[AWS re:Invent] AWS Moves Toward 'Chip Independence'... Unveils High-Efficiency Semiconductor Chips in Succession

After 'Trainium3' AI Chip, AWS Unveils 'Graviton5' CPU
25% Performance Improvement Over Previous Model
Adopts Latest Process Technology, Enhanced Power Efficiency
Graviton Now Powers Half of AWS Cloud Services

Amazon Web Services (AWS) is accelerating its so-called 'chip independence' by continuously introducing its own semiconductor chips. The company aims to create synergy by adding proprietary chips with improved power efficiency (performance per watt) to its cloud computing services, while also reducing dependence on existing chip manufacturers.

Dave Brown, Vice President of Compute and Machine Learning (ML) Services at AWS, unveiled AWS's self-designed central processing unit (CPU) for servers, 'Graviton5,' during the keynote speech at AWS's annual flagship event, 're:Invent 2025,' held on the morning of the 4th (local time). Brown emphasized, "Many customers have switched to Graviton, reducing costs and improving performance," and added, "It has delivered levels of performance and efficiency that were impossible with general-purpose hardware."

On the 4th (local time), the annual largest event of Amazon Web Services (AWS), 're:Invent 2025,' featured the announcement of 'Graviton5,' a central processing unit (CPU) for servers. Provided by AWS

Graviton is AWS's first proprietary CPU, introduced in 2018. With Graviton, AWS challenged Intel and AMD, which had previously dominated the server CPU market. The newly released Graviton5 offers 25% better performance compared to its predecessor, Graviton4. It uses the latest 3-nanometer (nm) process and features up to 192 cores per package. The L3 cache, a high-speed memory that stores frequently accessed data close to the processor, is now five times larger than before. A larger cache memory can increase the execution speed of CPU instructions.

What sets Graviton apart from traditional server CPUs is its 'efficiency.' Graviton is designed using the architecture of Arm, a British semiconductor design company. Arm's architecture consumes less power compared to the x86 architecture adopted by Intel and AMD. As a result, it delivers higher efficiency with lower power consumption and reduces operational costs.

Graviton has already become a threat to existing CPU manufacturers. As of last year, AWS reportedly used Graviton for about 50% of its servers dedicated to cloud computing services. Given that AWS is considered the world's largest cloud service provider and is expected to expand Graviton adoption further, demand for CPUs from traditional manufacturers is likely to decline.

Globally, companies such as Adobe, SAP, and Airbnb are utilizing Graviton-based cloud computing, while among major Korean conglomerates, Samsung Electronics is using Graviton-based servers. Since Graviton is custom-built by AWS, it can only be used through AWS cloud computing services.

Peter DeSantis, Senior Vice President of Utility Computing at Amazon Web Services (AWS), is delivering the keynote speech at re:Invent 2025 held on the 4th (local time). Provided by AWS

AWS is also challenging the artificial intelligence (AI) chip sector, which is virtually monopolized by Nvidia. On the first day of the re:Invent event, December 1st, AWS CEO Matt Garman unveiled 'Trainium3,' the latest model of the company's proprietary AI chip 'Trainium,' which boasts over four times the computing performance of its predecessor. Trainium is a chip specialized for AI model training.

Trainium also emphasizes cost-effectiveness, with Trainium3 reducing power consumption by about 40% compared to the previous model. According to AWS, using Trainium3 can cut AI model training and operation costs by up to half compared to using Nvidia graphics processing units (GPUs).

The keynote speech focused on 'infrastructure innovation.' Following Brown, Peter DeSantis, Senior Vice President of Utility Computing at AWS, introduced new features such as managed instances for 'Lambda,' a serverless computing service that eliminates the need to manage servers.

DeSantis stated, "The fundamentals of infrastructure-security, availability, resilience, agility, and cost-are more important than ever," adding, "AI is just the beginning. New architectures will open up new possibilities."

Meanwhile, during the keynote, Jaesung Lee, CEO of Twelve Labs, took the stage to introduce the video AI models 'Marengo' and 'Pegasus.' Twelve Labs has previously received support through 'AWS Activate,' AWS's startup support program.

Jaesung Lee, CEO of Twelve Labs, is giving a presentation at 're:Invent 2025' held on the 4th (local time). Provided by AWS