*Fortune* in Conversation with "Lightelligence": Why "Light" Is the Next Battleground in the AI Compute Race | Yunqi Partners

How much electricity does it take to run a large model? That question is moving from technical speculation into commercial ledgers.

Compute demand is still growing exponentially, and the physical limits of electrical computing are already in sight. At this year's GTC conference, NVIDIA unveiled its Feynman architecture, replacing metal interconnects between chips with optical fiber for the first time — boosting bandwidth 10x and slashing energy consumption. Light is entering the very core of AI infrastructure.

But before this consensus formed, some people had already been walking this path for years. Yunqi Capital angel+ round project "Lightelligence" was founded by two co-founders who met in high school and made a pact to build a tech company together. More than a decade later, they converged on the optical chip track, spent three years taping out the world's first 128×128 optical computing chip, and landed commercial orders.

How do they see the relationship between light and electricity? Where does commercialization speed come from? Where is the next compute battlefield? In this edition of "Yunqi Partners," we share highlights from Fortune magazine's recent conversation with Lightelligence co-founders Yinjiang Xiong and Tangsheng Cheng.

Key Takeaways

• Why light: 90% of AI computation is matrix multiplication and addition — light is inherently linear. This isn't a lucky match; it's physically destined.

• How they did it: Lightelligence prioritizes tape-out iteration, running two to three tape-outs per year, using feedback instead of prediction. Tangsheng Cheng had already demonstrated the first optical computing chip in his Oxford lab a decade ago; founding the company was about accelerating mass production, not starting from zero.

• How to win: NVIDIA no longer hypes single-card performance — cluster interconnect efficiency is the next battlefield. The industry is currently doing "electrical compute, optical interconnect," while Lightelligence is positioned at the intersection of both paradigms, working on optical computing itself and optical interconnect simultaneously.

• The starting point: The two co-founders met during a volunteer teaching trip in their final year of high school. When they talked about what they wanted to do later, the answer was: start a tech company. Ten years later, one shut down his U.S. startup, the other took leave from Oxford, and they returned to China together to build the company they'd once described.

The following content is republished from "Lightelligence"

Click the video for full details

---

Yang Anqi, Executive Deputy Editor of Fortune China New Media:

Explain what photonic computing is — what is its essence?

Yinjiang Xiong, Co-founder of Lightelligence:

Photonic computing is essentially a computing paradigm implemented on a novel semiconductor platform. Simply put, it uses photons to compute. Electrons, as fermions, can be trapped inside transistors, but light is more like water — more like a river. When we design photonic computing paradigms, it's like digging channels one by one, letting light flow like water. We place gates on these channels, controlling flow speed through their open and closed states. For light, intensity, phase, and polarization can all be modulated. Through modulation and the interactions between light beams, we achieve computation. The current application scenario for photonic computing is primarily AI inference.

Tangsheng Cheng, Co-founder of Lightelligence:

The computing logic is somewhat different. Photonic computing doesn't use 0-1 signals like traditional digital logic chips to perform computation; instead, it treats computation as an analog quantity. It's significantly stronger than binary computation.

Yang Anqi:

How should we understand this new paradigm of AI computing with light as the standard?

Yinjiang Xiong:

Looking at the industry's evolution, including NVIDIA's recent GTC, light is playing an increasingly critical role in data centers. There are many components here — photonic computing is the computing entity itself, plus interconnect media and data center transmission media, forming a complete system. At the current stage, everyone is working on "electrical compute, optical interconnect," and we have a footprint there too. What we do is called optical compute, though the interconnect still goes through electricity. The future computing paradigm will definitely be optical compute, optical interconnect. So Lightelligence stands at the crossroads of industry transformation, with both paradigm shifts iterating in our own hands.

Yang Anqi:

Will photonic and electronic computing be complementary in the future? Or is the ultimate state, as you said, optical compute and optical interconnect?

Tangsheng Cheng:

The future will definitely be optoelectronic co-design. Whether in computation, transmission, or storage, light's share will keep growing. Simply put, it's like how early chips were CPU-dominated, then GPUs emerged, and now large numbers of AI-specific chips have appeared. Even with all these architectures, CPUs still always hold a place for electronic chips. So I think the future will be a coexisting ecosystem — it's just that photonic computing can play a bigger role in the AI domain.

Yang Anqi:

Are your customers willing to make this migration from the previous computing paradigm to this light-based paradigm? When they weigh costs and risks, how do you convince them?

Tangsheng Cheng:

We need to build a completely reliable product and clearly understand customer needs — which performance improvements they most want to see. For example, we need to demonstrate significant advantages in compute power, power consumption, and latency. Our cost per token needs to be lower than others, and our tokens generated per second need to be higher too.

Yang Anqi:

What are your product plans for the next 3–5 years, and the longer-term 5–10 years?

Yinjiang Xiong:

We're currently working in two directions: one called optical compute, electrical interconnect, and another called electrical compute, optical interconnect. The 5–10 year plan is definitely all-optical computing systems — optical compute, optical interconnect. Future large-scale AI computing definitely won't be solved by a single chip or single board. The ultimate competition will be in cluster efficiency.

Tangsheng Cheng:

To build next-generation photonic computing chips and move toward all-optical computing, we're also exploring replacing silicon with glass as the substrate for photonic computing chips.

Yang Anqi:

I feel like Lightelligence resembles Intel's early story. Will we see that story replay in China?

Tangsheng Cheng:

From a macro trend perspective, the shift from copper to light is inevitable. Communications went from undersea cables to fiber-to-the-home. Now interconnects are going from electrical to optical. Future chip-to-chip interconnect will definitely use light. We believe the computing layer will also gradually move toward all-optical computing.

Yang Anqi:

You two could represent China's most advanced young-founded tech companies. If you wanted to communicate a value proposition to the world, what would it be?

Yinjiang Xiong:

We carry a mission every day: we represent the future of China's photonic computing, representing how to break through AI computing without taking the "NVIDIA plus TSMC" path.