Nvidia signs multiyear AI chip supply deal with Meta Platforms

Nvidia announced a multiyear agreement to supply Meta Platforms with millions of its current and upcoming AI chips.

The deal covers Nvidia’s Blackwell chips, future Rubin AI chips, and standalone processors like Grace and Vera.

Meta is developing its own AI chips and is in talks with Google about using its Tensor Processing Units.

Nvidia did not disclose the deal’s value, but Meta is believed to be among its largest customers, contributing significantly to Nvidia’s revenue. The announcement suggests Nvidia is maintaining a strong relationship with Meta amid its own chip development efforts.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Meta’s deal with Nvidia is just one piece of a much larger strategy

The announcement leaves out how broadly Meta is lining up AI chip options to avoid depending on one supplier.
Meta already runs its custom inference chip, the Meta Training & Inference Accelerator (MTIA), across its data centers, while a training chip for ranking and recommendations moves toward higher production volumes 1.
Meta has also talked with Google about a possible multi-billion dollar agreement for Google’s Tensor Processing Units (TPUs), starting with leased TPU capacity hosted at Google and later placing those chips in Meta data centers from 2027 2.
Alongside Nvidia’s Vera, Meta became the first customer for Arm’s “Phoenix” data center CPU, built to act as a head node (a coordinating CPU that manages and feeds work to attached chips) for accelerators 3.

The deal signals a shift from general-purpose to specialized CPUs in AI

Nvidia now pairs CPUs with GPUs, which pushes AI data centers away from common x86 processors sold by Intel and AMD.
Workloads such as reinforcement learning raise CPU needs for data prep and simulation environments 3.
Nvidia aims to sell integrated platforms where Arm-based Grace and Vera CPUs connect closely to its GPUs through a high-speed coherent link (a connection that keeps memory and data consistent between chips). Grace uses NVLink-C2C at 900 GB/s (bidirectional), while Vera is expected to lift NVLink-C2C bandwidth to 1.8 TB/s 4 5.
The approach can improve performance per watt, yet it can also narrow customer choice. Nvidia’s Blackwell GB200 systems match Blackwell GPUs with Grace CPUs, plus some rack-scale designs center on Grace instead of x86 CPUs 6.