Decart raises $300M to scale world models and real-time AI across Amazon, Nvidia, and Google chips

Decart raises $300M to put a real-time world model in front of Amazon's chips

Radical Ventures led the round, with Nvidia, Sequoia, Benchmark, Adobe, and Toyota alongside. Andrej Karpathy, Michael Eisner, and the Nintendo family are angels. Total raised is now over $450m. Decart, the AI research lab building real-time video and world models, announced on Monday it had raised $300m in new funding led by Radical Ventures. The round takes the two-year-old company's total raised past $450m, with Nvidia, Atreides Management, Valor Equity Partners, Adobe Ventures, Toyota Ventures, and eBay Ventures joining as new investors alongside returning backers Sequoia Capital, Benchmark, and Zeev Ventures. The angel list is the part that signals how Decart is positioning the company. Andrej Karpathy, the OpenAI co-founder and former Tesla AI head, is on the cap table alongside former Disney chief executive Michael Eisner, the Nintendo family, and gaming investor Moritz Baier-Lentz. \The mix is media, gaming, and infrastructure rather than pure software engineering. It maps to the constituencies whose use cases the company says its products are now being deployed against. Decart sells three things on its own framing. DOS is the Decart Optimization Stack, an inference and training platform that the company says runs across Nvidia GPUs, Google TPUs, and Amazon Trainium and delivers 1,600 tokens per second for agentic inference against an industry average it puts at around 200, plus full-HD video inference at up to 100 frames per second. Lucy is its 'world model for immersive experiences', responding to user input in under 30 milliseconds and now deployed across virtual try-on, live streaming, and dynamic in-video advertising. Oasis is the parallel product for physical AI, which the company has been pushing toward robotics and autonomous-systems customers since the original real-time Minecraft-style demo in October 2024 went viral. The Amazon partnership is the substantive commercial detail in the announcement. Decart describes itself as one of the first companies to deploy real-time AI models of this class and scale on AWS Trainium, with its Lucy2 model running on Trainium3. Nafea Bshara, vice-president of Amazon's Annapurna Labs, said in the statement that Lucy2 exceeds 80% Model FLOPS Utilisation, which, in his framing, means more of the chip's raw power is doing real, productive work. Decart's chief executive, Dean Leitersdorf, described world models as 'the key to moving AI from the virtual world into the physical world', arguing that language models 'fundamentally operate in text', and 'don't understand how the physical world behaves'. Decart's funding history makes the trajectory legible. The company closed a $32m Series A at a $500m valuation in December 2024, four months after its $21m seed; Fortune reported in August 2025 that Decart had raised $100m at a $3.1bn valuation. Today's round brings total funding above $450m. Leitersdorf and co-founder Moshe Shalev have been building the company since 2023; the Sequoia podcast appearance is the most accessible public articulation of their thesis that vertically integrated optimisation, rather than larger models, is the missing layer of the real-time AI stack. The investor mix follows a recognisable Nvidia-equity pattern. The chipmaker has, by TNW's running count, committed over $40bn of AI equity in 2026 alone, with most of the smaller positions following a model in which Nvidia takes a stake, the company signs a long-term GPU commitment, and some of the GPU revenue flows back to Nvidia as a return on the same equity. Decart, which says its DOS stack runs across all three major chip families, is a sharper test of that pattern: the company is, in equity terms, a Nvidia portfolio asset, and in customer terms also a public reference point for Amazon Trainium and Google TPU. The named end-markets are where the angel list points. Lucy is in production with retailers and streaming platforms for virtual try-on and dynamic advertising, on the company's own description; Oasis is being positioned for robotics and autonomous-vehicle simulation; the gaming and live-experience use cases sit underneath both. The wider category context, as the industry meets at events like Cannes and the conversation around AI-generated media moves from possibility to procurement decisions, is a market where real-time, low-latency inference at production scale is the bottleneck most existing video models have not solved. Decart did not disclose run-rate revenue, the new valuation, or the hyperscaler customers it describes as licensing DOS. The next public proof points are DOS 2.0, announced alongside the funding, plus Lucy 2.5 and Oasis 3, both expected in the coming weeks.

Decart raises $300M for its AI optimization software, world models - SiliconANGLE

Decart raises $300M for its AI optimization software, world models Artificial intelligence developer Decart.ai Inc. today announced that it has raised $300 million in funding at a nearly $4 billion valuation. Radical Ventures led the round with participation from Nvidia Corp., Adobe Ventures, Toyota Ventures and other institutional backers. They were joined by several prominent angel investors including OpenAI Group PBC co-founder Andrej Karpathy. Decart announced the raise alongside an update to DOS, one of its three software products. The platform helps AI developers speed up training and inference workloads. Decart says that it's generating "significant revenue" from DOS licensing agreements with cloud providers and AI labs. Typically, developers have to optimize a neural network for every single chip on which they plan to run it. The process can take months. Decart says that DOS compresses the workflow into a few weeks, which reduces the cost of optimizing a model for multiple chips. That cost reduction, in turn, makes it easier for companies to move their AI workloads across chips when requirements change. DOS 2.0, the new version of the platform that debuted today, enables AI agents to process over 1,600 tokens per second. That's eight times the industry average. Additionally, Decart says that the software enables world models to process up to 100 frames of high-definition video per second. DOS underpins the company's two other products: a pair of world models called Lucy and Oasis. Lucy takes a video stream as input and modifies the objects that the footage depicts in real-time. A department store, for example, could use it to power a smart mirror that lets shoppers virtually try on different clothes. An interior design firm, meanwhile, could record a video of a room and have Lucy generate different furniture combinations. Video models often have output quality limitations. One reason is that quality issues cascade across frames: a relatively minor mistake at the start of the video can lead to a significant decrease in visual fidelity later on. According to Decart, Lucy can catch such issues early and fix them automatically. The company equipped Lucy with that ability by showing the model its "own imperfect outputs" during training. According to Decart, another use case that the AI supports is generating training data for robots. An online retailer working to automate its warehouses could generate videos that show a robotic arm picking up different packages. Lucy can automatically generate a large number of parcel variations. Robot developers with more advanced requirements can combine the algorithm with Oasis, Decart's second world model. The latter AI is designed to generate three-dimensional environments such as simulated logistics warehouses.