doubleAI’s Maxwell rewrites running code and claims 38x faster pipelines

Tel Aviv startup doubleAI has launched Maxwell, which it calls the first “code execution model”, an AI that rewrites a program’s machine code while the program runs on CPUs and GPUs. The company also raised $220 million, according to the Wall Street Journal’s report on the launch.

A model that sits inside running software

According to doubleAI, Maxwell does not touch the source code. It observes a program as it executes and reshapes the underlying machine code to suit the hardware and the data it actually sees. The company says each change is formally proven to leave the program’s behavior unchanged, and that no human review is needed. Some faster versions depend on assumptions about the data. If those stop holding, a guard hands control back to the original code.

The company’s site shows examples that start from Python and PyTorch and produce low-level replacements in Rust, C and SIMD code. It reports these results from specific experiments, not guarantees for every workload:

• Data processing pipelines up to 38 times faster, and AI inference kernels up to 4.25 times faster. Press coverage of the launch cites the inference figure as more than 4.5 times.
• LTX-2.5 4K video generation on two Nvidia B200 GPUs cut from 250 seconds to 134 seconds, with output the company calls visually indistinguishable.
• Whisper-small transcription on an RTX 4090 cut from 106.23 to 49.82 seconds, with identical transcripts.
• NetworkX betweenness centrality cut from 61.44 to 17.50 seconds, with identical results.
• A drop-in cuGraph replacement that is 3.6 times faster on average, and Blackwell kernels that beat Nvidia’s PyTorch baselines on 90% of the SOL-ExecBench set.

These are the company’s own measurements and have not been independently verified.

Early access only, and the price is a cut of your savings

Maxwell is in controlled availability. The Maxwell site at heymaxwell.ai has a request form, but it says submitting it expresses interest only and promises no access, release date or acceptance. Later access may need a separate agreement, and no general availability date has been given.

On pricing, The Next Web reports, citing the Journal, that doubleAI will take a share of the compute costs it saves a customer. The percentage was not disclosed. Per that coverage, one early customer cut compute costs by 67%, and LTX, which builds open-weight multimodal models, said Maxwell doubled its inference speed.

Who stands to gain when compute is scarce

The pitch is aimed at anyone whose GPU bill is the largest line in the budget. The Journal, citing SemiAnalysis, says the hourly price of renting an Nvidia H100 on a one-year contract rose 60% over the past year. Co-founder and CEO Amnon Shashua told the paper that renting more compute is not a sustainable answer because prices keep climbing. In the company’s own words, “AI has changed how software is written, but not how it runs.”

Shashua expects the biggest beneficiaries to be labs with enormous compute bills, such as OpenAI and Anthropic, saying that even a 10% saving could be worth hundreds of millions of dollars to them. For smaller teams the practical question is different. A pay-from-savings model lowers the upfront risk, but the company has not said how savings are measured, which workloads qualify, or how the formal proofs extend beyond Python, which is what its formalization covers according to the site.

There is also an irony in the funding. Nvidia is among the investors, and doubleAI says it spent roughly half its money on 1,000 Nvidia B300 chips for its own use, per the Journal. A tool that squeezes more work out of existing chips is pitched as a way to ease the shortage, yet it was built on, and partly financed by, the same supplier.

A small lab with big backers

TNW reports that doubleAI is two years old, has about 40 employees and has raised $220 million in total, including a $200 million Series A co-led by Lightspeed and Bessemer Venture Partners. Other backers listed on its site include Nvidia, Dell Technologies Capital, Greenfield Partners, Pitango and BRM Group. Shashua is also the CEO of Mobileye, which said in July he would step down once a successor is named.

What remains unproven is how well the approach holds up outside the company’s chosen benchmarks. Maxwell’s headline numbers come from curated examples, the fee structure is undisclosed and access is limited to those the company admits. Independent tests on customers’ real workloads will show whether a 38x pipeline gain is typical or a best case.