Amazon launches the silicon challenge: Trainium3 arrives. 3nm efficiency to deflate the AI ​​cost bubble.

While the world scrambles to get its hands on Nvidia's incredibly expensive GPUs and investors anxiously scrutinize Big Tech's balance sheets weighed down by capital expenditures (Capex), Amazon is playing its biggest card. It's not a new language model, but the "engine" under the hood: AWS Trainium 3. It's 3, the size in nanometers; 3 is better than 4…

Presented at re:Invent, the new proprietary 3-nanometer chip isn't just an exercise in engineering style, but a profoundly economic and strategic move. The goal? Drastically reduce the costs of training and inference for Artificial Intelligence, democratizing access to resources that are currently the prerogative of a select few.

Amazon's Response to the "Chip Wars"

Let's be honest: the current AI development model, based on increasingly large and expensive GPU clusters, is becoming unsustainable for many. Amazon knows this and, with its trademark pragmatism, has decided to build its own infrastructure in-house.

Trainium3 is the heart of the newTrn3 UltraServers . These aren't just upgrades, but a generational leap designed for scalable efficiency. Here are the key data, stripped down from marketing, that 's of interest to anyone looking at productivity:

• Computing power: Up to 4.4 times faster than the previous generation (Trainium2).
• Energy efficiency: A 40% improvement . In an era where data centers consume as much energy as small nations, this is the figure that makes CFOs smile.
• Density: Each server hosts up to 144 Trainium3 chips, creating an integrated system capable of handling massive models.
• Networking: Reduced latency below 10 microseconds between chips, essential for Mixture-of-Experts (MoE) designs that require instantaneous data flow.

The economic impact: lower costs, more competition

The real news isn't so much in the FLOPS, but in the wallet. Customers like Anthropic (the creators of Claude, a rival to ChatGPT), but also emerging companies like Decart and Splash Music , are already using Trainium, reducing costs by up to 50% .

This creates an interesting ripple effect:

1. Independence from Nvidia: Amazon reduces its (and its customers') dependence on the de facto monopolist supplier, Nvidia, improving margins.
2. Lowering barriers to entry: If inference costs half as much, new business models (e.g., real-time generative video) emerge that were previously financially loss-making.
3. Infrastructure for everyone: With EC2 UltraClusters , AWS can connect up to a million chips. This is the firepower needed to train the foundation models of the future, no longer reserved only for those with unlimited budgets.

A Look Ahead: Trainium4 and the Openness to Nvidia

In a move we might call "technological realpolitik," Amazon has already previewed Trainium4 . The surprise? It will support Nvidia's NVLink Fusion technology . This means that Amazon, while competing on silicon, will allow its own chips to be mixed with Nvidia's in the racks. A hybrid and pragmatic approach: if you can't beat them entirely, make them work better on your platform.

Questions and Answers

What's new about Amazon's chip for AWS compared to the market? The main innovation is the 3nm manufacturing process , which guarantees superior density and energy efficiency. Unlike general-purpose GPUs (like those from Nvidia), Trainium3 is an ASIC, meaning a chip custom-designed specifically for AI. This eliminates unnecessary "fat" in the architecture, offering targeted performance (4.4x performance increase compared to the previous model) and, most importantly, halving operating costs for those who need to train large models.

Can it challenge Gemini and ChatGPT? Not directly, but indirectly, it's a thorn in their side. Gemini runs on Google's proprietary chips (TPU), while ChatGPT relies primarily on Microsoft/Nvidia infrastructure. Trainium3, however, is the weapon Amazon provides to Gemini and ChatGPT rivals (like Anthropic/Claude). By dramatically lowering training and inference costs for these competitors, Amazon fuels competition, forcing Google and OpenAI to stop resting on their current margins.

Is this an ASIC that follows the one Google made and sold to Meta? Yes, conceptually it's the same path. Google has had its own TPUs (Tensor Processing Units) for years, and Meta is developing its own MTIA chips. Trainium3 is Amazon's ASIC ( Application Specific Integrated Circuit ). The difference is that while Google uses TPUs primarily for its own use (and rents them to the cloud), Amazon is aggressively pushing Trainium as the market standard for all its AWS customers, seeking to break the de facto monopoly of Nvidia GPUs in the cloud.

The article Amazon launches the silicon challenge: Trainium3 arrives. 3nm efficiency to deflate the AI ​​cost bubble comes from Scenari Economici .

---

This is a machine translation of a post published on Scenari Economici at the URL https://scenarieconomici.it/amazon-lancia-la-sfida-del-silicio-arriva-trainium3-efficienza-a-3nm-per-sgonfiare-la-bolla-dei-costi-ai/ on Wed, 03 Dec 2025 12:00:11 +0000.