On November 16, 2025, the company disclosed its financing progress: The Pre-A and Pre-A+ rounds have been completed in stages, raising a cumulative total of several hundred million CNY. The Pre-A round was led by Shanghai Sci-Tech Innovation Group and Shanghai Future Industry Fund, with follow-on investment from the Pudong Angel Fund under Pudong Venture Capital; the Pre-A+ round was led by Matrix Partners China and funds under CICC Capital, with existing shareholders increasing their stakes. Prior investors include CAS Star, SMIC Capital, Suiyue Fund, Lishi Venture Capital, and Echip Semiconductor (希荻微). The capital is allocated to four key areas: R&D iteration, mass production delivery, talent acquisition, and securing top-tier customers.
The core objective for deploying this capital is ultimately to bring light closer to the chip.
In a rack filled with GPUs, chips must constantly communicate. This year operates at 800G, next year at 1.6T, and the following year at 3.2T. Passive direct-attach copper cables are reaching their physical limits; while active copper cables can extend distance, they also increase power consumption and cost.
The faster the signal runs, the shorter the distance it can cover. At the 1.6T generation, passive direct-attach copper cables within a rack typically operate only up to about one meter; beyond that, a retimer chip is required to reshape the signal. Each additional chip means more power consumption, higher latency, and another potential point of failure. With hundreds or even thousands of cables per rack, the implications are clear.
NVISION (英伟芯) is addressing this problem.
| Moving Light Closer
In current data centers, optical components mostly remain outside the rack on the panel level. Optical modules are plugged into rows on the switch faceplate; electrical signals exit the switching ASIC, travel tens of centimeters across the PCB to the faceplate, where the optical module converts them into light for transmission. This path is the bottleneck.
NVISION’s approach is to disassemble the optical module and move the light-emitting component—known in the industry as the "optical engine"—closer to the chip. The closer it moves, the shorter the path electrical signals must travel.
The industry categorizes these approaches by proximity. NPO (Near-Package Optics): The optical engine is moved adjacent to the switching ASIC, reducing the electrical path to 2–15 cm; in some linear drive solutions, this may eliminate the DSP chip used for signal compensation, saving both power and cost. CPO (Co-Packaged Optics): The optical engine and switching ASIC are integrated into a single package, further shortening the path. OIO (Optical I/O): Light is integrated directly into the chip die, allowing inter-chip communication via optics, keeping electrical signals almost entirely within the package and minimizing interconnect distance.
Moving components sounds simple, but implementation is complex. Once the optical engine enters the chassis, it sits adjacent to the hottest switching ASIC, requiring a redistribution of thermal budgets; whereas previously a failed module could be replaced individually, now if it fails inside the package, it may compromise the entire system.
The light source presents an even greater challenge.
Silicon itself does not emit light; photons on silicon photonics chips must be injected externally. If the light source is co-integrated with the package, it encloses a component with a finite lifespan; therefore, the more common approach is to keep the light source external as a pluggable laser, delivering light via fiber. However, since a single optical engine requires over ten optical channels, pulling ten separate fibers is impractical. Instead, a single fiber must carry multiple wavelengths simultaneously—this is why NVISION develops its own WDM light source chip. While outsiders might view it as a secondary component, without this "supporting actor," the "lead role" cannot perform.
| Key Milestones
The 1.6T DR8 silicon photonics chip features eight channels operating in the 1310 nm band, supporting 500-meter single-mode fiber transmission. Designed for pluggable optical modules, this represents the current revenue stream.
The 3.2T NPO silicon photonics chip offers sixteen transceiver channels, 112G per lane, flip-chip bonding, and alignment with the OIF 3.2T NPO standard. Note that "3.2T" refers to the system-level aggregate throughput, not simply the product of single-channel speed and channel count.
The 6.4T NPO/CPO optical engine integrates optical and electrical chips in a2.5D/3D advanced package, achieving 224G per lane and corresponding to 6.4T system-level throughput. According to the company, bandwidth density exceeds traditional optical modules by more than 5 times. This tier is the most difficult: it is not merely selling a device, but integrating into others' packaging ecosystems.
The WDM light source chip covers two bands (1310 nm and 1330 nm) with eight wavelengths each. The OIO optical engine remains in pre-research, representing the ultimate goal of this roadmap.
Based on current disclosures, the 1.6T DR8 is the primary focus for productization; while standards alignment and packaging roadmaps are defined for 3.2T NPO and 6.4T CPO, timelines for tape-out, sampling, and customer integration have not yet been detailed; OIO remains in pre-research. For a company that officially began operations only at the end of2024, how resources are focused amidst a broad product line will be more critical to monitor than the roadmap itself.
| Manufacturing Is the Real Test
Silicon photonics is not a new concept; it has been developed in laboratories for twenty years. The true difficulty lies in mass production.
NVISION’s strategy is "wafer-level heterogeneous integration." Since silicon does not emit light, compound semiconductors are used to supplement it: during the wafer stage, compound semiconductor epitaxial wafers are bonded to silicon photonics wafers before further processing. The challenge lies in the "bonding" process—different thermal expansion coefficients between materials create stress, leading to cracking or wavelength drift. These issues cannot be bypassed by design alone; they require iterative refinement of process windows.
Process windows reside within foundries and OSATs (Outsourced Semiconductor Assembly and Test providers). Consequently, half of this company’s work occurs not in its own lab, but on partners’ production lines: collaborating with leading domestic foundries and advanced packaging facilities to develop silicon photonics PDKs and packaging platforms. A PDK can be understood as a set of process guidelines and component libraries provided by the foundry to design firms, detailing capabilities and tolerances. This is not purchased off-the-shelf but co-developed through joint effort.
| Nie Hui’s Thirty Years
Nie Hui holds a Ph.D. in Electrical Engineering from the University of Texas at Austin. He worked at Lucent Bell Labs before joining Intel, where he led the world’s first low-cost silicon photonics project capable of mass production. That project shipped over 8 million fully integrated silicon photonics chips.
What does 8 million units signify? Not lab samples, but products installed in machines, sold to customers, and operating reliably in data centers for years. Many people domestically can get a chip to light up once; few possess the expertise to achieve yield rates suitable for mass production and costs acceptable for volume orders.
In December 2023, he registered the company in Xi’an, which officially commenced operations in December 2024. Xi’an possesses a strong foundation in the photonics industry and pilot lines. In March 2025, the Shanghai subsidiary was established in Zhangjiang, moving into the "Silicon Photonics Core Source" (硅光芯源) facility in September of the same year, with Suzhou serving as the industrialization hub. The company states that product timing must align with the 2027 window, requiring chips to be ready two years in advance.
Those 8 million chips were produced using Intel’s processes, equipment, and supplier ecosystem. Transitioning to the domestic supply chain is essentially starting from scratch again. This transition is the core work he has undertaken in recent years and is also the aspect most difficult for outsiders to assess.
| Conclusion
The company participates in China Mobile’s DORA Reconfigurable Optical Interconnect, OIF, and ODCC initiatives. On November 2, the company announced joining GCC (Global Computing Consortium) and plans to participate in formulating optical interconnect standards such as OPEN NPO. OPEN NPO is a project launched this year by the Open AI Infra community under GCC, aiming to establish the first domestic NPO optical interconnect MSA (Multi-Source Agreement); initial sponsors include China Mobile Research Institute, Huawei, JD Cloud, Baidu, Accelink, HGTECH, and over twenty other entities. NVISION currently states it "plans to participate" but is not yet listed among the initial sponsors. Essentially, this is where interfaces are defined; once interfaces are standardized, compatibility between modules and systems becomes clear.
In the company’s technology roadmap, OIO is the endpoint. When achieved, light will no longer enter and exit via panels but will travel directly within packages and between chips, rendering the term "optical module" obsolete.
NVISION is betting on the opening of the 2027 window. Its current efforts are aimed at standing at the door ahead of time.
Disclaimer: This article is compiled based on public disclosures and information provided by the company. It is for informational reference only and does not constitute investment advice, securities recommendations, or commercial commitments. Technical parameters, mass production schedules, and standard participation details mentioned herein may change; readers should verify independently and bear their own risks when making decisions.
Editor: Shuoshuo Xinyu - Xiaomingba