Two Delta Raises Tens of Millions With Ackman and Schmidt On Board

An Israeli AI startup with fewer than 10 employees 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.

The company, Two Delta, was founded in 2025 by Dvir Zeltzer and Dr. Eli David. It operates with what Calcalist describes as a high degree of secrecy and has disclosed almost nothing about its business beyond research papers published on its website. For a seed round of this size, that is a thin public record — the valuation, exact amount and deal terms remain undisclosed.

David is a veteran AI researcher and entrepreneur. Zeltzer brings a business and management background, including a senior role at Google.

What has Two Delta actually built?

The company is developing technology to accelerate AI models and improve how efficiently they use computing resources. Its most concrete public result is a collaboration with Nvidia. In October 2026, Two Delta published results showing its DCOCO technology accelerated inference for Nvidia's Nemotron 3.5 Super VL model.

DCOCO builds on speculative decoding, a known technique in which a smaller, faster model drafts token sequences that a larger model verifies in a single pass. The approach speeds generation, but the larger model rejects some proposed tokens. Two Delta's method attacks that weakness: it combines candidate-token embeddings with contextual information from the drafting model's hidden states, which draw on the target model's activations. It scores combinations of neighboring tokens in parallel to select more consistent sequences before verification.

Two Delta reported that DCOCO achieved average throughput of 2.03 times standard autoregressive decoding across 10 datasets covering mathematics, coding and general-purpose tasks, with reasoning disabled and greedy decoding used. Under the same configuration, it reported a 19.8% average throughput improvement over the competing DFlash method. The company also reported gains under different decoding settings and workloads of up to 32 concurrent requests. The evaluations ran on Nvidia DGX B200 hardware using the vLLM inference framework.

One caveat deserves emphasis: these figures come from Two Delta itself, not an independent assessment. The company says DCOCO preserves output quality while improving token-generation efficiency — a claim outside reviewers have not tested.

What does the EvoWeb benchmark reveal?

Two Delta has also published EvoWeb, a benchmark measuring how well AI coding agents build and maintain web applications as users repeatedly request changes. Unlike single-prompt tests, EvoWeb evaluates an application over 10 successive requests: one initial build followed by nine modifications.

The study tested 32 web applications. None of the evaluated models maintained full functionality throughout. The top performers, Fable 5 and Kimi K3, scored 0.694 and 0.684 respectively on a zero-to-one scale — under 70% for the best agents in the field.

The dominant failure mode was integrating new features with existing ones. Build failures stayed rare, but runtime errors climbed from 2% at the start of the process to 12% after 10 requests — a sixfold increase as applications grew more complex. In plain terms, agents can assemble software feature by feature but struggle to keep the connections between those features intact.

Why does a seed round this size carry risk?

Tens of millions of dollars for a company under two years old, with a single-digit headcount and no disclosed customers or revenue, is a bet on research pedigree rather than commercial traction. Ackman and Schmidt are effectively underwriting two published papers and a founding team's track record.

The Nvidia collaboration lends credibility, but the benchmark results — including Two Delta's own EvoWeb findings on how quickly AI-built software degrades — illustrate how early the underlying technology remains. Whether DCOCO's reported 2.03x throughput gain survives independent replication and commercial deployment will determine whether this seed round looks prescient or premature.