Artificial intelligence developer doubleAI today announced that it has raised $200 million in funding.
Lightspeed and Bessemer co-led the Series A round. They were joined by Nvidia Corp., Dell Technologies Inc. and several others. CTech reported that the investment values doubleAI at more than $1 billion.
Tel Aviv-based doubleAI, officially AA-I Technologies Ltd., has created an AI model called Maxwell that can automatically make software code more efficient. It also lends itself to a variety of other tasks. Maxwell broke a long-standing chess challenge earlier this month, while a doubleAI-developed agent made a new mathematical discovery around the same time.
Nvidia offers an open-source library called cuGraph that enables customers to harness its graphics cards for graph analytics. That’s a data science approach used in fields such as supply chain management. In March, doubleAI revealed that Maxwell had rewritten cuGraph to lower its hardware usage. The company says that the model sped up 55% of the algorithms in the library by a factor of two or more.
Maxwell went on to create an optimized version of pandas, a popular open-source data science toolkit. The AI improved the library’s performance on PDS-H, a query speed benchmark, by a factor of 38. A more recent project saw doubleAI use Maxwell to double the video generation speed of a world model called LTX-2.5.
The company plans to sell the model to enterprises interested in lowering their AI workloads’ hardware usage. According to doubleAI, Maxwell finds optimization opportunities by analyzing how a program interacts with the data that it processes and the underlying chips. The model supports both graphics processing units and central processing units.
Processors don’t run applications’ source code directly. Instead, they run a version of the source code called a binary. It’s a machine-generated translation that represents a program as a series of simple computing operations.
When developers find a way to make an application more efficient, they usually implement the optimization by modifying the program’s source code. Maxwell, by contrast, applies optimizations to a program’s binary. That removes the need for developers to manually make source code changes. Furthermore, doubleAI’s approach makes it possible to modify applications while they’re running in production.
The company’s technology also lends itself to other use cases. Earlier this month, doubleAI used an internal AI agent to make a new discovery about graphs, mathematical structures with an important role in software development. The agent paired its proof with a Lean file, a code snippet that makes it relatively simple to verify the finding.
In conjunction with today’s funding announcement, doubleAI revealed that Maxwell recently broke a nearly 50-year chess record. The model found a chess position, or arrangement of chess pieces, in which 63 moves give check. The previous record was 60 moves.
The Wall Street Journal reported that doubleAI will use half its funding round to buy 1,000 of Nvidia’s B300 Blackwell graphics cards. The chips should make it easier for the company to compete with OpenAI Group PBC and Anthropic PBC, whose latest AI models are also capable of automating research tasks.
A few days ago, the ChatGPT developer released 722 AI-generated math papers. A month earlier, an Anthropic model created a computer-verifiable version of a famous mathematical proof. The file comprises 13 million lines of code that would have taken a team of human mathematicians several years to write.