $10M-$30M rounds at $50M-$150M valuations. Significant attrition expected, only some will sustain growth.

What makes an AI startup defensible (the AI moat question):

Foundation model labs moats:

• Scale (largest models, most compute).
• Talent (top researchers).
• Infrastructure (training infrastructure).
• Brand (developer mindshare).
• Capital (ability to fund next training run).

AI infrastructure moats:

• Developer ecosystem (LangChain effect).
• Platform integrations.
• Open source community.
• Performance / cost advantage.

AI applications moats:

• Workflow integration (deeply embedded in customer workflow).
• Data flywheel (unique customer data improves AI over time).
• Distribution (customer relationships, partnerships).
• Domain expertise (legal AI requires legal knowledge).
• Brand and trust.
• NOT typically: raw model access (commoditizing).

The "AI wrapper" critique:

A common dismissal of AI startups: "you're just a wrapper around GPT-4." This is sometimes fair (the company has no defensible moat) and sometimes lazy (the company has built real workflow integration, data flywheel, or distribution that's not commoditized). See AI Wrapper for the nuanced view.

The economics question:

Foundation models: economics still being worked out. Training costs vs inference costs vs API revenue. Major capital expenditure for diminishing returns at scale.

Applications: economics depend on inference cost. A startup paying $0.10 per request via API needs to charge enough to cover that plus margins plus other costs. As inference costs drop (10-100x in 2023-2025), economics improve dramatically.

What founders should evaluate:

Which category fits: be honest about which category you're in.

Moat strategy: how will you defend against commoditization?

Inference cost trajectory: will your unit economics improve as costs drop?

Foundation model relationship: are you dependent on one provider? Strategy if they change pricing or terms?

Data moat: do you have unique data that improves your AI?

What founders get wrong: Treating "AI startup" as a single category and applying strategies that fit one type to another. The right discipline: clearly identify which category (foundation/infrastructure/application) you're in; build moats that match; design pitch and strategy accordingly.

Related: Foundation Model · Large Language Model · Generative AI · AI Moat · AI Wrapper · Data Flywheel

FAQ

What is an AI startup?
A company whose product depends on artificial intelligence or machine learning as a core differentiator. Three categories: foundation model labs (training largest models), AI infrastructure (tooling), AI application companies (products on top of foundation models).

What are the three categories of AI startups?
Foundation model labs (OpenAI, Anthropic, Google DeepMind): $100M-$10B+ capital, moat from scale and talent. AI infrastructure (Hugging Face, LangChain, Pinecone): $5M-$100M, moat from platform effects. AI applications (Cursor, Perplexity, Harvey): $2M-$50M, moat from workflow integration and data flywheel.

What makes AI startups defensible?
Foundation models: scale, talent, infrastructure, capital. Infrastructure: ecosystem, platform integrations, developer adoption. Applications: workflow integration, data flywheel, distribution, domain expertise. Raw model access is NOT typically a moat (commoditizing).

How much capital does an AI startup need?
Foundation model labs: $100M-$10B+ (training costs). Infrastructure: $5M-$100M (platform building). Applications: $2M-$50M (typical SaaS economics with inference cost layer). Significant variance by category and stage.