While Amazon is selling NVIDIA’s newest AI chips for its Amazon Web Services (AWS) AI infrastructure, it’s not letting go of its own silicon dreams either. In a span of a few weeks, Amazon showcased its dual-pronged chip strategy, using NVIDIA’s Blackwell architecture while simultaneously supporting its in-house Trainium and Graviton chips to challenge NVIDIA’s dominance in AI.
Amazon is acting as both the supplier and the substitute, fulfilling AI demand for customers while quietly chipping away at NVIDIA’s share of the future.
By training Claude Opus 4 entirely on its own chips, Amazon is proving the AI race isn’t exclusive to NVIDIA’s hardware.
A Growing Appetite for In-House Chips
Amazon’s chip ambitions aren’t new, but they are becoming far more public. As per a CNBC report, AWS is upgrading its Graviton4 CPU with 600 Gbps network bandwidth. The speed has been described by Ali Saidi, a distinguished engineer at AWS, as “fast enough to read 100 music CDs per second”.
However, that chip, built by Annapurna Labs in Texas, isn’t the main act. The key focus is on AI-centric Trainium2, which powers Project Rainier, a purpose-built supercomputer for Anthropic.
The Trainium2 chips were used to launch Claude Opus 4, the AI model from Anthropic, bypassing NVIDIA entirely. Gadi Hutt, senior director for customer and product engineering at AWS, reportedly said that Trainium2 is positioned to offer a cost-effective alternative to NVIDIA’s more expensive GPUs. While NVIDIA’s Blackwell may beat Trainium2 on raw performance, Hutt argued that AWS’ silicon wins on energy and cost efficiency. A third-generation chip, Trainium3, is already on the horizon, offering double the performance and 50% lower energy consumption.
Amazon appears to be investing more deeply in its silicon supply chain. Citing a Morgan Stanley report, an X user stated that Alchip is expected to be the exclusive foundry partner for AWS Trainium3 and possibly for Trainium4 as well.
Alchip’s 3nm chips, taped out successfully in May, are already set for mass production in 2026, with up to 6.5 lakh units forecasted. Notably, semiconductor solution provider Marvell has exited the high-stakes foundry race, leaving Alchip in near-total control of Trainium’s production roadmap. Alchip is also reportedly collaborating with Astera Labs on a potential 2nm Trainium4 chip, with decisions expected later this month.
While AWS is building small-scale lines for supplementary production, Alchip continues to hold nearly all turnkey foundry orders. Morgan Stanley expects this supply chain alignment to drive significant growth for Alchip, with estimated unit pricing of $3,000 per chip, excluding memory.
NVIDIA’s Blackwell Gets a Seat at the Table
Despite its in-house ambitions, AWS isn’t turning NVIDIA away. In fact, it’s offering customers the most powerful NVIDIA setup yet.
Recently, AWS announced the general availability of P6e-GB200 UltraServers, powered by NVIDIA’s Grace Blackwell Superchips. These instances feature 72 interconnected GPUs, delivering 360 petaflops of FP8 compute and over 13 TB of memory, ideal for trillion-parameter models and reasoning-intensive AI workloads.
For those with more modest needs or x86-specific software, AWS is also offering P6-B200 instances with eight Blackwell GPUs and Intel Xeon processors. Both configurations integrate tightly with AWS-managed services like SageMaker HyperPod and Amazon EKS, and are also available through NVIDIA’s DGX Cloud.
While the technology is cutting-edge, AWS is leaning heavily on its infrastructure chops. From its Nitro security architecture to new cooling systems and ultra-cluster networks, the goal seems to be to enable customers to scale without compromise.
“It’s about continuous innovation across multiple layers of infrastructure,” the blog post stated.
The Tightrope Strategy
What emerges is a hybrid approach. Amazon is eager to reduce reliance on NVIDIA for strategic control, pricing power, and platform lock-in reasons. However, it’s also pragmatic: NVIDIA remains essential to meeting today’s customer demands for the largest, most advanced AI models.
As more workloads shift to Trainium and Graviton, especially internal ones like Claude training, AWS builds its case for long-term silicon independence.
Yet, by retailing NVIDIA’s most advanced chips, Amazon is ensuring it stays relevant even if customers aren’t ready to switch.
With Alchip securing long-term foundry deals and AWS accelerating its silicon roadmap, the groundwork for a post-NVIDIA cloud infrastructure may be quietly falling into place.
The post Amazon is Selling NVIDIA’s Chips While Preparing to Replace It appeared first on Analytics India Magazine.
