Amazon.com and Meta Platforms Inc. have entered into a multi-billion dollar long-term agreement under which the social media giant will lease hundreds of thousands of Amazon's in-house developed ARM-based Graviton server CPUs. These processors will be deployed in Meta's large-scale new AI data centers to handle the massive artificial intelligence inference workloads generated by users of Facebook and Instagram. Amazon Vice President and Annapurna Labs co-founder Nafea Bshara stated that the multi-year deal grants Meta extended access to the Graviton series of data center server central processing units.

While large language models capable of text generation or massive-scale inference are typically trained and run on NVIDIA's AI GPUs or Google's TPU AI computing clusters, AI model developers and end-user platforms urgently require general-purpose central processors like Graviton for scheduling and coordination tasks. These include generating responses to queries after model training and managing AI agent workflows, a process generally termed "AI inference." "GPUs are useless without CPUs alongside them," Bshara commented.

The AI infrastructure race extends beyond GPUs and ASICs, with CPU technology becoming a critical component in the inference era. Amazon has deployed Graviton processors for most CPU requirements in its data centers, a significant achievement for a company once heavily reliant on Intel's hardware ecosystem. Graviton, Amazon Web Services' self-developed ARM-based server CPU, handles general-purpose computing, scheduling, data preprocessing/post-processing, service orchestration, and certain AI inference coordination tasks within AI data centers.

Amazon CEO Andy Jassy recently indicated that the company's data center chip business is progressing toward $20 billion in annual sales, with executives exploring direct sales of these chips to other technology firms like Meta and CoreWeave for long-term leasing or usage in their cloud server fleets—though these chips have so far been exclusive to Amazon's large cloud data centers.

The landmark Meta-Amazon agreement announced Friday represents the latest large-scale, long-term collaboration among major tech companies as the global technology sector races to secure sufficient AI computing resources—encompassing CPU, GPU, and ASIC processor clusters—to power new and developing AI models. OpenAI and Anthropic have increased their usage of Amazon's custom Trainium AI chips, Amazon's cost-effective ASIC alternative to NVIDIA's AI GPU computing architecture, which Amazon is actively marketing to AI developers including Meta.

Meta is pursuing a diversified chip procurement strategy for its expanding AI workloads, stating this approach maintains flexibility through partner diversification. The company has signed several major AI infrastructure supply agreements with chip leaders including NVIDIA and AMD, committing billions to acquire AI GPU solutions predominantly led by these two companies. Meta recently finalized another multi-billion dollar agreement to utilize Alphabet's custom TPU AI computing clusters and is heavily investing in developing its own AI chips to reduce costs and dependence on external chip suppliers. The company is currently developing four versions of its MTIA AI chip for training and inference applications and recently expanded its long-term partnership with Broadcom to assist in designing and manufacturing these AI chips.

The dawn of the AI agent era is driving unprecedented demand for data center CPUs. The rapid construction of AI data centers has pushed Intel's data center CPUs into short supply, with delivery times for some high-performance server CPUs extending to six months and prices rising approximately 10% this year. This supply-demand dynamic underpins the nearly 80% surge in Intel's stock price this year—after a prolonged slump—with shares hitting their highest level since 2000 last week. Intel's stock surged over 30% in premarket trading Friday following better-than-expected earnings results.

While early model inference primarily involved "single request-single generation" tasks where CPUs handled data movement, request routing, and basic scheduling as auxiliary control components, the AI agent and reinforcement learning era has transformed system loads. Workloads now involve complex cycles including task planning, tool invocation, sub-agent coordination, environmental interaction, state management, and result verification. This "orchestration layer" constitutes CPU-intensive tasks characterized by strong control flows, branching logic, system calls, and memory access—functions not efficiently handled by GPUs. Consequently, CPUs are evolving from supporting roles to becoming new bottlenecks determining system throughput, latency, and resource utilization.

Morgan Stanley's latest projections indicate that the AI agent explosion signifies a structural shift from pure computation to orchestration, potentially creating $32.5 billion to $60 billion in additional CPU market capacity by 2030, expanding the total addressable market for server CPUs to $82.5-$110 billion. TrendForce forecasts suggest CPU-to-GPU ratios in AI data centers may shift dramatically from traditional 1:4 to 1:8 ratios toward 1:1 to 1:2 configurations in the AI agent era.

For companies like Meta that process enormous volumes of AI agent interactions, recommendations, advertisements, content generation, and query responses daily, many tasks don't require expensive GPU involvement throughout. Large-scale adoption of high-density ARM architecture CPUs like Graviton—instead of Intel's x86 architecture—for peripheral inference workloads can reduce per-request costs, free GPUs for higher-value training/inference tasks, and improve total cost of ownership for computing clusters.

Arm Holdings has emphasized that AI data center expansion is making orchestration, data processing, and system control via low-power, high-efficiency ARM architecture CPUs critical bottlenecks. AWS's fifth-generation Graviton, which increases core count to 192 cores, reflects this rising demand for CPU density. Meta's latest agreement underscores how AI computing infrastructure competition is transitioning from "GPU-centric demand" toward heterogeneous systems integrating GPUs, custom AI ASICs, Arm/x86 data center CPUs, high-speed optical interconnects, and software stacks.