Amazon Web Services has scored a major win for its custom AWS Trainium accelerators after striking a deal with AI video startup Decart, with the partnership seeing Decart optimize its flagship Lucy model on AWS Trainium3 to support real-time video generation CNBC. Decart is achieving 4x faster inference for real-time generative video at half the cost of GPUs OpenAI, demonstrating that custom AI accelerators can challenge NVIDIA's dominance in computationally intensive generative AI applications.

Decart is essentially going all-in on AWS, making its models available through the Amazon Bedrock platform, allowing developers to integrate real-time video generation capabilities into almost any cloud application without worrying about underlying infrastructure CNBC. The company has obtained early access to the newly announced Trainium3 processor, capable of outputs of up to 100 fps and lower latency CNBC.

Lucy has a time-to-first-frame of 40ms, meaning it begins generating video almost instantly after prompt, and by streamlining video processing on Trainium, can match the quality of much slower, more established video models like OpenAI's Sora 2 and Google's Veo-3, generating output at up to 30 fps CNBC. By running Lucy on Trainium3, Decart hopes to improve current 30 fps outputs and generate live video at up to 100 FPS while reducing time-to-first frame to less than 40 milliseconds Thriveholdings.

Trainium3 UltraServers deliver up to 4.4x more compute performance, 4x greater energy efficiency, and almost 4x more memory bandwidth than Trainium2 UltraServers, with systems scaling up to 144 Trainium3 chips delivering up to 362 FP8 PFLOPs OpenAI. Built on 3-nanometer technology, each UltraServer delivers 362 FP8 PFLOPs with up to 20.7 TB of HBM3e memory, enabling massive models to train in weeks instead of months Yahoo Finance.

The partnership reflects broader industry movement toward custom AI accelerators as alternatives to NVIDIA GPUs. AI coding startup Poolside is using AWS Trainium2 to train its models with plans to use its infrastructure for inference as well, while Anthropic is hedging its bets by training future Claude models on a cluster of up to one million Google TPUs, and Meta Platforms is reportedly collaborating with Broadcom to develop custom AI processors CNBC. AWS claims Trainium and Google's TPUs offer 50-70% lower cost-per-billion-tokens compared to high-end NVIDIA H100 clusters Yahoo Finance.

Dean Leitersdorf, Decart co-founder and CEO, stated that Trainium3's next-generation architecture delivers higher throughput, lower latency, and greater memory efficiency, allowing the company to achieve up to 4x faster frame generation at half the cost of GPUs CNBC.

Leitersdorf emphasized that generative video is one of the most compute-intensive challenges in AI, and by combining Decart's real-time video models with AWS Trainium3, the partnership is making real-time video generation practical and cost-effective at scale Thriveholdings.

Anthropic's early adoption carries symbolic weight as Amazon holds an $8 billion stake in OpenAI's rival, yet chose Trainium for production workloads, with that endorsement signaling Trainium3 isn't experimental but production-ready and competitive with NVIDIA's flagship offerings Yahoo Finance. Yet NVIDIA's moat remains formidable, with CUDA becoming the industry standard for AI development, and switching to Trainium requiring rewriting code and retraining teams Yahoo Finance.

By generating high-fidelity AI video in real time, Decart says it can power use cases that simply weren't possible before, including live gaming where video clips can be incorporated into open-ended video games to generate environments based on player interactions, and social media applications where influencers can integrate AI video into live streams Thriveholdings.

For organizations spending millions monthly on AI infrastructure, Trainium3's economics are transformational, with the chip delivering over 5x more output tokens per megawatt than previous generations, directly slashing data-center power bills Yahoo Finance. Enterprises evaluating AI infrastructure strategies now face credible alternatives to NVIDIA-exclusive architectures, potentially reducing vendor lock-in risks. Amazon acknowledges reality by announcing Trainium4 will support NVIDIA's NVLink Fusion interconnect technology, enabling mixed deployments within the same racks Yahoo Finance.

The real question isn't whether Amazon can match NVIDIA's raw performance as Trainium3 already does, but whether cost and energy efficiency alone reshape a $50 billion+ AI chip market, or whether ecosystem lock-in and customer inertia keep NVIDIA entrenched Yahoo Finance. Decision-makers should monitor whether real-time video generation adoption validates custom accelerator economics across other computationally intensive AI applications. While ASICs aren't going to replace GPUs completely as flexibility of GPUs means they remain the only real option for general-purpose models, specialized workload optimization may fragment AI infrastructure markets CNBC.

Source & Date

Source: Artificial Intelligence News, AWS, Tech Startups, HPCwire, TechCrunch, Invezz
Date: December 3, 2025 (AWS re:Invent 2025