Meta and Amazon have reached a multi-year, multi-billion dollar cooperation agreement. The social media giant will use Amazon Cloud Technology’s tens of millions of Graviton chip cores to provide computing power support for its agents and other artificial intelligence projects. Both companies declined to disclose the specific amount of the agreement and the length of the cooperation. Amazon Vice President and Distinguished Engineer Nafiah Bushara revealed that the cooperation period is three to five years.
Bushala is also the co-founder of Annapurna Labs, Amazon Cloud Technology’s self-developed chip department. He said that most of these tens of millions of Amazon Cloud Technology Graviton chip cores will be deployed in the United States.
Currently, major technology giants and artificial intelligence laboratories are still competing for computing power resources to achieve their own artificial intelligence development goals. Analysts pointed out that Meta has successively reached cooperation with many chip companies such as NVIDIA, AMD, and ARM Holdings this year. This signing with Amazon Cloud Technology is one of them. It also highlights the rigid need for the development of artificial intelligence for the diversified layout of chips.
Amazon executive Bu Shala said: "As we all know, Meta has a wealth of choices in the supply chain, but in the end we selected our fifth-generation Graviton chip with a 3-nanometer process. The core consideration was its excellent cost performance."
This time Meta purchased Amazon Cloud Technology Graviton central processing unit (CPU). In the context of the artificial intelligence industry's long-term focus on graphics processing units (GPUs), central processing units were once marginalized.
With the rapid rise of intelligent technology, central processing units have once again become a popular computing product, and traditional CPU manufacturers such as Intel have also benefited from this industry trend.
Meta stated that this new cooperation reflects the company’s strategic thinking of diversifying the layout of infrastructure, and also confirms that a single chip architecture cannot efficiently adapt to all computing tasks.
Brandon Burke, director of semiconductor research at Futu Group, a market research institution, explained that the central processing unit can efficiently run various applications and provide data collaboration support for the graphics processor. In other words, when the agent performs various complex tasks, the CPU and GPU can form an efficient complementary match.
At the same time, the central processing unit is also the core hardware of the post-training optimization phase of the large language model. This phase mainly conducts targeted intensive training for the pre-trained model to adapt the model to specific business goals.
This is not the first time Meta and Amazon Cloud Technology have joined hands. Bushala introduced that the cooperation between the two technology giants can be traced back to 2016, but past cooperation has mainly focused on basic cloud services, the use of Amazon's Bedrock AI platform, and Meta's rental of Amazon Cloud Technology GPU computing clusters.
Burke mentioned that Meta spent more than US$2 billion to acquire artificial intelligence start-up Manus in December last year, which is enough to show its strategic ambition to delve into the field of intelligent agents, and this layout has also directly boosted the demand for CPU computing power in data centers. Manus Company focuses on the development of high-end intelligent agents that can perform complex tasks.
Burke bluntly said: "For leading cutting-edge artificial intelligence R&D institutions, there is almost no upper limit on the computing power demand of central processors."
In order to optimize its operational structure and share the huge investment in artificial intelligence research and development, Meta announced on Thursday that it will lay off 10% of its employees in May, with a layoff of approximately 8,000 people. In the first ten days of this month, the company released its new artificial intelligence model Muse Spark after a year, and said that it will continue to launch a number of new products in the future.
Futu Group analyst Burke believes that for Amazon Cloud Technology, this cooperation with Meta has effectively verified the practical value of its self-developed Graviton series CPU chips in artificial intelligence scenarios. Bushara said that after the implementation of this cooperation, Meta has officially become one of the five core customers of Amazon Cloud Technology Graviton chips.
Earlier this week, Amazon announced an additional US$5 billion investment in the artificial intelligence company Ansoupke. The cooperation content also includes the company's large-scale purchase of Amazon Cloud Technology Graviton CPU cores.
As early as 2018, Amazon Cloud Technology had started the research and development of self-developed chips, and launched the first-generation Graviton processor based on the Arm architecture that year.
However, Amazon Cloud Technology is not limited to the self-developed chip route. In March this year, the company reached a cooperation with Cerebras and plans to deploy the start-up's AI inference chips in its own data center to enrich the computing hardware matrix.