Broadcom’s custom‑silicon surge

Broadcom is having a very different artificial intelligence boom from Nvidia’s. Instead of mostly selling the same chip to everyone, Broadcom is getting paid to help a few giant customers build chips that fit their own data centers like a tailored suit. (broadcom.com) That showed up in the numbers on March 4, when Broadcom said artificial intelligence semiconductor revenue hit $8.4 billion in its fiscal first quarter of 2026, up 106% from a year earlier, and forecast $10.7 billion for the next quarter. Chief executive Hock Tan said the jump came from custom artificial intelligence accelerators and networking gear. (broadcom.com) A custom accelerator is a chip built for one customer’s workload instead of for the whole market. If a general-purpose graphics processing unit is a rental truck, a custom accelerator is a van built around one company’s exact route, cargo, and fuel budget. (cnbc.com) Meta is one clear example of why that model is growing. In March, Meta said its Meta Training and Inference Accelerator program is central to its artificial intelligence infrastructure plan, and it expects four new generations of those chips within two years. (about.fb.com) Meta also said those chips are aimed at concrete jobs inside Facebook, Instagram, and its generative artificial intelligence stack, including ranking, recommendations, and inference. Those are repetitive workloads at Meta’s scale, which is exactly where a specialized chip can beat a more flexible one on cost and power. (about.fb.com) Meta did not hide who helped build them. Its March post said the latest Meta Training and Inference Accelerator chips were developed in close partnership with Broadcom, tying Broadcom directly to one of the world’s biggest buyers of artificial intelligence compute. (about.fb.com) Broadcom says this is no longer a one-customer story. On its earnings call, Tan said the company’s custom accelerator business spans five customers, with Google’s seventh-generation Tensor Processing Unit already ramping in 2026 and multiple other programs moving through development. (fool.com) The latest investor note pushing the stock higher goes further and says Broadcom has now mass-deployed 3 nanometer custom silicon for Meta and ByteDance. That specific ByteDance claim is circulating in market commentary on April 8, but it has not been confirmed in public filings or a company statement the way Meta’s Broadcom partnership has. (markets.financialcontent.com) (about.fb.com) The 3 nanometer detail matters because smaller process nodes usually let chip designers pack in more transistors while cutting power per unit of work, but they also make manufacturing harder and more expensive. The reward is biggest for hyperscale companies that can spread that design cost across millions of identical tasks in giant server fleets. (markets.financialcontent.com) (broadcom.com) That is why Broadcom keeps talking about co-design, packaging, and networking in the same breath. A custom chip is only useful if it arrives with the memory, interconnect, and manufacturing flow needed to ship at volume, and Broadcom is selling that whole system rather than just a piece of silicon. (broadcom.com 1) (broadcom.com 2) Tan’s forecast now is that Broadcom’s artificial intelligence chip revenue in 2027 will be significantly above $100 billion. That is a huge bet that the biggest cloud and consumer internet companies will keep moving part of their artificial intelligence spending away from off-the-shelf chips and into hardware they control themselves. (cnbc.com)