A two-year-old Israeli unicorn is quietly stress-testing Amazon’s AI future
Decart’s real-time video model is running four times faster on Amazon’s new Trainium chip, turning the young startup into one of the tech giant’s largest AI customers.
While much of the tech world is preoccupied with whether Google’s TPU chips can meaningfully challenge Nvidia’s dominance in AI computing, Amazon quietly escalated the chip war last week. The launch of its new Trainium 3 chip didn’t generate the same buzz as Google’s recently unveiled TPU, used to train Gemini 3, nor did it boost Amazon’s stock. But for Israel, the announcement marks a uniquely significant moment.
For one, Trainium 3 is built on technology Amazon acquired a decade ago from Avigdor Willenz’s Annapurna Labs for what is now widely considered a bargain: $350 million. A decade later, Annapurna’s leadership, headed by Nafea Bshara, continues to spearhead Amazon’s AI chip development as the company attempts to stand toe-to-toe with Nvidia and Google.
But the more surprising development came from Israel’s rising AI star, Decart, one of the first companies to show performance gains running its model on Amazon’s chips instead of Nvidia’s.
Decart, founded two years ago by Dean Leitersdorf and Moshe Shalev, has built a real-time video-generation model that has propelled the company to “wonder-child” status, currently the only Israeli startup regularly appearing on global lists of the most promising AI model companies. The company has raised $156 million across three rounds, most recently at a $3.1 billion valuation.
Until recently, Decart ran entirely on Nvidia GPUs. But last summer, after an Amazon representative approached Leitersdorf at a conference, the company agreed to test Trainium. The results exceeded expectations. Weeks later, Leitersdorf was invited to demo the system at Amazon’s re:Invent conference, a stage typically reserved for industry giants, not two-year-old startups.
“The 26 days from Amazon’s invitation to going on stage were the busiest of my life,” Leitersdorf told Calcalist. “Everyone else gets a year to prepare; presentations are locked months in advance. We started 12 days before. At first, we thought there was no way we’d make it, but we did.”
He added that Amazon rejected anything that worked “only 99.9% of the time.” The live demo included 16 backup systems, covering scenarios as extreme as “the entire Amazon server farm going down,” with audiences unable to detect even a tenth-second delay.
Leitersdorf ultimately delivered a five-minute demo to millions of viewers, showcasing real-time video generation that transformed his on-screen appearance in an instant.
Decart’s technology is already gaining traction in gaming, where dynamic video generation creates uniquely personalized experiences. But Leitersdorf sees even broader commercial applications, particularly in robotics, enabling autonomous systems to train on infinite real-world scenarios, and in e-commerce, where users could try on clothing from home via full real-time video.
According to Leitersdorf, Decart has now become Amazon’s second-largest AI-chip customer, following Anthropic. Industry observers say that while Trainium’s hardware performance is strong, faster and more energy-efficient than Nvidia’s in some tasks, Amazon still trails Nvidia in software due to the massive advantage of CUDA.
For Decart, however, Amazon’s chips delivered an unexpected leap. “Our goal is to run video as fast as possible and at lower cost. Customers don’t care what chip we use,” Leitersdorf said. “Right now, Amazon lets us significantly reduce costs, and our video runs four times faster. The max we hit on GPU was 26 frames per second; on Trainium we’re approaching 101 frames per second. To the human eye, that’s barely visible, but in robotics or autonomous vehicles, it can be the difference between safety and an accident.”
Most of the AI world is working on Nvidia’s chips today, and that dominance is reflected in both the company’s business performance and its stock price. But Leitersdorf believes the future lies in tailoring chips to specific tasks. “I wouldn’t eulogize anyone yet, nor would I crown anyone a winner, but something is definitely happening,” he says. “The field is moving fast, and all we can say today is that the world will understand much more in three months.”
According to him, three months is now the longest meaningful forecasting window in the AI industry; looking a year ahead is already impossible. “Amazon’s stock didn’t rise after the Trainium announcement, but the market may react in a few months. Google, on the other hand, has already surpassed Microsoft in value, and that says a lot.”
