header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP

Amazon CEO: Chip Business Standalone Valued at $50 Billion, Trainium Gen4 Fully Booked

According to Omdia Beating monitoring, Amazon CEO Andy Jassy spent a considerable amount of time during the first-quarter earnings call discussing in-house chip development. He said that if the chip business were to be spun off and operated independently, selling chips to external parties like other chip companies, the annual revenue would reach $500 billion. According to him, Amazon's in-house chip business has now entered the top three in global data center chips, a feat he described as "unprecedented in speed."

Supply and demand status of the fourth-generation Trainium (Amazon's in-house AI training chip): Trainium2 is sold out; Trainium3 just started shipping early this year, with a 30% to 40% higher cost-performance ratio than Trainium2, almost fully booked; Trainium4 is about 18 months away from a large-scale launch, with more than half of its capacity already pre-booked. Jassy stated that Trainium has accumulated over $225 billion in revenue commitments from multi-year multi-billion-watt training contracts with Anthropic, OpenAI, and Uber, among other enterprise customers.

Jassy did some math: By extensively using Trainium for inference, annual savings of several billion dollars in CapEx compared to relying on external chips can be achieved, with an operational profit margin advantage of hundreds of basis points. Most of Amazon Bedrock (AWS's model inference service, used by over 125,000 customers, with nearly 80% of Fortune 100 companies using it) has been running on Trainium.

When asked if Trainium racks would be sold directly, Jassy mentioned the need to balance between meeting existing customer demands and external sales, but he said, "I expect we'll likely start selling in the coming years." AWS has a backlog of orders worth $364 billion, not yet including the billion-dollar deal announced by Anthropic last week.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish