Amazon launches a major challenge to Google and Nvidia: AI chip Trainium 3 is faster and more energy-efficient, with four Nova 2 models, pioneering "open training"

On December 2nd, Eastern Time, Amazon's cloud computing division AWS launched the next-generation artificial intelligence (AI) training chip Trainium 3 at the annual cloud computing event re:Invent, previewing the development plan for the next-generation product Trainium 4, intensifying its challenge against NVIDIA and Google’s dominance in the AI chip market. At the same time, it introduced the Nova 2 series models and new AI services, attempting to capture more market share in the fierce AI competition.

AWS announced that the Trainium 3 chip has recently been deployed in several data centers and will be available for customer use starting from that Tuesday. AWS Vice President Dave Brown stated, "As we enter early next year, we will begin to scale very, very quickly."

The speed of this chip's launch exceeds industry norms. Amazon released Trainium 3 about a year after deploying the previous generation accelerator, matching NVIDIA's promised pace of releasing new chips annually.

Amazon aims to attract companies seeking cost-effectiveness. The company claims that Trainium chips can power the intensive computations behind AI models in a more affordable and efficient manner than NVIDIA's market-leading graphics processing units (GPUs). Brown stated, "We are very pleased with Trainium's ability to achieve the right price-performance ratio."

On Tuesday, Amazon not only launched four new models but also introduced the industry's first "open training" service Nova Forge and released an agent service Nova Act focused on browser task automation. The newly launched Nova 2 family covers multiple areas including inference, multimodal processing, conversational AI, and code generation, emphasizing price-performance advantages.

Following the announcement of the new AI chip, Amazon's stock price approached $239, hitting a daily high, with an intraday increase of nearly 2.2%. Meanwhile, NVIDIA's stock price, which had risen 3.2% at the beginning of the day, further narrowed its gains, refreshing its daily low by the end of the early trading session, nearly erasing all gains. Its competitor AMD saw its stock price decline further, dropping nearly 1.7% intraday. By the close, Amazon was up over 0.2%, NVIDIA rose nearly 0.9%, and AMD fell nearly 2.1%.

Significant Performance Leap of Trainium 3

AWS claims that its fourth-generation AI chip Trainium 3 is the first AWS AI chip built on a 3-nanometer (nm) process, designed to provide optimal token economics for next-generation intelligent agents, inference, and video generation applications.

According to official data from AWS, the Trn3 UltraServer system equipped with the Trainium 3 chip has achieved significant improvements in both training and inference performance. Compared to the second-generation product, the new system's speed has increased by over four times, and memory capacity has quadrupled Each Trainium 3 chip provides 25.2 quadrillion floating-point operations (PFLOPs) of FP8 computing power, with memory capacity increased by 1.5 times to 144GB HBM3e compared to Trainium 2, and memory bandwidth improved by 1.7 times to 4.9TB/s. A fully configured Trn3 UltraServer can accommodate 144 chips, achieving a total computing power of 362 PFLOPs, and providing up to 20.7TB of HBM3e memory and 706TB/s of aggregated memory bandwidth.

The energy efficiency improvement is also remarkable. AWS states that the energy efficiency of the Trn3 UltraServer is 40% better than the previous generation, with a performance-to-power ratio improvement of 4 times. On the Amazon Bedrock platform, Trainium 3 becomes AWS's fastest accelerator, with performance improved by 3 times compared to Trainium 2, the number of tokens output per megawatt increased by more than 5 times, while latency levels remain comparable.

Brown stated in an interview, "We are very satisfied with Trainium's performance in terms of price and performance." The system can also scale to hundreds of thousands of chips in EC2 UltraClusters 3.0, and AWS claims its goal is to provide 1 million chips to AI startup Anthropic by the end of the year.

Trainium 4 will support NVIDIA interconnect technology

Amazon also previewed the upcoming Trainium 4 chip at the conference. This next-generation product will bring a significant leap in performance and will support NVIDIA's NVLink Fusion high-speed chip interconnect technology.

This compatibility means that AWS systems equipped with Trainium 4 will be able to interoperate with NVIDIA GPUs and scale performance while still using Amazon's self-developed low-cost server rack technology. This could help AWS attract large AI applications developed on NVIDIA GPUs to migrate to the Amazon cloud platform.

It is worth noting that NVIDIA's CUDA architecture has become the de facto standard supported by all mainstream AI applications. The support for NVLink in Trainium 4 may lower the technical barriers for these applications to transition to AWS.

Amazon has not disclosed a specific timeline for the release of Trainium 4. Based on past release patterns, it is expected that more information will be available at next year's re:Invent conference.

Software ecosystem remains a shortcoming

Despite strong hardware performance, Amazon's chips still face ecosystem challenges. Compared to NVIDIA, AWS chips lack a deep software library that helps customers deploy quickly.

Kevin Peterson, Chief Technology Officer of construction equipment autonomous driving company Bedrock Robotics, stated that while the company runs infrastructure on AWS servers, it still uses NVIDIA chips when building models for guiding excavators. "We need it to be powerful and easy to use," Peterson said, "and that is NVIDIA."

The main customers currently using Trainium chips are Anthropic, with these chips distributed across data centers in Indiana, Mississippi, and Pennsylvania AWS stated earlier this year that it has linked over 500,000 chips together to help Anthropic train its latest model and plans to provide 1 million chips to this OpenAI competitor by the end of the year.

However, Anthropic is also using Google's TPU (Tensor Processing Unit) and reached an agreement with Google earlier this year to obtain computing resources worth tens of billions of dollars.

Aside from Anthropic, Amazon has revealed very few other major customers, making it difficult for analysts to assess the actual effectiveness of Trainium. AWS stated that customers such as Japan's LLM Karakuri, Splashmusic, and Decart are using the third-generation Trainium chips and systems, significantly reducing inference costs.

AWS promises to allow developers to train and deploy without changing any model code through the AWS Neuron SDK and native PyTorch integration, attempting to narrow the gap with NVIDIA's software ecosystem.

Four Nova 2 Models Each Focus on Different Aspects

Amazon has launched four Nova 2 models designed for different application scenarios. Nova 2 Lite is a fast, economical inference model designed for everyday workloads, capable of processing text, images, and videos while generating text. In benchmark comparisons with competitors, this model outperformed or matched Claude Haiku 4.5 in 13 out of 15 tests and outperformed or matched GPT-5 Mini in 11 out of 17 tests.

Nova 2 Pro is Amazon's most intelligent inference model, capable of processing text, images, videos, and speech while generating text, suitable for highly complex tasks such as agent coding and long-term planning. This model performed better or matched Claude Sonnet 4.5 in 10 out of 16 benchmark tests, better or matched GPT-5.1 in 8 out of 16 tests, and better or matched Gemini 2.5 Pro in 15 out of 19 tests.

Nova 2 Sonic is Amazon's voice-to-voice model, unifying text and speech understanding and generation capabilities, supporting real-time, human-like conversational AI. This model has a context window of 1 million tokens, supports extended multilingual capabilities, and can seamlessly integrate with phone service providers such as Amazon Connect, Vonage, and Twilio.

Nova 2 Omni is the industry's first unified multimodal inference and generation model, capable of processing text, images, video, and speech input while generating text and images. This model can simultaneously handle up to 750,000 words, hours of audio, long videos, and hundreds of pages of documents, analyzing an entire product catalog, customer reviews, brand guidelines, and video library at once

Nova Forge Launches "Open Training" Model

Nova Forge is a pioneering service launched by AWS that allows enterprises to build their own customized versions of the Nova model—referred to by Amazon as "Novellas." This service introduces the "open training" model, providing customers with exclusive access to pre-trained, mid-training, and post-training Nova model checkpoints, enabling them to mix proprietary data with Amazon's curated datasets at every stage of model training.

This service addresses three major challenges faced by enterprises when embedding proprietary knowledge into AI applications: limited integration when customizing proprietary models, lack of access to original training data when continuing to train open-source weight models, or the high costs associated with building models from scratch.

In addition to model checkpoints and data mixing capabilities, Nova Forge offers three core features: the ability to train AI using the customer's own environment (referred to as reinforcement learning "gyms"), options for creating smaller and faster models, and a responsible AI toolkit for implementing safety controls.

Reddit is using Nova Forge to improve its content moderation system. Reddit's Chief Technology Officer Chris Slowe stated, "We are replacing multiple different models with a single, more accurate solution, making moderation more efficient. The shift from multiple specialized machine learning workflows to a unified approach marks a transformation in how we implement and scale AI on Reddit."

Companies such as Booking.com, Cosine AI, Nimbus Therapeutics, Nomura Research Institute, OpenBabylon, and Sony are also using Nova Forge to build their own models. Models created using Nova Forge can be deployed on Amazon Bedrock, enjoying the same enterprise-level security, scalability, and data privacy protections as other Bedrock models.

Nova Act Achieves Breakthrough in Browser Task Automation

Nova Act is a new service launched by AWS for building and deploying highly reliable AI agents that can perform operations in web browsers. This service is powered by a customized Nova 2 Lite model, providing the fastest and easiest path to building and managing automated browser task agents. Nova Act has achieved 90% reliability in early customer workflows and outperformed competing models in relevant benchmark tests.

Nova Act achieves breakthrough reliability by training the customized Nova 2 Lite model through reinforcement learning, running thousands of tasks in hundreds of simulated web environments. This training approach allows Nova Act to excel in user interface-based workflows, such as updating data in customer relationship management systems, testing website functionalities, or submitting health insurance claims Developers can start prototyping in minutes using a no-code platform with natural language prompts, then refine the proxies in familiar integrated development environments like VS Code, and finally deploy to AWS. Content built and tested locally by customers can scale in production environments and gain comprehensive management tools and monitoring through the Nova Act AWS console.

Amazon introduced that several companies have noticed the effectiveness of Nova Act. Among them, the startup Sola Systems, after integrating Nova Act, automates hundreds of thousands of workflows for customers each month, covering key business tasks such as reconciliation payments, coordinating freight, and updating medical records.

1Password uses Nova Act to reduce the manual steps for users to access login information, allowing automatic work across hundreds of different websites with just a simple prompt. Hertz has increased its software delivery speed by 5 times and eliminated quality assurance bottlenecks by automating end-to-end testing of its rental platform with Nova Act, reducing work that previously took weeks to just hours