Shashua’s doubleAI unveils AI models that optimize how software runs
Unlike coding agents that help developers write software, doubleAI’s models focus on how programs execute on CPUs and GPUs, aiming to improve performance without altering the original code.
After stepping down as CEO of Mobileye and leaving his roles at AI21, Amnon Shashua has unveiled the first product of his latest venture, AI company doubleAI. Shashua says the company has developed an entirely new category of AI designed to reduce computing costs by improving the efficiency of software execution. The company plans to charge customers based on the computing cost savings it delivers.
DoubleAI, which operated in stealth mode until recently, has raised $200 million in a Series A round co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with participation from Dell Technologies Capital, NVIDIA, Greenfield, Pitango and BRM. The company employs approximately 40 people. As previously reported by Calcalist, the funding round values doubleAI at more than $1 billion.
The company calls its technology “Code Execution Models,” or CEMs. It describes the models as operating “between the code and the silicon,” meaning between software and the hardware on which it runs. Unlike large language models and coding agents, which help developers write and modify code, CEMs are designed to optimize how software executes, identifying ways to improve performance without changing the original source code.
DoubleAI says its first model, Maxwell, continuously analyzes a running program, the hardware on which it operates and the data it processes. It then optimizes the underlying machine instructions to fit those conditions. The company says the system works with both CPUs and GPUs, and that each verified improvement can serve as the basis for further optimization as the program continues to run.
According to Shashua, one of the central challenges in developing the technology was ensuring that changes made to improve performance would not alter a program’s behavior. DoubleAI says Maxwell formally and mathematically verifies every optimization, establishing that the modified instructions preserve the program’s original behavior.
The company reports that Maxwell has made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster on real-world workloads. These figures have not been independently verified, and the company’s announcement does not provide sufficient detail about the underlying tests, hardware configurations or baseline performance to assess how broadly the gains can be replicated. Maxwell is now available for early access through the company’s website.
Yair Snir, managing partner at Dell Technologies Capital, which invested in doubleAI, said the fund had invested in many AI companies but had chosen to back a company developing its own models for the first time with doubleAI, leading its seed round.
“We saw that Amnon and his team possess a unique ability to approach complex problems and make rapid progress,” Snir said. “Every few weeks, they unveil another AI breakthrough, significantly improving the results achieved by some of the best teams in the world. They are now applying this same mindset to the challenge of computational efficiency, using an approach that no one else has taken to date. This is precisely the kind of innovation that Israel can and should lead.”
