Not just relying on GPUs! Meta injects billions into Amazon's Graviton; AWS's self-developed CPUs advance into the core of AI computing power.

```

Meta has reached a multi-year, multi-billion-dollar chip procurement agreement with Amazon, under which it will use Graviton CPU chips from Amazon Web Services to support its AI agent business. This move marks a subtle shift in the AI chip landscape, which has long been dominated by GPUs.

According to the agreement, Meta will rent tens of millions of AWS Graviton chip cores, most of which will be deployed within the United States. Neither company disclosed specific financial terms, but Amazon Vice President and Distinguished Engineer Nafea Bshara stated that the agreement's term is three to five years. This cooperation will make Meta one of the top five Graviton customers by scale on AWS.

This deal continues a recent series of major infrastructure investments by Meta. In the past few weeks, the company signed agreements with CoreWeave and Nebius, with the deals totaling as much as $48 billion, mainly for renting Nvidia GPU resources. Meanwhile, Meta announced on Thursday it would lay off about 8,000 staff in May—10% of its total workforce—to offset the persistently high costs of AI investment.

For AWS, this agreement further validates the commercial viability of Graviton chips for AI workloads. In the same week, Amazon added a $5 billion investment in Anthropic, and that agreement similarly includes provisions for Anthropic to use Graviton cores at scale. Amazon's stock rose nearly 2% in premarket trading.

Deal Details: Large Scale, Covering Core AI Agent Scenarios

According to information disclosed by both parties, Meta will use hundreds of thousands of Graviton physical chips under this agreement, totaling tens of millions of chip cores. Bshara noted that Meta had previously used Graviton on a small scale, but this is a leap in scale.

Meta is purchasing AWS's latest generation Graviton5—a 3nm process CPU chip based on Arm architecture. Bshara said that on the AWS EC2 platform, Graviton offers the best price-performance ratio among comparable compute options, while consuming about 60% less energy than similar products.

"Meta has extremely abundant options on the supply side, but they chose Graviton5 because of its balance between price and performance," said Bshara.

Meta’s head of infrastructure, Santosh Janardhan, stated: "Expanding to Graviton allows us to run the CPU-intensive workloads behind our AI agents at the performance and efficiency we need at our scale."

CPUs Regain Favor, AI Agents Drive Demand Reversal

In recent years, the AI compute narrative has been almost entirely dominated by GPUs, with CPUs relegated to a supporting role. However, with the rise of AI agents, this landscape is quietly changing.

Brendan Burke, semiconductor research director at market research firm Futurum Group, pointed out that CPUs and GPUs have good synergy in AI workloads—CPUs run specific applications and pass tasks to GPUs, with both types of chips together supporting AI agents in carrying out diverse tasks.

CPUs also play a key role in the “post-training” stage of large language models—that is, after pretraining, during the fine-tuning of models for specific objectives, CPUs carry out important computational functions.

Intel CEO Lip-Bu Tan also confirmed this trend at an analyst meeting on Thursday, saying that its Xeon server chips are currently in short supply. "For the past few years, high-performance computing narratives have revolved almost entirely around GPUs and other accelerators. But in recent months, we've seen clear signals that CPUs are once again becoming an indispensable foundation in the AI era," he said.

Burke believes that "for the top frontier AI labs, their demand for CPUs is almost limitless."

Meta’s Diversified Chip Strategy Continues to Expand

This collaboration with AWS is part of Meta’s ongoing diversified chip procurement strategy this year. In addition to the Graviton CPU agreement, Meta has also reached cooperation agreements this year with Nvidia, AMD, and Arm Holdings.

Meta said the new agreement reflects its diversified infrastructure approach and shows that no single chip architecture can efficiently handle all computing tasks.

Meta’s ambitions in AI agents also provide a direct driver for this demand. Last December, the company acquired AI agent startup Manus for more than $2 billion; Manus focuses on developing AI agent products capable of executing complex tasks. In early May, Meta also released its first new AI model in a year, Muse Spark, and said more models are in the pipeline.

Meta and AWS’s partnership dates back to around 2016 but previously primarily focused on core cloud services, use of Amazon’s Bedrock platform, and renting GPU clusters from AWS. Bshara said Meta began renting Nvidia GPUs from AWS as early as 2017.

AWS Strengthens Graviton’s Commercial Footprint

For AWS, landing Meta as a flagship client is an important milestone in the mass commercialization of Graviton chips. Bshara is a co-founder of chip company Annapurna Labs, which was acquired by Amazon in 2015 and has since become a core force in internal chip R&D at AWS.

Graviton has already attracted major enterprises such as Adobe, Apple, and Snowflake, and is further consolidating its market position in AI infrastructure. Bshara said, "Graviton is one of the main pre-training platforms for many foundational model companies, and Meta is a significant new client."

Meanwhile, AWS has not fully focused on self-developed chips in this field. In March, AWS announced a partnership with AI inference chip startup Cerebras, to deploy Cerebras’ inference chips in its data centers, further expanding the diversity of its chip ecosystem.

Burke believes that the agreement between Meta and AWS will further reinforce market recognition of Graviton's value in the AI era. "This is a case of significant validation," he said.

Risk Warning and Disclaimer ClauseThe market contains risks, and investment requires caution. This article does not constitute individual investment advice, nor does it take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their own specific situation. Investing accordingly is at your own risk. ```