Amnon Shashua has a new answer to the AI compute crunch: don't build more chips — make the ones you have work harder.

The Mobileye co-founder, who stepped down as the autonomous-driving company's CEO earlier this year, unveiled Monday the first product of his new venture, doubleAI: a new category of AI the Tel Aviv startup calls Code Execution Models, or CEMs. The first one, Maxwell, is now in early access, according to CTech.

A CEM lives inside a running program. While coding agents act on behalf of the user and produce source code, a CEM acts on behalf of the hardware: its inputs are the running program itself, the chip it executes on, and the data flowing through it, and its outputs are machine instructions meant only for the machine. The source code humans read never changes. Maxwell continuously rewrites the binary underneath the code, fitting it to the exact hardware, data, and usage conditions it encounters — with each proven improvement becoming the starting point for the next round of optimization.

The breakthrough, per Shashua, is trust. No human reviews changes made while code is running, and conventional AI models make mistakes — so Maxwell formally and mathematically proves every rewrite preserves the program's original behavior. "AI has changed how software is written, but not how it runs," Shashua said in the company's announcement. "They are the first AI that works for the machine, not the programmer."

The company claims Maxwell has made data pipelines more than 38x faster and AI inference kernels more than 4.5x faster on real-world workloads. Those figures have not been independently verified, and the announcement offers no detail on hardware, baselines, or test setups — treat the numbers as a direction, not a result.

That direction is what makes the timing interesting. The industry's answer to the compute crunch has been to build: gigawatt data centers, Amazon's plan to fund 12 nuclear reactors in Indiana. But power, not silicon, is now the binding constraint — Microsoft just walked away from Nscale's Loughton data center in the UK because the grid couldn't deliver enough electricity until the mid-2030s. doubleAI attacks the denominator instead: squeeze more work from hardware that already exists. Shashua told the Wall Street Journal that "the need for compute is rising faster than the ability to build GPUs and CPUs," and that better utilization means "something incredible in terms of savings." It's a pitch that lands harder with H100 hourly rental costs up 60% in the past year, per SemiAnalysis data cited by the Journal. doubleAI plans to charge customers based on the compute savings it delivers.

The startup raised $200 million in a Series A co-led by Lightspeed Ventures and Bessemer Venture Partners, with NVIDIA, Dell Technologies Capital, Greenfield, Pitango, and BRM participating; Calcalist previously reported the round values doubleAI above $1 billion, and it employs about 40 people.

What to watch: whether those speed claims replicate outside doubleAI's lab — and whether enterprises will let an AI rewrite binaries in production, mathematically proven or not.

— via CTech