CPO packs the optical engine into the switch chip package, incidentally doing something easily overlooked: moving the laser out. While silicon photonics modulators can move along with the switch chip, indium phosphide (InP) lasers cannot—they are sensitive to heat and stress, and must be replaceable if they fail. Therefore, they are housed separately in a pluggable External Light Source (ELS), using polarization-maintaining fiber to send light back to the modulators inside the package.
This change looks like a mere relocation, but it actually elevates the power requirements by an order of magnitude. In the pluggable era, a single laser only served its own module, and 70mW or 100mW was sufficient; in the CPO era, a single laser must split light to multiple links. The architecture presented by Coherent at ECOC 2026 features up to eight ultra-high-power (UHP) CW lasers in one ELS, with each laser splitting light to about four links. As the link budget is spread thinner, the fiber-coupled output power per chip must reach 400mW, and it must remain stable at an ambient temperature of 55℃ or even higher.
400mW has thus become a watershed. It is not simply scaling up a 100mW chip by four times, but rather pushing heat, wavelength, reliability, and manufacturing consistency to their limits simultaneously.
01. As Power Goes Up, Heat Hits the Limit First
The electro-optical conversion efficiency of InP lasers is roughly around 30%, with the rest turning entirely into heat. A 400mW optical output means the chip must dissipate over one watt of waste heat internally, while the ELS box is placed right next to the hot air channel of the switch.
The trouble is that this is a positive feedback loop: junction temperature rises → efficiency drops → more waste heat → junction temperature rises again. All packaging can do is lower the thermal resistance—chip die attach methods, heat dissipation paths, and interface materials; every step must be optimized through multi-physics electro-thermal coupled simulation. This is also why "self-heating - temperature drift - power degradation" is treated as a single chain in the industry, rather than three independent metrics.
The approach at the device level is to maximize conversion efficiency at high temperatures. In its technical presentation at ECOC, Coherent listed high output power, high efficiency, and low noise as the three parallel design goals for UHP lasers. The logic is straightforward: higher efficiency means lower system power consumption, and power consumption is the most sensitive variable in AI data centers.
02. Wavelength Must Not Drift, Noise Must Not Increase
CW lasers only emit light without modulation, but their wavelength and noise directly determine whether the link can be opened. Silicon photonics modulators are sensitive to wavelength. The wavelength of a DFB drifts with temperature by about 0.08–0.1 nm/℃, and the chip's own self-heating adds another layer. ELS units usually have no cooling or only partial cooling, meaning the chip must lock its wavelength on its own across a wide temperature range.
Noise is another hard metric. Relative intensity noise (RIN) and phase noise translate directly into bit error rate degradation; after a laser is split for four links, the noise budget per chip becomes even tighter. This is why both Lumentum and Coherent put "low noise" and "high power" in the same sentence—the two pull against each other.
03. No Shortcuts for Reliability at High Power
For long-term 400mW optical output, the most vulnerable part is the facet. The Catastrophic Optical Damage (COD) threshold of the facet determines whether the chip can survive tens of thousands of hours. The higher the power density, the more easily the facet degrades, requiring passivation and non-absorbing windows on the facet, which leaves a very narrow process window.
Structural choices also diverge here. Buried heterostructure (BH) can confine both carriers and the optical field simultaneously, making it the mainstream solution for high-power CW. However, the buried growth via secondary epitaxy is a notoriously difficult process—the industry treats it as a watershed to judge whether a company has entered the first tier. At ECOC, Coherent explicitly referred to its high-power CW as "high power buried heterostructure CW lasers," emphasizing that this product has been shipping millions of units per month in pluggable modules with field reliability records.
What domestic manufacturers often lack is not the metrics in the lab, but this history of reliability. Batch consistency, yield ramp-up, and long-term aging data can only be accumulated over time.
04. The Real Bottleneck is in the Epitaxy Reactors
Under the feet of all CW laser manufacturers lies the same floor: InP capacity.
Coherent CEO Jim Anderson was very direct during the earnings call on August 12, 2026—InP remains the company's primary internal constraint, and he specifically noted that assembly and test capacity is not the bottleneck; the ceiling lies in laser output. The company's InP laser production in the June quarter grew by about 80% year-over-year, and internal capacity will double year-over-year by the end of this quarter (a quarter ahead of the original plan), with plans to more than double again by the end of 2027; key inputs such as substrates have already been locked in. The 6-inch production lines in Texas and Sweden have higher yields for EML, CW lasers, and detectors than the older 3-inch lines, and the third 6-inch line in Zurich is scheduled to start production in the first half of 2027. Translated into cost terms: CFO Sherri Luther said the output of a 6-inch wafer is about four times that of a 3-inch wafer, at roughly half the cost.
Lumentum's solution is to buy factories. In March 2026, it acquired the former Qorvo semiconductor factory in Greensboro, North Carolina, planning it as a 240,000-square-foot manufacturing base fully adopting a 6-inch InP platform. It is currently operational and undergoing retrofitting, focusing on CW and UHP lasers, with production starting in 2028. Together with San Jose and Kaswell, it forms a "three-stage UHP capacity layout," of which about USD 1 billion comes from NVIDIA's strategic investment.
System manufacturers locking in capacity in advance is itself a pricing of the scarcity of this link: in March 2026, NVIDIA invested USD 2 billion each in Lumentum and Coherent, accompanied by multi-year purchase commitments and capacity usage rights worth billions of dollars.
05. Lumentum and Coherent: Two Routes, One Customer
Currently, only two overseas companies are publicly supplying the 400mW class, and they take different routes.
Lumentum takes the DFB+SOA integration route—the DFB is responsible for generating narrow-linewidth, low-noise seed light, while the SOA amplifies the power to over 400mW. The advantage is that the seed source and power stage can be optimized separately; the trade-off is a larger chip area, with the SOA eating up a portion of the output per wafer, which consequently restricts capacity ramp-up. It was the first to achieve UHP CW shipments; management denied rumors of delayed CPO deliveries at a communication meeting in July 2026, stating that normal supply to NVIDIA has resumed, UHP laser shipments will see a clear inflection point in the fourth quarter of fiscal 2026, and will scale up significantly in 2027.
Coherent takes the monolithic high-power DFB route, relying on the buried heterostructure structure to achieve high power directly on a single chip, coupled with a 6-inch InP platform. It delivered samples exceeding 400mW at 55℃ as early as September 2025; on September 21, 2026, it launched the PhotonLink platform, packaging InP lasers, VCSEL arrays, silicon photonics, isolators, specialty fibers, and detectors into an ELS; mass production of UHP CW and initial CPO production are both scheduled for the fourth quarter of 2026, with revenue ramping up starting from the fiscal quarter ending in December. The company claims to have over ten customer projects each for CPO and NPO, with NVIDIA as the anchor customer.
Broadcom is the third company off the list—it develops its own lasers, using field-replaceable ELSFP light source modules for the Tomahawk 6 switch, but has not disclosed the source of the chips.
06. Domestic Progress: Mass Shipments and Validation
The public statements from domestic manufacturers generally stop at the "validation" stage, but their respective positions vary significantly.
Yuanjie Keji's 70mW and 100mW CW lasers have been delivered in large volumes. In the first half of 2026, revenue reached CNY 925 million, a year-over-year increase of 351.40%, net profit attributable to shareholders was CNY 607 million, a year-over-year increase of 1212.20%, and data center revenue was CNY 774 million. But high-power tiers are a different story: the company clearly stated in its communication on August 30 that 120mW, 300mW, and 400mW CW laser chips are undergoing product validation and R&D in combination with customer needs. In other words, the realized profits come from the general tier used for pluggable silicon photonics modules, while the high-power models exclusive to CPO/NPO are still running at customers' sites. Capacity is being laid out simultaneously: in August, the company announced a planned investment of CNY 4.268 billion to build the Yuanjie Semiconductor Technology Industrial Park.
Shijia Guangzi is moving a bit faster. The 2026 semi-annual report states that the uncooled 100mW CW DFB laser for data center silicon photonics supporting has achieved mass shipments, and the commercial temperature 400mW CW DFB laser has achieved small-batch shipments. On August 4, the company further stated on the interactive platform that the 400mW products are mainly targeted at data centers, and small-batch shipments have currently been made, while promoting sample validation and introduction with multiple customers. The company also plans to raise no more than CNY 2.8 billion, of which CNY 1.456 billion will be invested in the industrialization of CW laser chips and COC.
San'an Guangdian has the widest product range in this tier. The company's official website discloses that BH structure CW laser chips across the full power range of 70mW to 200mW have completed customer validation and are shipped in mass volumes; R&D for the 400mW high-power CW chip for 3.2T optical switches is steadily advancing, with plans to deliver samples in the fourth quarter of 2026 and mass deliver in the first quarter of 2027. During the same period, 100G EML is shipped in mass volumes, 200G EML has completed reliability testing and started sample delivery, and 850nm 100G PAM4 VCSEL is in stable mass production. In terms of capacity, the existing monthly capacity of optical chips is 2,750 wafers, planned to be increased to 4,500 wafers, with capacity expansion equipment entering the factory in the fourth quarter of 2026.
Ligent Technologies, under Hisense, takes another path. The prospectus discloses that its 75mW high-power CW-DFB laser chips fully achieved scaled mass production and external sales in the first quarter of 2026, and the gross profit margin of the optical chip business increased from 32.7% in 2025 to 40.7% in the first half of 2026; it is developing 150mW and 400mW CW-DFB laser chips. Its characteristic is the full chain—integrating optical chip design and manufacturing, optical module packaging and integration, to optical terminal delivery, with self-produced chips directly supplying self-produced optical modules. It has achieved initial commercialization of ELSFP light source modules supporting NPO and CPO technologies, and delivered 3.2T NPO optical engine samples. The company was listed in Hong Kong in September 2026.
The CW lasers of Guangxun Keji also follow the structure of "low-end in delivery, high-end in validation": 70mW and 100mW CW are delivered stably in large volumes, mainly adapted for the external light sources of 800G silicon photonics modules; 120mW/300mW/400mW high-power CW is in the customer validation stage, with plans to deliver samples in the second half of 2026. The company has invested about CNY 6.75 billion to layout high-speed VCSEL, silicon photonics integrated chips, and EML technology, covering 800G/1.6T/3.2T, with large-scale capacity release falling in 2028.
Zhaochi Gufen, which crossed over from TV ODM and LED, is in a later position. The 2026 semi-annual report discloses that its sub-25G DFB optical chips applied to telecom networks have been mass-produced and shipped in large volumes, and the 70/100/200mW CW DFB chips applied to data centers have entered the testing and validation stage. Its approach is to utilize the existing 20 MOCVD equipment units compatible with both LED gallium arsenide and laser chip production, supplementing back-end equipment to build a laser chip production line, and flexibly allocating capacity between the two. The high-power series CW DFB and 50G EML are still under R&D, there is no public timetable for the 400mW tier, and it temporarily cannot enter the shipment column of this statistical table.
Looking at these companies together, the common situation domestically is very clear: lab metrics are approaching, mass production reliability data is still being accumulated, yield and batch stability are in the ramp-up period, and the expansion cycles for InP substrates, MOCVD equipment, and 6-inch processes are all measured in years.
07. The Remaining Question is When the Shortage Will End
Coherent's mass production of UHP CW and its first batch of CPO revenue are both scheduled for that time, and Lumentum's shipment inflection point is also in the same interval. Once overseas mass production lands, the price curve for 400mW and the validation cycles of domestic manufacturers will be repriced.
For domestic manufacturers, the window for this tier does not lie in whether technical validation can be passed, but in whether yield and capacity can be ramped up simultaneously after validation is passed. San'an Guangdian schedules the 400mW sample delivery for the fourth quarter of 2026, which exactly coincides with Coherent's mass production node—this coincidence itself shows that the real race is not about who can light up 400mW, but about who can deliver it stably, batch by batch.