Deep Dive: Optical Module Market

Powering next-gen connectivity: How optical modules enable fast, reliable data transfer, and What are the opportunities for investors?

Today, I’m exicited to share an in-depth analysis of the global optical module market, an industry I find particularly compelling due to its critical role in data center networks for the development of next-generation AI. Unlike my previous posts, which typically focus on insights into a specific public company, this piece offers a comprehensive overview of the entire industry and its up/downstreams, aiming to provide a foundational understanding of their dynamics before potentially diving into individual stock analyses in the future.

Equities relevant to this report include:

• Optical module makers: Zhongji Innolight (300308.SZ), Coherent (COHR), TFC Optical Communication (300394.SH), Fabrinet (FN), Eoptolink Technology (300502.SH), Lumentum (LITE)

• Component suppliers: Marvell (MRVL), Macom (MTSI), Broadcom (AVGO), Maxlinear (MXL), Semtech (SMTC), Yuanjie (688498.SH), Shijia (688313.SH)

• Foundries: GlobalFoundries (GFS), Tower Semiconductor (TLV: TSEM), TSMC (TPE: 2330)

• Cloud service providers: Nvidia (NVDA), Microsoft (MSFT), Google (GOOG), Amazon (AMZN), Meta (META)

Disclaimer: this research is for informational purposes only. This is NOT a recommendation to buy or sell securities discussed. Please do your ownwork before investing your money.

---

---

What are Optical Modules?

Optical modules are devices used in fiber optic communications to transmit and receive data through optical fibers. They convert electrical signals into optical signals and vice versa, enabling data transmission at high speeds with minimal signal loss. A typical optical module is composed of a transmitter, receiver, DSPs, and various other optical/ electronic components. The transmitter includes a laser diode for sending optical signals and a driver that controls the laser's operation, while the receiver consists of a photodiode (PD) that detects incoming optical signals and converts them to electrical current, along with a Transimpedance Amplifier (TIA) that amplifies the signal from the PD. Lastly, the DSPs enhance data transmission quality by mitigating signal distortions, enabling complex modulation, and implementing forward error correction. In terms of the BOM for a typical optical module, the driver and TIA account for 20%+ (each contributing equally at c.10%+), the DSP represents ~30%, and the remaining 40-50% is made up of other components.

Optical modules are typically defined by their bandwidth capacity, with the highest currently available being 800G modules. Next-generation 1.6T modules are expected to begin deliveries in Q4 2024, which will primarily be paired with Nvidia's GB200 chips (although GB200 can also work with 800G modules). The industry is moving towards using 1.6T modules primarily for model training, while 800G modules, which are currently heavily used for training, are expected to be more focused on inference in the future. 

Optical modules can also be categorized into single-mode versus multimode modules. Single-mode modules often use EML (Electro-absorption Modulated Lasers) as the lasers, which are better suited for longer-distance transmission but come with a higher cost. In contrast, multimode modules often rely on VCSEL (Vertical-Cavity Surface-Emitting Lasers); while VCSEL-based modules have limitations in transmission distance, they are more affordable and are commonly used in AI data center networks where servers and racks are positioned closer together.

---

Key Technology Roadmaps

In this section, we will explore several emerging technology solutions that the industry is currently developing. These advancements highlight potential future directions for the industry and can provide valuable insights for investors looking to navigate opportunities in the optical module supply chain.

1. Linear-drive Pluggable Optics (LPO)

LPO is a recent technology for optical modules designed to remove traditional DSPs from the modules by utilizing linear analog signal processing – in conventional optical modules, DSPs convert analog signals to digital, perform tasks such as signal equalization, forward error correction, and impairments compensation, and then convert digital signals back to analog for transmission; LPO simplifies this process by directly transmitting the analog signal produced by the driver over the optical link, using a linear analog drive to modulate the laser and eliminating the analog-to-digital and digital-to-analog conversions typically handled by a DSP.

A notable drawback of this approach is the reduced ability to perform the signal enhancement tasks that DSP typically does. To address this limitation, LPO leverages the DSPs and SerDes within the switches in the network to partially compensate for the signal degradation.

LPO maintains the traditional pluggable form factor, enabling modules to be hot-swapped, which provides flexibility for network upgrades or scaling without requiring system downtime. This differs from the CPO (Co-packaged Optics) technology that we will discuss later, which integrates optical chips with switch chips. LPO supports various transmission technologies such as EML and VCSEL, and is also compatible with Silicon Photonics (SiPh) technologies.

The advantages of the LPO solution include: 1) Lower power consumption: By eliminating DSPs, LPO reduces power consumption, as DSPs are a major power consumer in traditional modules – for instance, power consumption drops from 30W+ in a typical 1.6T module with DSP to around 10W in a 1.6T LPO module. 2) Lower latency: Conventional DSPs introduce ~200 nanoseconds of latency during the transmission as well as the reception of signals; in contrast, LPO modules have less than 5 nanoseconds of latency, a substantial reduction that is especially valuable for efficient large-scale AI training. 3) Lower cost of ownership: DSPs typically account for ~30% of the BOM of a conventional module; By eliminating the DSP, LPO can achieve ~20% cost savings compared to DSP-based methods

Test data shows that current LPO solutions can deliver reasonably close performance compared to traditional DSP-based modules, especially in scenarios with shorter connection distances (below a few hundred meters), though some minor technical challenges remain.

Potential disadvantages of LPO solutions include: 1) Weaker resistance to interference: In traditional systems, DSPs within optical modules take the main responsibility of managing signal interference; with LPO, this responsibility shifts to the DSPs in the switches, which must be enhanced to handle these tasks. However, even with stronger DSPs in switches, challenges like fiber disturbances, port contamination, and voltage fluctuations are hard to address. 2) Increased system maintenance complexity: Since signals remain analog between the optical modules and the first layer of switches, no signal monitoring can be built in between; this makes it difficult to identify and troubleshoot issues in the system, leaving LPO more suitable for larger organizations with the resources to manage such troubleshooting. 3) Shorter transmission distance: Without the DSPs to amplify and correct signals, LPO is much less effective for long-distance transmission, especially those beyond a few hundred meters; LPO is generally better suited for transmission between AI servers and the first layer of switches.

4) More challenging protocol compatibility: In traditional optical systems, DSPs in the optical modules modulate signals to ensure their compatibility with other optical modules; with LPO, analog signals are passed directly to the switches, requiring switches to be compatible with one another when transmitting data. This necessitates additional modulation and configuration at the switch level to ensure, for example, that Broadcom switches can communicate with Siemens switches. While this is not a major concern in Nvidia's IB networks (as Nvidia produces its own switches), Ethernet networks, with their various protocols and vendors, require more work to achieve compatibility. Initiatives like the Ultra Ethernet Consortium (UEC) are helping address this, but it may take 2-3 years for these compatibility issues to be largely resolved. 5) Later technology maturity: The 1.6T LPO technology is still in the testing phase and is expected to take 1-2 years to mature, with further work required on drivers, TIA, and switch SerDes; meanwhile, conventional 1.6T modules with DSPs are already ready for production, making LPO a little late to the “1.6T party”

Given these advantages and disadvantages, LPO remains an attractive option for optical networks, generating some client interest. The industry is projected to see a few hundred 800G LPO units deployed in 2024, with Nvidia being the primary customer. By 2025, the market could grow to 1-2 million units (still primarily 800G), with other CSPs such as Meta, Google, and Amazon also considering adoption. Based on our forecast of 12-15 million total 800G units in 2025 (details in later chapter), LPO’s market penetration next year is implied at high single digits (HSD%) to mid-teens percent.

In terms of pricing, 800G LPO modules have been sold at around $600 this year, which is cheaper than single-mode conventional optical modules priced above $700 but more expensive than multimode modules, which sell for around $500. Despite being pricier than multimode modules, LPO offers advantages such as relatively longer transmission distances and lower power consumption.

Looking ahead to the 1.6T node, the industry remains hopeful that 1.6T LPO can be achieved, based on theoretical calculations and early testing results. The challenge lies in the increased risk of signal degradation with higher bandwidths when DSP is not utilized, and there are still engineering obstacles to overcome. During this year’s OFC conference in early April, several companies also discussed the technical hurdles in developing 1.6T LPO solutions. Uncertainty remains for whether and when we will see the mass production 1.6T LPO.

This complexity in technological realization leads us to the next concept:

---

2. Linear Receive Optics (LRO)

LRO has recently entered the discussion as a middle-ground solution between traditional optical modules and LPO. It retains the DSP on the transmitting side while employing a linear design on the receiving side and eliminating the receiver DSP. This design reduces power consumption by removing the receiver-side DSP while offering better signal recoverability and interference reduction compared to LPO, thanks to the DSP remaining on the transmitting end.

This solution could be particularly attractive to smaller clients due to its better compatibility with existing networks and lower configuration complexity compared to LPO, while also providing power-and-cost-saving benefits through reduced DSP usage. In next-generation 1.6T transmission, LRO has an additional advantage over LPO by retaining DSP, which enhances its ability to mitigate signal interference, an even bigger challenge in next-generation transmissions.

---

3. Silicon Photonics (SiPh)

The concept: Silicon photonics (SiPh) is a technology that uses silicon as an optical medium to transmit data in the form of light. It fabricates optical components including waveguides, modulators, and detectors onto silicon wafers, enabling high-speed data transfer with reduced power consumption and less heat generation. SiPh modules also incorporate other components like CW lasers, drivers, TIA, and DSP using co-packaging or chiplet designs. The technology leverages the well established CMOS manufacturing infrastructure, providing cost-effective scalability for high-performance optical communication systems.

CW laser: SiPh uses CW laser on its transmitter side, which offers good affordability. A typical 800G/1.6T SiPh module requires 2-4 CW light sources, with power ratings of ~70mW (for 800G) and 100mW (for 1.6T) respectively, per CW light. One 70mW CW laser costs about $8-$10, bringing the total cost per module to $20-$40. In contrast, a more traditional EML-based 800G module requires eight 100G EML lasers, each costing $10+, resulting in a total cost of $80-$100, which is much higher than the CW-based SiPh solution. Additionally, CW lasers are produced using mature technology nodes with ample supply – companies such as Yuanjie (688498.SH), Shijia (688313.SH), and Lumentum ($LITE) can manufacture CW lasers in large volumes without problem, unlike the EML and VCSEL markets (for conventional modules) that are experiencing supply chain bottlenecks.

SiPh production: CMOS technology is widely used in SiPh production. Foundries like GlobalFoundries (GFS), AMF, Tower, and TSMC are capable of producing the SiPh chips. GFS, for example, runs the world’s only 12-inch SiPh production line, at 45nm nodes. AMF, headquartered in Singapore, focuses more on research but also has SiPh production capabilities in the thousands of units of wafers. Tower's SiPh production line has an annual capacity of ~10K pieces of (8-inch) wafers, each capable of yielding 400-600 dies for 800G SiPh,  or ~5 million total dies; to put this in perspective, the upper end of industry projections for 800G SiPh modules in 2025 is 4.5 million units – it appears that the production capacity in the industry is in good shape, at least in the short term.

Advantages of SiPh: 1) Lower power consumption: The higher integration of optical and electronic components in SiPh solutions leads to ~15% power savings. 2) Lower production costs: SiPh solution can potentially deliver 10-20% cost savings versus conventional methods due to three factors – first, module providers tape out the highly integrated SiPh modules themselves instead of assembling procured components from vendors; second, packaging costs are reduced due to higher level of integration and production costs are reduced due to more streamlined process; third, as mentioned before, CW lasers have cost advantage vs. current mainstream lasers such as EML. 3) Higher reliability: The tight integration of most optical and electronic components within SiPh module improves communication between components and improves the overall system reliability. 4) Higher compatibility: SiPh is compatible with many other advanced technology roadmaps; for instance, it works well with the LPO technology due to its ability to reduce signal degradation, with some 800G SiPh modules expected to use LPO as early as this year; SiPh is also a crucial step towards realizing CPO, which we will discuss later, a technology that has the potential of providing even greater integration and performance improvements in optical transmission.

Penetration potential: As the industry moves into the 800G and 1.6T eras, the need for lower power consumption, smaller module sizes, and reduced costs will become even more pressing, making SiPh an increasingly attractive option. In addition, while demand for 800G and 1.6T modules is substantial, the supply of EML and VCSEL, key laser types for more conventional modules, is currently insufficient to fully meet the high-end demand, creating an opportunity for SiPh modules to serve as a valuable alternative. Overall, I anticipate that SiPh will begin to make a meaningful contribution to the optical module market as early as the second half of 2024.

For 800G modules, industry expectations are for around 1 million SiPh module units (all 800G) being shipped in H2 2024. Considering a projected ~10 million total 800G module shipment in 2024, SiPh’s penetration is implied at ~10%. By 2025, that penetration is anticipated to grow to 20-30%, positioning SiPh as a significant contributor in the industry. A key challenge for SiPh mass production at the moment is improving yield rates. Traditional methods can achieve yields of over 90%, and for SiPh to be really competitive in pricing, it needs to reach at least 80-90% yield rates.

For 1.6T optical modules overall, the industry is currently focused on EML-based conventional solutions due to the technology's earlier maturity and readiness for production. Research on 1.6T SiPh began later, and its design is more complex, leading to a later production timeline. However, this doesn't mean SiPh lacks potential in the 1.6T space. In fact, 1.6T modules have even demanding requests for low power consumption and cost of ownership (CoW), areas where SiPh excels. Some industry experts predict that SiPh could eventually achieve 30-40% penetration in the 1.6T market once the technology matures, likely after the second half of 2025, which would represent another step up of its projected 20-30% penetration in 800G. One more positive sign for SiPh adoption is the technical bottleneck currently facing VCSEL-based multimode 1.6T module, which has the potential to be the cheapest 1.6T solution. With the timeline of 1.6T VCSEL solution still uncertain, SiPh would become the most cost-effective solution in the market.

SiPh module GPM: according to industry experts on Tegus, SiPh module gross margins are currently 30-40%. This margin benefits from SiPh’s up-to-20%-lower production costs discussed earlier. However, to drive customer adoption, SiPh modules are priced ~10% lower than EML-based single-mode solutions. Additionally, with SiPh still in the production ramp-up phase and experiencing lower yields, overall GPM advantage compared to comparable traditional modules is around 5 percentage points at the moment.

Industry players: In the SiPh module field, Zhongji Innolight (300308.SZ), Cloud Light (owned by Lumentum ($LITE)), and Coherent ($COHR) are key players. Zhongji Innolight appears to be a leader in the space; its 800G SiPh modules are already in the market and expected to take at least 2/3 of volume delivered in 2025. Cloud Light is another interesting player; it has gained some traction in the SiPh trend and is a key supplier to Google. Coherent (Finisar) has faced some delays due to recent client-requested additional testing; currently, its production is at a modest 1-2K units per month, but could ramp up quickly in the second half of 2024.

SiPh’s impact on EML: The potential rise in SiPh adoption could pose some challenges to EML’s volume growth, as SiPh offers a compelling solution especially for mid-and-short-distance transmission. However, a complete replacement of EML is unlikely, as it still holds a distinct advantage in long-distance transmission applications. I heard from the industry that some EML manufacturers are cautious about expanding their production capacity aggressively due to uncertainties surrounding future technology roadmap. Key EML manufacturers include Broadcom ($AVGO) and Lumentum ($LITE)

---

4. Co-Packaged Optics (CPO)

CPO is an emerging technology that takes even a step further from SiPh – aimed at integrating optical transceiver (main functional part of optical module) directly with electronic chips of the switches (such as ASICs or CPUs) within the same physical package, often using co-packaging techniques, and effectively eliminating the need for separate optical modules.

Through removing the need for external optical modules, CPO shortens the connection paths and enhances integration, leading to 30-50% reduction in power consumption (Broadcom claimed a 70% power savings for its latest Baily 51.2T ethernet CPO switch), lower latency, increased bandwidth, and faster data transmission, making it a high-performance solution ideal for large-scale, demanding AI data center networks.

Some argue that CPO represents the ultimate solution for optical networks, but the path to its widespread adoption is still long. One major challenge is that CPO requires a different supply chain structure than the current one, demanding major adaptation across the ecosystem. For instance, most communication in a CPO system occurs at the switch level, which typically operates within relatively closed ecosystems; it’s crucial that switch vendors ensure their protocols are compatible, enabling smooth communication between switches from different vendors. Moreover, in a CPO-driven world, switch chip producers are likely better positioned to produce the new integrated chips, rather than optical module producers. This raises questions about how the latter will adapt – will they resist or find ways to coexist with the new developments? Beyond these supply chain challenges, the CPO technology itself is still in early development stages. All these factors introduce uncertainty into CPO's future, despite its potential. Nvidia projected that the initial volume production for its CPO-based solutions could happen in 2026.

Which companies stand to benefit from the CPO technology roadmap? As previously mentioned, switch chip producers are likely to be key beneficiaries – companies like Broadcom ($AVGO), Marvel ($MRVL) and Nvidia ($NVDA) are actively developing CPO solutions, with production help from foundries like TSMC ($TSM) and GlobalFoundries ($GFS). Broadcom appears to have started early and invested considerable resources into the area; its 51.2T Bailly Ethernet switch, built on CPO design and the Tomahawk 5 architecture, has drawn interests from companies such as Meta and Tencent. Nvidia has been collaborating with TFC Optical Communications (300394.SZ) to develop its CPO solutions, aiming for initial volume production by 2026. Nvidia’s involvement as a switch producer gives it incentive and technical edge in producing the “all-in-one” chip. Regarding component vendors, I don’t think the shift toward CPO poses an immediate threat to driver/TIA manufacturers (Macom and Marvell), since these components are still needed in CPO systems, just that they will be integrated into switch chips rather than into optical modules​.

---

Volume and Price of Optical Modules

800G modules: Based on inputs from several experts on Tegus, I estimate that total shipments of 800G modules in 2024 will be around 10 million units. Of these, roughly half will be for multimode modules, where NVDA is the major buyer, while the other half will be single-mode, with key buyers including MSFT and GOOG. In 2025, 800G is expected to remain a mainstream solution, with projected shipments of 12-15 million units. The split between multimode and single-mode will likely remain roughly half-half with multimode possibly gaining a slight favor.

1.6T modules: Initial shipments are expected in Q4 2024, with ~300K units, mainly supplied by Zhongji Innolight to NVDA. Looking into 2025, the mainstream forecasts suggest 3-5 million units, with more conservative estimates at 2-3 million and most optimistic ones reaching 6 million. Over half of the volume is expected to be purchased by NVDA, to pair with its GB200, while GOOG may contribute 20-30%. GOOG’s primary focus in 2025 is still on 800G, however.

Currently, the primary bottleneck for 1.6T module adoption lies on the supply side, particularly with EMLs and DSPs. For instance, in the case of 200G EMLs used in 1.6T modules, current combined capacity of key suppliers Broadcom, Lumentum, and Mitsubishi are ~20 million units per year, falling short of 40 million units of EMLs demanded for 2025 according to the projected 5 million units 1.6T modules (* 8 channels). These companies handle EML production through an IDM model, which means scaling is not that easy. In addition, many of these companies are cautious about expanding too aggressively due to uncertainty surrounding future technology roadmaps, which we’ve discussed earlier. The industry is exploring solutions – some firms are converting 100G EML capacities to 200G, and new vendors will be entering the market in 2025. However, from a different angle, this challenge with EML supply potentially opens some opportunities for SiPh to step in and help bridge the gap in the latter half of the year, as we discussed earlier.

Optical modules to GPU ratio: One key approach to forecasting the demand of optical modules is by examining the number of total GPUs in data centers and applying a corresponding ratio of optical modules to GPUs. In current training networks, H100 chips are typically paired with 800G optical modules – in this setup, the ratio is approximately 1 H100 chip to 2.5 800G optical modules. Similarly, in the next generation networks, the ratio is 1 B200 chip 2.5 1.6T modules. This general guideline helps estimate the number of optical modules based on the number of GPUs.

However, several factors can influence this ratio. One factor is the number of network layers: as the network becomes more layered, the ratio increases. The 2.5 ratio is typical for current three-layer networks with tens of thousands of GPUs. However, for future networks with potentially five layers and hundreds of thousands of GPUs, the ratio could more than double due to increased complexity and higher bandwidth requirements.

Another consideration is the nature of the networks. The 2.5 ratio applies to training networks, but for inference networks, this ratio could drop significantly. In inference, most data transmission follows in the "north-south" direction – user data is sent to servers hosting the models, processed, and returned – without the need for extensive parallel communication across the network. This is in contrast to training networks, where data flows "east-west" and large amounts of information are exchanged between parallel GPUs. As a result, bandwidth requirements in inference networks tend to “converge” at the higher layers of the network. For example, if the lowest network layer requires 10T of bandwidth, the next layer up might only need 3.3T (would be 10T for training). Combined with the higher latency tolerance in inference networks, the optical module-to-GPU ratio can drop to as low as 0.5 in a pure inference network.

However, there are also some caveats to this 0.5 ratio in inference. First, this ratio applies when the same GPU-optical module pair referenced in training is used in inference, such as H100 GPU-800G optical modules or B200-1.6T. In a purely inference-based network, it’s likely that 800G modules would be sufficient to pair with B200 chips, which would push the ratio upward.

Additionally, the 0.5 ratio does not necessarily suggest that the demand for optical modules will decrease as networks transition from the current training-dominated structure to more inference in the future. At the moment, inference networks are being implemented alongside the training-predominant networks, serving as complementary additions. While inference networks typically leverage fewer optical modules, their ongoing expansion should still contribute to incremental increase in demand for modules.

Pricing: For 800G modules, the multimode versions are currently priced at over US$500. The industry is actively developing Gen-2 multimode modules, which are expected to reduce production costs, bringing prices down to $400-500 next year. Single-mode 800G modules, which are more expensive due to their more complex production structure and costly EML components, are currently priced at over $700, with limited room for price reduction next year. As for 1.6T modules, initial mass production is expected in Q4 2024, with an estimated starting price of around $2,000. However, as production scales up, the price is expected to drop to ~$1,500 next year.

---

Key Players to Pay Attention to in the Optical Module Supply Chain

1. Module Maker – Zhongji Innolight (300308.SZ)

Zhongji Innolight, founded in 2008, is the leading optical module provider in the market. The company is projected to capture 50-60% of the 1.6T module market in 2025 – it was the first to complete testing of 1.6T solutions with NVDA and expected to dominate shipments in Q4 2024; while other vendors are likely to enter the market in 2025, Innolight is still anticipated to retain at least 50% of the global market share.

Overall, I believe the company has a good chance of continuing to hold its leading position going forward. The optical module market is rapidly evolving with the development of AI. Clients in the sector increasingly demand high product quality, reliability, and stable production capacity, favoring industry leader Zhongji Innolight who has a strong R&D focus and robust production capabilities.

However, Zhongji Innolight also faces several challenges in its development. Emerging technology roadmaps, such as LPO and SiPh, provide new entry-point potential for competitors. Eoptolink, another Chinese player, has competitive offerings in LPO and SiPh and may enter NVDA’s supply chain in the coming years. Additionally, industry giants like Intel, Broadcom, and Marvell, which historically haven't competed directly with Innolight in optical modules, are now also trying to develop next-generation solutions in SiPh and CPO, leveraging their existing technological strengths. This could intensify competition and pose new challenges for Innolight moving forward.

Lastly, one of Innolight’s major clients, NVDA, is expected to shift a greater portion of its volume from third-party vendors (where Innolight holds a very leading position) to internally developed modules. While NVDA’s total volume will increase, which benefits Innolight, the company’s overall wallet share within NVDA is likely to decline as NVDA ramps up its in-house production.

---

2. Module Maker – Coherent ($COHR)

Coherent is another notable optical module provider. It has over 6 million units of annual production capacity, with the majority of production based in Wuxi, China, and a portion in Malaysia. The company had planned to shift a significant part of production to Malaysia in Q2 2024, but the transition faced challenges. Due to urgent orders from NVDA and Malaysian factories being unable to scale up production quickly enough, the company had to continue relying primarily on its Wuxi facility for production. The company has traditionally been weaker in single-mode technologies and relatively stronger in multimode, which is also reflected in more production capacity built for the multimode side. In terms of production costs, Coherent is at a disadvantage with costs estimated to be 10-20% higher than Zhongji Innolight.

Coherent's main customers include NVDA and GOOG. The company is expected to have ~20% wallet share in NVDA in 2024. However, regarding its supply to NVDA, issues with its single-mode 800G modules previously led to a halt in supply at the end of 2023, lasting through most of the first half of 2024. Production resumed in Q3 2024, but the supply volume will fall short of the originally planned amount. Regarding the 1.6T modules, Coherent aims to secure again 20% wallet share with NVDA in 2025, though it will face competition from Zhongji Innolight and NVDA’s internally developed solutions.

Coherent’s “habit of fumbling”: From my research on Coherent, I've noticed a troubling trend where the company recurrently fell short of expectations due to poor operations and management. In the 800G era, Coherent failed twice of its single-mode sample submissions to NVDA (as it tried to supply own drivers and detectors in its modules instead of using more reliable solutions from the industry), resulting in loss of a large amount of orders to Innolight. Now, in the 1.6T era, Coherent had originally planned to send samples to NVIDIA in Q2/Q3 of 2024, but the timeline has been pushed back to Q4. This delay puts them much behind Zhongji Innolight, which had already submitted its samples in Nov 2023, and leaves itself out of the running for the initial mass production set for Q4 2024.

On the production front, the attempted transition from Wuxi factories to Malaysia factories turned out to be a little chaotic. The Wuxi factory had preemptively laid off many workers in preparation for the production shift, only to rehire when the ramp-up at the Malaysia facility fell short of expectations. This back-and-forth added unnecessary costs and hurt the organization's overall efficiency.

Overall, these recurring issues highlight that Coherent, from an organizational standpoint, lacks operational discipline and efficiency. These deficiencies become even more pronounced and problematic in the rapidly evolving AI industry. As a result, investors should consider applying an appropriate discount to the firm’s future cash flow projections, given the reasonable likelihood that it may struggle to fully execute its plans.

In mid-2024, Coherent appointed Jim Anderson as the new CEO, who has experience from Intel, Broadcom, AMD, and most recently Lattice Semiconductor. So far, significant transformations within the company have yet to occur under Anderson’s leadership. The main change has been the acceleration of approval processes for high-end optical module projects, reducing timelines from previously 1-2 months to less than a month.

---

3. Customer – Nvidia ($NVDA)

Optical module usage: Currently, NVDA mainly uses 800G optical modules. It uses multimode (VCSEL-based) modules for short-term transmissions and single-mode (EML-based) for long distance. Currently, Close to 70% of the modules are multimode because most modules are used in its InfiniBand (IB) network where a big cluster of GPUs sit relatively close to each other. In addition, most of the modules are conventional modules, as newer technologies like LPO and SiPh are expected to be ready at a later time.

Going forward, the first batch of 1.6T modules (expected for Q4 2024 to Q2 2025) would still be based on more conventional technologies (EML-based). SiPh’s is projected for H2 2025 after its mass production becomes readier. The firm is also considering LPO solutions while there are still some technical challenges in achieving that technology – the company’s current plan is to have a couple of millions of 800G LPO modules in 2025.

Vendors: There are three major vendors for NVDA’s optical module procurement – Zhongji Innolight (300308.SZ), Coherent ($COHR), and TFC-Fabrinet. Innolight currently has over 50% of the wallet share and is a reliable supply chain partner for NVDA. Coherent (Finisar) accounts for 20%+ of procurement with more emphasis on the multimode solutions. The TFC-Fabrinet collaboration essentially represents NVDA’s in-house production, where TFC (300394.SZ) provides the optical transceivers and Fabrinet ($FN) handles the assembly. Fabrinet also acts as a performance benchmark for NVDA’s sample testing process – for third-party vendors to be approved, their solutions must outperform those of Fabrinet. Looking ahead, NVDA is also considering the inclusion of Eoptolink (300502.SH) into its supply chain, primarily for its self-produced modules, which we will explore further in the following sections.

Self-production: As noted earlier, NVDA’s current self-production rate, utilizing the TFC-Fabrinet solution, accounts for 15-20% of its total volume. The company aims to increase this penetration to over 50% by the mature stage of the 1.6T era (thinking 2026).

A key consideration in NVDA’s self-production efforts is the DSP, which makes up ~30% of a typical module’s BOM. NVDA is well-positioned to produce own DSP due to technological similarities with its GPUs and networking devices (such as NVLink and switches), giving the company access to all the IPs required for DSP design. However, according to an industry expert from Tegus, there are still issues with NVDA’s 1.6T DSP product, notably its power consumption, which exceeds 25 watts, making it unsuitable for mass production and widespread deployment at this stage. How fast NVDA can advance its DSP technology will impact the rate of self-production adoption.

In terms of production capacity, the TFC-Fabrinet collaboration currently has the ability to produce around 1 million units, but scaling further is challenging. To achieve over 50% penetration, NVDA will likely need additional help. The company is considering bringing in Eoptolink (300502.SZ) to handle some of the increased demand, effectively taking on a role similar to that of Fabrinet.

NVDA’s self-production efforts will also benefit Macom ($MTSI), which holds a significant wallet share for drivers and TIAs used in these modules. Currently, Macom is the sole supplier of these components for NVDA’s self-produced 1.6T modules. While Marvell ($MRVL) is equally strong in the driver/TIA space, its strategic focus on bundling its advantageous DSPs with drivers/TIAs means it is not suited for NVDA’s solution, where the latter’s own DSPs are heavily favored.

---

4. Driver/TIA Vendors

As previously mentioned, drivers (on the transmitter side) and TIAs (on the receiver side) are critical components for optical modules accounting for c.20%+ of the BOM. Consequently, the growth of the optical module market will naturally drive business of driver and TIA suppliers.

In the driver/TIA market, Marvell ($MRVL) and Macom ($MTSI) are the dominant players, especially in the high-end 400G+ segment. High-speed drivers and TIAs are technically complex to make, leading to a high level of industry consolidation. Semtech ($SMTC), once a strong player in the 100G era, has struggled to keep pace since the transition to 200G, and its 800G products have yet to pass NVIDIA's testing. Broadcom ($AVGO), a major provider of DSP, also has a weak presence in the driver/TIA market – analog chips have never been a core focus for the digital chip giant.

Between Marvell and Macom, the two companies are neck-and-neck in terms of product performance, and clients typically don't show a clear preference between the two. With similar product performance levels, customers focus on factors such as price, service, delivery time, and after-sales support. There is limited brand loyalty in this market, as it's relatively easy for clients to switch vendors of driver/TIA – as long as the products meet specifications for cost, performance, and power consumption, clients are open to considering alternatives.

However, in the LPO market, Macom holds a significant advantage, largely because Marvell, strong in DSP, tends to promote solutions that include DSP (while LPO does not include DSP). ~90% of the drivers and TIAs in LPO modules are currently supplied by Macom.

---

5. DSP Vendors

DSP is used to manage and optimize the transmission and reception of high-speed data signals. It handles complex tasks such as signal distortion mitigation, signal modulation, and error correction. It accounts for c.30% of a typical module’s BOM – for a 800G DSP, the price is currently around $80-90, while for a 1.6T DSP, the price could be over $150 when it reaches maturity stage next year.

Marvell ($MRVL) and Broadcom ($AVGO) are two major suppliers of DSPs, with Marvell holding the top position in the market. Due to its dominance, Marvell maintains rigid pricing and is usually reluctant to offer discounts. Broadcom's DSPs, on average, are 20-30% cheaper than Marvell's. However, both companies are facing some challenges with their 1.6T solutions, making DSPs one of the bottlenecks for the mass adoption of 1.6T optical modules. MaxLinear ($MXL) is another notable player, offering DSPs at about half the price of Marvell’s if it can achieve a meaningful mass production volume of at least 100K units per month.

Marvell's higher pricing reflects the greater ability of product differentiation in the DSP market compared to the driver/TIA market. Power consumption is critical in 800G/1.6T modules – subpar designs can lead to excessive power usage and heating issues. The technical challenge in achieving superior power performance creates an opportunity for leading player (Marvell) to charge premium prices.

Despite this, Broadcom and MaxLinear are also gaining some traction in the market lately. In the Gen-2 800G multimode segments aimed for low cost offerings, companies like Coherent are planning to shift from Marvell’s DSP to those from Broadcom or MaxLinear, giving the latter two an opportunity to increase their adoption.

From the client angles, Zhongji Innolight primarily sources DSPs from Marvell. Coherent also sources heavily from Marvell, with Broadcom/ Maxlinear potentially serving as second supplier with 20-30% share. Eoptolink is expected to source from Broadcom. A key development is NVDA, who traditionally relied on third-party vendors (Zhongji Innolight and Coherent) but has been increasingly shifting toward producing more of its own optical modules with internally developed DSPs. This in-house production is projected to account for over 50% of its future volume, potentially challenging Marvell’s business given the latter’s high reliance on Innolight and Coherent sales.