Meta Unveils Four MTIA Chip Generations and Extends Broadcom Partnership Through 2029 With One-Gigawatt Initial Commitment
Meta revealed a four-generation MTIA custom chip roadmap built with Broadcom on a 2nm process, committing to over one gigawatt of initial deployment as it accelerates its push to reduce reliance on Nvidia.
Overview
Meta Platforms disclosed an expanded custom chip program on April 14, revealing four generations of its Meta Training and Inference Accelerator and extending its partnership with Broadcom through 2029, according to SiliconANGLE. The deal commits Meta to an initial deployment of more than one gigawatt of MTIA-based computing capacity, with plans to scale to multiple gigawatts in 2027 and beyond.
Broadcom described the upcoming chips as the first custom AI silicon built on a 2-nanometer process node, according to SiliconANGLE. The announcement represents a significant deepening of Meta’s effort to build proprietary alternatives to Nvidia GPUs for AI workloads.
Four-Generation Roadmap
Meta’s MTIA lineup spans from the already-deployed MTIA 300 through three future generations on a roughly six-month cadence, according to Tom’s Hardware.
The MTIA 300 is already in production, handling ranking and recommendation training workloads. It delivers 1.2 petaflops of FP8 performance with 216 GB of high-bandwidth memory at 6.1 TB/s bandwidth, drawing 800 watts per module, as reported by Tom’s Hardware.
The MTIA 400, currently in lab testing, jumps to 6 petaflops of FP8 performance and 288 GB of HBM at 9.2 TB/s, with a 1,200-watt module power envelope. The MTIA 450, targeting early 2027 deployment, doubles HBM bandwidth to 18.4 TB/s while pushing FP8 performance to 7 petaflops. The MTIA 500, expected later in 2027, reaches 10 petaflops of FP8 performance with up to 512 GB of HBM capacity and 27.6 TB/s bandwidth, according to Tom’s Hardware.
From the MTIA 300 to the MTIA 500, HBM bandwidth increases 4.5 times and compute performance grows 25 times, per Tom’s Hardware. The MTIA 400, 450, and 500 are designed to share the same chassis, rack, and network infrastructure, simplifying deployment across data centers.
Chiplet Design and Software Compatibility
The MTIA chips use a chiplet-based architecture optimized primarily for inference rather than training, according to Tom’s Hardware. Meta has built native support for PyTorch, vLLM, and Triton into the chip’s software stack, including compatibility with torch.compile and torch.export.
Meta has already deployed hundreds of thousands of MTIA chips across its applications for inference on organic content and advertising workloads, per Tom’s Hardware.
Broadcom Partnership and Board Changes
Broadcom CEO Hock Tan will step down from Meta’s board of directors to transition into an advisory role focused solely on the custom chip strategy, according to Yahoo Finance. Tracey Travis, a board member since 2020, will also not seek re-election at the next annual shareholder meeting.
Tan used the announcement to push back on speculation about the program’s viability. “Contrary to recent analyst reports, Meta’s custom accelerator, MTIA roadmap is alive and well,” he said, as reported by SiliconANGLE.
Mark Zuckerberg framed the partnership in characteristically ambitious terms, stating that MTIA chips will “build out the massive computing foundation we need to deliver personal superintelligence to billions of people,” according to SiliconANGLE.
A Multi-Vendor Chip Strategy
The Broadcom extension is one piece of a sprawling procurement strategy. Meta has also committed to deploying six gigawatts of AMD GPUs, millions of Nvidia chips, custom Arm-designed processors, and billions in chip rentals from CoreWeave and Nebius, with the Nebius commitment alone reportedly valued at $27 billion, according to SiliconANGLE. Meta’s total capital expenditure commitment for fiscal 2026 exceeds $135 billion.
This follows earlier reporting by The Machine Herald on Meta’s multibillion-dollar deal to lease Google TPUs, which highlighted the company’s four-way silicon strategy spanning Nvidia, AMD, Google, and in-house designs.
Broadcom shares rose approximately 3.5 percent in extended trading following the announcement, while Meta shares were largely unchanged, according to Yahoo Finance.
What We Don’t Know
Meta and Broadcom have not disclosed the financial terms of the extended partnership. It remains unclear what fraction of Meta’s total AI compute the MTIA chips are expected to handle relative to Nvidia and AMD GPUs, or whether the six-month chip cadence can be sustained as designs move to the 2-nanometer node. The announcement also does not specify which foundry will fabricate the 2nm MTIA chips, though TSMC is the only manufacturer currently offering production at that process node.