Two Delta
Accelerating AI model inference and evaluating AI coding agents for web application maintenance.
Website: https://www.twodelta.com/
Cover Block
Public sources
| Field | Value |
|---|---|
| Name | Two Delta |
| Tagline | Accelerating AI model inference and evaluating AI coding agents for web application maintenance. |
| Headquarters | Israel [Calcalist Tech, October 2026] |
| Founded | 2025 [Dealroom News] |
| Stage | Seed [Calcalist Tech, October 2026] |
| Business Model | B2B |
| Industry | Deeptech |
| Technology | AI / Machine Learning |
| Geography | Middle East / North Africa |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2), Dvir Zeltzer and Dr. Eli David [Calcalist Tech, October 2026] |
| Funding Label | Seed [Calcalist Tech, October 2026] |
Links
Public sources
- Website: https://www.twodelta.com/
- GitHub: https://github.com/two-delta/DCOCO
- GitHub: https://github.com/two-delta/EvoWeb
Executive Summary
PUBLIC Two Delta is an Israeli AI infrastructure startup working on two tightly related problems: making large models run faster and testing whether AI coding agents hold up when web applications evolve over time, which deserves attention now because the company has paired an unusually early research footprint with reported seed backing from Bill Ackman and Eric Schmidt [Calcalist Tech, October 2026] [Two Delta] [Papers with Code]. Founded in 2025 by Dvir Zeltzer and Dr. Eli David, the company has stayed largely in stealth, with public disclosure limited to its website, research repositories, and a small amount of press coverage, a pattern that is common in frontier infrastructure bets but leaves investors with a narrow evidence base [Dealroom News] [Calcalist Tech, October 2026] [Two Delta].
The technical wedge appears to be real, if still early: Two Delta says its DCOCO speculative decoding method sped up NVIDIA Nemotron 3.5 Super by 2.03x versus vanilla decoding and by 19.8% versus DFlash at the same output quality, while its EvoWeb benchmark addresses a separate but adjacent need, measuring how coding agents perform across long-horizon, dependency-heavy web maintenance tasks rather than one-off code generation [Two Delta] [GitHub] [Papers with Code]. That combination matters because it suggests the company is not only chasing lower inference cost, but also trying to define evaluation standards around agent reliability, which could widen its product surface over time if enterprises demand both performance gains and better measurement [Calcalist Tech, October 2026] [GitHub] [Papers with Code].
On team quality, the public record is stronger for technical credibility than for company-building detail. Calcalist identifies Zeltzer and David as co-founders, while David's public profiles describe him as an AI researcher and entrepreneur focused on deep learning and evolutionary computation, with more than fifty published papers according to DeepAI; Zeltzer's background is less cleanly documented in primary reporting, though public secondary references tie him to prior senior finance and strategy roles and startup activity [Calcalist Tech, October 2026] [Forbes Technology Council] [DeepAI].
Commercially, the company is still best understood as a seed-stage B2B bet rather than a proven software vendor. Calcalist reported in October 2026 that Two Delta raised a seed round of undisclosed size described only as "tens of millions of dollars," with investors including Ackman and Schmidt, and the same report said the company had fewer than 10 employees, which implies substantial capital formation ahead of demonstrated go-to-market evidence [Calcalist Tech, October 2026]. Over the next 12 to 18 months, the key items to watch are whether DCOCO performance claims are validated beyond company channels, whether Nvidia remains a technical collaborator or becomes a stronger commercial signal, and whether EvoWeb develops into a recognized benchmark that gives Two Delta a durable position in agent evaluation rather than a single paper cycle [Two Delta] [Calcalist Tech, October 2026] [Papers with Code].
Company-stated, unverified -- This section relies materially on one independent news report, supplemented by company materials and public research repositories.
Taxonomy Snapshot
| Axis | Value |
|---|---|
| Stage | Seed |
| Business Model | B2B |
| Industry / Vertical | Deeptech |
| Technology Type | AI / Machine Learning |
| Geography | Middle East / North Africa |
| Growth Profile | Venture Scale |
| Founding Team | Co-Founders (2) |
| Funding | Seed |
How the Company Got Here
PUBLIC
Two Delta surfaced publicly in 2026, but the cleaner starting point is the company’s own footprint: its website states it is "quietly building something big," and presents the business as a stealth effort under the Two Delta name [Two Delta]. Secondary reporting traces the company’s founding to 2025 and places it in Israel, with Dvir Zeltzer and Dr. Eli David identified as co-founders [Dealroom News].
The public record is still thin, which matters more here than usual because the company is positioning around infrastructure-level AI performance rather than an easily observable application. What is visible on the company site is a research-led posture, including a research page and a contact flow that says the team is in stealth but willing to share more with selected counterparties [Two Delta]. No state filing or public legal entity detail was provided in the cited source set, so the overview remains limited to the operating brand and the limited chronology established by the website and published coverage [Two Delta] [Dealroom News].
The first clear milestone in the public timeline is the publication of research work under the Two Delta brand in 2026, including the company blog post on DCOCO and the broader research page on its site [Two Delta]. The second is the October 2026 press coverage reporting a seed financing backed by Bill Ackman and Eric Schmidt, which marked the company’s first broad public introduction beyond its own research materials [Calcalist Tech, October 2026].
Lightly corroborated -- Confirmed in part by the company website, with founding year and headquarters supported by secondary reporting rather than a public filing.
Product and Technology
Core product surface
MIXED Two Delta has disclosed a narrow but technically legible product surface: the company says it is working on technology that accelerates AI model inference and improves computational efficiency, while also publishing evaluation work for AI coding agents [Calcalist Tech, October 2026]. The public website remains deliberately sparse, describing the company as "quietly building something big" and inviting prospective counterparties to get in touch, but it does not add commercial packaging, deployment details, or customer-facing modules beyond the research trail [Two Delta].
The most concrete artifact is DCOCO, which Two Delta describes as its speculative decoding method for faster inference [Two Delta]. In a company blog post tied to NVIDIA Nemotron 3.5 Super, Two Delta reports that DCOCO delivered a 2.03x speedup over vanilla decoding and a 19.8% improvement over DFlash at the same output quality [Two Delta]. The supporting GitHub repository adds some technical context: DCOCO is compared against DFlash, Domino, and DSpark, uses evaluation prompts from DeepSpec and NVIDIA's SPEED-Bench, and is served on vLLM in the published setup [GitHub]. That is enough to establish a real research implementation, but not yet enough to establish production performance outside the disclosed benchmark environment [GitHub].
Evaluation and research tooling
MIXED The second visible strand is EvoWeb, which frames the company less as a single-feature inference startup and more as a team testing how capable coding agents behave over long, changing software tasks [Calcalist Tech, October 2026]. Two Delta's GitHub describes EvoWeb as the official implementation of a benchmark for coding agents on evolving web applications, and Papers with Code characterizes it as a testbed for evaluating and improving agents in long-horizon, multi-request development workflows [GitHub] [Papers with Code].
The published findings matter because they point to a specific technical pain point rather than a generic agent benchmark. According to Papers with Code, later requests with deeper or more numerous dependencies are solved less accurately, and runtime errors become more frequent over successive requests [Papers with Code]. Taken together, the public record suggests Two Delta is operating at the intersection of inference efficiency and agent reliability, with NVIDIA named as a technical collaborator in the DCOCO work but no verified public evidence yet of a broader commercial product suite, paid deployments, or announced roadmap beyond these research assets [Calcalist Tech, October 2026] [Two Delta].
Lightly corroborated -- This section relies on one independent publisher report and multiple company-controlled or repository sources, with partial corroboration from Papers with Code and GitHub.
Where the Demand Sits
PUBLIC
The market matters now because inference efficiency and agent reliability have moved from research curiosities to operating constraints, as enterprises deploy larger models and begin testing software agents in real workflows [Calcalist Tech, October 2026] [Papers with Code].
Public evidence on Two Delta’s exact addressable market is thin, so the cleanest approach is to frame the opportunity through adjacent markets rather than force a precise TAM. On the inference side, the company is developing technology to accelerate AI models and improve computational efficiency, with public work centered on DCOCO and a collaboration around NVIDIA’s Nemotron 3.5 Super [Calcalist Tech, October 2026] [Two Delta]. That places it inside the AI infrastructure and model serving stack, where budget is often justified by lower latency, better throughput, or reduced compute cost rather than by net-new software line items. On the agent side, EvoWeb targets evaluation of coding agents on evolving web applications, which aligns more closely with developer tools, application testing, and software maintenance workflows than with generic foundation model research [Calcalist Tech, October 2026] [GitHub] [Papers with Code].
The immediate demand driver is straightforward: as model usage expands, inference becomes a recurring expense and a user-experience bottleneck. Two Delta’s own disclosed results claim DCOCO speeds up NVIDIA Nemotron 3.5 Super by 2.03x over vanilla decoding and 19.8% over DFlash at the same output quality [Two Delta]. Those figures are company-published, so they should be read as directional until independently replicated, but they still point to the commercial logic of the category. If an acceleration layer can preserve output quality while reducing tokens-per-second cost or waiting time, the buyer conversation is legible across model providers, enterprises running internal inference, and platform teams supporting AI features.
A second tailwind comes from the shift in enterprise attention from one-shot code generation to longer-horizon software maintenance. EvoWeb is described as a benchmark for coding agents on evolving web applications, and its reported analysis shows that later requests with deeper or more numerous dependencies are solved less accurately, while runtime errors become more frequent over successive requests [GitHub] [Papers with Code]. That matters because it surfaces a gap between impressive demo performance and production reliability. If that gap persists, evaluation infrastructure may become a required spend category for teams selecting, tuning, or governing coding agents.
The substitute markets are broad enough to matter. For inference acceleration, adjacent solutions include model optimization frameworks, serving stacks such as vLLM, and alternative decoding methods, including the DFlash, Domino, and DSpark approaches referenced in Two Delta’s public materials [GitHub]. For coding-agent evaluation, substitutes include internal benchmark suites, QA automation, CI testing infrastructure, and broader developer productivity tools. In practice, Two Delta may need to fit into existing engineering workflows rather than create a fully standalone budget line, which can slow adoption but also widen the integration surface if the tooling proves credible.
Macro and regulatory forces are less direct here than in heavily regulated vertical AI markets, but they still shape demand. Rising scrutiny of AI system reliability and enterprise pressure for measurable performance can support both sides of Two Delta’s work: acceleration when compute budgets tighten, and evaluation when governance standards get stricter. The public record does not yet show vertical compliance positioning or sector-specific certifications [Calcalist Tech, October 2026] [Two Delta]. For now, the more relevant macro variable is whether enterprise AI spend continues to move from experimentation into cost-accountable production environments.
| Market lens | Relevance to Two Delta | Evidence |
|---|---|---|
| AI inference infrastructure (analogous market) | DCOCO is positioned as an inference acceleration method aimed at higher efficiency and faster decoding | [Calcalist Tech, October 2026] [Two Delta] |
| Model serving and optimization tools (adjacent market) | Public materials reference vLLM and comparisons to DFlash, Domino, and DSpark, suggesting participation in a performance-tooling layer rather than only model development | [GitHub] |
| Coding agent evaluation (emerging market) | EvoWeb targets testing of coding agents on evolving web applications and measures degradation across successive requests | [GitHub] [Papers with Code] |
| Developer tools and software maintenance (adjacent market) | The benchmark is relevant where teams need to assess whether agents can safely maintain real applications over time | [Calcalist Tech, October 2026] [Papers with Code] |
The table shows a company sitting across two markets that are early but commercially legible: inference optimization, where ROI can be framed in cost and latency terms, and coding-agent evaluation, where the wedge is trust and measurable reliability. The absence of third-party sizing data prevents a hard market model here, but the category signals are strong enough to justify continued attention.
Lightly corroborated -- Section relies on one independent news report for company positioning and public primary sources for research scope; no third-party market-sizing report was available.
Competitive Landscape
MIXED Two Delta appears to sit at the intersection of two crowded but still unsettled lanes, inference optimization for frontier models and evaluation tooling for AI coding agents, where the immediate alternatives are less direct company peers than a mix of open research methods, infrastructure stacks, and model vendors that can absorb adjacent features into broader platforms [Calcalist Tech, October 2026] [GitHub] [Papers with Code].
On the inference side, the public record points to a research-first posture rather than a packaged platform. Two Delta says DCOCO builds on DFlash and compares against Domino and DSpark, while its evaluation setup uses prompts from DeepSpec and NVIDIA's SPEED-Bench and serves through vLLM [GitHub]. That matters because the practical competitive set here includes not only startups selling faster inference, but also open-source serving frameworks, model labs optimizing their own stacks, and cloud or chip vendors that can ship speed improvements as part of a larger bundle. Nvidia is the named technical collaborator in public reporting, and that gives Two Delta some visibility, but it also underscores how much bargaining power still sits with the platform owner rather than with an external optimization layer [Calcalist Tech, October 2026].
On the coding-agent evaluation side, EvoWeb places the company in a different arena, one closer to benchmarks and developer tooling than to core model acceleration. The repository and paper summary describe EvoWeb as a benchmark for coding agents operating across evolving web applications and repeated feature requests, with observed degradation as dependency chains deepen and requests accumulate [GitHub] [Papers with Code]. The incumbents in that lane are not clearly named in the sourced material, but the substitute products are easy to define from the evidence: internal eval harnesses built by model labs, benchmark suites maintained by open research communities, and application teams testing agents against their own repos. That makes this a distribution problem as much as a technical one. A benchmark can shape developer behavior if it becomes a standard reference point, but it can also remain academically useful and commercially peripheral.
The edge visible today is concentrated in talent and research timing, and both look real but perishable. Calcalist identifies the founders as Dvir Zeltzer and Dr. Eli David, and public profiles tie David to a long AI research track record while Zeltzer's background appears to span strategy, finance, and startup activity rather than a conventional infrastructure-sales path [Calcalist Tech, October 2026] [elidavid.com] [DeepAI] [LinkedIn]. The combination could help in two ways: credible technical work that earns attention from model builders, and access to influential networks that matter in a small market for high-end AI infrastructure. Still, there is no public evidence yet of proprietary deployment data, embedded customer workflows, or a commercial channel that would make the position self-reinforcing. Research edges in inference tend to decay quickly when methods are published, reimplemented, or folded into upstream frameworks.
The clearest exposure is to better-capitalized actors that control the model, hardware, or serving layer. Nvidia is a useful example from the public record because it is simultaneously a collaborator and a potential absorber of value if acceleration techniques become part of the default stack [Calcalist Tech, October 2026] [Two Delta]. vLLM is another relevant pressure point because Two Delta's own DCOCO materials rely on it for serving, which suggests dependence on an ecosystem layer the company does not own [GitHub]. In that sense, the strongest competing force may not be a named startup at all. It may be the tendency for performance improvements and eval frameworks to migrate toward open infrastructure or platform-native tooling, leaving a standalone company to prove it can keep discovering gains faster than the stack commoditizes them.
The most plausible 18-month scenario is bifurcation between research relevance and product capture. Nvidia is the most credible winner if inference optimization remains a feature of the broader hardware-plus-model stack, because it already owns developer mindshare, benchmark context, and distribution into production environments [Calcalist Tech, October 2026]. Two Delta is the most plausible loser if DCOCO and EvoWeb remain respected technical artifacts without translating into a repeatable commercial wedge, particularly since the company is still described as stealthy and sub-10 employees in the only named press coverage [Calcalist Tech, October 2026]. The upside case exists if its methods become important enough that platform owners or model labs prefer partnership over internal replication, but the public evidence today is stronger on technical promise than on defensible market control.
Lightly corroborated -- This section relies on one independent reported source for company positioning and funding context, with additional corroboration from company and repository materials for technical claims [Calcalist Tech, October 2026] [GitHub] [Papers with Code] [Two Delta].
Opportunity
Upside case
PUBLIC The prize here is unusually large if Two Delta can turn early research into product, because inference efficiency sits on the cost line of every serious AI deployment and coding-agent evaluation sits on the reliability line of every enterprise software team experimenting with autonomous development [Calcalist Tech, October 2026] [GitHub] [Papers with Code].
The clearest upside is not that Two Delta becomes a general AI company. It is that the company becomes a specialized infrastructure layer for two pain points that are getting more expensive, faster inference and more trustworthy agentic software work. Public evidence is still thin, but it is pointed. Calcalist Tech reports that the company is focused on accelerating AI models and improving computational efficiency, and separately on evaluating AI coding agents for evolving web applications [Calcalist Tech, October 2026]. On the technical side, the company has put its name on DCOCO, which it describes as a speculative decoding method that improved NVIDIA Nemotron 3.5 Super by 2.03x over vanilla decoding and 19.8% over DFlash at the same output quality [Two Delta] [GitHub]. On the evaluation side, EvoWeb is positioned as a benchmark for long-horizon web-application maintenance, and Papers with Code summarizes its central finding plainly: as requests compound and dependencies deepen, coding-agent accuracy falls and runtime errors rise [Papers with Code]. That combination matters because it gives Two Delta a plausible route to matter in both the economics of model serving and the governance of agent quality, rather than only one of the two.
The path to reachability is still narrow, but it is visible. The Nvidia-linked work is the strongest public signal that the research may travel beyond a lab artifact, because it ties the company to a named model ecosystem and a measurable performance claim, even if no commercial deployment has been disclosed [Calcalist Tech, October 2026] [Two Delta]. The founder pairing also fits the problem shape: Calcalist Tech identifies Dvir Zeltzer and Dr. Eli David as co-founders, and public profiles describe David as an AI researcher and entrepreneur with a long publication record in deep learning [Calcalist Tech, October 2026] [elidavid.com] [DeepAI]. Seed backing from Bill Ackman and Eric Schmidt, as reported by Calcalist Tech, does not prove product-market fit, but it does suggest the company should have the time and network to push a technically demanding wedge into market [Calcalist Tech, October 2026].
| Scenario | What happens | Catalyst | Why it's plausible |
|---|---|---|---|
| Inference control plane | Two Delta turns DCOCO from a research result into a serving layer used by model builders and enterprise AI teams to cut latency and compute cost in production. | A deeper Nvidia ecosystem relationship or broader support around vLLM-based serving. | DCOCO is already framed as a decoding method for faster inference, and the public repo notes evaluation and serving work built on vLLM while comparing against other decoding approaches such as DFlash, Domino, and DSpark [Two Delta] [GitHub]. |
| Agent reliability standard | EvoWeb becomes a benchmark that frontier model labs and enterprise buyers use before shipping coding agents into maintenance workflows. | Wider citation, benchmark adoption, or use in vendor evaluation by model providers and platform teams. | Two Delta has already published EvoWeb, and Papers with Code describes it as a benchmark for coding agents on evolving web applications with clear degradation patterns across successive requests, which maps directly to a live buyer concern around agent reliability [Calcalist Tech, October 2026] [Papers with Code] [GitHub]. |
| Dual-stack optimization platform | The company combines acceleration and evaluation into a workflow where customers both run agents more cheaply and measure whether those agents remain dependable over long task chains. | Product packaging that links model-serving improvements with benchmark-driven QA for agentic development teams. | The public record already shows both technical surfaces under one roof, inference acceleration through DCOCO and agent evaluation through EvoWeb, which creates a credible foundation for a combined platform thesis even before monetization is visible [Calcalist Tech, October 2026] [Two Delta] [Papers with Code]. |
What compounding could look like is straightforward even if it is not yet proven in public. A first win in inference optimization can create a distribution loop into model providers, inference stacks, and enterprise AI teams that care about latency and cost on every request. If those deployments also generate trust in Two Delta's evaluation tooling, the company can sell not only speed but confidence, which is often the harder purchase in enterprise AI. The public signs of that flywheel are early rather than mature: DCOCO is connected to a named Nvidia model and built in an ecosystem that references vLLM, while EvoWeb addresses a recurring failure mode in long-horizon agent tasks, namely that errors accumulate as tasks evolve over time [Two Delta] [GitHub] [Papers with Code]. If that pairing becomes productized, each successful deployment could improve the next sale by lowering compute cost and sharpening the evidence that an agent remains usable after repeated changes.
The size of the win is best framed through scenario logic rather than a straight-line projection. There is no cited market-size figure in the available record, and no direct public comparable is named in the source set, so valuation math would be too loose to treat as analysis. Even so, if the "Inference control plane" scenario plays out and Two Delta becomes a recognized software layer around model-serving efficiency for major AI deployments, the company could plausibly support a multibillion-dollar outcome (scenario, not a forecast), because it would sit in a high-frequency, compute-linked part of the AI stack where small technical gains can translate into large budget impact [Calcalist Tech, October 2026] [Two Delta]. If the "Dual-stack optimization platform" scenario lands, the upside expands further, because the company would then touch both the cost side and the reliability side of enterprise AI operations. That remains an ambitious case, but it is not detached from the evidence currently on the page.
Lightly corroborated -- Relies primarily on Calcalist Tech reporting plus company and repository materials, with partial corroboration from Papers with Code and founder-profile sources.
Sources
Public sources
[Calcalist Tech, October 2026] Bill Ackman and Eric Schmidt back secretive Israeli AI startup Two Delta | Ctech | https://www.calcalistech.com/ctechnews/article/rkpvzkuhmx
[Two Delta] Two Delta · Intelligence is about to change hands, into yours. | https://www.twodelta.com/
[Papers with Code] EvoWeb: A Benchmark for Coding Agents on Evolving Web Applications (arXiv:2610.16402) | https://paperswithcode.co/paper/2610.16402
[Dealroom News] Ackman and Schmidt back secretive Israeli AI startup Two Delta in seed round | https://dealroom.co/news/165191-ackman-and-schmidt-back-secretive-israeli-ai-startup-two-delta-in-seed-r/
[GitHub] GitHub - two-delta/DCOCO | https://github.com/two-delta/DCOCO
[Forbes Technology Council] Eli David | Researcher and Entrepreaneur - Bar-Ilan University | Forbes Technology Council | https://councils.forbes.com/profile/Eli-David-Researcher-Entrepreaneur-Bar-Ilan-University/b5a70384-2c13-4600-bb28-25c8e1a22490
[DeepAI] Eli David | DeepAI | https://deepai.org/profile/eli-david
[elidavid.com] Dr. Eli David | AI Researcher and Entrepreneur | https://www.elidavid.com/
[LinkedIn, 2026] Alice Clarke - Google | LinkedIn | https://www.linkedin.com/in/clarkealice/
Articles about Two Delta
- Two Delta's AI Acceleration Bet Wins Backing From Ackman and Schmidt — The stealth startup, with a team under 10, is publishing research on model acceleration and coding agent evaluation while raising tens of millions.