Analysis
Tel Aviv-based doubleAI, officially AA-I Technologies Ltd., has raised $200 million in a Series A round co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with Nvidia and Dell Technologies joining alongside several other unnamed investors, the company announced Monday, according to SiliconANGLE. Israeli outlet CTech reported the round values doubleAI above $1 billion.
doubleAI's core product is an AI model called Maxwell, designed to make existing software run faster without a developer rewriting the source code. Rather than editing source directly, Maxwell optimizes the compiled program binary -- and it can keep modifying an application even while it runs in production. It supports both GPU and CPU workloads, and the company is targeting enterprises trying to cut hardware costs on AI workloads they already run.
What Maxwell Has Actually Shown So Far
The company's disclosed results are specific: Maxwell reportedly rewrote parts of Nvidia's cuGraph library, speeding up 55% of its algorithms by at least 2x; improved the pandas data-science toolkit's performance on the PDS-H benchmark by 38x; and doubled the video-generation speed of the LTX-2.5 world model. In a separate stunt, an internal doubleAI agent reportedly found a new position in a decades-old chess puzzle, breaking a nearly 50-year-old record for the longest forced-checkmate sequence -- 63 moves, versus the prior record of 60.
Those are the kind of benchmark headlines that get attention, but they're also self-reported by doubleAI rather than independently verified by a third party, and a chess puzzle record is a parlor trick more than a production workload. The real test is whether Maxwell's gains hold up on the messy, varied codebases enterprise customers actually run -- not the curated examples a company picks for a funding announcement.
According to the Wall Street Journal, doubleAI plans to spend roughly half the new $200 million on 1,000 of Nvidia's B300 Blackwell GPUs. That's an unusual move for a startup whose pitch is reducing GPU dependence -- the company appears to need significant compute itself to keep training and validating Maxwell, even as it sells efficiency gains to customers trying to cut their own GPU bills.
The article names OpenAI and Anthropic as competitors in the sense that their newest models can also automate research-style tasks, though neither is a direct code-optimization rival in the way doubleAI's Maxwell is positioned. doubleAI has not disclosed its founding date, founders' names or prior funding history.
What to watch: whether doubleAI publishes third-party-verified benchmarks beyond its own curated examples, and how it balances buying its own GPU fleet against pitching customers on needing fewer GPUs.