AI ventures: the global race to fund foundation models, agents, infrastructure, and coding tools

What it is

The AI ventures beat tracks private capital flowing into four technology sub-sectors: foundation-model labs, AI agents, AI infrastructure, and AI coding and developer tools. Together these four segments absorbed the majority of global venture capital deployed in 2025 and 2026. Tracking them matters because the capital is directional: it signals which countries control the next generation of AI capability, which compute infrastructure becomes strategically critical, and which governments respond with export controls, sovereign investment, or antitrust action. The beat sits at the intersection of technology and geopolitics. A funding round in San Francisco can tighten a bottleneck in Tokyo or raise a parliamentary question in Brussels.

History

The modern AI venture cycle started with the November 2022 release of ChatGPT by US company OpenAI. Foundation-model labs, which had attracted roughly US$1.5 billion in total private capital through 2021, drew more than US$20 billion in 2023 alone as Anthropic (United States), Google DeepMind (United Kingdom), and Mistral (France) raised large rounds. The GPU shortage of 2022-2023 created the AI infrastructure segment: developers who could not source Nvidia chips on time began renting them from neoclouds, which became their own venture category. The AI coding sector crystallized in 2023 when US company Anysphere released Cursor, an AI-first code editor that grew to US$4 billion in annualized revenue by mid-2026. AI agents became an enterprise procurement category in 2024 when vertical specialists in United States legal, customer service, and software engineering reached nine-figure revenues. Stanford HAI's 2026 AI Index documents a 127.5% jump in global private AI investment from 2023 to 2025, with generative AI alone growing over 200%.

Current state

As of mid-2026, all four sub-beats are in simultaneous rapid movement. Foundation-model labs continue to raise at billion-dollar-plus valuations: Mirendil, a 20-person team of ex-Anthropic and ex-OpenAI researchers based in the United States, raised US$200 million at approximately US$1 billion in June 2026 at the seed stage, co-led by Andreessen Horowitz and Kleiner Perkins with Nvidia. AI infrastructure recorded its largest neocloud round yet when Together AI raised US$800 million at US$8.3 billion in July 2026, led by Saudi Arabia's Aramco Ventures; Baseten closed US$1.5 billion at US$13 billion after reaching one billion inference calls per day across 87 clusters. In AI coding, SpaceX agreed in June 2026 to acquire Anysphere for US$60 billion (see SpaceX以600亿美元收购Cursor开发商Anysphere，创史上最大初创企业收购纪录), the largest acquisition of a venture-backed startup on record. AI agents drew the largest cohort at Y Combinator's Spring 2026 Demo Day, where 196 startups competed and leading companies commanded seed valuations above US$175 million.

Relationships

The four sub-beats are tightly coupled. Foundation-model labs supply the reasoning engines that agent and coding-tool companies run on; the same Anthropic Claude and OpenAI GPT-4o models appear inside products across all three application-layer segments. AI infrastructure companies are the compute layer beneath them all. Nvidia is the common thread: it supplies the H100 and H200 chips that neoclouds rent, holds equity in several (including Together AI), and competes through its own DGX Cloud. Groq raised US$650 million in mid-2026 to rebuild as a pure inference cloud after Nvidia's US$20 billion licence-and-hire deal stripped its founding team and chip assets. Sovereign capital from Saudi Arabia, Singapore, and the United Arab Emirates entered all four segments in 2025-2026, raising questions about foreign ownership of US compute. PitchBook-NVCA data show AI and machine learning deals captured 65.6% of global VC deal value in 2025, roughly US$222 billion of US$339 billion, with half of all dollars in just 0.05% of deals.

What to watch

Energy supply is the hard constraint. The IEA documents global data-centre electricity demand rising 17% in 2025, with AI-focused sites up 50%, projecting total consumption to approximately double to 950 TWh by 2030. US-China export control rules determine which chips reach non-US companies and which neoclouds can serve Asian customers. Stanford's 2026 AI Index notes that the Foundation Model Transparency Index has fallen as the most capable models reveal less about training data and parameters, concentrating capability in a handful of US labs. Watch for regulatory responses in the European Union and India, and for whether the foundation-model sub-segment consolidates around two or three survivors or sustains ten-plus labs into 2027.