Another challenger! Amazon (AMZN.US) brought Trainium3 to join the AI chip Three Kingdoms. Citi: Compatible with Nvidia's strategy is very flexible

The Zhitong Finance App learned that soon after Google announced the expansion of overseas sales of self-developed TPU chips, Amazon (AMZN.US) joined the battle team with its own Trainium 3 chip. At the Amazon Web Technology Conference held recently, Amazon released two key developments around the Trainium chip family — the Trainium 3 chip has achieved “official full use”, and also announced the Trainium 4 chip with better performance. Both have achieved significant breakthroughs in computing power, energy efficiency, and compatibility, directly targeting the core requirements for large-scale implementation of generative AI.

Amazon's move is thought to be another giant trying to challenge Nvidia's GPU after Google. Citi pointed out in a research report released later that in the context where both Microsoft and Google are speeding up the layout of self-developed AI chips, iterations of the Trainium series have helped Amazon maintain its lead in the “self-developed computing power ecosystem.”

Part.01 Trainium 3: Commercially Used “Computing Power Multiplier”

As the main product of the current Trainium family, the core advantages of the Trainium 3 chip focus on the two dimensions of “performance improvement” and “cost optimization”. The specific parameters compared to Trainium 2 are as follows:

Computing power: 4.4 times higher than Trainium 2, which can support the efficient operation of more complex generative AI models (such as big language model inference and multi-modal processing).

Energy efficiency: The energy efficiency ratio has been increased by 4 times, which means that with the same computing power output, the customer's energy consumption costs can be reduced by 75%, which is in line with the core demand of enterprises to “reduce costs and increase efficiency” of AI deployments.

Memory bandwidth: The memory bandwidth has been increased by nearly 4 times, effectively solving the bottleneck of data transmission in large models and reducing delays in model training and inference.

Commercial progress: It is now officially fully usable. Customers can directly access it through Amazon Cloud Services without additional hardware infrastructure.

Part.02 Trainium4: Compatible with Nvidia interconnect technology

Amazon simultaneously revealed the development progress of the Trainium4 chip. It is expected to become the next generation AI computing power core. The key expected performance indicators are as follows:

Performance: The performance is expected to reach 6 times that of Trainium 3, which can support training and inference of very large parametric models (such as trillion-level models with large parameters).

Memory configuration: The memory bandwidth is increased by 4 times and the memory capacity is doubled, further breaking the high requirements of large models for storage and data transmission.

Ecological compatibility: Specially designed to support “Nvidia NVLink Fusion chip interconnection technology”. This compatibility means that Trainium4 can form collaborative computing power with Nvidia GPUs to meet the customer's “hybrid architecture deployment” needs and avoid the limitations of a single chip ecosystem.

Notably, before introducing the Trainium series of chips, the CEO of Amazon Web Technology specifically emphasized the close partnership with Nvidia — this statement was viewed by Citi as an important sign of its chip strategy: not seeking “complete replacement,” but rather providing customers with more flexible computing power choices through “self-developed chip+ecological collaboration.”

Part.03 Trainium family deployment exceeds one million

In addition to the release of the new chip, Amazon also revealed the overall deployment and production capacity of the Trainium family. The data shows that it has formed the dual advantage of “large-scale implementation+rapid expansion of production”, laying the hardware foundation to meet the needs of generative AI.

Deployment scale: Over 1 million chips have been launched to build a huge computing power network

Up to now, Amazon has deployed more than 1 million Trainium chips in data centers around the world. These chips are widely used in customers' AI model training, inference, and cloud-native computing scenarios, making them currently publicly owned.

Production capacity climbing: Trainium 2 expands at a record rate

As the predecessor of Trainium 3, Trainium2's production capacity climbed significantly faster than all previous AI chips. Citi pointed out in the report that Trainium2's production capacity expansion rate is 4 times that of Amazon's previous AI chips. This efficiency means that it can quickly meet customer needs for middle and high-end AI computing power and avoid business delays caused by hardware shortages.

Looking at the overall pace, the Trainium family has formed a “Trainium 2 base (meeting low to medium computing power requirements), Trainium 3 main force (supporting large-scale AI deployment), and Trainium 4 outlook (targeting future high computing power scenarios)”, covering the hierarchical computing power needs of different customers.

Part.04 Attaching great importance to Trainium chip iteration

Combining the progress and overall business of the Trainium series chips, Citi clearly stated in the report that the technical breakthrough and large-scale deployment of Trainium chips is one of the core supporting factors for Amazon to achieve 23% year-on-year revenue growth in 2026 and maintain 20% + growth expectations until 2027. The specific logic includes three points:

Reduce customer AI deployment costs

The high energy efficiency ratio of Trainium 3 and the large-scale deployment of Trainium 2 can directly reduce customers' AI computing power costs — Citi believes this will attract more small and medium-sized enterprises and traditional industry customers to shift generative AI projects from “proof of concept” to “commercial implementation”, which in turn will drive the growth in demand for AWS core cloud services.

Make up for shortcomings in computing power infrastructure

The number of generative AI proof of concept projects was high in 2025, but some customers were unable to implement them on a large scale due to “insufficient computing power” or “excessive cost.” The commercialization of Trainium 3 and the preview of Trainium 4 mean that Amazon will provide a more adequate and cost-effective computing power supply in 2026, which can effectively handle this backlog of demand and become a new engine for revenue growth.

Consolidate competitive advantage in the cloud market

In the context of Microsoft Azure and Google Cloud both accelerating the layout of self-developed AI chips, the iteration of the Trainium series helped AWS maintain its lead in the “self-developed computing power ecosystem” — Citi Analysts believes that the performance advantages and ecological compatibility of Trainium chips (such as supporting Nvidia technology) will enhance customer stickiness to AWS and further consolidate its leading position in the global cloud market.