Stealth AI Startup Two Delta Secures Seed Funding and Accelerates Nvidia Models
Stealth AI startup Two Delta raised millions in seed funding from Bill Ackman and Eric Schmidt. The company unveiled DCOCO, a technology accelerating Nvidia Nemotron 3.5 Super.
The stealth artificial intelligence startup Two Delta, founded in 2025, has successfully closed a multi-million-dollar seed funding round backed by prominent investors, including Pershing Square founder and CEO Bill Ackman and former Google CEO Eric Schmidt. The company was co-founded by Dvir Seltzer and Dr. Eli David. Two Delta is developing technology designed to accelerate AI model inference and optimize computing resource utilization.
Acceleration Technology and Nvidia Collaboration
In October 2026, the company published the results of a joint collaboration with Nvidia, detailing a method developed to accelerate Nvidia's Nemotron 3.5 Super AI model. The proprietary technology, designated DCOCO, aims to increase token generation throughput per processing step without degrading output quality. The method builds upon speculative decoding techniques, wherein a smaller, faster draft model generates token sequences subsequently verified by a larger model.
"DCOCO leverages contextual data and candidate token scoring to select more consistent sequences, significantly increasing accepted tokens per verification cycle," research data indicates.
According to experimental data across ten datasets, DCOCO achieved an average throughput 2.03 times higher than standard decoding under deterministic processing conditions. Compared to the DFlash method, DCOCO delivered a 19.8% average performance improvement, maintaining efficiency even under heavy concurrent workloads.
EvoWeb and AI Agent Evaluation
Two Delta also published a research paper titled EvoWeb, evaluating the capacity of autonomous AI agents to build and maintain software applications over extended development cycles involving cumulative user modifications. Unlike static benchmarks evaluating single-prompt application generation, EvoWeb assesses multi-step processes where consecutive requests build upon existing codebases.
Key research findings indicate that none of the evaluated models maintained complete application integrity throughout the test lifecycle. Leading models, including Fable 5 and Kimi K3, achieved a score of approximately 0.69 out of 1. Major difficulties arose when integrating new features with pre-existing architecture, while runtime error rates escalated from 2% initially to 12% after ten sequential modification requests.
