Key Takeaways
- • Israeli AI startup Two Delta has raised tens of millions of dollars in a seed round, with backers including Pershing Square founder and CEO Bill Ackman and former Google chairman Eric Schmidt, according to Calcalist.
- • Founded in 2025 and employing fewer than 10 people, Two Delta operates with a high degree of secrecy and has disclosed little beyond research papers published on its website.
- • Two Delta reported that its DCOCO technology achieved average throughput 2.03 times that of standard autoregressive decoding in tests across 10 datasets, and 19.8% higher throughput than the competing DFlash method; these figures are company-reported and not independently verified.
- • The company's EvoWeb benchmark found that none of 32 evaluated web applications maintained full functionality across 10 successive AI-agent requests, with runtime errors rising from 2% to 12%.
Summary
Israeli artificial intelligence startup Two Delta has raised tens of millions of dollars in a seed round backed by investors including Bill Ackman, founder and CEO of Pershing Square, and former Google chairman Eric Schmidt, according to Calcalist. The company was founded in 2025 by Dvir Zeltzer and Dr. Eli David and has fewer than 10 employees. David is described as a veteran AI researcher and entrepreneur, while Zeltzer brings a business and management background, including a senior role at Google. The company operates with a high degree of secrecy and has disclosed little about its business beyond research papers published on its website.
The startup is developing technology designed to accelerate AI models and improve the efficiency with which they use computing resources. In October 2026, Two Delta published results from a collaboration with Nvidia in which its DCOCO technology was used to accelerate inference for Nvidia's Nemotron 3.5 Super VL model. DCOCO builds on speculative decoding, in which a smaller model drafts tokens that a larger model verifies, and seeks to produce drafts the larger model is more likely to accept. The company reported average throughput of 2.03 times standard autoregressive decoding across 10 datasets covering mathematics, coding and general-purpose tasks, and a 19.8% improvement over the DFlash method under the same configuration. These evaluations, run on Nvidia DGX B200 hardware using the vLLM framework, are company-reported rather than independent assessments.
Two Delta has also published EvoWeb, a benchmark that evaluates how well AI coding agents build and maintain web applications through an initial build followed by nine feature changes. Of 32 web applications tested, none of the evaluated models maintained full functionality throughout the process. The top performers, Fable 5 and Kimi K3, scored 0.694 and 0.684 on a zero-to-one scale. The main difficulty was integrating new features with existing functionality, with runtime errors rising from 2% at the start to 12% after 10 requests. The findings point to a broader challenge for AI coding tools as they move beyond generating initial software versions.
The startup is developing technology designed to accelerate AI models and improve the efficiency with which they use computing resources. In October 2026, Two Delta published results from a collaboration with Nvidia in which its DCOCO technology was used to accelerate inference for Nvidia's Nemotron 3.5 Super VL model. DCOCO builds on speculative decoding, in which a smaller model drafts tokens that a larger model verifies, and seeks to produce drafts the larger model is more likely to accept. The company reported average throughput of 2.03 times standard autoregressive decoding across 10 datasets covering mathematics, coding and general-purpose tasks, and a 19.8% improvement over the DFlash method under the same configuration. These evaluations, run on Nvidia DGX B200 hardware using the vLLM framework, are company-reported rather than independent assessments.
Two Delta has also published EvoWeb, a benchmark that evaluates how well AI coding agents build and maintain web applications through an initial build followed by nine feature changes. Of 32 web applications tested, none of the evaluated models maintained full functionality throughout the process. The top performers, Fable 5 and Kimi K3, scored 0.694 and 0.684 on a zero-to-one scale. The main difficulty was integrating new features with existing functionality, with runtime errors rising from 2% at the start to 12% after 10 requests. The findings point to a broader challenge for AI coding tools as they move beyond generating initial software versions.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.