Bill Ackman and Eric Schmidt back secretive Israeli AI startup Two Delta
The company has raised tens of millions of dollars in seed funding as it develops technology to accelerate AI models. Its research also examines why AI coding agents struggle to keep applications working as users add features.
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 has learned.
Founded in 2025, Two Delta operates with a high degree of secrecy and has disclosed little about its business beyond research papers published on its website. The company was co-founded by Dvir Zeltzer and Dr. Eli David and is developing technology designed to accelerate AI models and improve the efficiency with which they use computing resources.
David is a veteran AI researcher and entrepreneur, while Zeltzer brings a background in business and management, including a senior role at Google.
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. The method is designed to increase the number of tokens a model can generate in each verification step while preserving output quality.
DCOCO builds on a technique known as speculative decoding. In this approach, a smaller, faster model drafts a sequence of tokens that a larger model verifies in a single pass. The process can speed up generation, but some proposed tokens may be rejected. The challenge is to produce drafts that the larger model is more likely to accept.
Two Delta’s method seeks to improve those proposals by combining candidate-token embeddings with contextual information from the drafting model’s hidden states, which are informed by the target model’s activations. It scores possible combinations of neighboring tokens in parallel, helping select more consistent sequences before the larger model verifies them.
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. Under the same configuration, it reported an average throughput improvement of 19.8% over the competing DFlash method. The company also reported throughput gains under different decoding settings and workloads involving up to 32 concurrent requests. The evaluations were conducted on Nvidia DGX B200 hardware using the vLLM inference framework.
These are results reported by Two Delta, rather than an independent assessment of the technology. The company says DCOCO improves the efficiency of token generation while preserving the target model’s output quality.
Two Delta has also published EvoWeb, a benchmark designed to evaluate how well AI coding agents build and maintain web applications as users repeatedly request new features and modifications. Unlike benchmarks that evaluate an agent’s performance on a single prompt, EvoWeb tests an application over 10 successive requests: an initial build followed by nine changes.
The study tested 32 web applications and found that none of the evaluated models maintained full functionality throughout the process. The top performers, Fable 5 and Kimi K3, received scores of 0.694 and 0.684, respectively, on the study’s scale of zero to one.
The main difficulty was integrating new features with existing functionality. Although build failures remained rare, runtime errors increased from 2% at the beginning of the process to 12% after 10 requests, according to the study.
The results highlight a challenge for AI coding tools as they move beyond generating initial versions of software. An agent may be able to build an application and add features successfully, yet still struggle to preserve the connections between those features as the application becomes more complex.
