Broadcom (AVGO) is still considered the “best-in-class” in the custom application specific integrated circuit market, but Marvell (MRVL), Arm (ARM), and perhaps surprisingly, Qualcomm (QCOM), are making inroads, Evercore said.
“AVGO mentioned as both best-in-class in IP and execution as well as the priciest,” analyst Mark Lipacis wrote in a note to clients. “MRVL is viewed as highly capable and believed to have AWS’s Trainium 4, some of Trainium 3 and MSFT’s Maia. ARM (ARM) is also competing aggressively for CPU, DPU, and XPU programs, and QCOM is viewed as making progress in the AI-ASIC market, with its chips running small language models in Meta’s Ray Bans and other AI wearables.”
Delving deeper, Broadcom is considered the “premium, one-stop shop” for ASICs, given that it has leading intellectual property in both analog and digital and is the best at execution, Lipacis added. The company also has strength in its high-speed I/O offering, and its experience working with Taiwan Semiconductor (TSM) and its deep, technical bench are both assets, Lipacis explained. Conversely, its high cost (margins are 10% higher than Marvell) and its lack of flexibility are considered negatives, he added.
For Marvell, its flexibility compared to Broadcom is considered a major strength, Lipacis added, notable, given its intellectual property assets are not as strong as Broadcom’s. However, it does have advantages over its Taiwanese competitors, notably its “superior front-end circuit design capabilities,” Lipacis explained.
It also has two major customers—Amazon (AMZN) and Microsoft (MSFT)—with its Amazon work improving, as the defect rate may have been cut by 65% when it took over work on Trainium 2. It is also seen as “indispensable” for Microsoft’s Maia program. (Microsoft released a new Maia AI accelerator last month.)
Arm, on the other hand, is working to “aggressively” compete for development programs at the major hyperscalers and show off its capabilities in CPUs, XPUs and data processing units. It’s also believed to be working with OpenAI (OPENAI) on a custom CPU or DPU and Meta (META) on a custom CPU, Lipacis said.
Lastly, Qualcomm (QCOM), which has made some progress in the competitive ASIC market, has shown that its offerings are capable of running small language models on “most devices,” Lipacis said.