TSMC collaborates with Marvell to create an unparalleled "2nm + Silicon Photon" project that captures the global market share of ASIC chips.

TSMC collaborates with Marvell to set a foothold for customized ASIC chips in the AI era, from 2nm processes to silicon photonics. (Background: The U.S. chip blockade strikes hard! WSJ: Samsung and TSMC face technical exemptions cancellation in China, U.S. equipment stocks sound the alarm.) (Background: TSMC's Arizona factory ships its first wafers! NVIDIA's AI chips need to return to "Taiwan's hometown" for packaging.) Artificial Intelligence (AI) is expanding at an unprecedented speed, and cloud service providers are shifting from general-purpose GPUs to customized application chips (ASIC) to reduce power consumption and costs. At this critical moment of changing lanes for overtaking, TSMC, with which we are familiar, has announced a deepened collaboration with high-speed network chip giant Marvell, targeting processes below 3nm and next-generation silicon photonics technology. TSMC's announcement has effectively dropped a bomb in the AI chip market, invisibly forecasting an explosive growth for ASICs, giving a feeling of an invincible declaration. The scale of ASICs is accelerating, pushing the computation and energy efficiency of large AI model training to the limit, prompting data centers to seek more specialized and energy-efficient solutions. According to Fortune Business Insights, the market size for ASICs could reach 22.78 billion USD by 2025, challenging 36.8 billion USD by 2032, with a compound annual growth rate of about 7.1%. Compared to GPUs, customized ASICs have significant advantages in power consumption, unit cost, and heat dissipation, becoming the new favorites in cloud and edge computing. This point should be well understood by those with experience in crypto mining. Marvell's layout results have emerged rapidly in recent years. The company secured its first ASIC order from AWS in 2021 and announced a partnership with Meta to design AI chips in October 2024. CEO Matt Murphy has even named the "third" cloud customer order is about to be secured: We will soon complete our third ASIC contract with another major cloud company. TSMC x Marvell: Advanced Processes + Silicon Photonics TSMC holds over 60% of the global foundry market share and leads its peers in mass production of advanced processes. Currently, TSMC has multiple 12-inch, 8-inch, and 6-inch factories in Taiwan, with an annual production capacity of approximately 17 million 12-inch wafers. Marvell announced that its flagship AI ASICs will directly utilize TSMC's processes below 3nm, and even cross into the 2nm node. For TSMC, this long-term contract can further solidify the load of advanced processes and bring stable demand for CoWoS high-end packaging. The spotlight for investors is on silicon photonics technology: traditional electrical signals face bandwidth bottlenecks and thermal burdens. TSMC has introduced the COUPE (Compact Universal Photonic Engine) architecture, stacking electrical bare chips and photonic bare chips to complete optoelectronic integration within the package. The silicon photonics solution is expected to complete verification by 2025 and enter CoWoS mass production in 2026, with bandwidth expected to increase tenfold while significantly reducing latency and power consumption. Broadcom leads, Marvell targets growth gaps In the ASIC battlefield, Broadcom still holds the top spot with a market share of 55%-60%, but Marvell, leveraging flexible acquisitions and cloud customer relationships, is rapidly catching up, with a current market share of about 15%. 36Kr reports estimate that Marvell's AI-related revenue will exceed 1.5 billion USD in 2024 and reach up to 2.5 billion USD in 2025. The company has also revised its total addressable market for customized AI chips (TAM) in 2028 from 43 billion USD to 55 billion USD. The key to Marvell's rapid expansion lies in its focus on cloud ASIC design and collaborative research models with customers. Broadcom emphasizes platform integration to provide customers with one-stop solutions, while Marvell specializes in high-speed networking, storage, and data center switching chips, further strengthening its IP portfolio through acquisitions such as Cavium, Avera, and Innovium, creating a differentiated path. From a broader perspective, AI application landing is forcing the semiconductor industry chain into competition in "computing power, networking, and packaging". TSMC is laying out in both processing and packaging fields, while Marvell captures high-speed switching and network interface technology, with both companies flowing together, likely defining the next generation of AI chip standards. This collaboration not only rewrites the landscape of the ASIC market but also intensifies the deep reliance of cloud giants on TSMC. Related Reports The U.S. chip blockade strikes hard! WSJ: Samsung and TSMC face technical exemptions cancellation in China, U.S. equipment stocks sound the alarm. Apple AI shifts to "automated design chips", with M6 and A20 chips expected to enhance energy efficiency. The article "TSMC Teams Up with Marvell to Dominate Global ASIC Chip Market with '2nm + Silicon Photonics' Plan" was first published in BlockTempo, the most influential blockchain news media.