--- title: "The AI arms race is not just about GPUs! Meta teams up with Amazon Graviton to challenge x86 hegemony" type: "News" locale: "en" url: "https://longbridge.com/en/news/284011961.md" description: "Meta has reached a multi-billion dollar agreement with Amazon to lease the Graviton series ARM architecture central processing units for its newly built AI data centers to meet the AI inference needs of social media users. Amazon Vice President Nafea Bshara stated that this agreement will enable Meta to use Graviton CPUs in the long term, emphasizing the importance of CPUs in AI inference. In recent years, Amazon has primarily used Graviton processors in its data centers, gradually reducing its reliance on Intel hardware" datetime: "2026-04-24T13:30:03.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/284011961.md) - [en](https://longbridge.com/en/news/284011961.md) - [zh-HK](https://longbridge.com/zh-HK/news/284011961.md) --- # The AI arms race is not just about GPUs! Meta teams up with Amazon Graviton to challenge x86 hegemony According to Zhitong Finance APP, Amazon (AMZN.US), the American cloud computing and e-commerce giant, has reached a multi-billion dollar long-term agreement with Facebook's parent company Meta Platforms Inc. (META.US). The social media giant will lease hundreds of thousands of Amazon's self-developed ARM architecture general-purpose data center server CPU chips for its large-scale AI data centers under construction, to meet the massive artificial intelligence inference workloads of users on Facebook and Instagram. Nafea Bshara, Amazon's Vice President and co-founder of the company's Annapurna Labs chip division, stated in an interview that this multi-year agreement will grant Meta long-term access to the Graviton series of data center server-level central processing units. Artificial intelligence large language models capable of generating text or performing massive inference workloads are typically trained and run using NVIDIA-dominated AI GPUs or Google TPU AI computing clusters. However, AI large model developers and users of B-end or C-end AI application platforms also urgently need general-purpose central processors (i.e., data center server CPUs) like Graviton at large AI data centers to perform various scheduling and coordination tasks, particularly generating responses to queries and managing AI agent workflow processes after model training, a process commonly referred to as "AI inference." "Without a CPU, a GPU is useless," Bshara stated. **The AI Arms Race Beyond GPU/ASIC: Meta Partners with Amazon, CPU Becomes a Key Piece in the Inference Era** In recent years, most of the CPUs deployed in Amazon's data centers have been Graviton central processors; this is a significant achievement for a company that once heavily relied on Intel (INTC.US) hardware. Graviton is a general-purpose server CPU developed by Amazon's AWS cloud computing division, primarily responsible for general computing, scheduling, data preprocessing/postprocessing, service orchestration, and some AI inference-related scheduling and coordination tasks in AI data centers. Amazon CEO Andy Jassy recently stated that the company's data center chip business is firmly moving towards achieving $20 billion in sales within a year, and executives are considering actively selling these Amazon-developed data center-related chips for long-term leasing or on-demand use by other tech companies like Meta and CoreWeave in their cloud computing server clusters—so far, these chips have only existed in Amazon's large cloud computing data centers. The significant deal announced on Friday between Meta and Amazon represents the latest long-term and large-scale collaboration among major tech companies; currently, the global tech industry is competing to secure sufficient AI computing power related CPU/GPU/ASIC AI processor device clusters to drive new and developing future AI large models OpenAI and Anthropic have stated that they are increasing their use of Amazon's self-developed Trainium AI chips; Trainium is an alternative AI computing infrastructure solution developed exclusively by Amazon AWS, designed to benchmark core performance metrics against NVIDIA's AI GPU computing power system with a high cost-performance ratio AI ASIC technology route. The company has begun actively marketing its series of self-developed AI chip computing systems, including Trainium, to OpenAI, Anthropic, and Meta. Meta has taken extensive measures to acquire chips for its increasingly large AI workloads, stating that this is to diversify its partners and maintain flexibility. The company has signed a series of large AI computing infrastructure supply agreements with chip giants such as NVIDIA and AMD. Meta has agreed to spend billions of dollars to purchase AI GPU computing infrastructure solutions led by NVIDIA and AMD. The company has also recently signed another multi-billion dollar agreement to use TPU AI computing clusters exclusively developed by Google, a subsidiary of Alphabet Inc. Meta is also heavily investing in the development of its own AI chips to help reduce costs and decrease reliance on third-party chip giants. The company is currently developing four versions of the MTIA AI chips for AI training/inference purposes and has recently announced an expansion of its long-term deep collaboration with Broadcom, which will assist Meta in designing and manufacturing these AI chips. **The Era of AI Agents Has Arrived, Data Center CPU Demand Soars** The construction of AI data centers is in full swing, driving Intel's data center CPUs into a state of supply shortage. The delivery time for some of Intel's most in-demand high-performance server CPUs has been extended to as long as six months, and the prices of these high-performance server-grade CPUs aimed at data centers have generally increased by 10% this year. This is why Intel, a chip manufacturer whose stock price has been sluggish for a year and a half, has seen its stock price soar over 80% this year, and the stock price of this veteran chip giant reached its highest point since 2000 last week, reflecting the underlying bullish logic. After announcing all-around better-than-expected results on Friday morning Beijing time, Intel's stock price surged over 30% in pre-market trading on Friday. In the early stages, large model inference was primarily based on "single request - single generation," with CPUs mainly handling data transfer, request routing, and basic scheduling, serving as a typical auxiliary control component. However, with the advent of AI agents and reinforcement learning, system loads have evolved from simple forward inference to complex closed loops that include task planning, tool invocation, sub-agent collaboration, environmental interaction, state management, and result verification. The aforementioned "orchestration layer" is essentially a CPU-intensive task characterized by strong control flow, strong branching judgment, strong system calls, and strong memory access, which cannot be efficiently replaced by GPUs. Therefore, CPUs are transitioning from being the "supporting role" to becoming the new bottleneck that determines system throughput, latency, and resource utilization Morgan Stanley's latest forecast data shows that the explosion of intelligent agents marks a structural shift from computation to orchestration, leading to an additional market space for CPUs of $32.5 billion to $60 billion by 2030, and significantly expanding the total addressable market (TAM) for server-level CPUs to a range of $82.5 billion to $110 billion. A forecast report from TrendForce indicates that in the era of AI agents, the CPU:GPU ratio may be re-evaluated from the traditional AI data center ratio of 1:4 to 1:8, to a range of 1:1 to 1:2. For companies like Meta, which handle vast amounts of AI agents, recommendations, advertisements, content generation, and query responses daily, many tasks do not require expensive GPUs to be involved throughout; leveraging high-density ARM architecture like Graviton instead of Intel x86 architecture CPUs for inference service peripheral loads can reduce the cost per request, free up GPUs for higher-value training/inference tasks, and improve overall cluster TCO. Arm has also emphasized that the expansion of AI data centers is making orchestration, data processing, and system control on the low-power, high-efficiency ARM architecture CPUs a critical bottleneck, while AWS's fifth-generation Graviton, which increases the core count to 192 cores, reflects this rising demand for CPU density. The latest deal led by Meta also highlights that the competition for AI computing infrastructure is shifting from a "single demand center for GPUs" to a heterogeneous system of GPU + self-developed AI ASIC + Arm/x86 data center-level CPUs + high-speed optical interconnect systems for data centers + software stacks ### Related Stocks - [META.US](https://longbridge.com/en/quote/META.US.md) - [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md) - [SMH.US](https://longbridge.com/en/quote/SMH.US.md) - [IGV.US](https://longbridge.com/en/quote/IGV.US.md) - [METW.US](https://longbridge.com/en/quote/METW.US.md) - [XDAT.US](https://longbridge.com/en/quote/XDAT.US.md) - [DTCR.US](https://longbridge.com/en/quote/DTCR.US.md) - [XSW.US](https://longbridge.com/en/quote/XSW.US.md) - [SOXX.US](https://longbridge.com/en/quote/SOXX.US.md) - [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md) - [FBL.US](https://longbridge.com/en/quote/FBL.US.md) - [CLOU.US](https://longbridge.com/en/quote/CLOU.US.md) - [XSD.US](https://longbridge.com/en/quote/XSD.US.md) - [AMZU.US](https://longbridge.com/en/quote/AMZU.US.md) ## Related News & Research - [BUZZ-Street View: AI demand boosts Intel, but capacity limits loom](https://longbridge.com/en/news/283982891.md) - [Meta’s loss is Thinking Machines gain](https://longbridge.com/en/news/284053848.md) - [Anthropic Commits $100B to AWS as Amazon Bets Up to $25B on AI Partner. Is AMZN a Buy Now?](https://longbridge.com/en/news/284008126.md) - [Meta will now allow parents to see the topics their child discussed with Meta AI](https://longbridge.com/en/news/283816465.md) - [Meta is tracking employee keystrokes on Google, LinkedIn, Wikipedia as part of AI training initiative](https://longbridge.com/en/news/283739838.md)