Table of Contents
Amazon is making waves with Trainium3, but already attention is moving to Trainium4
In sum – What we know:
- Trainium4 is on the way – Amazon has just announced Trainium3, but already it’s readying the release of its next-gen AI chip, Trainium4.
- Optimized for speed – Trainium4 offers a 6x performance boost at FP4 precision, and a 3x performance boost at FP8, relative to Trainium3.
- Interoperability is key – The chip is designed to be interoperable with chips from others, like Nvidia, making it easier to adopt.
Amazon Web Services used its re:Invent 2025 conference to unveil a wave of AI infrastructure announcements, including the debut of its Trainium3 chip and the broader AI Factories initiative. But tucked into the news cycle was a tease of Trainium4, the next generation of AWS’s custom AI silicon.
The preview signals AWS’s intent to iterate quickly in the accelerator market, even as Trainium3 has only just arrived. Here’s what we know about Trainium4 so far.
Technical details
AWS has shared limited specifics about Trainium4’s architecture, but the performance targets offer some insight into its ambitions. Notably, Amazon has said that the chip will offer a 6x performance improvement at FP4 precision and a 3x performance boost at FP8 over Trainium3, suggesting that it’s being optimized for inference workloads where lower-precision formats can accelerate throughput without sacrificing accuracy.
Trainium4 will also get memory improvements. Amazon says it will offer 4x the memory bandwidth of Trainium3, with 2x the memory capacity. That should equate to around 288 GB of memory capacity, and around 19.6 TB/s of memory bandwidth, based on Trainium3 specs.
Trainium4 will be designed to be easy to adopt, through interoperability. Specifically, Amazon has said that the chip will support Nvidia’s NVLink Fusion interconnect. Given Amazon’s big push into AI Factories, which are essentially on-prem AWS data-centers, this could be very valuable.
A multi-prong strategy
The Trainium4 preview is part of a broader push by AWS to establish its own silicon as a credible alternative to third-party accelerators. At re:Invent, the company announced that its AI Factories would combine Trainium chips with NVIDIA accelerated computing platforms, positioning the two as complementary rather than mutually exclusive.
That dual approach reflects AWS’s broader strategy to offer customers the flexibility to choose between proprietary and third-party hardware, while steadily improving its own chips to capture a larger share of AI workloads over time. For AWS, custom silicon also offers margin advantages. By reducing reliance on external suppliers, the company can better control costs and pricing for AI-intensive services, a key consideration as demand for inference and training capacity continues to surge.
What remains unknown
Despite the early tease, much about Trainium4 remains unclear. AWS has not disclosed specific architectural details, such as the process node, or supported model types. The exact release timeline is also unconfirmed, though the announcement suggests the chip is far enough along in development to warrant a public preview.
Pricing structures have not been discussed, nor have detailed performance benchmarks beyond the comparison to Trainium3. It’s also unclear how Trainium4 will be positioned relative to NVIDIA’s latest accelerators, which continue to dominate the AI infrastructure market.
What is clear is that AWS views custom AI silicon as a long-term strategic priority. With Trainium4 already on the horizon, the company is signaling that it intends to keep pace with, or outpace, the rapid evolution of AI hardware across the industry.
