Nvidia, Dell back $200M round for AI model developer doubleAI

Artificial intelligence developer doubleAI announced it has raised $200 million in funding. Lightspeed and Bessemer co-led the Series A round. Nvidia Corp., Dell Technologies Inc. and several other investors also participated. The investment values doubleAI at more than $1 billion.

doubleAI, based in Tel Aviv, has developed an AI model called Maxwell. Maxwell can automatically make software code more efficient and can be used for a variety of tasks. The company demonstrated Maxwell's capabilities by rewriting Nvidia's cuGraph library, reducing hardware usage and speeding up 55% of algorithms by a factor of two or more. Maxwell also optimized a popular data science toolkit, pandas, improving performance by a factor of 38.

The AI model was used to double the video generation speed of a world model called LTX-2.5. doubleAI plans to sell the model to enterprises looking to lower their AI workloads' hardware usage. Maxwell finds optimization opportunities by analyzing program interactions with data and underlying chips, supporting both graphics processing units and central processing units.

doubleAI's technology allows for optimizations to be made to applications while they're running in production. Recently, an AI agent created by doubleAI made a new discovery about graphs, mathematical structures with a significant role in software development. The agent's proof, paired with a Lean file, enabled relatively easy verification of the finding.

In conjunction with the funding announcement, doubleAI revealed that it will use half of the investment to purchase 1,000 of Nvidia's B300 Blackwell graphics cards. This move positions doubleAI to compete with AI developers like OpenAI Group PBC and Anthropic PBC, whose latest models can automate research tasks. OpenAI's ChatGPT recently released 722 AI-generated math papers, while Anthropic's model created a computer-verifiable version of a famous mathematical proof.

The proof's file, consisting of 13 million lines of code, would have taken a human mathematician team several years to write.