doubleAI's Maxwell claims 2x to 38x software speedups, with no outside checks yet

doubleAI, the Tel Aviv startup co-founded by Mobileye's Amnon Shashua, has launched Maxwell, a model it describes as a code execution model. Its pitch is that it can make software run faster by rewriting the compiled binary rather than the source code, and that it can do this while applications are running in production. Enterprises that want to cut the hardware bill for their AI workloads are the intended buyers.

The numbers are eye-catching. doubleAI says Maxwell sped up 55% of the algorithms in Nvidia's cuGraph library by at least two times, and boosted a pandas build by 38 times on the PDS-H query benchmark. It also says the model doubled the video generation speed of the LTX-2.5 world model, and it broke a chess record with a 63-move check sequence. Every one of those figures comes from doubleAI itself. No customer has reported a bill, a latency number or a production result.

The benchmark-to-business gap is the catch. A 38x gain on a query benchmark is not the same as 38x less spend on your data pipeline, and a library that runs faster in a lab is not yet a line item anyone has verified. The company has $220 million in backing from Nvidia and Dell, and a reported valuation above $1 billion, which means it has every reason to show big numbers early.

Still, the idea is worth watching. If binary-level optimization holds up outside the demo, it could matter to any team paying a lot to run AI on GPUs. Until someone outside doubleAI runs Maxwell on their own workload and publishes the bill, treat it as a promising pitch with a very good scoreboard.