Meta News

By MSB | April 2026

In a move reflecting the rapid evolution of the artificial intelligence ecosystem, Meta has signed a deal to use millions of CPUs designed by Amazon Web Services (AWS), marking a significant shift in an industry historically dominated by GPUs.

The agreement—considered multi-million dollar and spanning several years—positions Meta as one of Amazon's largest client of Graviton chips, signaling a deep transformation in how modern AI systems are built and scaled.

Over the last decade, AI infrastructure has been strongly associated with GPUs, especially Nvidia's, due to their ability to accelerate machine learning model training.

However, the agreement between Meta and AWS introduces a key nuance: CPUs are regaining prominence, but in a different role.

The chips used—AWS Graviton, based on ARM architecture—are not primarily designed for model training, but for executing post-training tasks (post-training) such as:

• Real-time inference
• Running autonomous agents
• AI service orchestration
• Code generation and reasoning

Meta will use tens of millions of Graviton cores for these workloads, showing that modern AI no longer relies on a single type of hardware.

The move may seem counterintuitive, but it responds to a technical reality:

• Training: requires massive parallelism → GPUs
• Inference and Agents: require efficiency, low latency, and flexibility → CPUs

In advanced AI systems—especially those called agentic AI—workloads are more dynamic:

• Real-time decision making
• Accessing multiple data sources
• Coordination between processes

In this context, CPUs offer key advantages:

• Better management of sequential tasks
• Lower energy consumption
• Lower cost per operation
• More granular scalability

Some estimates indicate that this type of workload can run with up to 60% less energy consumption on optimized CPU architectures.

Another determining element is the global limitation of GPUs, which has forced major tech companies to diversify their infrastructure.

• Nvidia continues to dominate the market
• But demand exceeds supply
• Costs are extremely high

This has driven a clear industry strategy:

👉 decoupling the AI stack across multiple types of hardware

Meta, in particular, already maintains parallel agreements with Nvidia, AMD, and other vendors, confirming a trend toward heterogeneous architectures.

For Amazon, this agreement represents a key strategic victory.

AWS has spent years investing in proprietary chips:

• Graviton (CPU) → general and inference workloads
• Trainium → AI training
• Inferentia → optimized inference

The goal is clear: reduce reliance on third parties (like Nvidia) and improve the economic efficiency of the cloud.

Amazon's chip business already generates multi-million dollar revenue and is becoming a central pillar of its AI strategy.

For Meta, the deal responds to a structural necessity:

• Increasingly large models
• Massive use in products (Facebook, Instagram, WhatsApp)
• Expansion into intelligent agents and automation

This implies enormous pressure on infrastructure:

• Low latency for billions of users
• Controlled operating costs
• Ability to scale dynamically

Using optimized CPUs allows Meta to run AI