Amazon is in talks to sell its in-house AI chips, AWS Trainium, for use in other companies' data centers, a move framed as an expansion of its effort to cut into Nvidia's dominance of the AI semiconductor market.
AWS · Custom Silicon
Amazon May Sell Its Trainium AI Chips Outside Its Own Cloud
AWS is in talks to deliver full racks of its in-house Trainium accelerators directly to other companies' data centers — sharpening its challenge to Nvidia's grip on AI chips.
$20B+
Annual scale of Amazon's in-house silicon unit
500K
Non-Nvidia chips in Anthropic's Rainier cluster
~30%
Better price-performance vs. Nvidia GPUs (Trainium2)
Trainium3 vs. Trainium2 — MXFP8 throughput
The new generation doubles low-precision compute per chip.
Three generations of Trainium
Trainium2
Sold out
~30% better price-performance vs. Nvidia GPUs
Trainium3
Early 2026
2× MXFP8 throughput · 144GB HBM3e @ 4.9TB/s · ~362 PFLOPs per UltraServer (up to 144 chips) · MoE support
Next-gen (T4-class)
2027
Drawing strong interest from customers
Early customer reception
50%
SplashMusic cut training time & cost
40%
Poolside expected cost savings on future training
Databricks
Adopted for training & serving its own models
TRAINIUM'S EDGE
Advantages in total cost of ownership and memory bandwidth. Neuron SDK runs PyTorch, vLLM and Hugging Face code largely unchanged — lowering the migration barrier.
THE COUNTERPOINT
Trainium trails Nvidia on some performance metrics, and Nvidia's CEO argues rival platforms have not demonstrated superiority.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…