Recent analysis from Wall Street firm Wedbush Securities suggests that Israel-based semiconductor foundry leader Tower Semiconductor (TSEM.US) may be undervalued, with its announced large-scale expansion of a major plant in Japan positioning it for significant growth in AI infrastructure.
The expansion of the Uozu factory in Japan is seen as a key factor that could help the company exceed its current, seemingly conservative financial outlook. "Given TSEM's previously announced plan to expand the Uozu large-scale semiconductor plant—which we believe will roughly double TSEM's silicon photonics capacity to about $1 billion per quarter—we believe buy-side investors and at least some top-tier sell-side research forecasts have already adjusted to incorporate the new revenue stream," wrote Wedbush senior analyst Matt Bryson in a client note.
"For example, our latest forecast for 2028 is already slightly above the company's updated guidance. Nevertheless, TSEM's outlook still appears conservative, as it assumes only about 60% utilization for high-performance optical products in 2028, creating favorable conditions for TSEM to surpass the revised outlook. Furthermore, we believe TSEM is expected to bring more capacity online in 2029 that will deliver strong incremental earnings, which is new and additional information. Overall, we view this news as significantly positive for the company's fundamental prospects and its stock," the analyst added.
Expansion Details and Financial Impact
Tower Semiconductor plans to invest approximately $3 billion to expand its 300mm mega-fab in Uozu, Japan. The facility is expected to support the production of both 300mm silicon photonics and silicon germanium chips, as well as cutting-edge advanced packaging operations. The first production line under this dual-track expansion plan is anticipated to be ready for operation in the fourth quarter of 2027.
Following this expansion, the company has raised its 2028 financial outlook to $3.6 billion in revenue and an operating profit of $1.2 billion, up significantly from previous expectations of $2.8 billion and $750 million, respectively.
Strategic Positioning in the AI Ecosystem
Unlike TSMC, which competes on the most advanced digital logic process nodes, Tower Semiconductor is a specialty analog semiconductor foundry. Its competitive edge stems from customized, special-process platforms including silicon photonics (SiPho), silicon germanium (SiGe), RF-SOI, RF-CMOS, high-performance analog chips, power management ICs, image sensors, mixed-signal chips, and MEMS. It serves over 300 customers across communications infrastructure, automotive, industrial, medical, consumer electronics, and aerospace and defense markets.
The company's moat is built more on process IP, customer qualification cycles, and the scarcity of specialty capacity than on leading-edge transistor scaling. The core objective of the Japan expansion is to ramp up 300mm silicon photonics, silicon germanium, and advanced optical packaging capabilities. This strategic move upgrades Tower Semiconductor from a traditional analog foundry to a supplier of optical interconnect infrastructure for AI data centers.
The Rise of Silicon Photonics and Advanced Packaging
As GPU cluster sizes grow, system bottlenecks are shifting from single-chip computing power to inter-chip bandwidth, network power consumption, and photoelectric conversion efficiency. Silicon photonics handles high-speed optical transmission, silicon germanium is suited for high-speed, low-noise analog and RF devices, and advanced optical packaging enables the high-density integration of optical chips, lasers, and electronic chips.
Tower Semiconductor has already deployed its silicon photonics platform for 1.6T data center optical modules, explicitly targeting AI infrastructure and next-generation optical networks. From an investment perspective, this positions the company as a high-growth potential "pick-and-shovel" play in the AI compute chain's optical interconnect segment.
The underlying bottleneck for AI is migrating from transistor count per chip to data movement capability. Training and inference of large models require thousands or even larger-scale accelerators to work together, but data transmission between GPUs and HBM, between chips, and between server racks is constrained by copper interconnect limitations in distance, signal integrity, bandwidth density, and rapidly rising power consumption.
Silicon photonics replaces some high-speed electrical signal transmission with light and, through co-packaged optics (CPO), places optical engines near switch or compute chips, shortening signal paths, increasing bandwidth density, and lowering power per bit. AI chip leader Nvidia has positioned CPO as a critical networking technology for scaling million-GPU AI factories.
Advanced packaging is no longer just a final "encapsulation" step but a system-level architecture platform extending Moore's Law. It can overcome single-die reticle size and yield limits, integrating CPUs, GPUs, I/O chips, analog devices, and multiple HBM stacks manufactured on different process nodes via 2.5D or 3D methods, building a "supercomputer-in-a-package" with lower latency, higher bandwidth, and better cost.
Implications for Investors
For investors, the next phase of alpha generation in AI semiconductors will not come solely from advanced process nodes but will spread to core segments of the AI compute supply chain, including CoWoS and 3D packaging, silicon photonics foundry services, optical engines and lasers, HBM/DRAM/NAND memory chips, packaging substrates, hybrid bonding, testing, and liquid cooling.
The true moat for leading chip manufacturers will be mass-production yield, customer qualification, system co-design capabilities, and capacity delivery execution—not merely possessing a conceptual technology.
Comments