Meta and Broadcom extend AI chip deal to 2029, prepare 2nm MTIA

Meta has locked in one of the key pieces of its AI infrastructure for years to come: the company has extended and expanded its agreement with Broadcom on custom MTIA chips through 2029. This is no longer a one-off project but a long product line spanning several processor generations. The new program starts at more than 1 gigawatt of compute, and the deal is described as the first phase of a sustained multi-gigawatt rollout. For the market, it signals that Meta is ready to build its own hardware stack at hyperscaler scale rather than limit itself to buying general-purpose accelerators.

The expanded partnership covers several generations of MTIA, the company's own line of AI processors. The main technical detail is the move to a 2nm process. If the plans are carried out as announced, these will be the first custom AI chips of this class at 2nm.

For Meta, this matters for more than performance. A finer process node usually means a better balance of transistor density, power consumption and speed, and in AI workloads that directly affects the cost of each request, model inference speed and data center efficiency. As the power draw of AI clusters grows, this parameter is becoming almost as important as peak compute.

Broadcom is the key partner in this arrangement for design and for scaling the solution. Meta has the money, the infrastructure and huge demand from its own products, from recommendation systems and advertising to generative services and internal AI tools. But turning the ambition of "we make our own chip" into a multi-year production program is extremely hard without a strong semiconductor partner. That is why extending the deal to 2029 looks like an attempt to lock in the roadmap early, reduce exposure to market volatility and synchronize chip development with the growth of compute clusters.

The scale at which the new phase begins is telling in its own right: more than 1 GW of compute capacity. In corporate announcements such figures often sound abstract, but in the context of AI infrastructure they mean not lab tests or a limited pilot but industrial deployment. The wording "first phase of a sustained multi-gigawatt rollout" hints that Meta is planning for growth years ahead and is preparing not one successful chip but a repeatable platform. That matters especially now, as the largest tech companies compete not only for the best models but for control of the entire AI supply chain, from the data center to the silicon.

Meta's logic is clear. General-purpose accelerators remain critically important, but the more a company spends on training and serving its own models, the stronger the incentive to optimize part of the workload for itself. Custom chips make it possible to tailor the architecture to specific types of operations, manage power consumption better and potentially lower total cost of ownership in large clusters.

For products at Meta's scale, even a small efficiency gain on a single server or a single inference pipeline turns into huge savings across the whole network of data centers. It is also a strategic lever in negotiations with outside suppliers: once you have a working in-house line, dependence on other companies' roadmaps drops. This is why large platforms increasingly invest in their own silicon alongside purchases on the open market.

The takeaway is simple: the AI race stopped being only a race of models a long time ago. The extension of the Meta-Broadcom partnership to 2029 shows that the next level of competition is control over the compute base, energy and the pace of infrastructure deployment. If Meta really does bring several generations of 2nm MTIA to multi-gigawatt scale, it will strengthen the trend toward vertical integration among Big Tech and split the market even more sharply between those who buy someone else's hardware and those who design their critical chips themselves.

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