Meta teams up with Broadcom to build multi-gigawatt MTIA chips

Meta is preparing to take its custom AI chip game to a whole new level.

The social media giant just announced an expanded multi-year partnership with Broadcom to co-develop multiple generations of its MTIA (Meta Training and Inference Accelerator) custom silicon.

Meta recently revealed plans to roll out four new generations of MTIA chips over the next two years to run ranking, recommendation algorithms, and generative AI workloads. This new pact is designed to supercharge that timeline.

1 gigawatt and a boardroom shift

The agreement is enormous: Meta's initial deployment commitment exceeds 1 gigawatt of power capacity, serving as the first phase of a planned multi-gigawatt rollout.

Broadcom isn't just acting as a manufacturer. Under the deal, Broadcom will collaborate with Meta across chip design, advanced packaging, and its high-bandwidth Ethernet networking, all built on Broadcom's custom XPU platform.

The partnership also triggers an executive shuffle. Broadcom President and CEO Hock Tan will step down from Meta's board of directors, where he has served for the past two years, moving instead into an advisory role focused specifically on Meta's custom silicon strategy and infrastructure roadmap.

Mark Zuckerberg noted that the massive compute buildout is meant to help deliver real-time AI and personal superintelligence to billions of users worldwide.

Why it matters

Nvidia dominates the headlines, but the hyperscaler custom silicon race is quietly entering its industrial phase.

Meta has been one of the biggest buyers of standard GPUs, but relying solely on commercial off-the-shelf silicon for massive scale recommendation and high-volume inference is brutally expensive. By teaming up with Broadcom, Meta gets top-tier custom networking and packaging without having to reinvent decades of physical design IP from scratch.

For Broadcom, this cements its role as the undisputed king of custom hyperscale ASICs alongside its storied Google TPU partnership. Locking in gigawatt-scale commitments proves that custom inference hardware is no longer an internal science experiment, but the core foundation of big tech infrastructure.

We think this shows Meta is playing a disciplined long game on compute costs. Relying on commercial GPUs for daily consumer-facing inference doesn't scale sustainably. The real test, however, will be Meta's compiler and software stack: shipping four chip generations in two years is an incredible hardware cadence, but keeping developer tools aligned with that silicon turnover will be the hard part.