Last week, CSC Financial released a report on the completion of the IPO counseling program for NetForward.
Such milestones occur every year in the A-share market and are not exactly news in themselves. However, what NetForward does is worth highlighting—it develops network interconnect chips, which are the core components that transport data around the clock in switches, servers, and data centers.
| Why This Chip Company Deserves a Second Look
Over the past two years, computing power has been repeatedly mentioned, but most people discuss GPUs and models. Few have noticed that when thousands of GPUs need to collaborate to train a single model, network bandwidth and latency are the real bottlenecks. No matter how powerful the GPU is, if data cannot be fed in and gradients cannot be transmitted out, the entire card is just an expensive ornament.
This is why NVIDIA was willing to pay a premium to acquire the networking business (Mellanox/InfiniBand), and why Broadcom alone captures the lion's share of the global Ethernet switch chip market.
AI computing infrastructure construction is precisely one of the most certain capital expenditure directions in China over the next three to five years. Underlying interconnect chips have shifted from "optional" to "mandatory," with an unprecedented demand for domestic substitution—the demand curve in this track is steadily rising.
| The "Critical Choke Point" Behind an 8Tbps Chip
Back to NetForward.
Their flagship ES8000 series integrates 160 lanes of 56G SerDes on a single chip, providing an 8.0Tbps switching capacity and supporting up to 20 400GE ports. The numbers look impressive, but true experts will focus on one word: SerDes.
SerDes refers to the "transceivers" on a chip responsible for high-speed data transmission and reception. Running a 56Gb/s signal through an extremely narrow channel causes signal integrity issues to worsen exponentially—losses in PCB materials, reflections in vias, and textures in the packaging substrate all become fatal noise.
Developing a single SerDes is not difficult; the challenge lies in integrating 160 lanes of 56G SerDes into a single switch chip and ensuring their stable simultaneous operation under a full configuration of 20×400GE. In the category of commercial Ethernet switch chips, there are only a handful of domestic companies capable of mass-producing 56G SerDes and a complete switching architecture stably. There are no shortcuts for this; it is built on time, the number of tape-outs, and the experience of the engineering team.
Therefore, from an investment or strategic perspective, a company's SerDes capability largely determines the depth of its technological "moat." This is a hard metric that cannot be faked.
| What Is Truly Scarce Is the "Full-Set Capability"
What is even more intriguing is NetForward's product portfolio.
Unlike many peers that only focus on standalone switch chips, it simultaneously covers four major categories: Ethernet switch chips, PHY (Physical Layer), smart NICs, and PCIe Switch. Simply put, from server NICs to switch ports and intra-chassis GPU interconnects, it aims to develop and is shipping products for all of them.
Why emphasize this? Because when enterprises procure network infrastructure, their biggest fear is a "patchwork" supply—different chips from different vendors, making compatibility, consistency, and fault localization a nightmare. Companies that can provide full-stack solutions and self-develop core components naturally enjoy higher customer stickiness and pricing power.
According to public data, its annual shipments of LAN switches over the past two years have exceeded 5 million units, mid-tier chips over 1 million units, and high-end chips above 100G over 400,000 units—low-end for volume, mid-tier for mainstream adoption, and high-end for setting benchmarks. All three tiers have real products running on the network. This capability structure, which "achieves high volume in the consumer segment while establishing a strong presence in the high-end segment," is relatively rare among domestic network chip manufacturers.
| Where Does It Truly Fit in the AI Narrative
In AI data center Scale-out scenarios, which involve multi-rack cluster interconnects, 8Tbps Ethernet switch chips belong to the mid-to-high-end commercial tier, with larger capacity players like 25.6T above them. The more realistic main battleground for the ES8000 is the mid-to-high-end commercial market, including telecom operators, enterprise network TOR (Top of Rack), and industrial Ethernet.
What is more worth tracking for NetForward on the AI front is actually the PCIe 6.0 Switch and the Alink Switch designed for GPU Scale-up. The former addresses high-speed interconnects between GPUs, CPUs, and storage within servers/chassis, while the latter targets Scale-up networks between GPUs. If these products scale up in volume as scheduled, it will truly step into the core position of AI computing interconnects.
| Where Does It Stand in the Landscape
Zooming out to the entire landscape of domestic network chips:
Huawei and ZTE follow the path of "self-developed for self-use + whole machine bundling," with chips primarily feeding their own equipment; Centec Communications is already listed on the STAR Market and is a leading player in mid-to-high-end commercial Ethernet switch chips; Hygon and H3C enter the interconnect protocols (HSL, Lingqu) from the CPU/whole machine side to build ecosystems; while companies like NetForward are positioned more as independent third-party full-stack interconnect chip suppliers—neither bound to a single equipment manufacturer, yet capable of providing complete components from ports to switching.
The unique value of this position lies in the fact that when telecom operators, industry customers, and white-box switch manufacturers want a "multi-vendor strategy" to diversify risks and reduce costs, independent third parties are the most favored option.
| Concluding Remarks
A company that builds the "heart of the network" has reached the doorstep of going public, backed by the slow, meticulous work of a decade spent polishing a single chip.
For executives from the buyer's perspective, the significance of such companies is not just "having another bidder." A more practical implication is that in the second half of the computing power arms race, the independent and controllable nature of the network, much like the chip self-sufficiency rate, will carry increasingly heavier weight in the evaluation of enterprise supply chain resilience.
The 8Tbps ES8000 is a footnote to its firm foothold in the mid-to-high-end commercial Ethernet market; what truly determines whether it can reap the dividends of AI Scale-up is whether the PCIe 6.0 Switch and the UALink-compatible Alink Switch can scale up in volume as scheduled.
Disclaimer: This article is for informational exchange purposes only and does not constitute any investment advice. The stock market involves risks, and investment requires caution.