Two Delta, a secretive Israeli AI startup, has raised tens of millions of dollars in a seed round involving Bill Ackman and Eric Schmidt, according to Calcalist's October 22, 2026 report. Founded in 2025 by Dvir Zeltzer and Dr. Eli David, the company has fewer than 10 employees. The report did not give an exact funding amount or valuation. Ackman is the founder and CEO of Pershing Square, while Schmidt previously served as Google's chairman.
The company develops technology to accelerate AI model inference and use computing resources more efficiently. It has also published technical work on DCOCO, a method developed in collaboration with NVIDIA, and EvoWeb, a benchmark for coding agents that repeatedly modify web applications. The two projects reveal some of its work, but do not provide a full account of its products or business model.
Funding before a full business picture
According to Calcalist, Two Delta has disclosed little about its operations beyond the technical publications on its website. That allows its development objectives and reported investors to be described, but does not establish sales, customer numbers or commercial product readiness. Ackman and Schmidt's participation is not a measure of technological performance either, which requires assessment under defined usage conditions.
The Israeli angle is clear: a small local company founded only in 2025 is raising capital from prominent international investors to work on AI's computing layer. Rather than presenting a new consumer application, it focuses on how existing models produce results. The potential implication is more work from the same infrastructure, although integration and operational steps separate an experimental result from customer savings.
The investment must also be distinguished from the NVIDIA collaboration. The funding report names Ackman and Schmidt among the investors. The technical publication describes work with NVIDIA to accelerate its model. That does not mean NVIDIA participated in the round, bought the technology or adopted it across all its products; no such announcement accompanied the published details.
DCOCO improves the draft before verification
DCOCO operates in inference, the use of an already trained model to produce an answer. One way to accelerate it is speculative decoding: a small, fast model proposes a sequence of tokens, and the larger model checks that proposal in one pass. Tokens are information units processed by a model and are not necessarily whole words. Accepting several proposals together can save sequential work.
A fast proposal, however, is not necessarily a good one. If parts of the sequence do not fit together, the larger model may reject them, reducing the benefit. DCOCO aims to improve compatibility between candidates before verification. The company describes combining token representations with contextual information and scoring neighboring combinations in parallel to choose a more coherent sequence.
Put simply, the method tries to increase how much output the larger model can accept in each pass without making draft preparation too slow. The larger model remains responsible for verification. The company says the process preserves the target model's output quality. That is a claim about execution, not an announcement that the model has acquired new knowledge or become better at understanding questions.
What the 2.03 times figure represents
Two Delta reports that in its own tests on NVIDIA's Nemotron 3.5 Super VL, the method achieved average throughput 2.03 times that of standard autoregressive decoding. These are company published results without independent confirmation. The figure applies to specific experimental conditions: thinking disabled, greedy token selection and one request at a time. It is not a promise to double speed in every application.
The publication describes tests on 10 datasets covering mathematics, coding and general tasks, using NVIDIA DGX B200 hardware and the vLLM serving framework. Those settings are part of the result. Changing the model, hardware, answer length or number of concurrent requests can change the outcome. The figure is therefore useful as a result in a particular configuration, not a constant multiplier for every service.
Throughput also needs a precise reading: it measures output produced per unit of time. It is not identical to the total time a user waits for an answer, which sometimes includes additional stages. Likewise, more tokens per second do not automatically translate into an equivalent price reduction. Service costs also depend on supporting resources, system utilization and the agreement with the computing provider.
For an Israeli team considering such technology, the practical question is how it performs on its own workloads and tasks while maintaining the required quality. The funding and headline number provide context, but cannot replace compatibility with an existing system. It is important to distinguish the efficiency objective from benefits already measured for a customer, for which the company has not provided detailed commercial figures here.
EvoWeb looks beyond the first version
The second project, EvoWeb, asks a different question: can a coding agent keep an application working as a user continues requesting additions and changes? Two Delta's benchmark includes 32 web applications, each undergoing 10 consecutive requests: an initial build followed by 9 changes. It therefore also tests connections between new features and what has already been built.
The distinction from a single request benchmark is highly practical. An application can build successfully and then break when an agent adds a feature dependent on existing data or behavior. A continuing test examines whether changes fit together, not merely whether each addition looks convincing on its own. These are the results and definitions of the company's benchmark, not an absolute ranking of every coding tool on the market.
In practice, maintenance testing should also exercise older capabilities after a change. Code building without an error does not guarantee that a user action will complete correctly. That is the distinction between technically preparing an application successfully and having it function during use.
For developers in Israel, the connection between the projects is the need to measure systems under conditions closer to continuing use: both how efficiently output is generated and whether software keeps working after changes. Two Delta presents a technological direction and substantial funding for a company of its size. Its commercial operations remain sparsely disclosed, and the acceleration figures remain company results within the experiments it described.