Amazon (AWS) is in talks to sell its in-house AI accelerator, AWS Trainium, directly to other companies' data centers, it emerged on June 18, 2026. The move positions the chip as a cheaper alternative in an AI compute market dominated almost entirely by Nvidia.
June 18, 2026 · Amazon Web Services
Amazon Moves to Sell Its Trainium AI Chips — Taking the Fight Directly to Nvidia
In early-stage talks to supply its in-house training accelerator to outside data centers, Amazon would convert a cost-cutting internal tool into a standalone challenger to Nvidia's dominance — backed by a chip business approaching a $50B annual run rate.
~$50B
Estimated annual run rate of the chip business as a standalone unit
40%
Lower power for equivalent work on Trainium3 (3nm)
30–40%
Claimed price-performance lead over GPU-based instances
Reported cost savings vs GPU workloads
Customer & case-study figures — taller column means greater reported reduction
40%
Poolside
50%
SplashMusic
80%
Reported peak
Trainium2 → Trainium3: generation jump
Trainium2
96GB HBM memory
Trainium3
144GB HBM3e · 3nm
The Opportunity
OpenAI, Anthropic, Uber & Databricks already use Trainium via AWS
Strong, low-cost option for cost-focused training workloads
Diversifies revenue and expands market share against Nvidia
The Challenges
Neuron SDK has a steep learning curve vs CUDA
Gaps in library compatibility and ecosystem maturity
Can lag GPUs in peak performance and flexibility
Where it stands
Talks are early — focused on training-oriented Trainium and international customers with localized data center demand. No customers, pricing, release timing, or deal terms have been finalized.
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…