Amazon is in talks to sell its in-house Trainium AI chips directly to third-party companies, not only through AWS. If realized, the move would mount a direct challenge to Nvidia, which dominates the market for AI training chips.
June 18, 2026 · AWS
AWS Weighs Selling Its Trainium AI Chip Direct — A Shot at Nvidia
Amazon is exploring selling its in-house Trainium training chip to third parties, not just through its cloud. Talks are early — no customers or timing named — but standalone, the chip business would be a ~$50B-a-year operation.
~$50B
Annual revenue of the chip business, standalone basis
4.4×
Trainium3 peak performance vs. prior generation
144GB
HBM3e memory per chip · 4.9TB/s bandwidth
Cost: the biggest draw vs. Nvidia H100
Reported Trainium2 savings against comparable H100 setups — cheaper is the whole pitch.
Trainium2 · training
~60%
~40% cheaper
Trainium2 · inference
~20%
up to 80% cheaper
The case for Trainium
Strong cost edge for cost-sensitive workloads
Used at scale by Anthropic for Claude; OpenAI, Uber, Databricks, Poolside also on board
Trainium2 fully subscribed — demand outstrips supply
40% lower power, Neuron SDK runs PyTorch / vLLM with little change
The caution flags
Some startups report weaker performance, latency vs. H100
Tight access limits, service interruptions reported
Supply skewed to large customers; little for smaller users
Nvidia's CUDA ecosystem remains a high barrier
Roadmap & open questions
Trainium3
Late 2025 · up to 144 chips, ~362 PFLOPs
→
Direct sales?
Early-stage talks for third parties
→
Trainium4 · ~2027
Nvidia NVLink Fusion support studied
A common bottleneck remains: manufacturing reliance on TSMC. Real-world evaluations after any direct sale have yet to appear.
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…