General Intuition x CoreWeave

New York-based AI research lab spun out of Medal (the gaming clip-sharing platform) that trains [Foundation](/companies/foundation) models for physical AI using action-labeled video game data. The company builds both action models (which decide what actions to take) and world models (which predict outcomes of actions), betting that gameplay clips with embedded controller inputs provide a scalable training signal for embodied AI. Its models have been demonstrated controlling both virtual game agents and physical quadrupedal robots.

CoreWeave is the largest of the "neocloud" GPU operators, a category of specialist infrastructure providers built to rent Nvidia accelerators to AI labs and enterprises at hyperscale. Founded in 2017 in New Jersey as Atlantic Crypto, it pivoted from Ethereum mining into AI infrastructure and by 2025 was operating roughly 250,000 GPUs across 32 data centers in the United States and Europe. Its March 28, 2025 IPO on Nasdaq under ticker CRWV raised $1.5 billion and made it the first pure-play GPU cloud to trade publicly.\n\nWhat distinguishes CoreWeave from the rest of the neocloud pack is its depth of concentration with the two entities that matter most in AI infrastructure: Nvidia, which is both a strategic investor and its preferred launch partner for new silicon generations (H100, H200, GB200/B200), and a small set of anchor tenants that account for the vast majority of its revenue. Microsoft alone drove more than 60% of 2024 revenue, and a five-year OpenAI contract signed in 2025 is worth roughly $12 billion. That anchor-tenant model let CoreWeave scale faster than competitors like Lambda or Crusoe, but also gives it customer-concentration risk that public filings flag prominently.\n\nThe company reported $5.13 billion in 2025 revenue against a $1.2 billion net loss, reflecting the capital intensity of financing GPU fleets ahead of tenant ramp. Its 2025 acquisition of Weights & Biases for roughly $1.7 billion pushed CoreWeave up the stack from raw compute rental toward the MLOps tooling layer, followed by smaller tuck-ins of OpenPipe and Monolith AI. Positioning is unambiguously hyperscale and training-first: long-term reserved capacity contracts with a handful of the largest AI buyers, rather than the on-demand long-tail served by cheaper mid-market neoclouds.