THE ESSENTIALS IN 30 SECONDS

  • doubleAI raised a $200 million Series A co-led by Lightspeed and Bessemer Venture Partners, bringing total funding to $220 million.
  • Its Maxwell model rewrites machine code of running programs on CPUs and GPUs without touching the source code, the company says.
  • doubleAI plans to charge a share of customers' compute savings; Maxwell is in early access.

As the price of AI compute keeps rising, one of the better-funded bets of October is not on building more chips but on wasting fewer cycles. Tel Aviv-based doubleAI announced on 19 October a $200 million Series A and the launch of Maxwell, a model designed to make existing software run faster on the hardware companies already have.

The round

The Series A was co-led by Lightspeed and Bessemer Venture Partners, with Nvidia, Dell Technologies Capital, Greenfield, Pitango and BRM also taking part, according to The Next Web. The company has raised $220 million in total. SiliconANGLE, citing Israeli outlet CTech, reported a valuation above $1 billion; the company has not confirmed a figure. About half of the money has gone to buying 1,000 Nvidia B300 GPUs for doubleAI's own use, according to reports citing The Wall Street Journal.

doubleAI is roughly two years old and employs about 40 people. Its co-founders include CEO Amnon Shashua, also chief executive of Mobileye, and Shai Shalev-Shwartz.

What Maxwell does

The company calls Maxwell a "code execution model". Instead of editing source code, it observes a program while it runs and rewrites the underlying machine code to fit the specific hardware and data, on both CPUs and GPUs. doubleAI says each change is formally checked to preserve the program's behaviour, and that a guard falls back to the original code if the conditions behind an optimisation stop holding.

The company's performance claims include data pipelines running more than 38 times faster, AI inference kernels more than 4.5 times faster, and one early customer cutting compute costs by 67%. Video model developer LTX reportedly saw inference speed double. These figures come from doubleAI and its customers and have not been independently benchmarked.

Open questions

Maxwell is available in early access only. doubleAI intends to charge customers a percentage of the compute savings it generates but has not disclosed that percentage. Enterprises will likely want to see how the formal verification holds up on large, messy production codebases, and how easily the tool fits into existing deployment and audit processes. The bet is clear, though: with GPU rental prices climbing, software that squeezes more from current hardware has a ready market.