The most expensive line of code is the one you never have to write. For the team at doubleAI, the most expensive instruction is the one that runs inefficiently on the hardware you already own. Their product, Maxwell, doesn't ask engineers to rewrite their Python or C++. Instead, it watches a program execute, analyzes the hardware, and rewrites the machine instructions on the fly to fit that exact workload [heymaxwell.ai, retrieved 2026]. The company claims the result can be a near-instant doubling of inference speed or a two-thirds cut in compute costs, all without touching the source code [The Next Web, October 2026].
A Wedge Into the Execution Layer
Most developer-focused AI tools operate at the code generation or editing layer. doubleAI's bet is a level deeper, in the space between compiled software and the silicon it runs on. The company calls its systems "Code Execution Models" (CEMs) [Calcalist Tech, October 2026]. Maxwell's approach is to perform a continuous, formal optimization of the binary instructions being sent to the CPU or GPU. The company says it formally proves every optimization is equivalent to the original program, a critical guarantee for production systems where correctness is non-negotiable [heymaxwell.ai, retrieved 2026].
The early reported results are striking, though they come from a limited set of early users. Video generation company LTX reported that Maxwell doubled its model-inference speed, cutting 4K video generation time from 250 to 134 seconds per clip [The Next Web, October 2026] [LinkedIn, Cristian-George Farauanu, retrieved 2026]. Another, unnamed early customer saw its compute costs drop by 67% [The Next Web, October 2026]. The company also claims Maxwell outperformed NVIDIA's PyTorch baselines on 90% of a standard kernel benchmark after one day of search [LinkedIn, Cristian-George Farauanu, retrieved 2026].
The Mobileye Mafia's Second Act
The ambition is matched by a founding team with a proven track record in building and scaling deep tech. Co-founder and CEO Amnon Shashua was previously the CEO of Mobileye, the autonomous driving company he co-founded [The Wall Street Journal, October 2026]. Co-founder and CTO Shai Shalev-Shwartz is an AI researcher who also held senior leadership roles at Mobileye [Calcalist Tech, October 2026]. They are joined by a cohort of co-founders including Yoav Levine, Or Sharir, Noam Wies, and Gal Beniamini, forming a technically dense founding team [The Next Web, October 2026].
This pedigree has attracted a tier-one investor syndicate. The company raised a $25 million seed round in May 2025, followed by a $200 million Series A in November 2025 co-led by Lightspeed Venture Partners and Bessemer Venture Partners [Calcalistech, October 2026]. That round reportedly valued the company at approximately $1 billion [Calcalistech, October 2026]. Strategic investors include NVIDIA and Dell Technologies Capital, signaling belief in the product's infrastructure-level potential [SiliconANGLE, October 2026].
| Round | Date | Amount | Lead Investor(s) | Reported Valuation |
|---|---|---|---|---|
| Seed | May 2025 | $25M | Dell Technologies Capital, BRM Group, Pitango | Not disclosed |
| Series A | November 2025 | $200M | Lightspeed Venture Partners, Bessemer Venture Partners | ~$1 billion |
| Table: doubleAI's funding history. Total disclosed raise is $220 million [The Wall Street Journal, October 2026]. |
The Technical Breakdown and Scale Risks
From an engineering perspective, Maxwell's promise rests on a few technical pillars. It operates as a runtime optimization layer, requiring no code changes, which lowers the adoption barrier. Its performance claims are against already-optimized pipelines; the LTX gains, for instance, came on top of a hand-written, end-to-end compiled inference stack [doubleai.com/research/maxwell-ltx-2-5-4k-in-half-the-time, retrieved 2026]. The planned business model aligns incentives perfectly: doubleAI intends to charge customers a fee proportional to the compute savings it generates [Calcalist Tech, October 2026].
The sober assessment, however, must consider what could go wrong at scale. The technology is inherently complex, and maintaining formal equivalence proofs across the vast matrix of hardware, software libraries, and workload types is a monumental engineering challenge. Performance gains may not be uniform; a 67% reduction for one workload does not guarantee similar results for all. Furthermore, the commercial model, while elegant, depends on precise, auditable measurement of cloud compute savings, which could lead to friction with customers over attribution. The company had about 40 employees as of October 2026 [Calcalist Tech, October 2026], a relatively small team for the scope of the problem they are tackling.
The Road to Enterprise Adoption
The next twelve months will be about moving from impressive proofs-of-concept to hardened, scalable enterprise deployment. The hiring page lists openings for deep learning engineers, research engineers, and DevOps roles, indicating a build-out of the core platform team [doubleai.com/careers, October 2026]. The key milestone to watch will be the announcement of a second or third named enterprise customer beyond LTX, particularly one in a different vertical like financial modeling or large-scale data processing. Landing a design win with a major cloud provider would be a transformative partnership, embedding Maxwell's optimization directly into the infrastructure layer.
For now, doubleAI has secured the capital, the team, and the early technical validation to make a serious run at a fundamental problem in AI infrastructure. The bet is that the next frontier of efficiency isn't in bigger models or more chips, but in smarter use of the cycles we already have.
Sources
- [The Wall Street Journal, October 2026] Nvidia-Backed Startup doubleAI Launches Model to Pare Cost of Compute | https://www.wsj.com/cio-journal/nvidia-backed-startup-doubleai-launches-model-to-pare-cost-of-compute-b780f215
- [The Next Web, October 2026] doubleAI’s Maxwell rewrites code as it runs to cut compute costs | https://www.thenextweb.com/news/doubleai-maxwell-code-execution-model-compute-costs
- [Calcalist Tech, October 2026] Shashua’s doubleAI unveils AI models that optimize... | https://www.calcalistech.com/ctechnews/article/sk11xpc73ge
- [SiliconANGLE, October 2026] Nvidia, Dell back $200M round for AI model developer doubleAI | https://siliconangle.com/2026/10/19/nvidia-dell-back-200m-round-for-ai-model-developer-doubleai/
- [heymaxwell.ai, retrieved 2026] Maxwell product page | https://heymaxwell.ai
- [doubleai.com/careers, October 2026] Careers page | https://www.doubleai.com/careers
- [doubleai.com/research/maxwell-ltx-2-5-4k-in-half-the-time, retrieved 2026] Research case study | https://www.doubleai.com/research/maxwell-ltx-2-5-4k-in-half-the-time
- [LinkedIn, Cristian-George Farauanu, retrieved 2026] Post on Maxwell performance benchmarks | https://www.linkedin.com/posts/cristian-george-farauanu-activity
- [Calcalistech, October 2026] Funding round reporting | https://www.calcalistech.com/ctechnews/article/sk11xpc73ge