Meta Secures Millions of AWS Graviton Chips
Amazon revealed on Friday that Meta will deploy millions of its AWS Graviton processors to support rising AI operations. The agreement marks a key success for Amazon's in-house silicon. Meta, known for developing large language models like Llama, faces surging compute needs as it advances AI capabilities across its platforms.
The Graviton series consists of ARM-based central processing units focused on general computing duties. These differ from graphics processing units, which dominate model training phases. AWS positions its newest Graviton iteration to manage intensive AI tasks effectively.
AI Agents Drive Demand for CPU Power
After initial training, AI systems shift toward agent-based applications. These involve heavy computation for activities like instant reasoning, code generation, web searches, and orchestrating complex, multi-stage processes. Such demands favor CPUs over GPUs in many cases. Amazon states its latest Graviton processors address these AI-specific requirements directly.
This evolution reflects broader trends in AI deployment. Companies build foundation models on GPUs from leaders like Nvidia. Once deployed, inference and agent operations often require versatile, cost-efficient CPUs. AWS Graviton chips, introduced publicly around 2018 with subsequent generations, offer strong performance per dollar, appealing to large-scale users.
Meta Returns Spending to AWS
The new arrangement directs more of Meta's budget toward AWS, away from rivals such as Google Cloud. In August of the previous year, Meta entered a six-year pact worth $10 billion with Google Cloud. Before that shift, Meta relied mainly on AWS, supplemented by Microsoft Azure services.
Amazon chose to disclose the Meta news immediately following the conclusion of Google Cloud Next. That event featured Google's updates to its custom AI processors. The timing underscores competitive dynamics among cloud giants, each pushing proprietary hardware.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Amazon's Trainium Faces High Demand
Amazon also produces the Trainium chip, a specialized accelerator for both model training and inference stages. Despite its training-focused name, Trainium handles post-training prompt processing too. Earlier this month, Anthropic, creator of the Claude AI models, locked in a massive commitment. Anthropic pledged $100 billion over 10 years for AWS capacity, emphasizing Trainium usage. In exchange, Amazon committed an additional $5 billion investment, raising its total stake in Anthropic to $13 billion.
Anthropic, founded by former OpenAI executives, has grown rapidly with its safety-focused AI approach. This deal secures much of Amazon's Trainium supply for years, highlighting supply constraints in AI hardware.
Graviton Challenges Nvidia's Vera
Meta's adoption validates AWS Graviton as a contender against Nvidia's Vera CPU. Both are ARM architectures tailored for AI agent workloads. Nvidia markets Vera and related systems directly to businesses and cloud operators, including AWS itself. AWS, however, provides Graviton access solely through its cloud infrastructure.
Amazon CEO Andy Jassy addressed this rivalry in his recent annual letter to shareholders. He criticized Nvidia and Intel offerings, stressing customer demand for superior price-to-performance in AI. Jassy vowed to capture contracts based on those merits. Such statements increase expectations on Amazon's chip development group, which continues to innovate amid fierce market pressures.
Amazon Web Services leads the cloud market with over 30% share, per industry trackers. Its custom silicon strategy, spanning Graviton CPUs and Trainium accelerators, aims to reduce reliance on third-party vendors while cutting costs for clients. Meta, with its vast user base on Facebook, Instagram, and WhatsApp, generates enormous data volumes that fuel AI advancements.