Nvidia, Dell back $200M round for AI model developer doubleAI
SiliconANGLE Maria Deutscher
doubleAI just raised $200 million from Lightspeed, Bessemer, Nvidia and Dell. Its model rewrites software binaries to cut hardware use, and even cracked a chess record.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
doubleAI has landed a $200 million Series A, with Lightspeed and Bessemer co-leading the round. Nvidia and Dell also joined in, and CTech says the deal puts the Tel Aviv company at a valuation above $1 billion.
The startup, officially AA-I Technologies Ltd., is built around a model called Maxwell. The pitch is simple enough to grasp and unusual enough to matter: instead of asking developers to rewrite source code by hand, Maxwell works on binaries, the machine-made version of a program. That means the model can hunt for efficiency gains without waiting for engineers to ship a fresh code change.
The company says Maxwell has already shown that idea can pay off. In March, it rewrote Nvidia’s cuGraph library and cut hardware use. DoubleAI says 55% of the algorithms in the library ran at least twice as fast, and a later project used Maxwell to create an optimized version of pandas that improved performance on PDS-H by a factor of 38. It also says Maxwell doubled the video generation speed of LTX-2.5.
There’s a research angle here too, not just a software-optimization pitch. Earlier this month, a doubleAI agent made a new mathematical discovery and paired the result with a Lean file for verification. Around the same time, Maxwell broke a nearly 50-year chess record by finding a position with 63 moves that give check, beating the old mark of 60.
The money is going toward scale as much as product. The Wall Street Journal reported that half the round will buy 1,000 Nvidia B300 Blackwell graphics cards. That gives doubleAI more firepower as it tries to sell Maxwell to enterprises that want lower AI hardware costs, and to compete with OpenAI and Anthropic, which have also been showing off models that can do research-like work.
My take — AI-written commentary, not fact-checked reporting
This is the cleanest kind of AI pitch: save money, then brag about the math later. The real tell is the hardware buy — 1,000 Nvidia chips are never purchased by believers in “software efficiency” alone. Everyone in AI says they want smarter models; doubleAI is betting companies mostly want cheaper ones.
Read more about this at: SiliconANGLE