Two Delta secures seed backing and reports faster inference with DCOCO
The AMW Read
The seed round adds an inference optimization entrant to the infrastructure player map, but undisclosed funding terms and company-reported benchmarks limit evidence of broader market impact.
Two Delta secures seed backing and reports faster inference with DCOCO
Israeli startup Two Delta has raised tens of millions of dollars in seed funding from investors including Bill Ackman and former Google chairman Eric Schmidt, according to Calcalist. Founded in 2025 by Dvir Zeltzer and Dr. Eli David, the company develops technology to accelerate AI models. In an October research collaboration with Nvidia, it reported that its DCOCO method delivered 2.03 times standard autoregressive decoding throughput on Nvidia’s Nemotron 3.5 Super VL model, averaged across 10 datasets with reasoning disabled and greedy decoding enabled.
Two Delta enters the AI infrastructure market through inference software optimization. DCOCO builds on speculative decoding, in which a smaller model drafts tokens for a larger model to verify, and seeks to improve the consistency of those drafts. Under the same configuration, Two Delta reported 19.8% higher throughput than DFlash. The distinction matters commercially: a gain against standard decoding establishes one baseline, while the smaller advantage against another acceleration method is more relevant to buyers comparing optimized serving options. The tests used Nvidia DGX B200 hardware and vLLM; the results are company-reported, without an independent assessment.
For builders and investors, the concrete diligence step is to test DCOCO against an optimized serving baseline under their own decoding settings and concurrency requirements before translating throughput into cost savings. Two Delta also published EvoWeb, which tested coding agents across an initial application build and nine successive changes. None of the evaluated models preserved full functionality throughout, highlighting a separate evaluation question: whether successful feature additions leave the existing application working.