Meta is doubling down on custom AI hardware to keep up with the accelerating demand across its apps and services, and it’s doing it with a major multi-year effort alongside Broadcom. The two companies will co-develop several generations of next-gen AI silicon, built around Meta’s MTIA program, to power everything from AI training to real-time generative AI inference at massive scale.
Meta’s MTIA chips are part of a broader shift toward purpose-built “XPUs,” a term commonly used for accelerators that combine multiple IP blocks to target specific AI workloads more efficiently than one-size-fits-all solutions. Meta recently outlined four MTIA chips designed for different roles, spanning research and training, general-purpose acceleration, and GenAI inference. Together, they show how aggressively the company is optimizing performance, memory bandwidth, and scale-out connectivity for modern AI systems.
Here’s the MTIA lineup Meta detailed, along with the core specifications it shared:
MTIA 300 is aimed at R&R training, with an 800W module TDP, 216GB of HBM memory, and 6.1 TB/s of HBM bandwidth. For performance, it lists 1.2 PFLOPs FP8/MX8 and 0.6 PFLOPs BF16. Its scale-up domain size is 16, with up to 1 TB/s unidirectional scale-up bandwidth and 200 GB/s unidirectional scale-out bandwidth.
MTIA 400 targets general workloads, moving up to a 1200W module TDP, 288GB HBM, and 9.2 TB/s memory bandwidth. It lists 12 PFLOPs for MX4, 6 PFLOPs FP8/MX8, and 3 PFLOPs BF16. It supports a scale-up domain size of 72 and 1.2 TB/s unidirectional scale-up bandwidth, with 100 GB/s unidirectional scale-out bandwidth.
MTIA 450 is tuned for generative AI inference, with a 1400W module TDP, 288GB HBM, and a substantial 18.4 TB/s HBM bandwidth. Performance figures include 21 PFLOPs MX4, 7 PFLOPs FP8/MX8, and 3.5 PFLOPs BF16. Like the MTIA 400, it lists a scale-up domain size of 72, 1.2 TB/s unidirectional scale-up bandwidth, and 100 GB/s scale-out.
MTIA 500 is also focused on GenAI inference and pushes even further: a 1700W module TDP, 384–512GB HBM, and a massive 27.6 TB/s of HBM bandwidth. It’s listed at 30 PFLOPs MX4, 10 PFLOPs FP8/MX8, and 5 PFLOPs BF16, with the same 72 scale-up domain size, 1.2 TB/s unidirectional scale-up bandwidth, and 100 GB/s unidirectional scale-out bandwidth.
What changes now is the scale and the depth of the partnership. Meta says it will work with Broadcom not only on chip design, but also on advanced packaging and high-bandwidth networking—three areas that increasingly determine real-world AI accelerator performance once you start building giant clusters. Broadcom’s Ethernet and XPU platform technologies are positioned as key building blocks for linking Meta’s expanding compute fleets while maintaining the throughput required for training and serving large models.
A central detail is the power and deployment commitment: the first phase targets an AI ecosystem that exceeds 1 gigawatt. That’s not a single chip spec—it’s an indicator of how much data center capacity Meta intends to bring online using this custom silicon approach. Meta describes this as the beginning of a sustained, multi-gigawatt rollout over time, signaling that this is meant to be a recurring, multi-generation hardware roadmap rather than a one-off project.
Meta has also made it clear it intends for its custom MTIA accelerators to compete with commercially available alternatives, and it’s aiming for a yearly multi-product cadence to keep pace with its internal AI needs. In practical terms, that suggests a fast iteration cycle where new MTIA designs arrive regularly, each optimized for evolving training and inference requirements, memory demands, and cluster networking realities.
Broadcom leadership framed the agreement as the start of a multi-generation roadmap designed to support rapid growth over the coming years, highlighting its strength in AI networking and its platform for building custom accelerators. Meta’s CEO Mark Zuckerberg emphasized the goal of using this custom compute foundation to deliver “personal superintelligence” experiences at global scale, starting with more than 1GW of deployments and expanding to multiple gigawatts over time.
The takeaway is straightforward: Meta is betting that the most efficient way to scale AI for billions of users is to control more of the stack—from silicon architecture and packaging to the networking fabric that ties entire AI clusters together. With Broadcom as a long-term co-development partner, Meta is positioning MTIA as a cornerstone of its AI infrastructure strategy for years to come.
