TradingKey - As AI computing power demand expands from model training to inference and AI agent applications, major cloud service providers are accelerating the deployment of custom AI chips such as ASICs, driving demand growth across chip design, advanced process nodes, and packaging and testing.

In the industry chain, Broadcom (AVGO) and Marvell (MRVL) participate in custom ASICs and related chip platforms, respectively; TSMC (TSM) provides advanced-process wafer foundry services and advanced packaging; while Amkor (AMKR) focuses on OSAT and advanced packaging and testing. So, for investors, besides primary supplier Broadcom, which among Marvell Technology, TSMC, and Amkor is more worthy of attention?

What Is a Custom AI Chip?

AI custom chips are application-specific integrated circuits (ASICs) developed for specific customers, models, or computing tasks, which can optimize performance, energy efficiency, and cost per unit of compute based on specific workloads.

In August 2026, TrendForce raised its forecast for global AI server shipment growth from approximately 28% to nearly 31%. The firm previously projected that ASIC servers would account for 27.8% of global AI server shipments in 2026, with shipment growth outpacing that of GPU servers.

Some market research firms estimate that Broadcom holds a substantial share of the custom AI accelerator design service market. As Google (GOOGL), Amazon (AMZN), Microsoft (MSFT), and Meta (META) expand the deployment of their self-developed chips, investments in AI computing power are extending to segments such as chip design, wafer foundry, and advanced packaging, drawing market attention to the order and revenue growth potential of relevant companies.

Marvell Technology (MRVL): Major Player in Custom AI Chips

Marvell Technology is one of the key custom AI chip partners worth watching alongside Broadcom. The company has long been involved in custom chip projects for major cloud service providers and continues to expand its collaboration with hyperscalers.

In 2026, Marvell further expanded its custom AI product partnership with Google, with a business scope covering AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory computing products, further covering the AI infrastructure surrounding the Google TPU ecosystem.

From a financial perspective, Marvell's data center revenue in the latest quarter reached approximately $2.17 billion, up 46% year-over-year and accounting for about 79% of total company revenue. The company stated that AI-related orders remain strong and expects revenue growth to accelerate further in subsequent quarters.

Regarding the custom chip business, management expects it to more than double in FY2028 and potentially surpass $10 billion in FY2029. As of early September, MRVL's share price stood at around $212, with some market statistics showing an average analyst target price of about $285, representing a potential upside of approximately 35%.

On the risk front, investors still need to pay attention to customer concentration, advanced process nodes and tape-out costs, as well as competitive pressures in the custom AI chip market from rivals like Broadcom.

In addition, the timeline for revenue realization from the new Google partnership is also worth monitoring; Marvell management previously stated that large-scale revenue contributions are expected to become more pronounced only in FY2029.

TSMC (TSM): Core Segment in Wafer Foundry and Advanced Packaging

TSMC is not primarily engaged in ASIC chip design, but as a leading global semiconductor foundry, it serves as an important manufacturing and advanced packaging partner for many AI chip companies.

In the second quarter of 2026, TSMC's revenue reached $40.2 billion, up 33.7% year-over-year to a new quarterly high; HPC business accounted for 66% of revenue, and demand for advanced nodes such as 2nm and 3nm continued to grow. Specifically, 2nm and 3nm accounted for 3% and 30% of wafer sales, respectively, totaling 33% combined.

As AI accelerators evolve toward larger chip sizes and higher HBM integration, the importance of advanced packaging continues to rise. TSMC's CoWoS demand is driven by the AI boom, and related capacity continues to expand but remains tight, making it a key link to monitor in the AI accelerator supply chain.

As of early September, TSM's stock price stood at around $417, with market statistics showing an average analyst price target of approximately $554, representing a potential upside of about 33%. Major risks include high capital expenditures, investments in capacity expansion for advanced nodes and advanced packaging, and geopolitical factors.

Amkor (AMKR): US-Headquartered OSAT Leader

Amkor is a major global OSAT company and the world's largest U.S.-headquartered packaging and testing provider. In the second quarter of 2026, the company's net sales reached $1.9 billion, up 26% year-over-year, with diluted earnings per share at $0.70.

In June 2026, Amkor signed a 10-year long-term agreement with TSMC, focusing on building advanced packaging and testing capacity in Arizona, U.S. In July, the company announced a multi-year strategic partnership with Nvidia, under which Nvidia will provide advance payments to support Amkor in expanding its U.S. advanced packaging capacity, with the agreement reaching $1.5 billion.

Sustained growth in demand for AI and HPC drove Amkor's computing business to record quarterly revenue. Management stated that average capacity utilization across its overall manufacturing network rose from the 50%-plus range to the 70% range in the first half of 2026, further reaching the high-70% range in the second quarter, with certain technology platforms operating near or at full capacity.

As of the close on September 3, AMKR stock was priced at $46.94. Market statistics show that the average analyst price target is about $75 to $76, representing potential upside of approximately 60%.

Regarding risks, the new Arizona facility may face depreciation and startup cost pressures during the initial capacity ramp-up phase. In addition, the relocation of certain SiP products and equipment from South Korea to Vietnam may pose execution challenges, including equipment transfer, requalification, and capacity ramp-up.

Comparison of Investment Attributes of Three Companies

Conclusion: Grasp Division of Labor Differences in the ASIC Industry Chain

From the perspective of its current position in the industry chain, Marvell's earnings growth elasticity is relatively higher, particularly as its custom chip business enters an accelerated growth phase, which is expected to become a major growth engine for the company over the next few years; however, revenue realization from certain large customer projects will still take time.

TSMC holds the most solid position in the industry chain, benefiting from growing demand for advanced nodes, HPC, and advanced packaging, giving it relatively higher business certainty; Amkor, on the other hand, represents a high-elasticity advanced packaging target, possessing significant earnings elasticity amid sustained growth in AI packaging demand, though new capacity expansion and ramp-up also introduce higher uncertainty.

The three companies occupy key links in the ASIC industry chain, namely design, wafer fabrication, and packaging and testing. Rather than simply judging who can replace GPUs, it is better to focus on the investment opportunities brought by different technological routes and supply chain segments as the AI chip market continues to expand.

For investors, decisions should still be made by combining valuation levels, the speed of order realization, earnings growth certainty, and individual risk tolerance.

Find out more