0

The multi-gigawatt custom silicon deal will power Meta’s next-generation AI models and infrastructure

In sum – what we know:

• A massive compute commitment – Meta and Broadcom are rolling out a multi-gigawatt expansion of custom MTIA chips, deeply integrated with Broadcom’s Ethernet networking.
• Leaping to 2nm silicon – Broadcom aims to deliver the industry’s first 2nm AI compute accelerator, optimizing power and performance for Meta’s massive inference workloads.
• Aligning with new AI models – The hardware push coincides with Meta’s strategy shift to more compute-efficient, closed-source AI models like the newly unveiled Muse Spark.

Meta and Broadcom are taking their partnership to the next level when it comes to AI chips. The two companies have announced a new partnership to co-develop multiple generations of Meta’s Training and Inference Accelerator (MTIA) custom silicon. 

The initial commitment alone clocks in at over 1 gigawatt of compute capacity, which is just the first phase of what both companies are calling a sustained multi-gigawatt rollout. There’s no specified ceiling on the overall scope, either.

Under the hood, everything is built on Broadcom’s XPU (Custom Accelerator) platform, which tightly couples logic, memory, and high-speed I/O into a cohesive, optimized system purpose-built for AI workloads. Broadcom is also supplying the networking backbone for Meta’s AI clusters — high-radix switches, optical connectivity, PCIe switches, and high-speed SerDes capabilities, all running on advanced Ethernet technologies. This networking piece is arguably just as important as the chips themselves. As AI clusters scale, the interconnect fabric between accelerators increasingly becomes the bottleneck, and Broadcom’s Ethernet stack is engineered to handle scale-up (within racks), scale-out (across nodes), and scale-across (network capacity) demands.

On the deployment front, these MTIA accelerators will drive workloads across Meta’s core products, including WhatsApp, Instagram, and Threads — powering ranking, recommendations, and real-time generative AI features. Meta has positioned the partnership as foundational to delivering what it calls “personal superintelligence” to billions of users worldwide, though that particular framing deserves a healthy dose of skepticism given how loosely “superintelligence” gets tossed around in this industry.

The transition to 2nm process nodes

The most technically consequential detail in this announcement is that Broadcom will deliver the industry’s first 2nm AI compute accelerator for Meta. If that timeline holds, it would mark a genuine leap ahead of the current state of play in custom AI silicon. Most custom-built AI chips currently run on 4nm or 5nm process nodes.

Smaller process nodes generally deliver better transistor density, improved power efficiency, and higher performance per watt. For Meta’s specific use case, where the MTIA chips are optimized for inference, recommendations, and low-precision workloads, efficiency gains from a smaller node translate directly into cost savings and throughput improvements at massive scale.

That said, bleeding-edge process nodes have historically come with lower yields and steeper per-wafer costs in their early production stages. TSMC’s 2nm (N2) process is still relatively nascent, and whether Broadcom and Meta can hit volume production at the required scale and economics is uncertain. The multi-generation, multi-year structure of the partnership suggests both companies understand this and are planning for iterative improvement rather than going all-in on a single tape-out.

Broadcom’s platform approach also matters a great deal here. A faster chip on a better node doesn’t buy you much if the surrounding system can’t keep pace. Because Broadcom’s involvement spans chip design, advanced packaging, and networking, the partnership has a full-stack character that goes well beyond simply shrinking transistors — and that kind of system-level integration is where a lot of the real performance gains will come from.

Meta’s big AI push

As part of the announcement, Broadcom President and CEO Hock Tan is stepping down from Meta’s Board of Directors — where he served for two years — and into a dedicated advisory role focused on Meta’s custom silicon roadmap and infrastructure investments. The move deepens the relationship between the two companies while neatly sidestepping the governance complications that come with a major vendor’s CEO sitting on a customer’s board.

Meta has been pushing on AI for years now, but it hasn’t always kept pace with the other hyperscalers and other AI companies, especially when it comes to the quality of its models. Google’s Gemini, ChatGPT, and Claude have long been considered ahead of Meta’s Llama models — though Meta has been at the forefront of open-source model development.

But there are real signs that Meta’s AI strategy is shifting into a more competitive gear. The company recently unveiled Muse Spark, a new model out of Meta Superintelligence Labs under Chief AI Officer Alexandr Wang. Unlike Meta’s earlier Llama models, Muse Spark is closed-source and represents a significant architectural departure. It will feature a “Contemplating Mode” that orchestrates multiple reasoning agents in parallel — a different approach from the “think longer” test-time scaling strategies used by competitors like GPT-5.4 and Gemini Deep Think. The model was also built as natively multimodal from pre-training, rather than having vision capabilities grafted on after the fact, and it incorporates what Meta calls Visual Chain-of-Thought reasoning. Most relevant to the Broadcom deal, Muse Spark was designed around extreme token efficiency, leveraging a technique called thought compression to hit frontier-level performance with dramatically less compute.

Whether Muse Spark genuinely competes with the best from OpenAI and Google in real-world deployment — not just on benchmarks — remains an open question. But pairing a legitimately competitive model with a massive custom silicon pipeline at 2nm is a credible strategy. If Meta can build inference-optimized chips at scale on a leading-edge node and match them with models engineered from the ground up for compute efficiency, it could establish a meaningful cost advantage in serving AI to billions of users.

Of course, custom silicon programs are expensive, complex, and notoriously slow to pay off. Google has been iterating on TPUs since 2015 and still leans on NVIDIA hardware for many workloads. Meta is placing a big bet here, and the Broadcom partnership gives it a credible path forward — but executing on multi-gigawatt custom silicon deployments across 2nm and beyond is the kind of challenge that plays out over years, not quarters.