Two funding rounds and hundreds of millions in capital raised—LUXSiLiCON was established only two and a half years ago. Its financing pace is considered fast for the hard-tech sector.
3D optoelectronic packaging, procurement of mass-production wafers, and the optical I/O team. These three dimensions point in the same direction: transforming "can be made" into "can be afforded." The former is a design challenge; the latter is a manufacturing statistical problem. Between them lies a yield curve.
| OCI Defines Not the Direction, But the Problem
Denali is often described as "China's first optoelectronic integrated transceiver chip compliant with the OCI MSA standard." The emphasis here is not on "first," but on "compliant."
OCI is a physical-layer specification launched in March 2026 by six industry giants: AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI. It strictly defines wavelength grids, modulation formats, packaging tolerances, and external light source metrics. The O-band is divided into A/B groups, with bidirectional transmission over a single fiber; each wavelength operates at 53.125 GBaud using NRZ modulation, with four wavelengths forming one lane. The specification explicitly mandates micro-ring modulators on silicon photonics chips—marking the first time micro-rings have been written into the main body of an MSA.
Denali achieves 8-wavelength multiplexing and 3.2 Tb/s per single chip, doubling the density compared to OCI Gen1 while maintaining architectural homology. This engineering feat carries significant weight, meaning every subsequent metric will be measured against others' standards.
The position of domestic enterprises within this system is quite specific. The standard belongs to the buyers; capacity and PDKs are held by others. What can be earned is earlier delivery and fewer iterations under the same specifications.
| There Is Only One Ledger for Micro-Rings
Debates about micro-rings often go off track. The only factor truly determining success or failure is composite yield.
A single 8-wavelength transceiver chip involves multiple resonant devices connected in series at both transmitting and receiving ends. Yield is an exponential function, not additive. Small differences in single-ring yield, amplified through series connection, become system-level chasms. Shanghai Securities News cited industry experts stating that early yields were less than 10% under 180nm silicon photonics processes. Mass production conditions only gradually matured with the adoption of 12-inch wafers, sub-40nm processes, and optical corrections. LUXSiLiCON CEO Dong Jingjing also confirmed this milestone.
The solution to yield lies not in design, but in process and production control. The entry of industrial capital essentially moves the question of "whether it can be delivered" to the table of those who can make decisions. While its specific investment logic has not been disclosed, its positioning itself highlights where the challenges lie.
| The Temperature Issue Is More Complex Than "0.1°C"
The statement "a 0.1°C change causes wavelength shift" is widely circulated, originally appearing in reports by Shanghai Securities News.
Silicon's thermo-optic coefficient dictates that resonant wavelength drift with temperature is a physical fact. The 3dB bandwidth of micro-rings is very narrow; if detuning exceeds half the linewidth, modulation depth collapses. Compensation relies solely on heaters. However, heater power consumption is not negligible in the system ledger. Academic literature indicates that the power consumption for thermal tuning of micro-rings ranges from several milliwatts per nanometer to tens of milliwatts per Free Spectral Range (FSR), depending on process and design. Denali involves hundreds of rings working in coordination, meaning the system-level power consumption for wavelength control will be significantly higher than that for modulation itself.
Conversely, the modulation energy consumption of micro-rings is inherently extremely low, with academic papers reporting sub-fJ/bit levels. Power consumption for wavelength control can be one to two orders of magnitude higher than modulation power—this is why marketing claims of "10x reduction in energy consumption" typically account only for modulation, excluding wavelength locking. In CPO scenarios, this calculation becomes even more difficult: optical engines are placed adjacent to GPUs consuming hundreds of watts, causing thermal shocks due to load jumps, requiring heaters to dissipate heat while simultaneously locking the wavelength.
This is not a defect that software patches can bypass; it is a physical problem that must be solved with hardware cost at the device level.
| Two Solutions for Yield
Overseas players follow a "pay-as-you-go during service life" model: per-ring factory calibration, runtime thermal control, and DSP compensation, spreading costs across daily electricity bills and maintenance in data centers. Nvidia's CPO route is deeply tied to TSMC, with its yield strategy relying more on runtime measures rather than one-time pre-shipment adjustments.
A different approach has emerged domestically. In September this year, Guanglian Xinke released a maskless programmable lithography machine, using laser annealing to reduce micro-ring wavelength dispersion on 12-inch wafers. They claim full-wafer inspection, adjustment, and verification take 5 hours, boosting yield from less than 10% to 99%. The logic is converting OPEX to CAPEX: pay once for adjustment, then require only minimal energy for maintenance.
There is no industry consensus on which path is cheaper, nor have either reached million-ring scale. But the direction is clear: whether micro-rings can be mass-produced depends on who controls wavelength consistency in the manufacturing stage, not whose design specs look better.
| Implementation Matters More Than Parameters
LUXSiLiCON's disclosed progress includes NPO interconnection innovations with a leading domestic AI company's 10,000-card cluster, providing single-card 6.4T CPO interconnection solutions for core data center switch manufacturers, and supplying NPO socket modules for leading high-performance interconnect chip companies. The socket form factor implies maintainability and replaceability, avoiding the CPO weakness of "one failure requires entire unit replacement," and represents a form easier to scale domestically at present.
The commonality is that all occur within the OCI framework and target replicable large-scale scenarios. If a design company could only achieve impressive numbers in its own laboratory without securing such order structures, it would not survive.
| The Remaining Challenge Is Time
Tape-out verification proves "the design can be manufactured on a specific platform," but does not prove "it can be mass-produced at commercial conditions with acceptable yield and cost." The company has not disclosed yield rates, per-chip costs, production capacity scale, or customer onboarding timelines.
The industry timeline is shifting.
GlobalFoundries' customer tape-outs for OCI-compliant silicon photonics are proceeding throughout 2026, with mass production expected to start in 2027. Bernstein judges that cloud providers will continue to rely primarily on pluggable optical modules between 2026 and 2028, with CPO entry into Scale-up scenarios potentially delayed until after the second half of 2028. Nomura points out that technical challenges beyond single-channel 400G actually leave a longer window for 1.6T.
The window period is measured in years, not months. What LUXSiLiCON is seizing now is not orders, but the position of being "among the first to make micro-rings mass-producible": locking down process collaboration first, accumulating failure mode libraries first, and securing wafer capacity first.
| Conclusion
There is no definitive conclusion on the endgame of optical interconnects. Pluggables, NPO, CPO, and OIO each occupy a segment of the distance, with micro-rings being the most intricate and difficult-to-manage device among them. Choosing them means accepting the premise of "extremely high ceiling, extremely low floor."
Hundreds of millions bought not certainty, but time to continue solving the problem. The question set by OCI is already defined—within the standard, making hundreds of temperature-sensitive rings stable enough for shipment. The answer is not in PPTs, but in the yield curve.
Disclaimer: This article is based on publicly available information. Data not disclosed by the involved companies has not been estimated. Estimates in the text are converted using general industry parameters to illustrate magnitude, not actual measurement data. Opinions belong to the author's analysis and do not constitute investment advice or procurement basis. Data current as of October2026.
Editor: Shuo Xin Yu - Xiaoming Ba