Mobileye CEO's doubleAI Launches Maxwell, a 'Code Execution Model' That Cut One Customer's Compute Bill 67 Percent

A Tel Aviv startup called doubleAI wants to create a new category of AI that does not write software for people — it makes software already running on expensive hardware faster. The company has launched Maxwell, its first "code execution model" (CEM), which sits inside running programs and works autonomously to improve how code performs, according to a Wall Street Journal report on Monday. The company says Maxwell lowered one early customer's compute spending by 67 percent, and broader release is slated before the end of the month.

The distinction from coding assistants is structural. Large language models help humans write software; CEMs are embedded in running systems and continuously optimize execution — hunting for faster algorithms, better scheduling, tighter kernels — without waiting for a developer to file a ticket. Co-founder and CEO Amnon Shashua, who is also CEO of Mobileye, said verification was the main obstacle, and that doubleAI found a way to verify its optimizations formally and mathematically rather than by eyeballing benchmarks.

The early evidence is customer-reported, not independently audited. LTX, a maker of open-weight multimodal AI models, says Maxwell made inference for its own models twice as fast. "We measure GPU times, and GPU times have a certain market price at every point in time," LTX co-founder and CEO Zeev Farbman told the Journal. doubleAI cites a 67 percent compute-cost reduction for one early customer. All three figures come from the vendors involved.

The business model follows the savings: doubleAI bills customers a share of the compute cost it eliminates. Shashua's example: if the company brought a customer's spending from $10 million down to $6 million, doubleAI would collect a portion of the $4 million saved; he did not disclose the percentage. He expects the biggest training and inference spenders — companies like OpenAI and Anthropic — to benefit most once Maxwell is widely available, noting that a 10 percent savings could be worth hundreds of millions of dollars to the largest AI labs.

The macro argument is that compute efficiency is becoming the highest-leverage line item in AI. The hourly price to rent NVIDIA H100 capacity under a one-year contract is up 60 percent from a year earlier, according to consultancy SemiAnalysis. "More hardware is being built, but the need for compute is rising faster than the ability to build GPUs and CPUs," Shashua told the Journal. "If you can utilize the hardware better, then you can achieve something incredible in terms of savings."

The company is unusual in how it financed its own credibility. doubleAI has raised $220 million in total from Lightspeed Venture Partners, Bessemer Venture Partners, Nvidia and Dell, and put about half of that money toward buying 1,000 Nvidia B300 chips for its own computing needs. "You cannot rely just on renting compute," Shashua said. "Prices are going up." The two-year-old company has about 40 employees.

If the category sticks, the interesting question is where CEMs end and the rest of the systems stack begins: compiler teams, GPU kernel libraries, scheduler optimizations and inference servers have all chased the same wasted cycles for years, with human engineers writing the improvements. A model that finds them autonomously — and can prove the results correct — turns those teams' output into a subscription. Incumbent chip vendors will not sit still: Nvidia's own software ecosystem is the first place a "make every GPU faster" capability would land.