Meta, Broadcom extend AI chip deal to 2029

Meta and Broadcom said they have extended their partnership to design Meta’s in-house AI accelerators through 2029, with Meta committing to an initial deployment of 1 gigawatt of its training and inference chips, and more over time based on Broadcom technology.

Meta also said in a filing that Broadcom CEO Hock Tan told Meta last week he will not stand for reelection to its board after joining in 2024, and director Tracey Travis will leave after serving since 2020.

Meta first introduced its MTIA chips in 2023 and added four new versions in March, as large tech firms build custom ASICs to reduce reliance on Nvidia and AMD GPUs for AI data centers.

🔗 Source: CNBC

🧠 Food for thought

Implications, context, and why it matters.

Partnership offers a path to less Nvidia dependence

Broadcom’s XPU platform supports work on custom AI accelerators over multiple generations 1.
Broadcom also supplies networking gear such as Ethernet switches plus optical connectivity products to scale Meta’s AI data center clusters and ease bottlenecks 1.
Meta’s MTIA chips focus on inference and recommendation workloads, not general-purpose acceleration 2.
This approach casts Broadcom as a partner for hyperscalers (large cloud and internet companies with massive data centers) that want less reliance on general-purpose GPUs by using workload-optimized silicon, which can support longer-term co-design revenue 3.

AI hardware market splits into two lanes

The agreement matches a trend where large tech companies build a second lane of custom silicon alongside general-purpose GPUs from Nvidia and AMD 4.
These companies tune hardware to specific jobs, training some models on GPUs while running high-volume inference on lower-cost chips such as MTIA 5.
This shift lines up with expectations that inference may surpass training as the main AI workload over time, which lifts demand for cost-efficient hardware 5.
More hardware variety raises the stakes for the software layer, including PyTorch and Triton, which help AI models run across more hardware environments 6.