Decart

Decart is a vertically integrated AI research and infrastructure company founded in late 2023 by Israeli engineers Dean Leitersdorf, Moshe Shalev and Orian Leitersdorf. CEO Dean Leitersdorf and CPO Moshe Shalev met in Israel Defense Forces intelligence/Unit 8200; Dean completed a Technion computer-science doctorate by age 23, then a record, while younger brother and chief scientist Orian later completed an electrical/computer-engineering doctorate at the Technion at about 22 and became the institution's youngest doctorate recipient. Decart's thesis is that AI progress is constrained as much by cost, latency and hardware utilization as model capability. It therefore develops both low-level infrastructure and models designed around that infrastructure. DOS—the Decart Optimization Stack—spans hardware-aware model design, kernels, memory, proprietary compilers, training and inference optimization, and is intended to run across NVIDIA GPUs, Google TPUs and AWS Trainium rather than lock customers to one accelerator. On top, Lucy performs real-time video/world editing, including style/character/object changes, virtual try-on, product placement and live advertising; Oasis generates interactive environments conditioned on user or agent actions, evolving from a 2024 Minecraft-like demo into a physical-AI/robotics/autonomous-driving simulator. The current platform exposes real-time and asynchronous image/video APIs and SDKs with pay-as-you-go prices, while enterprise DOS and volume arrangements are custom. Decart emerged from stealth on October 31, 2024 with $21m from Sequoia Capital and Zeev Ventures/Oren Zeev and the Oasis demo, created with Etched. It announced a $32m Benchmark-led Series A on December 19, 2024, with Sequoia and Zeev participating, at an estimated $500m-plus valuation. Oasis reached a company-reported one million users in three days, although reviewers found the initial low-resolution model unstable and prone to forgetting scene layout. Decart already said enterprise GPU-optimization contracts were generating millions of dollars and that it became profitable within months; these are company statements, not audited accounts. An August 7, 2025 $100m round from Sequoia, Benchmark, Zeev and new investor Aleph valued it at $3.1bn and brought cumulative funding to $153m. Fortune reported the company had spent less than $10m of investor capital at that point. In May 2026 Decart raised another $300m led by Radical Ventures at a reported/estimated nearly $4bn valuation, bringing exact announced funding to $453m. New investors included NVIDIA, eBay Ventures, Adobe Ventures, Toyota Ventures, Atreides Management and Valor Equity Partners; existing Sequoia, Benchmark and Zeev rejoined, alongside angels Andrej Karpathy, Michael Eisner, Nintendo-family members and Moritz Baier-Lentz. Amazon became a strategic customer and NVIDIA both investor and infrastructure partner; Decart also works with Comcast/NVIDIA on sub-35ms edge AI and with ElevenLabs on live characters. The company claimed less than $10m of total investor capital had been spent as of the May round and reported roughly 100 employees by August 2026, up from 35 at end-2024 and 60+ in August 2025, across Tel Aviv, northern Israel, San Francisco and New York. Its current site claims up to 100x efficiency, sub-30ms world-model inference and full-HD video at 100fps; those are first-party performance claims requiring workload-specific validation. As of August 17, 2026, Anthropic is in advanced but unsigned talks to acquire Decart for roughly $6bn (some Israeli reporting says $6-7bn and an NVIDIA rival bid), mainly for inference/performance and chip-optimization capability. Reuters and Bloomberg reporting stress that the deal could still fail. Decart therefore remains an active private company; the rumored price is not an exit or current confirmed valuation. Industry estimates suggest the three founders still own about 64% combined, but neither cap table nor round dilution is public. The business has two interlocking revenue engines: enterprise licenses/usage for optimization infrastructure and consumption/enterprise fees for models/APIs, with strategic cloud/hardware partnerships providing distribution. It targets hyperscalers, AI labs, accelerator vendors, media/advertising/commerce platforms, game and social developers, robotics/autonomous systems and model builders requiring deterministic low-latency inference. Competitive risks are severe: NVIDIA/Triton/TensorRT, Google/AWS software, Modular, Together, Fireworks, Run:ai/Deci, other inference platforms and world-model/video labs can compress pricing or duplicate capability. Hardware portability claims are technically hard; optimization gains may be workload-specific and require continuous support as chips/models change. World models still struggle with spatial persistence, physics, long-horizon consistency and controllability; failures matter much more for robotics/driving than entertainment. Video transformation brings deepfake, consent, biometric/privacy, political misuse and IP risks. Oasis' first demo was trained on Minecraft gameplay footage and explicitly was not authorized or endorsed by Microsoft/Mojang; no lawsuit was found, but training and output resemblance create unresolved copyright/trademark exposure. Decart's terms place third-party rights/consent obligations on users, allow use of content to improve and market the platform, and acknowledge non-unique, inaccurate, biased or offensive outputs. The $4bn last priced valuation and prospective $6bn bid reflect exceptional expectations relative to disclosed millions—not billions—of revenue. No material Decart-specific breach, enforcement or litigation was found in the manual review.

Founding story

Leitersdorf and Shalev met while serving in Israeli intelligence/Unit 8200 and saw that AI systems were too expensive and slow for persistent real-time experiences. They formed an AI lab in late 2023 to optimize the compute stack first, then use that advantage to build real-time models. Dean's younger brother Orian joined the founding group as chief scientist, adding unusually deep systems/academic training.

Business model

Vertically integrated frontier AI lab selling enterprise optimization infrastructure and usage-based real-time video/world-model APIs, with strategic hardware/cloud partnerships and consumer/product demonstrations driving adoption.

Enterprise licensing and usage contracts for DOS training/inference optimization; pay-as-you-go per-second/per-generation Lucy/Oasis/image APIs; enterprise volume contracts; possible strategic co-development. Revenue, gross margin and contract concentration are not disclosed.

Traction

Optimization contracts reportedly produced millions in revenue and profitability within months; team grew from 35 at end-2024 to 60+ by Aug 2025 and about 100 by Aug 2026; Oasis claimed 1m users in three days; company had spent under $10m of investor money by Aug 2025 and still claimed under $10m in May 2026; Amazon is a strategic customer, NVIDIA a partner/investor, and Comcast is testing edge applications. Current ARR, customer count, retention, margins and platform usage are undisclosed.

Latest developments

May 2026 $300m financing at nearly $4bn brought total capital to $453m and added NVIDIA, Adobe, eBay, Toyota, Atreides, Valor and prominent angels; Amazon became strategic customer. June-July releases added Oasis 3 and Lucy 2.5, while Comcast/NVIDIA work targeted sub-35ms edge inference. On Aug 13-17, 2026 Reuters/Bloomberg and Israeli press reported advanced Anthropic talks around $6bn (some reports $6-7bn and an NVIDIA competing offer). No transaction was signed by the review cutoff.