doubleAI gets $200M for Maxwell, a model for fast code

doubleAI, a company in Tel Aviv, announced a $200 million investment. CTech reports a value of more than $1 billion. The company will sell a model that makes software use less hardware. It says the model made one data tool 38 times as fast on one query test. It gave no external test of this result.

How it works

Maxwell changes the binary of a piece of software. A binary is the software code that chips read. Developers change no source code, and Maxwell can change software while it works in production. It works with GPUs and CPUs, 2 types of chips.

The numbers

Lightspeed and Bessemer lead the investment. doubleAI gives these results:

• cuGraph from Nvidia: 55% of algorithms 2 or more times as fast.
• pandas on the PDS-H query test: 38 times as fast.
• LTX-2.5 video model: 2 times as fast.

What is not known

All results come from the company. The source gives no price and no customer. The Wall Street Journal reports that half of the money will buy 1,000 Nvidia B300 chips.

Insight: Nvidia pays for a tool that cuts use of its chips

doubleAI used its model to cut the hardware use of an Nvidia library. Nvidia backs doubleAI, and half of the money will buy 1,000 Nvidia chips. On 13 October, Nvidia opened the code of its software that gives GPUs to AI tasks. This reads as a sign that Nvidia thinks software which uses less hardware will make AI use and the use of chips increase. Tell each vendor to show a before and after test.