At this year's re:Invent 2025 conference, AWS not only showcased its powerful performance...Graviton5, the fifth-generation self-developed processorFurthermore, and unusually once again, Apple TV has been invited to endorse AWS. Regarding AWS's plans for custom chip development and its views on the wave of ASIC (Application-Specific Integrated Circuit) development, AWS Vice President and Distinguished Engineer, and also...Annapurna Labs, IsraelIn an interview, co-founder Nafea Bshara offered some unique insights, stating that he does not believe ASIC will be the sole solution driving the next wave of AI, and emphasized that providing diverse options and meeting different deployment needs is the key.
Apple was the first to adopt Graviton 5, proving the value of customized chips.
Amazon began its journey after acquiring Annapurna Labs in 2015.Self-developed chipDevelopment. The Graviton5 previewed this time uses TSMC's 3nm process, with a 25% performance improvement over the previous generation. Nafea Bshara stated that through customized chips, AWS can receive customer feedback more directly and quickly make optimized responses at the hardware level.
Apple's experience is a prime example. As a major Graviton user, Apple leverages its iCloud service, which supports 10 billion daily accesses, for Swift coding development. Nafea Bshara points out that this proves AWS's custom chips can meet the stringent performance and cost requirements of leading tech companies.
Furthermore, AWS is currently the only provider specifically for Apple's development environment.Dedicated solutions AWS, a cloud provider of (Amazon EC2 Mac instances), makes it easier for developers to build macOS environments in the cloud, a service not offered by other cloud providers. This highlights AWS's ability to provide solutions for different application deployment needs.
Instead of following the path of Google TPU, it emphasizes "general applicability" over "specialization".
Facing the active development of Google CloudTPUs specifically designed for their own frameworks such as TensorFlowNafea Bshara offered a different perspective. He argued that AI is evolving rapidly, and hardware design should not simply follow specific software or model architectures, emphasizing that "we will not create custom hardware for specific software."
He cited recent popular generative AI applications as an example, stating bluntly that "Google's TPUs struggle to handle workloads like Decart that require real-time rendering."
Nafea Bshara emphasized that while AWS will make custom chip design adjustments for new model architectures (such as mainstream frameworks like PyTorch), the core principle remains to maintain versatility and flexibility. Therefore, while promoting its self-developed Trainium and Inferentia chips, AWS will never reduce its cooperation with NVIDIA, AMD, or even Intel.
"Providing more choices is always a good thing for customers," Nafea Bshara said.
Closely connected to Taiwan's supply chain: TSMC, Foxconn, and Wistron included
When discussing hardware manufacturing, Nafea Bshara also emphasized its close partnership with the Taiwanese supply chain.
• Chip manufacturing:AWS's self-developed chips use TSMC's advanced process technology, but Nafea Bshara also remains open, stating that it does not rule out the possibility of cooperating with Samsung or Intel on wafer foundry in the future to ensure capacity and technology diversity.
• Server manufacturing:In terms of system level and server rack assembly, AWS maintains long-term and in-depth partnerships with Taiwanese ODM companies such as Foxconn and Wistron to jointly build the hardware infrastructure for instances that support global cloud computing.
In summary, AWS's chip strategy is not about "replacing" existing chip giants, but about filling the gaps in cost-effectiveness and energy efficiency that general-purpose hardware cannot meet by developing its own chips. At the same time, it maintains an open ecosystem, giving customers the greatest freedom of choice between NVIDIA's extreme performance, AMD's high cost-effectiveness, and the optimal integration of AWS's self-developed chips.
