One notable gap in Meta's announcement: no generative AI training chip, lending weight to reports that its most ambitious custom silicon project has been shelved for now.
In a recent announcement, the Magnificent Seven tech giant Meta Platforms Inc (NASDAQ:META) unveiled four customized artificial intelligence (AI) chips. This comes on the heels of semiconductor design behemoth Broadcom Inc (NASDAQ:AVGO) earnings report, where CEO Hock Tan made it a point to specifically address Meta.
Now, Meta is giving a wink back to Broadcom in a development that has clear positive implications for AVGO. Still, there are other negative factors worth noting. What does this news mean for Broadcom going forward?
For some time, market watchers and analysts have generally believed that Meta is one of Broadcom's custom AI processor buyers. However, Broadcom itself has never referred to Meta in this capacity during its earnings calls, until now. In Broadcom's Q1 2026 call, Tan said, "Contrary to recent analyst reports, Meta's custom accelerator MTIA road map is alive and well. We're shipping now."
MTIA, which stands for Meta Training and Inference Accelerator, is a custom chip lineup developed in collaboration with Broadcom. Tan's statement came after reports that Meta halted the development of its most advanced custom AI training chip, codenamed Olympus.
Notably, Meta specifically mentioned Broadcom in its custom chip press release. The company said, "Meta Training and Inference Accelerator (MTIA), our family of homegrown AI chips developed in close partnership with Broadcom, has remained and will continue to be an important part of Meta's AI infrastructure strategy."
Markets already generally accepted this partnership, but the fact that both companies are specifically acknowledging this now removes any doubt.
The title of Meta's post, "Four MTIA Chips in Two Years: Scaling AI Experiences for Billions," provides direct support for the bullish scenario Broadcom laid out in its earnings. Hock Tan noted that the majority of its customers are moving to develop two custom chips with AVGO a year, the exact pace of development that the post describes. This lends legitimacy to Tan's statements and the idea that Broadcom is deepening its relationships with customers.
Meta is using MTIA for a variety of purposes. This includes training and inference on its ranking and recommendation (R&R) models. Training is the process of developing more intelligent models, while inference is the process of deploying those models to answer questions and perform tasks.
However, it is also important to note that Meta's announcement did not include a GenAI training chip. This adds weight to reports that Meta is rolling back Olympus development. Meta's Chief Financial Officer, Susan Li, also recently made a statement that aligns with this at the Morgan Stanley Technology Conference. She said Meta "expects" and is "hopeful" that it can expand its use of custom silicon to train AI models "eventually."
This is a negative for Broadcom, which would presumably also be co-developing Olympus withMeta. However, it is good to see that Meta isn't saying its GenAI training chip ambitions are dead; it still wants to develop them over time. Still, the timeline for Broadcom generating significant revenue through that project may have just become much longer.
Overall, Meta's relationship with Broadcom is now fully substantiated and appears to be growing significantly outside of GenAI training.
Importantly, many expect inference to overtake training as the dominant AI workload in the coming years. McKinsey predicts that inference will grow by a compound annual rate of 35% over the next five years, accounting for over half of AI compute by 2030.
This supports Broadcom's outlook, as its relationship with Meta -- especially when it comes to inference -- is clearly deepening.