Meta and Broadcom announced a multi-year strategic partnership on Tuesday aimed at the joint development of multiple generations of Meta’s custom silicon chips, known as MTIA. The partnership includes an initial rollout exceeding 1 gigawatt of capacity, with ambitious plans to scale this range to several gigawatts sustainably through 2029.
This deal marks a major deepening of the current cooperation between the two tech giants and establishes Broadcom’s advanced engineering platform as a fundamental backbone for Meta’s AI computing infrastructure, which is expanding rapidly to meet growing global demand for complex data processing.
Technical and logistical agreement details
Under this strategic agreement, Broadcom will provide advanced technology including chip design, advanced packaging, and Ethernet-based networking solutions. These joint efforts aim to support current and future versions of the MTIA chip family, which are custom training and inference accelerators developed by Meta specifically for its services.
The MTIA chip lineup is built primarily to handle inference workloads and recommendation algorithms across Meta’s various apps and services. As technology evolves, newer models like MTIA 450 and MTIA 500 will expand to include complex generative AI inference tasks, with full-scale deployment of these advanced models scheduled to begin by 2027. This step is essential for running the large language models that power new interactive features in the company’s social media apps.
Leadership vision for the future of computing
Commenting on this important partnership, Meta CEO Mark Zuckerberg said: “Meta is collaborating with Broadcom across chip design, packaging, and networking to establish a massive computing base essential for delivering personalized superintelligence to billions of users globally.” This vision underscores Meta’s clear ambition to integrate advanced AI capabilities into the daily lives of its users.
For his part, Broadcom CEO Hock Tan described the initial rollout of MTIA chips as “just the beginning of a sustainable roadmap spanning multiple generations.” Tan highlighted the pivotal role his company plays in providing AI networking and designing custom accelerators tailored to each company’s exact technical needs.
Meta’s hardware diversification strategy
This partnership reflects a broader trend by Meta to diversify its AI hardware and gradually move away from traditional commercial processors. Last March, Meta revealed the development of 4 models of MTIA chips: 300, 400, 450, and 500, in close partnership with Broadcom. Some of these chips are already operating successfully in the company’s live production environment.
The company stated that the MTIA 450 chip delivers performance “significantly higher than current leading commercial products” regarding generative AI inference operations. Meta confirmed its current capability to produce a new custom chip approximately every 6 months, reflecting exceptional flexibility and speed in the hardware development cycle. This technological shift not only helps improve performance efficiency but also significantly reduces the massive operational costs associated with purchasing off-the-shelf electronic components from external suppliers.
Capabilities limits and continued reliance on Nvidia
Despite this impressive progress in internal innovation, the MTIA chip lineup still faces some technical limitations. Previous reports from earlier this year indicated that Meta had halted work on its most ambitious custom training chip, codenamed Olympus. This means the company will continue, for now, to rely heavily on Nvidia GPUs for large-scale model training tasks.
Model training requires processing vast amounts of data, a task where Nvidia remains the undisputed leader. In contrast, inference is limited to using those models after training to generate content or answer queries. Therefore, the current focus of Meta’s custom chips is on the inference side, which constitutes the bulk and ongoing part of daily usage for its services.
Broadcom as a key architect in the tech landscape
Meta’s latest deal adds to a continuous series of major agreements struck by Broadcom in custom silicon. Earlier this month, the company disclosed in an official financial filing that it signed a long-term agreement to develop and supply future generations of tensor processing units (TPUs) for Google, in a deal extending through 2031.
Furthermore, Broadcom expanded its partnership with leading AI firm Anthropic to provide approximately 3.5 gigawatts of TPU-based computing capacity, starting in 2027. The market reacted positively to these moves, with Broadcom shares rising significantly following the announcement of these consecutive partnerships, first reported by CNBC.
