Tel Aviv-based doubleAI has raised $200 million in a Series A round co-led by Lightspeed and Bessemer, SiliconANGLE reports. Nvidia and Dell Technologies took part alongside other investors. CTech reported that the round values the company, formally AA-I Technologies Ltd., at more than $1 billion.
The company's product is a model called Maxwell, which automatically makes software run more efficiently. The company says it finds savings by studying how a program interacts with its data and with the chips underneath, and it supports both GPUs and CPUs. The target customer is an enterprise that wants its AI workloads to consume less hardware.
The technical distinction is where the changes land. Developers usually apply optimizations by editing source code. Maxwell instead modifies the compiled binary, which removes the manual rewrite step. According to the report, this also makes it possible to alter applications while they run in production.
The performance figures come from doubleAI. In March it said Maxwell rewrote Nvidia's open-source cuGraph library for GPU graph analytics and sped up 55% of its algorithms by a factor of two or more. An optimized version of the pandas data science toolkit reportedly improved on the PDS-H query benchmark by a factor of 38. A later project doubled the video generation speed of a world model called LTX-2.5.
The company also pointed to research-style outputs. An internal agent made a new discovery about graphs and supplied a Lean file for verification. Maxwell found a chess position with 63 checking moves, against a prior record of 60 that had stood for nearly 50 years.
The Wall Street Journal reported that doubleAI will spend half of the round on 1,000 of Nvidia's B300 Blackwell GPUs. That purchase is meant to help it compete with OpenAI and Anthropic, whose latest models can also automate research tasks. The report cites recent examples from both: OpenAI released 722 AI-generated math papers, and an Anthropic model produced a computer-verifiable proof file of 13 million lines.
Why it matters: If binary-level optimization performs outside doubleAI's own demonstrations, it could reduce the hardware needed for some workloads without changes to source code. The benchmark gains are company-reported, though, and the source gives no independent validation or customer deployment data.
TwinStack automated analysis, based on reporting by SiliconANGLE. Check the original for full details.