Xinguangjie closed three funding rounds in four months, racing into NPO/CPO to break through the compute interconnect power wall—but the real battle over yield and mass production has only just begun.
Uploading image...
When a semiconductor startup founded in April 2026 and only officially launching in May crosses seed, angel, and angel+ rounds in just four months, secures hundreds of millions in real capital, and simultaneously announces that its first NPO (near-packaged optics) optical engine is being fast-tracked into validation with a leading domestic GPU giant and a core intelligent computing center, what the optical communications and computing capital circles smell is a pungent mix of extreme geopolitical computing anxiety and a generational breakout in optoelectronic co-packaging.
At a time when traditional pluggable optical modules are still shipping at scale, the primary market is injecting capital into a tiny entity at near-rocket speed. What this tears open is not only a desperate self-rescue by computing clusters at the critical point of physical power consumption and the retreat of copper before optics, but also a high-risk sprint around the underlying protocols and packaging yield of the next generation of co-packaged optics (CPO).
The Physical Bottleneck of Computing Interconnect: The Critical Point Where Copper Retreats and Optics Advance
As AI large models race toward parameter counts in the tens of trillions, the networking scale of computing clusters is being pushed to a terrifying magnitude of ten thousand or even one hundred thousand cards. Yet the physical limits of Moore's Law are slamming the brakes on this arms race: on traditional printed circuit boards (PCBs), copper traces face extremely severe signal attenuation, crosstalk, and exponentially exploding heat dissipation power when transmitting ultra-high-frequency electrical signals.
Although optical modules have completed optoelectronic conversion between racks, within the final tens of centimeters of "electrical distance" between the chip and the optical module, power consumption and heat generation have already approached the physical ceiling.If the optical engine cannot be brought as close as possible to the computing core, future computing racks will no longer be constrained by the chip's computing power itself, but will be literally "burned to death" by the heat they emit.Upgrading optoelectronic conversion from traditional pluggable to NPO and ultimately to CPO has become the only path to breaking through the power wall of computing interconnect.
Piercing Through the Business Registration Card: The Contrast Between Foreign-Funded Structure and a Tiny Entity
Following the traces left in business registration to penetrate the organizational structure and capital foundation of this optical upstart, its extremely agile and special entity form stands out clearly in Tianyancha data. Tianyancha business registration data shows that Shanghai Xinguangjie Intelligent Technology Co., Ltd. was founded in April 2026, with Liu Yu as the legal representative, registered capital of approximately RMB 19.4724 million, and enterprise type explicitly marked as a limited liability company (foreign-invested, non-sole proprietorship).
In the tag library accumulated by Tianyancha, this hard-tech enterprise with hundreds of millions in financing commitments is still currently in the "small and micro enterprise" startup classification range.
The combination of foreign-invested status and registered capital in the tens of millions reflects that this team built an extremely mature cross-border structure and technical background from the very beginning. In the field of high-end photonic integration, purely domestic startup teams are rarely able to directly complete closed-loop validation with leading GPU manufacturers within just four months of founding.The business registration characteristics of a foreign-funded entity most likely point to a team with core R&D DNA from top North American chip companies or multinational optoelectronics giants, entering entrepreneurship with mature technical blueprints and process accumulation.
And the frenzied pace of three financing rounds in four months further exposes the fear of missing out among primary market investors.
Capital Frenzy and Industrial Minefields: The Hidden Yield War Under the NPO/CPO Roadmap
Facing the pressing advance of Nvidia and Broadcom in CPO architecture, the hunger for high-speed interconnect in China's domestic GPU autonomous computing chain has risen to the level of strategic security.Investment institutions no longer follow the slow-paced due diligence of traditional semiconductor projects—"first tape-out, then small-batch sampling, then a new round next year"—but instead adopt an aggressive playbook of "securing a position first, adding capital monthly, and using money to hammer out R&D acceleration."
The roadmap put forward by Xinguangjie—"enter via NPO, mainstream CPO, OIO layout"—appears to balance near-, medium-, and long-term commercial considerations, but in reality every step treads along the edge of industrial manufacturing minefields.
NPO places the optical engine on a dedicated substrate close to the computing chip. Although its engineering implementation difficulty is lower than CPO, which is directly integrated on the same packaging substrate, under the domestic semiconductor manufacturing environment it still faces process hurdles such as high-density fiber pigtail alignment and coupling, ultra-high-precision micro-assembly (flip-chip), and severe thermal expansion mismatch under high and low temperatures.
The more fatal hidden reef is the chain-reaction yield penalty brought by optoelectronic integration: under the traditional model, a broken optical module can be hot-swapped and replaced at any time; once the optical engine is packaged together with high-end GPUs or switch chips costing tens of thousands or even hundreds of thousands of yuan, a single failure of the optical chip can mean the direct scrapping of an entire expensive main chip.
At present, the first NPO optical engine has entered the "technical validation stage" with a leading GPU manufacturer. In the cold language of the semiconductor industry chain, it is still vastly far from true commercial mass production. Technical validation only means that basic eye diagrams and bit error rates have been run through on a rigorous test bench in a temperature-controlled laboratory. Moving into a real intelligent computing center will require enduring brutal interrogation under tens of thousands of hours of continuous high load, complex vibration, and extreme heat dissipation conditions.
Domestic dependence on foreign sources for high-end laser epitaxial wafers (InP/GaAs), high-frequency modulators, and silicon photonics foundry processes is likewise a sword of Damocles hanging over such startups, ready to fall at any time.
Conclusion: The Ultimate Showdown Between Hot Money and Yield
This myth of hundreds of millions in optoelectronic wealth creation in Shanghai's Pudong sends the most authentic signal to the entire semiconductor and optical communications track: the physical bottleneck at the foundation of global computing is giving rise to the most aggressive architectural restructuring in decades.But behind the frenzy of hot money pouring in densely and PowerPoints quickly cashing out into high valuations, the final outcome of any optoelectronic integration must still be judged by the placement yield of automated production lines and real heat dissipation power data.
If the high packaging cost cannot be driven down through to the commercial break-even point in real intelligent computing racks, then no matter how dense the halo of financing, it may at any time suffer a cyclical backlash when faced with the severe test of industrialization.