Israeli artificial intelligence startup Two Delta has raised tens of millions of dollars in seed funding, attracting prominent investors including Bill Ackman, founder and CEO of Pershing Square, and Eric Schmidt, former chairman of Google. Founded in 2025, the company focuses on improving AI model performance while reducing computational demands. The funding comes as Two Delta advances research into faster AI inference and the reliability of autonomous coding agents.
Funding and Company Background
Two Delta was established by Dvir Zeltzer and Dr. Eli David, combining expertise in business management and artificial intelligence research. Zeltzer previously held a senior position at Google, while David has extensive experience as an AI researcher and entrepreneur. The startup operates largely in stealth mode, publicly disclosing limited information beyond its technical research.
With fewer than ten employees, Two Delta is developing technologies intended to make advanced AI systems faster and more computationally efficient. Its work addresses growing demand for optimized AI infrastructure as increasingly sophisticated models require substantial processing resources. The company has not disclosed the precise funding amount or detailed commercial plans.
DCOCO Technology and Nvidia Collaboration
In October 2026, Two Delta released research conducted in collaboration with Nvidia, demonstrating its DCOCO technology for accelerating inference on Nvidia's Nemotron 3.5 Super VL model. The approach aims to increase token generation throughput without compromising output quality. It builds on speculative decoding, a technique designed to reduce the computational overhead associated with generating AI responses.
Speculative decoding allows a smaller, faster model to propose multiple tokens before a larger model evaluates them together. However, rejected predictions can reduce the efficiency gains achieved through this process. DCOCO addresses this limitation by improving how candidate token sequences are selected before verification.
The technology combines token embeddings with contextual information from the drafting model's internal representations, informed by the larger model's activations. It evaluates neighboring token combinations in parallel to identify sequences more likely to receive approval. This approach seeks to increase the number of accepted tokens during each verification cycle.
Reported Performance Improvements
Two Delta evaluated DCOCO across ten datasets covering mathematics, programming, and general-purpose tasks. Under testing conditions with reasoning disabled and greedy decoding enabled, the technology achieved average throughput 2.03 times that of conventional autoregressive decoding. It also demonstrated a reported 19.8% improvement over the competing DFlash approach under comparable conditions.
The experiments used Nvidia DGX B200 hardware and the vLLM inference framework, including workloads involving up to 32 simultaneous requests. Additional performance gains were reported across different decoding configurations and operating conditions. These findings remain company-reported results rather than independently verified performance assessments.
EvoWeb Benchmark Examines AI Coding Reliability
Beyond inference optimization, Two Delta introduced EvoWeb, a benchmark evaluating how effectively AI coding agents maintain web applications through repeated development requests. The framework examines ten consecutive interactions, beginning with application creation and continuing through nine modifications. This approach assesses whether AI systems can preserve existing functionality while incorporating new features.
Testing across 32 web applications revealed significant limitations in maintaining software reliability throughout successive changes. None of the evaluated models preserved complete functionality across the entire testing sequence. The highest-performing models, Fable 5 and Kimi K3, recorded scores of 0.694 and 0.684, respectively, on a scale of zero to one.
The research identified feature integration as a major challenge for autonomous coding systems. Although application build failures remained relatively uncommon, runtime errors increased from 2% initially to 12% after ten requests. These findings highlight the difficulties AI agents face when maintaining increasingly complex software environments.
Two Delta's seed financing and recent research developments position the startup within two important areas of artificial intelligence: inference efficiency and autonomous software development. Its DCOCO technology targets faster model execution, while EvoWeb provides a framework for evaluating the reliability of AI-generated applications. Although further independent validation is needed, both initiatives reflect ongoing efforts to improve the performance, efficiency, and dependability of advanced AI systems.
Source: Calcalist