The explosion of AI data centers is driving a profound shift in Taiwan’s semiconductor industry. Companies traditionally associated with smartphones, Ethernet, displays, or PC interfaces such as MediaTek, Realtek, Himax, and Elan Microelectronics are increasing their investments in optical communications, a market that could enter a much more aggressive deployment phase starting in 2027.
The reason is fundamentally physical. As AI clusters connect tens of thousands of GPUs and accelerators, transporting all this data through electrical links consumes more energy and becomes more challenging. The industry is responding with 800G and 1.6T transceivers, SerDes of 112G/224G, silicon photonics, Linear Pluggable Optics, and increasingly, Co-Packaged Optics (CPO). TrendForce specifically positions 2027-2028 as the critical window for large-scale deployment of CPO, after the start of commercial production in 2026.
Key takeaways in 30 seconds
- TrendForce expects CPO to move from early production in 2026 to mass procurement during 2027-2028.
- MediaTek is already developing CPO, SerDes, and active optical cables with MicroLED for data centers.
- Airoha, a MediaTek subsidiary, aims to triple its revenue from high-speed optical transceiver chips in 2026.
- Realtek is working on 224 Gbps PAM4 SerDes IP targeted at data centers, enterprise cloud, and private cloud.
- Himax and FOCI are preparing CPO solutions of 1.6T, 3.2T, and eventually 6.4T, with volume acceleration expected from 2027.
- Elan has invested in PETA Optronics to combine EIC circuits with photonic PIC circuits to enter high-speed AI servers and data centers.
- The main bottleneck is no longer just GPUs: interconnection, energy per bit, and bandwidth density have become critical elements of AI factories.
2027 may be the year when optics are no longer optional
The evolution of communications within data centers has followed a relatively predictable trajectory for decades. More speed meant successive generations of Ethernet and improved SerDes, with electrical connections maintained as long as possible and fiber resorted to for longer links.
AI is disrupting this balance.
A cluster with thousands of accelerators needs to move enormous amounts of data between GPUs, memory, switches, and racks. The individual speed of each accelerator matters, but so does how long it waits for data from other nodes.
At 200 Gbps per lane and above, electrical losses and the energy required to compensate for them become increasingly problematic. Major players like Broadcom, Nvidia, Marvell, and others are bringing optics closer to the ASIC.
TrendForce considers that CPO has already entered its commercialization stage, with 2026 marking the first volume production year. Meanwhile, 2027 and 2028 are seen as the decisive window for broader adoption.
But CPO won’t be the only technology in play.
LPO, improved pluggable modules, NPO, and emerging solutions like XPO will coexist for years. In fact, DIGITIMES suggests that certain XPO solutions could reach mass production in 2027 even before CPO fully matures.
MediaTek no longer wants to be just a mobile chip company
MediaTek is arguably the clearest example of this transformation.
While still a major global provider of SoCs for smartphones, it is increasingly dedicating resources to cloud infrastructure and custom ASICs for AI.
Its expansion into optical communications has been underway for several years now.
In 2024, it introduced an ASIC platform integrating electrical and optical links within the same design, utilizing its own SerDes and Ranovus Odin optical engines. This configuration combined eight electrical links of 800G and another eight optical links of 800G. MediaTek claimed this integration could reduce system power consumption by up to 50% compared to existing solutions.
At Computex 2026, it took another step: displaying CPO technology capable of reaching up to 400 Gbps per fiber and new MicroLED developments for data center interconnections.
MediaTek and Microsoft aim to replace some lasers with MicroLEDs
One of the most striking developments stems from a collaboration between MediaTek and Microsoft Research.
They developed an Active Optical Cable based on MicroLED that replaces the conventional approach of a few high-speed laser channels with numerous MicroLED channels operating at lower speeds.
MediaTek asserts that this design can scale to 800 Gbps and beyond within standard formats like QSFP and OSFP. It also estimates up to 50% less power consumption compared to VCSEL-based AOCs, mainly due to the elimination of complex DSPs.
The system integrates on a single CMOS chip the SoC logic, gearbox, MicroLED drivers, and TIA amplifiers. The MicroLED and photodetector arrays are directly bonded to the CMOS via heterogeneous integration.
This demonstrates that the interconnection battle is not limited solely to CPO.
The industry is simultaneously exploring new light sources, DSPs, near-package optical links, and different methods of integrating electronics and photonics.
Airoha could become a key piece for MediaTek
MediaTek also relies on Airoha Technology to target the communications market.
In May, Airoha announced it expects revenues from chips used in high-speed transceivers to triple in 2026, driven by AI infrastructure build-out.
It sells PAM4 DSP solutions for 400G and 800G transceivers and is developing higher-speed generations.
Airoha aims to launch in 2026 new DSPs capable of transmitting 100 Gbps per channel and is already working on 200 Gbps technology, essential for future 1.6T modules.
The company does not believe CPO will immediately replace traditional optical modules.
Its view is that pluggable transceivers with DSPs will remain vital for scale-out connections during this decade, while CPO could be especially attractive for scale-up and extremely dense systems.
This coexistence will likely remain a key feature of the market through 2027.
Realtek enters a critical component: SerDes
Realtek is mainly known for Ethernet, audio, Wi-Fi controllers, and other connectivity chips embedded in millions of PCs and devices.
However, its expertise in digital communications is directly transferable to the new market.
Industry sources in Taiwan indicate that Realtek is developing 224 Gbps PAM4 SerDes IP aimed at data centers, enterprise infrastructure, and private clouds.
A SerDes — Serializer/Deserializer — converts parallel data into high-speed serial streams and vice versa at the receiver.
It is one of those low-profile yet essential building blocks for high-performance communication systems.
At 224 Gbps per lane, eight lanes can support roughly 1.6 Tbps before accounting for protocol overheads.
The next generation is already moving toward 448 Gbps per lane. Broadcom demonstrated 448 Gbps electrical signaling developments at DesignCon 2026, as industry continues to finalize specs for new-generation interfaces.
Realtek doesn’t need to produce lasers to benefit from the optical boom
An important aspect of this transition is that participating in optical communications doesn’t necessarily mean manufacturing all photonic components.
A system can be divided into:
ASIC → SerDes → driver/DSP → PIC → modulator/laser → fiber → photodetector → TIA/DSP
Opportunities exist at virtually all these layers.
Realtek can mainly provide electronics and communication processing, partnering with photonics specialists or module manufacturers.
This division of labor is a key reason why Taiwan is well-positioned in the market.
The island already has chip design, foundries, advanced packaging, optical component manufacturers, and assembly companies.
Himax takes a very different approach
Himax’s case is particularly interesting because its traditional business is more related to display controllers than networking.
Its advantage lies in wafer-level optical technologies.
The company has been developing Wafer Level Optics (WLO) and nano-imprint processes for years, now applying these to Co-Packaged Optics.
Its key partner is FOCI Fiber Optic Communications.
Both companies confirmed in early 2026 that their first-generation CPOs are in validation with customers and are working towards mass production readiness during the year.
Himax later provided more concrete figures.
Himax already talks about CPOs up to 6.4T
The initial generation developed with FOCI supports 1.6T and 3.2T bandwidths.
Himax expected to begin small shipments in late 2026.
But the second generation is even more significant.
Known as Gen 2, it targets 6.4T bandwidth and was in the final stages of customer validation for AI data center applications.
Himax states that 2026 will mainly be a preparation year for production, with limited volumes, while ramp-up is expected to accelerate in 2027.
This forecast aligns closely with TrendForce’s 2027-2028 window for large-scale CPO adoption.
Himax has increased its financial stake in FOCI
The collaboration isn’t just technological.
In 2026, FOCI raised capital by NT$3.16 billion for R&D, equipment, and CPO production readiness.
Himax participated in this round after having invested in the company in 2023 and 2024.
As of May 2026, Himax held approximately 5.36% of FOCI, a stake the company valued at around $156 million based on market prices.
For a company still heavily reliant on display controllers, CPO offers a diversification opportunity into a market with potentially higher prices and margins.
Himax has even stated that it expects CPO to be a significant contributor to revenue and profit in the coming years.
Elan also aims to move beyond traditional PCs
Elan Microelectronics is pursuing a different diversification strategy.
Best known for touch controllers, touchpads, and biometric sensors, in June it announced that semiconductors for optical communications are now one of its three new priority markets alongside Agentic AI PCs and unmanned vehicles.
To enter this space, Elan invested in U.S.-based PETA Optronics and became one of its significant corporate shareholders.
This partnership combines two parts of a photonic system:
- PETA contributes expertise in Photonic Integrated Circuits (PIC).
- Elan provides design in Electronic Integrated Circuits (EIC).
The goal is to jointly develop integrated optoelectronic solutions for AI servers and high-speed data centers.
PIC + EIC: the new chip frontier
Understanding the distinction between PIC and EIC is key to grasping why many traditional semiconductor companies are now venturing into optics.
EIC processes electronic signals.
PIC manipulates photons using elements like modulators, waveguides, multiplexers, or photodetectors.
A modern communications platform needs both.
The main challenge is bringing them closer so that the electrical-to-light conversion occurs almost right next to the processor.
This minimizes electrical path length and therefore reduces losses, power consumption, and the need for compensation.
CPO advances this philosophy by integrating optical engines within the same package as the ASIC switch.
Why is everyone eyeing 2027?
There isn’t a single date when the industry will switch from copper to fiber.
The transition is gradual.
However, several roadmaps are beginning to converge around 2027 as a commercial inflection point.
TrendForce expects large-scale CPO purchases in 2027-2028.
Himax foresees accelerating the volume of its platform with FOCI during 2027.
Airoha anticipates that 100G per lane products will give an additional boost to optical communications in 2027, while preparing for 200G transitions.
Various suppliers expect initial commercial productions of new optical architectures and higher-density modules to begin in this period.
Nvidia is accelerating the entire ecosystem
Furthermore, there is a key buyer capable of shifting the schedules across the entire supply chain: Nvidia.
The expansion of its rack-scale systems has made networking a core part of its architecture.
It has long driven Ethernet and InfiniBand switches with silicon photonics and CPO, inspiring PIC, laser, packaging, fiber, DSP, and optical component providers to increase capacity.
This influence extends beyond Nvidia.
Google, Microsoft, Meta, Amazon, and other hyperscalers are building clusters whose bandwidth demands outpace the ability to maintain purely electrical links.
Fiber also faces its own bottlenecks
However, transitioning from electronics to photonics doesn’t eliminate manufacturing issues.
It simply changes them.
TrendForce currently identifies three major challenges to bringing CPO into mass production:
- Fiber alignment precision;
- Thermal management;
- Testing and validation costs.
The consultancy also warns of high supplier concentration in areas like silicon photonics wafers, InP substrates, and Fiber Array Units.
Indium phosphide is especially critical since it’s used to produce certain lasers and active components.
Market expansion is pushing the transition from four-inch to six-inch InP wafers, but substrate availability and process maturity remain potential limitations.
Copper isn’t going away anytime soon
It would be a mistake to interpret the growth of photonics as the immediate end of copper.
Marvell demonstrated at OFC 2026 a connection based on LR SerDes of 224G capable of transmitting over a 2.5-meter electrical channel using co-packaged copper solutions.
Copper still offers significant cost and power advantages for relatively short distances.
The future architecture is likely to be hybrid.
Within the package and for shorter distances: electrical interfaces.
Within the rack: copper, AEC, CPC, and progressively optical links depending on density.
Between racks: increasing use of optical links.
Between data centers: high-speed fiber.
The boundary between these zones will shift as each generation pushes speeds higher.
From 800G to 1.6T and then to 3.2T
The most visible change in the coming years will be increasing transceiver capacities.
800G is already being deployed in large facilities.
1.6T will be the next major commercial generation.
Followed by 3.2T.
This jump requires a gradual migration from 100G per lane toward approximately 200G per lane.
At OFC 2026, demonstrations of 1.6 Tbps using four PAM4 channels at 448 Gbps each were shown using 3 nm CMOS SerDes and silicon photonics, indicating that the technical roadmap is looking well beyond current commercial implementations.
Not all these manufacturers will benefit equally
MediaTek, Realtek, Himax, and Elan are all aligned with this trend, but their strategies vary significantly.
MediaTek aims to combine ASIC, SerDes, CPO, and new technologies like MicroLED.
Realtek leverages its historical expertise in connectivity ICs and SerDes.
Himax utilizes its optical know-how and WLO, partnering with FOCI to directly participate in CPO development.
Elan has chosen to partner with PETA Optronics to combine electronics and PIC circuits.
This broadening of the optical market reflects its expanding scope beyond traditional transceiver manufacturers.
Taiwan’s ambition to replicate its semiconductor success with photonics
Finally, there is a strategic consideration.
Taiwan possesses nearly all the elements needed to foster a complete photonic ecosystem: fabless designers, TSMC, specialized foundries, OSATs, module manufacturers, precision optics, and a large server industry.
SEMICON Taiwan 2026 even created a dedicated Silicon Photonics pavilion, covering PICs, electro-optic integration, advanced packaging, CPO, and optical communication modules.
The opportunity lies in connecting all these elements.
Taiwan’s electronics manufacturing tradition has been built on specialization: one company designs, another fabricates, another encapsulates, and a system integrator assembles.
Photonics can follow a similar model.
2027 could mark the start of a new semiconductor era
During the first phase of the AI revolution, the spotlight was primarily on GPUs and HBM.
The second phase is revealing new bottlenecks.
Building an extremely fast accelerator is less useful if it spends more and more time waiting for data from other chips.
Hence, data movement is becoming as strategic as processing.
The market around SerDes of 224G, DSP, PIC, EIC, silicon photonics, CPO, AOC, lasers, and optical packaging could become one of the next major chapters in AI infrastructure.
MediaTek, Realtek, Himax, and Elan have seemingly arrived at this conclusion independently.
While 2026 is about validation, initial production, and capacity building, 2027 increasingly looks like the moment many of these technologies will reach volume-scale business.
