Will AWS buy TPUs from Google? Breaking down the AI chip supply chain war
Dec 3, 2025
Key Points
- AWS launches Tranium 3, a custom AI chip claiming 4x faster performance and 50% cost savings versus GPUs, positioning the company to challenge Nvidia's dominance in AI infrastructure.
- Nvidia's profit margin exploded from 16% to 56% between 2023 and 2025, spurring hyperscalers and AI labs to diversify chip suppliers through deals with Google, AMD, and Broadcom.
- AWS remains strategically unaligned with either the Google or Nvidia coalition, instead offering customers choice by selling competitors' chips alongside its own and launching high-speed links to Google Cloud.
Summary
AWS is muscling into Nvidia's territory with Tranium 3, a custom AI chip that the company claims is 4x faster than its previous generation and can reduce the cost of training and operating AI models by up to 50% compared with equivalent GPU systems. The chip is produced by Annaperna Labs, which Amazon acquired a decade ago for $350 million. What looked like a small bet at the time has paid off as custom silicon became critical to competing in AI infrastructure.
When asked directly whether AWS would purchase Google's chips, CEO Matt Garman dodged the question and said AWS is excited about Tranium's potential and believes in optimizing every layer of that stack. The case for why AWS might eventually buy from a rival rests on Bezos's customer-obsessed philosophy, not competitor obsession. Amazon already sells Nvidia GPUs alongside its own chips. If a customer wants TPU access through AWS infrastructure, the company has shown willingness to provide it. AWS just announced a separate partnership with Google Cloud that lets customers establish private high-speed links between the two platforms in minutes instead of weeks.
The broader dynamic reflects an emerging coalition against Nvidia's dominance. Nvidia's revenue grew from $27 billion in 2023 to $60 billion in 2024 to $130 billion in 2025, while its net profit margin expanded from 16% to 56%. That $56 billion in annual profit creates an obvious incentive for hyperscalers and AI labs to diversify their chip suppliers and commoditize the accelerator market. Meta is in talks with Google for billions in TPUs. OpenAI has struck deals with AMD and Broadcom. Google, Broadcom, Celestica, Lumenum, and TTM Technologies form what analysts call the Google complex. OpenAI aligns more with the Nvidia-SoftBank-Oracle-AMD-Microsoft-Coreweave coalition. Amazon remains strategically unaligned with either group, a major player that fits neatly into neither.
Decart, a $3.1 billion video generation startup, offers one concrete win for Tranium. The company's founders spent months trying to train its Lucy application for real-time video rendering on competing chips, including Nvidia processors, before getting early access to Tranium 3 from AWS. The breakthrough came quickly. That success story matters, but analyst Zephyr Post notes the broader competitive gap is actually widening. Tranium's 667 teraflops in BF-16 precision lags TPU V6E and falls further behind as Nvidia releases Blackwell and newer architectures like Ruben. Post expects Google's Ironwood chip to gain market share quickly on throughput and total cost of ownership, suggesting Tranium's niche advantage may be temporary.
AWS doesn't need to buy TPUs to compete. It needs to offer customers choice within its own platform. The high-speed Google Cloud link is the proof of concept. Tranium will likely thrive in specialized workloads like real-time video generation where the chip's design aligns with application needs, not as a general-purpose Nvidia replacement. Whether that narrow moat is enough depends on execution and developer adoption.