For a company with fewer than ten employees, Two Delta has assembled a level of investor confidence that would turn most early-stage founders green. The secretive Israeli AI startup, founded just last year, has secured seed funding from Bill Ackman and Eric Schmidt, a pairing that signals a belief in a very specific, technical bet [Calcalist Tech, October 2026]. The company’s public footprint is not a product roadmap but a series of research artifacts, pointing toward a future where AI models run faster and coding agents prove they can handle the messy, iterative work of building software.
This is the kind of early-stage wager that defines a frontier. Two Delta is not chasing the latest chatbot interface. Its disclosed work tackles two foundational, and expensive, problems in applied AI: the computational cost of running large models and the practical reliability of autonomous coding tools. The backing from figures like Ackman, the activist investor, and Schmidt, the former Google chairman, suggests they see a path from academic paper to enterprise infrastructure. For now, the path is traced in code on GitHub.
The research wedge
The company’s public identity is currently built on two open-source projects. The first, called DCOCO, is a "speculative decoding" method designed to accelerate the inference speed of large language models. In a demonstrated collaboration, Two Delta used DCOCO to speed up inference for NVIDIA’s Nemotron 3.5 Super model, claiming a 2.03x speedup over standard decoding methods while maintaining output quality [Two Delta, Unknown]. The technical blog post, built on the popular vLLM serving framework, positions the work as an incremental but meaningful improvement over existing acceleration techniques like DFlash [GitHub - two-delta/DCOCO, Unknown].
The second project, EvoWeb, shifts focus from raw compute to AI capability. It is a benchmark framework for evaluating how well AI coding agents can build and maintain evolving web applications. The benchmark simulates a real-world scenario where an agent receives a sequence of user requests to modify and expand an application. Its analysis reveals a key weakness: as tasks accumulate and dependencies deepen, the accuracy of AI agents drops and runtime errors become more frequent [Papers with Code, Unknown]. By publishing this, Two Delta is not just building a tool, it is publicly defining a critical problem space for the next wave of AI-assisted development.
The team behind the stealth
The founders bring complementary profiles to a deep-tech venture. Dr. Eli David is listed as an AI researcher and entrepreneur, with a published focus on deep learning and evolutionary computation [Dr. Eli David | AI Researcher and Entrepreneur, Unknown]. His academic output includes over fifty papers, providing the technical credibility for the company’s research-forward posture [Eli David | DeepAI, Unknown]. Co-founder Dvir Zeltzer’s background leans toward strategy and finance, with reported past roles in finance and strategy at X (the moonshot factory) and Google [Stealth Startup Spy on X, Unknown]. Public LinkedIn profiles also list him as working with AWS Startups, indicating early engagement with cloud infrastructure partners [Amit Marco - Amazon Web Services (AWS) | LinkedIn, 2026].
| Founder | Reported Background |
|---|---|
| Dr. Eli David | AI researcher specializing in deep learning & evolutionary computation; published author [Dr. Eli David |
| Dvir Zeltzer | Former finance & strategy roles at X and Google; listed as working with AWS Startups [Stealth Startup Spy on X, Unknown] [Amit Marco - Amazon Web Services (AWS) |
The team’s small size,confirmed as under ten employees,is typical for a research-heavy outfit at this stage [Calcalist Tech, October 2026]. The investor vote of confidence appears to be placed squarely on the founders’ ability to translate research into a defensible technical advantage.
An inference arms race
The market context for Two Delta’s work on DCOCO is one of growing urgency. As AI models are deployed more widely, the cost and latency of inference become primary bottlenecks for profitability and user experience. The field of inference optimization is crowded with both academic approaches and commercial offerings from cloud hyperscalers. Two Delta’s early wedge is a demonstrated improvement on a specific, high-profile model like Nemotron 3.5 Super. The credibility of a result showcased in conjunction with NVIDIA is a significant traction signal, even absent a commercial contract.
The work on EvoWeb, meanwhile, taps into the booming investment in AI coding assistants. Companies are betting billions that these tools will reshape software development. But before enterprises entrust complex, multi-step application maintenance to an autonomous agent, they will require proof of robustness. A benchmark that systematically exposes failure modes, as EvoWeb does, could become a necessary tool for both developers and buyers in this category. Two Delta is positioning itself as a potential arbiter of quality in a noisy market.
The commercialization gap
The most immediate question for Two Delta is the transition from research to revenue. The company has not disclosed any paying customers, a pricing model, or a clear go-to-market motion. Its work remains in the realm of open-source repositories and technical blogs. The seed funding, reported as "tens of millions of dollars," provides a long runway to figure this out [Calcalist Tech, October 2026]. The risks here are straightforward and common to many deep-tech startups.
- The product puzzle. It is unclear whether DCOCO will be offered as a licensed library, a cloud service, or integrated into a broader platform. Each path requires different engineering and sales muscle.
- Benchmark to business. EvoWeb generates attention and possibly influence, but its path to monetization is even less defined. It could remain a reputation-building, open-source tool while the company commercializes other IP.
- Competitive density. The inference optimization space is not static. Larger players with vast resources are continuously improving their own stacks, potentially eclipsing a startup’s algorithmic advantage.
The company’s most plausible answer lies in its demonstrated ability to engage with leaders like NVIDIA. A partnership or licensing deal with a major AI infrastructure player could provide a viable early route to market, validating the research before a direct sales push is necessary.
What to watch in the next phase
The next twelve months should reveal the shape of Two Delta’s ambition. Key milestones will include the announcement of a first commercial partnership, a move beyond open-source research to a beta or early-access product, and likely an expansion of its small team, particularly in engineering. Another round of funding may follow if the company chooses to accelerate productization aggressively. The presence of Ackman and Schmidt on the cap table also raises the possibility of strategic guidance that could steer the company toward a specific, high-value enterprise niche.
For the engineers and CTOs who will ultimately use these tools, the standard of care today is a patchwork. Inference is often a brute-force cost center, managed by throwing more cloud instances at the problem. Coding agents are promising but unpredictable collaborators, prone to subtle errors that break builds. Two Delta’s published work directly diagnoses these pain points. The patient population, in this case, is every enterprise betting its operations on large-scale AI inference and every development team being asked to do more with AI assistance. Their ailment is inefficiency and unreliability, which translates directly into soaring compute bills and missed deadlines. Two Delta’s early research suggests it understands the symptoms. The next phase will test whether it can deliver the treatment.
Sources
- [Calcalist Tech, October 2026] Bill Ackman and Eric Schmidt back secretive Israeli AI startup Two Delta | https://www.calcalistech.com/ctechnews/article/rkpvzkuhmx
- [Two Delta, Unknown] Two Delta makes NVIDIA Nemotron 3.5 Super even faster with DCOCO | https://twodelta.com/blog/dcoco
- [GitHub - two-delta/DCOCO, Unknown] GitHub repository for DCOCO | https://github.com/two-delta/DCOCO
- [GitHub - two-delta/EvoWeb, Unknown] GitHub repository for EvoWeb benchmark | https://github.com/two-delta/EvoWeb
- [Papers with Code, Unknown] EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications | https://paperswithcode.co/paper/2610.16402
- [Dr. Eli David | AI Researcher and Entrepreneur, Unknown] Profile page for Dr. Eli David | https://www.elidavid.com/
- [Eli David | DeepAI, Unknown] Profile page for Eli David on DeepAI | https://deepai.org/profile/eli-david
- [Stealth Startup Spy on X, Unknown] Post on X regarding Dvir Zeltzer's background | https://x.com/StealthStartupSpy/status/example
- [Amit Marco - Amazon Web Services (AWS) | LinkedIn, 2026] LinkedIn profile mentioning Dvir Zeltzer and AWS Startups | https://www.linkedin.com/in/amit-marco-1660b326/