Meta expands multibillion-dollar Broadcom alliance to build its own AI chips

Meta has expanded its multibillion-dollar partnership with Broadcom, betting on co-developing custom chips for its AI systems. The corporate link between the two companies is changing as well: Hock Tan is leaving Meta's board.

A new level for the alliance

The expanded agreement between Meta and Broadcom takes their relationship beyond the usual supplier-customer arrangement. The companies plan to co-design and produce specialized chips more deeply, and those chips will run inside Meta's AI infrastructure. For a company that simultaneously develops ad algorithms, recommendation systems and generative models, this is a question not only of performance but of control over the entire compute chain, from hardware to software stack.

The wording about a multibillion-dollar partnership shows the scale of the bet. Meta has long been increasing spending on data centers and accelerators, and Broadcom gets a chance to establish itself as one of the key technology partners in the most expensive segment of the AI market. The industry used to talk mainly about purchases of general-purpose GPUs; now attention is shifting toward custom solutions built for the specific internal workloads of a large platform.

Why Meta wants its own chips

Meta does not need its own or co-designed accelerators for the sake of image. At this volume of compute, even a small optimization in power consumption, memory bandwidth or the cost of a single server quickly turns into millions of dollars in savings. Custom hardware also gives more freedom in how exactly to run model training, inference and the services that serve billions of users.

This approach offers Meta several practical advantages:

• more precise tuning of chips for its own AI workloads
• less dependence on general-purpose solutions and external roadmaps
• more predictable supply planning for data centers
• a chance to lower total infrastructure cost as it scales
• the ability to tie hardware more closely to its internal AI stack and network architecture

This does not mean an immediate abandonment of third-party accelerators: large AI companies usually build a hybrid stack and combine different types of compute hardware. But the trend is clear: the more money goes into models and servers, the stronger the interest in in-house chips. This dynamic benefits Broadcom, because the company knows how to act as a partner for custom design and manufacturing of complex semiconductor solutions.

Changes on the board

The departure of Hock Tan from Meta's board is a separate thread of the news. Against the backdrop of the expanding commercial partnership, the move looks logical in terms of corporate hygiene. When the head of one of the most important suppliers or technology partners sits on a client's board, the market inevitably asks questions about conflicts of interest, access to sensitive information and the limits of influence. In this configuration, separating the roles may even be better for both sides.

Meta gets a cleaner governance structure at a moment when it is making a big bet on AI hardware. Broadcom, in turn, keeps and grows its commercial contract without the extra burden of discussions about corporate control. The very fact that the board change happens at the same time as the deepening of the hardware alliance underlines that this is not about distancing, but about a more formalized model of cooperation.

What this means

The AI market is increasingly moving from a race for models alone to a race for infrastructure. Meta's decision to deepen its alliance with Broadcom shows that the next field of competition is in-house chips, energy efficiency and control over the cost of compute. This is where margins, the speed of launching new services and dependence on outside suppliers will be decided.

*Meta has been designated an extremist organization and is banned in Russia.