Is Takata Semiconductor (TSEM.US) undervalued? Wedbush says silicon photon production expansion will open up room for AI infrastructure growth

The Zhitong Finance App learned that the latest research report released by Wedbush Securities, a well-known investment agency on Wall Street, shows that TSEM.US (TSEM.US), a leader in semiconductor manufacturing and foundry headquartered in Israel, announced a large-scale expansion of its large-scale semiconductor factory in Uozu, Japan, which will largely help it achieve the company's current conservative performance outlook.

“Given that TSEM has previously announced plans to expand its Uozu large-scale semiconductor factory — which we believe will roughly double TSEM's silicon photonic chip production capacity to about $1 billion per quarter — we believe that the forecasts of buyers' investment institutions and at least some of the top seller research institutes have been adjusted to include additional revenue projects.” Matt Bryson, a senior analyst at Wedbush, wrote in a report to clients.

“For example, our latest forecast for 2028 is already slightly higher than the company's updated performance guidance. Despite this, TSEM's outlook is still conservative because it assumes that the capacity utilization rate of high-performance optical products will be only about 60% in 2028, which creates favorable conditions for TSEM to surpass the revised outlook. Furthermore, we believe that TSEM is expected to start producing more production capacity in 2029 that will provide strong incremental returns. This information is new and additional. Overall, we think this news is significantly positive for the company's basic outlook and its stock.”

Takata Semiconductor (TSEM) will spend approximately $3 billion to expand its large 300mm fab in Uozu, Japan. It is expected that the plant will also support the production of 300mm silicon photonics and silicon-germanium-related chip products, as well as support cutting-edge advanced packaging business. The first production line in the dual-track expansion plan is expected to be ready for operation in the fourth quarter of 2027.

With this expansion of production, Takata Semiconductor raised its 2028 performance outlook to $3.6 billion in revenue and the latest operating profit of $1.2 billion. Previously, expectations were only $2.8 billion and $750 million, respectively.

Gaota Semiconductor is not competing with TSMC for the most advanced manufacturing process, but is flooded with card AI data

Tower Semiconductor (TSEM) is a specialty analog semiconductor foundry headquartered in Israel. It is not centered around the most advanced digital logic process like TSMC. Its competitiveness mainly comes from customer-specific process platforms, including silicon photons (SiPhO), silicon on silicon germanium (SiGe) and RF insulators, RF complementary metal-oxide semiconductors, high-performance analog chips, power management chips, image sensors, mixed-signal chips, and microelectromechanical systems, serving communication infrastructure, automotive, industrial, medical, consumer electronics, and aerospace and defense markets, with more than 300 customers.

Gaota Semiconductor's commercial moat comes more from process intellectual property rights, customer certification cycles, and scarcity of characteristic production capacity, rather than simply relying on transistor size leadership.

The core purpose of this expansion of production in Japan is to focus on adding 300 mm silicon photons, silicon germanium, and advanced optical packaging capabilities. Among them, the first production expansion path will transform the original Arai plant into a 300mm silicon photonics and advanced optical packaging base and support Uozu Fab 7, which is expected to be fully put into operation in the fourth quarter of 2027; the second path is to build a new 300mm plant near Fab 7, with the goal of increasing the production capacity of silicon photons and silicon-germanium several times, and making significant incremental contributions starting in 2029. According to the company's latest documents, the maximum scale of the project is about 3 billion US dollars, and it has received 1 billion US dollars in subsidies from the Japanese government.

In terms of specific semiconductor manufacturing and foundry strategies, this is a key expansion for Gaota Semiconductor to upgrade itself from a traditional analog foundry to an artificial intelligence data center optical interconnection infrastructure provider. As GPU clusters expand, system bottlenecks are shifting from single-chip computing power to inter-chip bandwidth, network power consumption, and optoelectronic conversion efficiency; silicon photons are responsible for high-speed optical transmission, silicon-germanium is suitable for high-speed, low-noise analog and RF devices, and advanced optical packaging determines whether optical chips, lasers, and electronic chips can be integrated at high density. Gaota has used the silicon optical platform for 1.6T data center optical modules and is clearly aimed at artificial intelligence infrastructure and next-generation optical networks.

The specific investment level of the AI computing power industry chain makes it an “optical interconnection seller” with high growth elasticity in the AI computing power chain, but the return on production expansion still depends on customer order fulfillment, capacity climbing, and utilization rates. It cannot be inferred from capital expenditure of 3 billion US dollars alone that profits will inevitably grow at the same time.

Moore's Law fades out of sight, silicon photons break through the “copper wall”, and advanced packaging rewrites the hashrate chip value chain

The bottleneck at the bottom of artificial intelligence is shifting from the number of transistors in a single chip to data handling capabilities. Large model training and inference require thousands or even larger accelerators to work together, but data transmission between GPUs and HBM, chip-to-chip, and cabinet-to-cabinet is constrained by rapid increases in copper interconnection distance, signal integrity, bandwidth density, and power consumption; as the energy consumption of data movement begins to approach or even exceed the computation itself, it is impossible to independently solve the system efficiency problem. Silicon photons use light instead of some high-speed electrical signal transmission, and place the optical engine near the switching chip or computing chip through co-packaged optics, thereby shortening the signal path, increasing bandwidth density, and reducing power consumption per bit.

Nvidia, the leader of AI chips, has positioned co-packaged optics as an important network technology for expanding million-level GPU artificial intelligence factories. Its Spectrum-X photon switch claims to provide 1.6 Tbps of bandwidth per port, about 3.5 times more energy efficiency, and 10 times more network resilience, indicating that “optical interconnection” is being upgraded from communication devices to the core infrastructure that determines the effective computing power of clusters.

Advanced packaging is no longer a “shell process” at the end of chip manufacturing; it is a system-level architecture platform that continues Moore's Law: it can break through the size and yield limits of a single wafer mask, closely integrate CPUs, GPUs, input/output chips, analog devices, and multiple HBMs manufactured in different processes in 2.5D or 3D to build an “in-package supercomputer” with lower latency, higher bandwidth, and better cost. TSMC CoWOS is undertaking the integration of logic chips with HBM, and its 5.5x mask size CowOS-L is scheduled to be certified in 2026, and larger platforms are also being promoted. At the same time, CouPe technology integrates silicon optical chips and electronic control chips through SoIC, and further introduces CoWOS co-package optics, indicating that advanced packaging and silicon photons will eventually merge into a “computing-storage-network” integrated platform.

For investors, the next phase of AI semiconductor excess alpha revenue will not only come from advanced manufacturing processes, but will also spread to the core areas of the AI computing power industry chain such as CoWOS and 3D packaging, silicon optical wafer foundry, optical engines and lasers, HBM/DRAM/NAND storage chips, package substrates, hybrid bonding, testing, and liquid cooling; the real moat for chip manufacturing giants will be mass production yield, customer certification, system collaborative design and production capacity delivery capabilities, rather than just having a certain conceptual technology.