Key Takeaways
- • Israeli AI startup Two Delta has raised tens of millions of dollars in a seed round backed by Bill Ackman, founder and CEO of Pershing Square, and former Google chairman Eric Schmidt, according to Calcalist.
- • Founded in 2025 and co-founded by Dvir Zeltzer and Dr. Eli David, Two Delta operates with a high degree of secrecy and has disclosed little beyond research papers published on its website.
- • Two Delta's DCOCO technology accelerates inference for Nvidia's Nemotron 3.5 Super VL model, with the company reporting 2.03 times the throughput of standard autoregressive decoding and a 19.8% improvement over the DFlash method. These are company-reported results, not independent assessments.
- • Two Delta's EvoWeb benchmark found that none of 32 web applications maintained full functionality across 10 successive AI coding agent requests, with top scores of 0.694 (Fable 5) and 0.684 (Kimi K3) on a zero-to-one scale.
- • Runtime errors in the EvoWeb study rose from 2% at the start of the process to 12% after 10 requests, highlighting integration of new features with existing functionality as a key weakness for AI coding tools.
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, Calcalist reported. The company was founded in 2025 and operates with a high degree of secrecy, disclosing little about its business beyond research papers published on its website. It was co-founded by Dvir Zeltzer, who brings a background in business and management including a senior role at Google, and Dr. Eli David, a veteran AI researcher and entrepreneur.
The company 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 it used its DCOCO technology 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 in a single pass, and seeks to improve draft quality by combining candidate-token embeddings with contextual information from the drafting model's hidden states.
In tests across 10 datasets covering mathematics, coding and general-purpose tasks, Two Delta reported that DCOCO achieved average throughput 2.03 times that of standard autoregressive decoding when reasoning was disabled and greedy decoding was used, and a 19.8% average throughput improvement over the competing DFlash method under the same configuration. The evaluations ran on Nvidia DGX B200 hardware using the vLLM inference framework. The company itself acknowledges these are its own reported results rather than an independent assessment.
Two Delta has also published EvoWeb, a benchmark that tests how well AI coding agents build and maintain web applications across an initial build followed by nine feature changes. Across 32 web applications, 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 respectively. Build failures stayed rare, but runtime errors rose from 2% to 12% after 10 requests, with the main difficulty being integration of new features with existing functionality.
The company 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 it used its DCOCO technology 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 in a single pass, and seeks to improve draft quality by combining candidate-token embeddings with contextual information from the drafting model's hidden states.
In tests across 10 datasets covering mathematics, coding and general-purpose tasks, Two Delta reported that DCOCO achieved average throughput 2.03 times that of standard autoregressive decoding when reasoning was disabled and greedy decoding was used, and a 19.8% average throughput improvement over the competing DFlash method under the same configuration. The evaluations ran on Nvidia DGX B200 hardware using the vLLM inference framework. The company itself acknowledges these are its own reported results rather than an independent assessment.
Two Delta has also published EvoWeb, a benchmark that tests how well AI coding agents build and maintain web applications across an initial build followed by nine feature changes. Across 32 web applications, 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 respectively. Build failures stayed rare, but runtime errors rose from 2% to 12% after 10 requests, with the main difficulty being integration of new features with existing functionality.
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.