doubleAI Launches Maxwell AI to Accelerate Running Code and Cut Compute Costs
The startup’s new model rewrites machine code in real time to speed up software execution without altering the original source code.
Key points
- doubleAI announced Maxwell on Monday, an AI model that optimizes running software to reduce computing power usage.
- The system reportedly accelerated data pipelines by over 38 times and AI inference kernels by more than 4.5 times.
- The company plans to charge customers a percentage of the compute savings they achieve using the technology.
doubleAI, a Tel Aviv-based startup co-founded by Amnon Shashua, launched Maxwell on Monday. The company opened early access to the new AI model, which operates inside running software to optimize performance and reduce computing power requirements. Shashua serves as the chief executive of the two-year-old firm, which has raised $220 million in total funding.
Maxwell represents a new category of AI that the company calls code execution models. Unlike coding agents that write source code for developers, Maxwell monitors programs while they run. It rewrites the underlying machine code to better fit the specific hardware and data patterns it observes. The original source code remains completely unchanged during this process, allowing the optimization to work on both CPUs and GPUs.
Technical Mechanism and Performance
The company states that Maxwell formally proves each change leaves the program’s behavior unchanged, ensuring no human review is needed for the modifications. If a faster version relies on conditions that stop holding, a guard mechanism hands control back to the original code. According to doubleAI, Maxwell made real-world data pipelines more than 38 times faster. It also accelerated AI inference kernels by more than 4.5 times.
LTX, a builder of open-weight multimodal models, reported that Maxwell doubled the speed of its model inference. Zeev Farbman of LTX told The Wall Street Journal that halving GPU time immediately translates to lower serving costs and improved margins. The startup also told the WSJ that it cut one early customer’s compute costs by 67%.
Market Context and Strategy
doubleAI plans to charge customers a share of what they save on compute, though Shashua did not specify the exact percentage. He expects AI labs with large compute bills, such as OpenAI and Anthropic, to benefit most. Shashua told the WSJ that even a 10% saving could be worth hundreds of millions of dollars for these entities.
The launch comes as renting AI computing power becomes more expensive. The WSJ reported, citing SemiAnalysis, that the hourly price for Nvidia H100 chips on one-year contracts has risen 60% in a year. Shashua stated that companies cannot rely solely on renting compute because prices are increasing. The startup spent roughly half its funding on 1,000 Nvidia B300 chips for its own needs.
What should you do?
Security and technology professionals should monitor early access programs for potential stability issues when using real-time code optimization tools. Organizations evaluating such AI models should verify that the formal proof mechanisms adequately protect against unintended behavioral changes in critical production environments. Teams should assess their current compute costs to determine if the percentage-based pricing model offers a viable return on investment compared to standard hardware rentals.
What to do and how to stay safe: doubleAI
- Review the formal verification methods used by the AI model to ensure it guarantees unchanged program behavior during optimization.
- Analyze current compute expenditure to calculate whether the potential savings justify the adoption of a revenue-share pricing model.
- Test the optimization tool in non-production environments to observe how guard mechanisms handle changing data conditions.
- Monitor industry reports on hardware rental prices to compare the long-term cost of owning optimized infrastructure versus renting capacity.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does Maxwell differ from standard coding AI agents?
Standard coding agents write source code for people, whereas Maxwell watches running programs and rewrites the underlying machine code to fit the hardware, leaving the source code unchanged.
What performance improvements has doubleAI reported for Maxwell?
The company stated that Maxwell made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster on real-world workloads.
How does doubleAI plan to charge for Maxwell?
The startup plans to charge customers a share of the compute savings they achieve, rather than a fixed subscription or licensing fee.
