Breaking

AWS Custom AI Chips Gain Enterprise Adoption as Cheaper Nvidia Alternative for Inference

June 17, 2026 at 16:02 EDT

Infra & Chips
Sales & Business
Foundation Models

Custom AI chips designed by Amazon's AWS — Inferentia and Trainium — are gaining enterprise adoption among companies seeking cheaper inference without relying entirely on Nvidia. Marketed on price-performance, they are increasingly positioned as an alternative to Nvidia GPUs for generative AI workloads.

AWS Custom Silicon · The Nvidia Alternative

Amazon's Own AI Chips Start Winning Over Developers

As Nvidia GPUs stay scarce and pricey, AWS Trainium and Inferentia are gaining ground with firms chasing cheaper inference — claiming up to 80% lower cost than running on Nvidia H100.

80%

Cheaper inference vs Nvidia H100, developers report

4x+

Trainium3 compute vs its predecessor

144

Chips bundled per Trainium3 UltraServer

40%

Lower power draw than the previous generation

Customer cost cuts switching off GPUs

Reported savings on AWS inference chips — taller column means deeper cut.

80%

Leonardo.ai
Inferentia2

80%

Finch Computing
Inf1 vs GPU

83%

Tomofun
BLIP inference

Dataminr
throughput / $

Inferentia · Inference-focused

Up to 4x throughput vs gen 1
Up to 10x lower latency (Inf2)
Up to 70% lower cost vs gen 1

Trainium3 · Train + Inference

Up to 144 chips per UltraServer
30-50% better price-performance vs Nvidia
HBM3e + NeuronLink/EFA for large clusters

Why firms are switching

Scarce, costly Nvidia GPUs push cost- and availability-conscious teams toward AWS silicon — with native PyTorch/TensorFlow support via the Neuron SDK and flexible pricing.

The catch

Nvidia keeps an edge through its mature CUDA ecosystem and tuning for certain workloads. Amazon's standing will be tested as inference demand grows.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.