Meta and Broadcom extend AI chip contract through 2029, prepare 2-nm MTIA

Meta locked in one of the key elements of its AI infrastructure for years to come: the company extended and expanded its agreement with Broadcom on custom MTIA chips through 2029. This is no longer a one-time project, but a long product line spanning multiple generations of processors. The initial capacity of the new program exceeds 1 gigawatt of compute, and the deal itself is named as the first phase of a sustainable multigigawatt rollout.

For the market, this is a signal that Meta is ready to build its own hardware stack at hyperscaler scales, rather than limiting itself to purchasing generic accelerators. The expanded partnership covers multiple generations of MTIA — the company's proprietary line of AI processors. The key technical detail is the transition to 2-nm process technology.

If plans are executed as announced, these will be the first custom AI chips of this class on 2 nm. For Meta, this matters not only as a question of performance. A finer process technology usually means a better balance between transistor density, power consumption, and speed, and in AI workloads this directly affects the cost of each request, the speed of model inference, and the efficiency of data centers.

Against the backdrop of growing power consumption of AI clusters, this parameter becomes almost as important as peak computational power. Broadcom in this scheme acts as a key partner in design and scaling the solution. Meta has the money, infrastructure, and massive demand from its own products — from recommendation systems and advertising to generative services and internal AI tools.

But turning the ambition "we're building our own chip" into a multiyear serial production program without a strong semiconductor partner is extremely difficult. That's why extending the deal through 2029 looks like an attempt to lock in the roadmap early, reduce dependence on market volatility, and synchronize chip development with the growth of compute clusters. The scale from which the new phase begins is particularly telling: more than 1 GW of computational power.

In corporate announcements, such figures often sound abstract, but in the context of AI infrastructure they mean no longer laboratory tests and limited pilots, but industrial deployment. The phrasing about "the first phase of a sustainable multigigawatt rollout" hints that Meta is laying groundwork for years to come and preparing not one successful chip, but a repeatable platform. This is especially important now, when the largest tech companies are competing not only for better models, but for control over the entire AI creation chain — from the data center to the silicon.

Meta's logic is clear. Generic accelerators remain critically important, but the more the company spends on training and maintaining its own models, the stronger the incentive to optimize part of the load for itself. Custom chips allow for more precise architecture tailoring to specific types of operations, better power management, and potentially lower total cost of ownership in large clusters.

For products at Meta's scale, even modest efficiency improvements at the level of a single server or inference pipeline translate into enormous savings across the entire data center network. Plus it's a strategic lever in negotiations with external suppliers: when you have a working proprietary line, dependence on others' roadmaps becomes lower. That's why large platforms increasingly invest in their own silicon alongside purchases in the open market.

The conclusion here is simple: the AI race stopped being just a race of models long ago. The extension of Meta and Broadcom's partnership through 2029 shows that the next level of competition is control over the compute base, energy, and the pace of infrastructure deployment. If Meta truly delivers multiple generations of 2-nm MTIA at multigigawatt scale, it will reinforce the vertical integration trend among Big Tech and further divide the market between those who buy others' hardware and those who design mission-critical chips for themselves.