Meta and Broadcom Scale AI Ambitions with Multi-Gigawatt Custom Chip Deal
Meta Platforms and Broadcom have announced a sweeping expansion of their strategic partnership to co-develop custom AI silicon through 2029. The deal centers on the development of Meta’s Meta Training and Inference Accelerator (MTIA) chips, signaling a massive commitment to building the compute foundation necessary for “personal superintelligence.”
The Scale of the Partnership
The agreement includes a significant initial commitment to deploy 1 gigawatt of Training and Inference Accelerators. This represents the first phase of a larger, sustained multi-gigawatt rollout. According to Broadcom, the MTIA chips will be the first AI silicon to utilize a 2 nanometer process, pushing the boundaries of chip efficiency and performance.
Broadcom’s role extends beyond simple design; the company is partnering with Meta across three critical pillars:
- Chip Design: Co-developing multiple generations of custom silicon.
- Advanced Packaging: Ensuring the physical architecture supports massive workloads.
- Networking: Utilizing Broadcom’s advanced Ethernet technologies to enable high-bandwidth networking across Meta’s expanding AI compute clusters.
“Meta is partnering with Broadcom across chip design, packaging, and networking to build out the massive computing foundation we necessitate to deliver personal superintelligence to billions of people.”
— Mark Zuckerberg, Meta Co-founder and CEO
Strategic Shift: The XPU Platform
The partnership leverages Broadcom’s XPU platform, which is specifically designed for creating custom AI accelerators. By using this platform, Meta can optimize its AI infrastructure across several silicon generations, matching specific accelerators to particular workloads to balance performance and total cost of ownership.
Broadcom CEO Hock Tan emphasized the vitality of the MTIA roadmap during a March earnings call, noting that the company is shipping now and plans to scale to multiple gigawatts in 2027 and beyond.
Corporate Governance: Hock Tan Departs Meta Board
Alongside the technical announcement, Meta revealed a change in its leadership structure. Broadcom CEO Hock Tan has decided not to stand for reelection to Meta’s board. Tan joined the board in 2024 and will depart after two years of service, having notified Meta of his decision last week.
Key Takeaways
- Timeline: The partnership extends the design of custom AI accelerators through 2029.
- Capacity: Initial deployment starts at 1 gigawatt, scaling to multiple gigawatts.
- Technology: MTIA chips will be the first AI silicon to use a 2nm process.
- Scope: Covers design, packaging, and high-bandwidth Ethernet networking.
Frequently Asked Questions
What is MTIA?
MTIA stands for Meta Training and Inference Accelerator. It is Meta’s purpose-built custom silicon optimized for inference and recommendation at scale, powering AI across all of Meta’s apps and services.
Why is the 2 nanometer process significant?
A 2nm process allows for smaller transistors, which generally leads to better power efficiency and higher performance compared to previous generations of silicon.
How does this impact Meta’s AI strategy?
By co-developing its own silicon with Broadcom, Meta reduces its reliance on off-the-shelf hardware and can tailor its chips specifically for its generative AI and recommendation workloads.