Amid continuously rising demand for artificial intelligence computing power, Broadcom (AVGO.US) is accelerating the commercialization of its next-generation stacked AI custom chips. The company stated that, based on this advanced packaging technology, it expects to sell at least one million chips by 2027, with potential revenue reaching billions of dollars. Harish Bharadwaj, Vice President of Marketing for Custom Chips and Products at Broadcom, explained in an interview that this sales forecast is based on the company's developed 3.5D/3D stacking design, which tightly stacks two chip dies to significantly increase data transfer speeds between chips and deliver higher computing performance at lower power consumption. This technology, refined over approximately five years, has now entered a substantive implementation phase. Fujitsu has become the first customer for this technology. Bharadwaj revealed that Fujitsu is currently testing engineering samples and plans to mass-produce chips based on this stacking technology later this year for use in data centers, with potential future expansion into supercomputers. Fujitsu's chip is manufactured by Taiwan Semiconductor Manufacturing Company (TSM.US), utilizing a fusion packaging process that combines its 2-nanometer and 5-nanometer process chips. Broadcom indicated that the projected cumulative sales of one million chips by 2027 are not limited to the Fujitsu project alone but also include multiple stacked chip designs developed for other clients. Bharadwaj stated, "Virtually all of our customers are now adopting this technology." Unlike some competitors in the market, Broadcom does not design complete AI chips directly but works closely with customers, participating in the entire process from early architecture to physical implementation. Its partners include Google (GOOG.US, GOOGL.US) and OpenAI's self-developed processors. Broadcom's engineering team is responsible for translating conceptual designs into chip layouts that can be manufactured by foundries like TSMC. Driven by this model, Broadcom's AI chip business is growing rapidly. The company anticipates that its AI chip-related revenue will double year-over-year in the first fiscal quarter of this fiscal year, reaching $8.2 billion. With its customized, high-efficiency solutions, Broadcom has become a significant competitor to NVIDIA (NVDA.US) and AMD (AMD.US) in the AI computing space. From a technical perspective, Broadcom's stacking solution allows customers to flexibly combine different process nodes, with TSMC performing the fusion packaging of the upper and lower chips during the manufacturing stage. The company is developing more products based on this technology, planning to deliver two additional related chips in the second half of this year and conduct engineering validation for three more products by 2027. Furthermore, Broadcom engineers are exploring more advanced configurations, aiming to achieve complex designs involving up to eight sets of dual-chip stacks. On Thursday, chip stocks declined broadly, with Broadcom (AVGO.US) falling over 3%, Advanced Micro Devices (AMD.US) dropping over 2.4%, Taiwan Semiconductor Manufacturing Company (TSM.US) declining over 1.7%, and NVIDIA (NVDA.US) decreasing over 2.5%.
Comments