Broadcom, Marvell to 'dominate' custom AI chip market, J.P. Morgan says
Jun. 04, 2024 8:34 AM ET
By: Chris Ciaccia, SA News Editor
 da-kuk
Broadcom (NASDAQ: AVGO) and Marvell Technology (NASDAQ: MRVL) should "dominate" the high-end of the fast-growing custom chip market, investment firm J.P. Morgan said on Tuesday.
The research firm now expects the high-end of the application-specific integrated circuit, or ASIC, market to hit between $20B and $30B — up from a prior view of $20B to $25B — and grow at a 20% compound annual growth rate, due largely to generative artificial intelligence.
"Custom ASICs are a strong alternative for compute accelerators vs using off-the-shelf merchant solutions (Nvidia H100/B200) as cloud/Hyperscale titans (Google, Meta/Facebook, Microsoft, Amazon, etc.) continue to look for more differentiation, better performance, lower power consumption and lower overall silicon costs," analyst Harlan Sur wrote in an investor note.
Sur believes that Broadcom is working with Google ( GOOG) ( GOOGL) on its next-gen TPU v7, which should secure the partnership between the two companies and revenue for Broadcom through 2026 and 2027. Google unveiled the sixth-generation of its tensor processing unit chips at its developer conference last month.
The major hyperscalers, including Google, Meta Platforms ( META), Microsoft ( MSFT) and Amazon ( AMZN) have all announced their own custom silicon programs, partially to reduce reliance on companies like Nvidia ( NVDA). As such, J.P. Morgan believes Broadcom and Marvell will benefit, with Broadcom having roughly 60% of the custom ASIC market and Marvell accounting for 15%.
"For Broadcom, we believe the team could drive ~$11B+ in AI revenues this year (Google/Meta AI ASICs + AI networking) drive by accelerating orders at Google (Broadcom/Google co-designed the last 6 generations of its Google's TPU family of AI processors) and with Meta set to become a multi-billion dollar opportunity over the next two years as Meta builds out their compute, networking, and storage architecture to support its next-gen AI compute infrastructure," the analysts wrote.
Marvell could generate between $1.6B and $1.8B in AI revenue from ASICs and networking this year and between $2.8B and $3B next year, as it works with Amazon on its 5 nanometer Tranium chip and Google on its 5 nanometer Axion ARM CPU chip, the analysts added.
Marvell is also believed to be working with Amazon on its Inferentia chip and Microsoft on its Maia chip. |
