If 2023 was the year the world discovered generative AI, 2024-2025 has been the year it wrote increasingly large cheques to build it. The AI startup ecosystem has bifurcated into three layers: foundation model companies racing to build the most capable general-purpose AI systems, infrastructure companies supplying the compute and tools to run those systems at scale, and application-layer companies building products on top of the models. Each layer has its own economics, competitive dynamics and risk profile.
Layer 1: Foundation Models — The Billion-Dollar Club
The foundation model layer has become one of the most capital-intensive industries in history. OpenAI’s $6.6 billion funding round at a $157 billion valuation dominates the headlines, but it is not alone. Anthropic, OpenAI’s primary competitor, raised a total of $7.3 billion across multiple rounds, with Amazon committing $4 billion and Google $2 billion — both structured as convertible notes rather than equity, reflecting the strategic importance of AI to both companies’ cloud businesses. xAI, Elon Musk’s AI company, raised $6 billion in May 2024 at a $24 billion valuation, with a stated ambition to build “truth-seeking” AI as an alternative to what Musk characterises as “woke” models from competitors. Mistral, the French challenger, raised approximately $650 million at a $6 billion valuation, positioning itself as Europe’s answer to American AI dominance. Cohere, focused on enterprise AI rather than consumer chatbots, raised $500 million at a $5.5 billion valuation.
The economics of the foundation model business are punishing. Training runs for frontier models cost an estimated $100-500 million in compute alone, and the models must be retrained regularly to stay competitive. The API pricing war has driven costs down dramatically — GPT-4 API prices fell by approximately 80% between its launch and early 2025, and open-source models like Meta’s Llama 3 and Mistral’s models provide capable alternatives at near-zero licensing costs. The foundation model companies are betting that scale wins: that the most capable models, despite their cost, will be indispensable to enough enterprises that the economics will eventually work. That bet has not yet been proven.
Layer 2: AI Infrastructure — Selling Picks and Shovels
If the foundation model companies are the gold miners of the AI rush, the infrastructure companies are selling the picks and shovels — and they have been the more reliably profitable bet. CoreWeave, originally a crypto mining operation that pivoted to GPU cloud computing, raised $1.1 billion in 2024 at a $19 billion valuation, riding the insatiable demand for NVIDIA H100 GPUs. Its revenue grew from roughly $30 million in 2022 to an estimated $2 billion in 2024, an extraordinary trajectory fuelled by AI companies willing to pay premium prices for GPU access.
Lambda Labs, another GPU cloud provider, raised $500 million. Together AI, which provides a platform for running and fine-tuning open-source models, raised $200 million. Pinecone and Weaviate, both vector database companies that enable AI models to store and retrieve information efficiently, each raised over $100 million. The common thread is that AI needs enormous amounts of specialised infrastructure — compute, storage, data pipelines, monitoring, security — and building that infrastructure is a capital-intensive but increasingly proven business model. NVIDIA’s data centre revenue, which exceeded $60 billion in fiscal 2025, is the ultimate validation: the infrastructure layer is where real money is being made.
Layer 3: AI Applications — The Promise and the Uncertainty
The application layer is where most AI startups live, and it is the most uncertain. For every success story — like Harvey, the AI legal assistant that raised $100 million at a $1.5 billion valuation, or Cursor, the AI code editor that has become essential to thousands of developers — there are dozens of companies struggling to find product-market fit in a market where the underlying technology (the LLM) is a commodity and switching costs are low.
The fundamental challenge for AI application startups is defensibility. If your product is a wrapper around GPT-4 or Claude, what stops a competitor from building the same wrapper, or the model provider from offering the same functionality natively? OpenAI’s launch of custom GPTs — user-configurable versions of ChatGPT — was widely seen as a threat to the hundreds of startups building specialised chatbots. The surviving application-layer companies share a common trait: they own proprietary data, deep integrations with customer workflows or specialised domain expertise that cannot be replicated by a generic model. AI startups that lack one of these moats are vulnerable.