Lightelligence Deep Research

The second stage was the "product puzzle period" from 2021 to 2023. PACE was released in 2021. The company's website described it as its first-generation fully integrated photonic computing platform, built around a 64x64 optical matrix multiplication unit, with more than 12,000 photonic devices on the optical chip, a 1GHz system clock, and 150ps oMAC compute latency. The PACE page provides these product metrics and the optoelectronic conversion path. In 2023, the company also released Hummingbird, Photowave, and other products, expanding the story from "optical computing" to "on-chip optical networking, inter-chip optical networking, and PCIe/CXL optical interconnect." A June 2023 Lightelligence press release described Hummingbird as an oNOC processor for AI workloads. That product moved the company from standalone matrix computing toward interconnect architecture. Photowave was released in August 2023 for PCIe/CXL composable data-center architecture, and the company said it used optical interconnect to unlock infrastructure scalability and disaggregation.

The third stage, from 2024 to 2025, was "a shift of center of gravity toward optical interconnect and scaled delivery." This is the inflection point this report values most. In 2024, the company generated RMB 60.2 million of revenue, including RMB 47.0 million from optical interconnect, mainly from Scale-up products reaching mass production. By 2025, optical interconnect revenue rose to RMB 84.3 million, with Scale-up OCS and new Scale-out revenue becoming growth sources. The prospectus attributes 2025 growth to sales of Scale-up OCS, Scale-out products, and increased product sales. In July 2025, the company also demonstrated an xPU-CPO prototype system with a domestic GPU partner, aiming to package optical engines closer to GPUs/xPUs. The prospectus describes this prototype as a validation node for co-packaged optics technology. In other words, the company itself is moving from "photonic-computing replacement chips" toward "AI cluster interconnect infrastructure."

The fourth stage is the "listing and secondary-market scarcity rerating" in 2026. The company globally offered 13.795215 million H shares in April 2026 at a final offer price of HK$183.2. The Hong Kong public offering was subscribed 5,784.70 times and the international offering 53.83 times, with net proceeds of about HK$2.3769 billion. The allotment results announcement fully discloses the offer price, subscription multiples, and net proceeds. On its first trading day, the stock closed at HK$886, up 383.6% from the offer price, and Sina Finance reported first-day market cap of about HK$81.5 billion. The share price then fell sharply, but remained significantly above the offer price as of June 4, 2026. The core tension in this stage is no longer "whether it can list," but whether the secondary market has priced scarcity too optimistically.

Longitudinal Financial Review

The most direct signal from Lightelligence's financial statements is that revenue has begun to show productization, while losses and R&D intensity remain far above revenue. Revenue was RMB 38.2 million in 2023, RMB 60.2 million in 2024, and RMB 106.4 million in 2025. The prospectus discloses 76.7% revenue growth from 2024 to 2025. In absolute terms, this is not a hardware company that has already crossed the commercialization threshold. It is a company that has just begun converting individual orders, cluster deployments, and optical interconnect products into revenue.

Revenue quality matters more than total revenue. In 2025, optical interconnect revenue was RMB 84.3 million, or about 79.2% of revenue. Optical-computing revenue was RMB 22.1 million, or about 20.8%, with sales driven by the PACE series, the first optical-computing cluster deployment based on PACE 2, and LTSimulator. The prospectus explicitly states that optical-computing product sales increased from RMB 3.0 million in 2024 to RMB 20.2 million in 2025. This split supports two judgments. First, optical computing is no longer only papers and demonstrations. Second, it is still not the company's main revenue source. Treating Lightelligence simply as a "photonic-computing GPU substitute" would overstate the current degree of validation.

Gross margin reminds investors not to look only at revenue growth. Overall gross margin fell from 53.5% in 2024 to 39.0% in 2025, while optical interconnect gross margin fell from 49.1% to 35.1%. The prospectus attributes this to higher raw material, consumable, and outsourcing costs required for Scale-up EPS and intelligent transceivers. This indicates scaled delivery does not necessarily improve gross margin by default. At least in the early stage, customer customization, switch integration, testing and validation, and yield ramp can absorb much of the apparent profitability of high-end hardware.

The expense base still looks more like a research institution than a mature company. R&D expenses in 2023, 2024, and 2025 were RMB 279.8 million, RMB 352.1 million, and RMB 479.0 million respectively, equal to 731.8%, 584.9%, and 450.4% of revenue. The prospectus discloses that all R&D expenses were expensed and not capitalized. This is investment in the technology moat, but it is also a continuing financing requirement for shareholders. Unless revenue quickly steps up, R&D intensity will keep pressuring the profit path.

There is also an accounting factor in the loss that is easy to misread. The company lost RMB 1.3424 billion in 2025, including RMB 843.5 million of fair-value loss on financial instruments issued to investors, mainly driven by valuation increases. The prospectus identifies this non-operating item as an important reason for the expanded loss. Therefore, the RMB 1.342 billion should not be interpreted entirely as operating cash burned. But the loss also cannot be dismissed, because adjusted losses and operating expenses still far exceed revenue.

On cash, before the IPO the company had RMB 209.1 million of cash and cash equivalents at the end of 2025, plus RMB 102.2 million of bank deposits with maturities over three months. The accountant's report discloses these two cash-like assets. IPO net proceeds of about HK$2.3769 billion significantly improved the funding buffer, and the allotment results announcement discloses the net proceeds amount. But this money is not profit. It merely gives the company a 2 to 3 year validation window. If revenue in 2026 to 2027 cannot step up from RMB 100 million level to several hundred million or even RMB 1 billion level, the market will reprice the company as a high-cash-burn hard-tech business.

Share Price and Valuation History

Lightelligence's capital-market history is short, but changes in primary-market valuation are informative. The prospectus shows that the Series A post-money valuation in 2018 was about RMB 256 million, the post-A+/A2 valuation in 2020 was about RMB 884 million, the post-B/B+ valuation in 2021 exceeded RMB 3.2 billion, the post-C1-C3 valuation in 2024 was about RMB 6.279 billion, and the post-C4 valuation in 2025 was RMB 7.8 billion. The Pre-IPO investment table in the prospectus is the first-hand source for this valuation timeline. This valuation line shows that primary markets were trading the technology platform and scarcity of domestic AI silicon photonics, not profitability.

The IPO issue valuation moved another step higher. Based on the HK$183.2 offer price and roughly 91.968 million total shares at listing, the implied market cap at issue was about HK$16.85 billion. Including the exercise of the over-allotment option after the stabilization period, share capital was about 94.037 million shares. The company's May 2026 announcement stated the end of the stabilization period and disclosed the number of over-allotment shares. This was roughly another doubling from the Series C4 post-money valuation of RMB 7.8 billion in 2025 to the listed issue valuation, meaning the market had already assigned premiums for "listed scarcity" and the "AI optical interconnect window."

The first trading day after listing was a typical IPO supply-demand imbalance. The final offer price was HK$183.2, and the stock closed at HK$886 on the first day. AASTOCKS reported that the stock closed at HK$886 for the full first day, up 383.6% from the listing price. The allotment results announcement also warned that because ownership was highly concentrated among a small number of shareholders, even a small amount of H-share trading could cause substantial price volatility. This warning appears directly in the announcement summary. The first-day price therefore should not be treated as fundamental value. It only indicates market supply-demand and theme heat.

By the base date, the share price had fallen from the first-day high but remained far above the offer price. EODHD daily data showed a June 4, 2026 close of HK$569.5. Based on about 94.037 million total shares, the implied market cap was about HK$53.5 billion. That market cap was roughly several hundred times 2025 revenue of RMB 106.4 million. Third-party quote pages in late May also showed a price-to-sales ratio in the several hundreds. The Webull 01879 page showed a market cap of HK$53.55 billion and a P/S ratio of 454.8x at a May 29 price of HK$616. Even though the base-date price was slightly lower, valuation remained in a zone that requires rapid execution.

Business Model and Moat

What Lightelligence really sells is not the concept of "light," but the delivery of photonic devices, electronic control, packaging, software, and system integration to AI compute customers. The optical interconnect line includes Scale-up EPS, Scale-up OCS, intelligent transceivers, NexusBench, and a roadmap toward NPO/CPO. The optical-computing line includes PACE, Gazelle, OptiHummingbird, PACE 2, and others. The company product page lists PACE 2, Gazelle, Risewave, Photowave, and the LightSphere supernode solution. Customers theoretically buy it to solve bandwidth, latency, power consumption, topology reconfiguration, and resource utilization problems in GPU clusters, rather than to replace GPUs across all workloads.

The first layer of technology moat comes from three core technologies: oMAC, oNOC, and oNET. The PACE page explains that inputs are fetched from SRAM, converted by DACs into analog signals, modulated into optical vectors, passed through an optical matrix to produce output optical vectors, and finally converted back to the digital domain through detectors, TIAs, and ADCs. This shows that PACE is a hybrid optoelectronic system rather than an all-optical computer. PACE 2 upgrades to a 128x128 configurable optical matrix, more than 40,000 photonic devices, and a PCIe Gen4 x16 interface. The official website calls it the second-generation optoelectronic accelerator computing card. These technologies are indeed complex and difficult for ordinary optical module vendors to replicate quickly.

The second moat layer comes from system integration and customer validation. The prospectus discloses that in Q4 2025, the company obtained Scale-up OCS orders from a major communications/ICT supplier and an AI computing chip developer, totaling about RMB 87.3 million. Of this, 56% was recognized as revenue in 2025 and 44% was expected to be recognized in 2026. These orders are stronger evidence of commercialization than a trade-show demonstration. The company also said that as of April 13, 2026, its Scale-up solutions had achieved commercial deployment with 51 partners. The prospectus discloses this figure in the commercialization section. However, many customers remain anonymous, so investors cannot independently assess deployment scale or repeat-purchase strength.

The third moat layer may come from the autonomy needs of China's AI infrastructure. In October 2022, the U.S. BIS imposed new export controls on advanced computing and semiconductor manufacturing. The BIS press release lays out restrictions on advanced computing chips and supercomputing uses in China. In October 2023, it further strengthened restrictions on advanced computing semiconductors and manufacturing equipment, and the BIS announcement explains the expanded scope. If Lightelligence can provide optical interconnect components that do not rely on the most advanced process nodes, it may obtain a strategic position in domestic GPUs and AI clusters.

The moat also has obvious gaps. First, the optical-computing software ecosystem is far weaker than NVIDIA CUDA, and weaker than mature AI frameworks' default support for GPUs/ASICs. Although the company says PACE 2 supports ONNX, OpenCL, PyTorch, and TVM, and the OFC 2026 press release describes the PACE 2 software ecosystem this way, "support" is not the same as a mature developer ecosystem. Second, optical interconnect competitors include more than similar startups. They also include Broadcom, Marvell, NVIDIA, Ayar Labs, Celestial AI, and other players that can package optical I/O and switch ASICs directly into system roadmaps.

On management, Yichen Shen is founder, chairman, executive director, and CEO, and received a PhD in applied physics from MIT in 2016. Huaiyu Meng is co-founder, executive director, and CTO, and received a PhD in EECS from MIT in 2018. Long Wang serves as COO and previously worked at Zeku, Unisoc, and Qualcomm. The directors and senior management biographies in the prospectus disclose these backgrounds. The team has a strong technical background, but public-company governance is only beginning to be tested. The company has not yet gone through cycles of public-market guidance, delivery against expectations, equity-incentive dilution, and investor communication.

Industry and Cycle Analysis

The largest industry fact is that AI clusters are pushing interconnect from supporting role to protagonist. As GPU compute becomes stronger, data movement between nodes becomes more of a bottleneck. When training and inference require more GPUs/TPUs in large-scale clusters, copper distance, pluggable optical module power consumption, switch latency, and topology flexibility all become constraints on system throughput. When Broadcom released its 51.2Tbps CPO platform at OFC 2024, it said pluggable optics could account for more than 50% of system power in systems of 51.2Tbps and above, while its CPO solution could save more than 70% of optical interconnect power. Broadcom's press release provides this power comparison. This explains why, after Lightelligence shifted from optical computing to optical interconnect, its commercial evidence became stronger.

Market space should be viewed in three layers. The first layer is overall data-center CapEx and power constraints. The IEA says global data-center investment has nearly doubled since 2022 and was about US$0.5 trillion in 2024. This is a macro base, not Lightelligence's obtainable market. The second layer is datacom optical modules and optical components. Cignal AI expects 2025 datacom optical component revenue to exceed US$16 billion and grow more than 60%. That forecast is mainly driven by 400G/800G shipments. The third layer is NPO/CPO, OCS, PCIe/CXL optical interconnect, and optical-computing accelerator cards. TAM estimates for this layer vary widely, and forecasts of several billion or tens of billions of dollars should not be mechanically applied to Lightelligence's valuation.

Lightelligence sits at the intersection of the "technology iteration cycle" and the "AI capital expenditure cycle." In an upcycle, the most beneficial variables are domestic GPU cluster expansion, AI customers' willingness to pay a premium for optical interconnect, and Scale-up OCS/NPO/CPO moving from demonstrations to standard configurations. In a downcycle, the most fragile variables are customers delaying capital expenditure, projects failing to move from letters of intent to purchase orders, and customers choosing traditional Ethernet/InfiniBand or self-developed interconnect routes to reduce risk.

Geopolitics is both opportunity and constraint. The opportunity is that domestic AI infrastructure needs to reduce reliance on U.S. high-end GPUs and networking ecosystems. The constraint is that the company's EDA, tape-out, packaging, testing, and customer deployments may still be affected by export controls. The prospectus discloses that the latest PACE 2 tape-out service was completed in June 2024. Since June 2024, the company has not conducted substantive follow-up cooperation with the relevant supplier F and has shifted subsequent projects to foundry partners in mainland China. This shows export controls are not an abstract risk, but have already affected supply-chain arrangements. If such constraints lengthen validation cycles or slow yield ramp, Lightelligence's R&D advantage will be eroded by time costs.

Horizontal Competitor Analysis

Lightelligence has no perfectly comparable listed peer. Direct competitors are mostly private companies. Lightmatter works on both photonic computing and Passage optical interconnect. In October 2024, it completed a US$400 million Series D round at a US$4.4 billion valuation, bringing cumulative financing to US$850 million. BusinessWire disclosed this financing and valuation. Ayar Labs works on in-package optical I/O and CPO, and completed a US$500 million Series E round in 2026 at a valuation of about US$3.75 billion, with cumulative financing of about US$870 million. Berkeley BWRC reposted Ayar Labs' financing news. Celestial AI's Photonic Fabric was acquired by Marvell for initial consideration of about US$3.25 billion and maximum consideration of US$5.5 billion. Marvell's press release discloses the transaction structure and scale-up connectivity goal. These cases show that optical interconnect is valued by capital markets, but they also show that high valuations are usually tied to clear customer pull and system roadmaps.

Compared with Lightmatter, Lightelligence's strengths are China localization and listed liquidity. Its weaknesses are still-small disclosed revenue, high customer anonymity, and greater difficulty entering the global ecosystem. Lightmatter's Passage is more explicitly aimed at AI data-center interconnect, and at the time of financing the company said proceeds would prepare Passage for large-scale deployment in partner data centers. DCD's coverage of the round emphasized the mass deployment goal. Lightelligence's LightSphere/Scale-up OCS route is similar, but publicly verifiable customers are fewer, and the Hong Kong share price already implies a market cap close to or above some private leaders.

Compared with Ayar Labs and Celestial AI, Lightelligence is more of a "broader product surface" company. It makes optical-computing cards, optical interconnect, OCS, and NPO/CPO prototypes. But a broad product line also means resource dispersion and a longer validation path. The key after Marvell's acquisition of Celestial AI is not the "photonic concept"; it is Marvell's ability to embed Photonic Fabric into custom XPU and scale-up switch ecosystems. Marvell says its chiplets will be co-packaged with custom XPUs and scale-up switches. To prove comparable value, Lightelligence also needs evidence of adoption by system platforms, not only technical specifications.

Compared with indirect alternatives such as NVIDIA, Broadcom, and Marvell, Lightelligence's disadvantages are ecosystem and capital, while its advantage is focus. NVIDIA has released its Spectrum-X Photonics/CPO switch roadmap and described it as targeting million-GPU AI factories. NVIDIA's press release ties CPO to its AI factory networking roadmap. Broadcom has switch ASICs, optical DSPs, silicon photonics, and a CPO platform. Marvell has strengthened AI scale-up connectivity through Celestial AI. These companies will not directly replace Lightelligence in every China scenario, but they will define global standards and customer expectations.

Compared with the traditional optical communications chain such as Zhongji Innolight, Eoptolink, Coherent, and Lumentum, Lightelligence is positioned closer to computing protocols and system architecture. Zhongji Innolight and Eoptolink benefit from 800G/1.6T optical module volume growth, while Coherent and Lumentum benefit from lasers, optical devices, and optical module components. Lightelligence is trying to place optical interconnect into GPU/xPU supernodes and CXL/PCIe memory expansion layers. This niche is scarcer, but harder to validate. Ordinary optical modules have clearer shipment volumes and customer specifications. Lightelligence's OCS/NPO/CPO products need to enter customer system architectures, increasing sales-cycle and compatibility risk.

Using listed-company valuations as reference, Lightelligence's current pricing is far away from mature chains. StockAnalysis shows that EV/Sales multiples for AI hardware and interconnect companies such as NVIDIA, Broadcom, AMD, and Marvell are mostly around 20 to 30x. NVIDIA's statistics page lists its EV/Sales metric, Broadcom's statistics page lists its EV/Sales metric, and Marvell's statistics page lists its EV/Sales metric. Lightelligence's current market cap implies several hundred times 2025 revenue. That can only be explained by exponential revenue growth over the next several years, not by ordinary peer multiples.

Current Fundamentals and Bull-Bear Divide

The strongest part of current fundamentals is that optical interconnect already has real revenue and orders. Scale-up OCS began contributing revenue in 2025. The company said the product began mass production in Q4 2025 and expects that as production volume increases, manufacturing processes mature, and yield improves, unit cost will decline and optical interconnect gross margin will improve. The prospectus presents this operating assumption in its gross-margin explanation. If this assumption plays out, Lightelligence has a chance to move from "project-based hard-tech company" to "AI cluster interconnect supplier."

The weakest part is customer concentration and order certainty. The top five customers accounted for 84.6%, 87.0%, and 78.9% of revenue in 2023, 2024, and 2025 respectively. The largest customer accounted for 31.1%, 31.8%, and 40.6%. The prospectus customer section discloses this concentration. This means the seemingly rapid 2025 growth may still have been driven by a small number of customers and projects. If the largest customer delays deployment or shifts to self-developed or other suppliers, both revenue and gross margin will fluctuate.

There are three bull arguments. First, AI cluster interconnect is a real bottleneck, and optical interconnect is being advanced by major companies such as NVIDIA, Broadcom, and Marvell. This is not only Lightelligence's story. Second, Lightelligence has three technology lines, oMAC, oNOC, and oNET, giving it both optical-computing optionality and more realistic optical interconnect revenue. Third, IPO proceeds are now in place, giving the company time to continue R&D for PACE 3, NPO/CPO, and Scale-up OCS volume growth.

There are also three bear arguments, and they are more immediate today. First, 2025 revenue was only RMB 106.4 million, while the secondary market prices the company at several hundred times sales, leaving very little valuation tolerance. Second, gross margin fell from 53.5% in 2024 to 39.0% in 2025, indicating that product complexity, outsourcing costs, and delivery validation can suppress economies of scale. Third, optical computing has not yet proven large-scale commercialization. PACE/PACE 2 are better understood as accelerators for specific scenarios and ecosystem exploration, rather than substitutes for general-purpose GPUs.

Valuation Analysis

Valuation should start with factual anchors. The final primary-market Series C4 post-money valuation was RMB 7.8 billion. The prospectus Pre-IPO table discloses the Series C4 pre-money valuation of RMB 7.5 billion and post-money valuation of RMB 7.8 billion. The IPO offer price of HK$183.2 implied an issue market cap of about HK$16.85 billion. The allotment results announcement discloses the final offer price and share capital at listing. The first-day closing market cap was about HK$81.5 billion, and the base-date market cap was about HK$53.5 billion. In other words, even after the pullback, the secondary market remained more than about 2x above the issue valuation and about 5x above the final primary-market valuation.

Peer multiples give a sharper conclusion. Mature AI hardware/interconnect companies such as NVIDIA, Broadcom, and Marvell mostly trade at revenue multiples around 20 to 30x. Marvell's StockAnalysis page provides EV/Sales statistics. High-growth interconnect stocks such as Astera Labs can command higher valuations. StockAnalysis shows Astera Labs in a higher-growth-stock EV/Sales range. But Lightelligence's current revenue multiple is several hundred times, and the company is not yet profitable. "Peers are also expensive" is not enough to conclude that Lightelligence is reasonably valued.

Absolute valuation can only be scenario-based; pretending precision would be misleading. My three price ranges are as follows:

The base-date price of HK$569.5 was above the top of the bull-case range, so the margin-of-safety conclusion is: none. Even assuming Lightelligence's revenue grows rapidly over the next 3 years, the current price requires revenue, gross margin, customer diversification, and capital-market style to all favor the bulls at the same time. If profit remains negative over the next 3 years and revenue only rises from RMB 100 million to RMB 500 million to 800 million, the current market cap will still look expensive. If revenue enters the RMB 2 billion to 3 billion range and gross margin recovers to 45% to 55%, then the current valuation may be reinterpreted.

The ideal buy price is set below HK$100, not because the company is worth only HK$100 per share, but because the largest risk for early-stage hard tech is that valuation can quickly return to financing anchors if the path is disproven. The holdable price range is HK$150-260, and the clearly overvalued range is above HK$460. The current price is classified as clearly overvalued. Is it worth waiting for a better price? Yes. The opportunity cost of waiting is that a major customer order announcement could create short-term upside, but that does not change the current lack of margin of safety.

Risk Analysis

The first risk is commercialization risk, with medium-high probability and high impact. Observable indicators include 2026 optical interconnect revenue, Scale-up OCS/NPO/CPO orders, customer count, and single-customer revenue share. If 2026 revenue remains around RMB 100 million to 200 million, or the largest customer continues to account for more than 40%, the market will again view the company as a project-based company rather than a platform supplier.

The second risk is gross-margin risk, with medium probability and high impact. Overall gross margin fell to 39.0% in 2025, and optical interconnect gross margin fell to 35.1%. The prospectus explains why Scale-up EPS and intelligent transceivers had higher costs. If subsequent orders require more customized switches, outsourcing, testing, and customer on-site support, revenue growth may not translate into profit improvement.

The third risk is technology-route risk, with medium probability and high impact. PACE 2, PACE 3, and the optical-computing ecosystem remain early. Optical computing may stay for a long time in dedicated acceleration, research, optimization, and niche scenarios, rather than becoming general-purpose AI inference infrastructure. LightOn once promoted optical processing units, but by the time it listed on Euronext in 2024, it looked more like a generative AI software company. Euronext describes LightOn as a generative AI software solutions company. This precedent is a reminder that optical-computing companies often pivot toward interconnect or software.

The fourth risk is geopolitical and supply-chain risk, with medium probability and medium-high impact. The company needs coordination across EDA, tape-out, packaging, testing, optical devices, and customer system integration. Any link affected by export controls, supplier restrictions, or domestic-substitution yield drag could delay revenue recognition. BIS advanced-computing restrictions in 2022 and 2023 have already shown that policy can directly reshape AI hardware supply chains. The 2023 BIS announcement strengthened restrictions on advanced computing semiconductors and semiconductor manufacturing equipment.

The fifth risk is valuation and liquidity risk, with high probability and high impact. The allotment announcement already warned that concentrated ownership could cause large price swings from small trading volumes. This warning appears in the allotment results announcement summary. The current price remains high relative to the offer price. If free float increases after the stabilization period, early investors' lock-up expiry expectations heat up, or hard-tech sentiment cools, the share price may correct before fundamentals do.

Catalysts and Tracking Indicators

Positive catalysts include: Scale-up OCS winning new large purchase orders; NPO/CPO prototypes entering customer validation with named partners disclosed; PACE 2 or PACE 3 receiving independent third-party benchmarks in real AI inference or optimization tasks; 2026 revenue exceeding RMB 500 million with gross margin recovering; largest-customer revenue share falling below 30%; and changes in U.S. export controls instead accelerating Chinese customers' adoption of domestic optical interconnect solutions.

Negative catalysts include: 2026 revenue below RMB 300 million; optical interconnect gross margin remaining below 35%; orders mainly coming from letters of intent rather than purchase contracts; delays in PACE 3 R&D or the software ecosystem; major customers canceling or delaying supernode deployment; and lock-up pressure resonating with a style rotation away from technology stocks.

The tracking dashboard is as follows:

Zen Horizon Cross-Sectional Summary

Longitudinally, Lightelligence has genuinely proven three things. First, it can turn MIT-level photonic-computing research into multiple generations of hardware products. Second, after finding that pure optical-computing commercialization was slower, it could shift its focus toward optical interconnect, which is closer to AI data-center pain points. Third, it could complete financing and listing within China's hard-tech environment and Hong Kong's Chapter 18C window. None of these was easy, and they show the team has more than papers and concepts.

But it has not yet proven the most important fourth point: repeatable, scalable, profitable delivery of large-scale AI interconnect systems. 2025 revenue was only RMB 106.4 million, and scaled delivery has just begun. Gross-margin decline shows that system complexity and delivery costs remain high. Customer concentration shows revenue quality has not yet diversified. The next chapter of the longitudinal story is not the release of an even larger optical matrix. It is continuous customer purchasing, falling costs, and software and system integration that do not become bottlenecks.

Cross-sectionally, Lightelligence's greatest advantage is scarcity. It combines optical computing, optical interconnect, NPO/CPO prototypes, and the domestic China AI infrastructure narrative. The Hong Kong market has almost no second comparable name. Its greatest weakness is also scarcity. Because there is no perfectly comparable listed company, the market can easily replace "verifiability" with "uniqueness" and write five years of future success into the price all at once.

There are four bull points. First, AI data-center interconnect bottlenecks are real, not marketing language. Second, Lightelligence's optical interconnect revenue is already stronger than optical-computing revenue, showing that the company has found a more realistic commercial entry point. Third, IPO proceeds improved the cash buffer, supporting continued R&D and customer validation. Fourth, geopolitics may make domestic AI clusters more willing to try independent optical interconnect solutions.

There are also four bear points. First, current valuation far exceeds revenue scale and profitability. Second, customer concentration and anonymous customers make order quality difficult to verify externally. Third, the optical-computing route has not proven it can replace the GPU ecosystem and should be viewed as dedicated acceleration and interconnect enhancement. Fourth, gross margin declined while revenue grew, indicating scaled delivery may be more expensive than investors expect.

Pre-mortem scenario one: in 2027, new Scale-up OCS orders fall short of expectations, major customers delay domestic GPU supernode expansion, company revenue reaches only RMB 300 million to 400 million, gross margin remains below 35%, and the market rerates the company from an "AI silicon photonics platform" to a "project-based optical interconnect supplier." At that point, the revenue multiple could compress from several hundred times to 50-80x. Even if the technology still has value, the share price could return to HK$150-250.

Pre-mortem scenario two: PACE 3 is delayed, NPO/CPO customer validation progresses slowly, Broadcom, Marvell, NVIDIA, and large domestic companies' self-developed routes dominate mainstream interfaces, and Lightelligence can deliver only in a small number of customized scenarios. If technology-stock valuations contract at the same time, the share price could fall back near or even below the IPO offer price. That would not necessarily mean the company failed. It would be the common path where good technology fails to become a good stock.

Final company profile: fundamental quality is medium-low, growth is high but unverified, moat is moderately strong, financial resilience is medium-low, management's technical credibility is high, valuation attractiveness is low, and risk level is high. It is suitable for high-risk hard-tech thematic investors, not ordinary long-term investors who emphasize margin of safety.

Investment rating: Avoid. One-sentence investment thesis: the technology deserves tracking, but the current price has overdrawn commercialization. The ideal buy price is below HK$100; the holdable price range is HK$150-260; the clearly overvalued price is above HK$460. The current HK$569.5 is clearly overvalued. If reassessed in the future, the target holding horizon should be at least 3 to 5 years. At present, 12-month thematic trading should not replace fundamental judgment. This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.

Key Data Table

Research Uncertainty

First, most customer names are anonymous, and public information cannot independently determine the real scale, repeat-purchase probability, or bargaining power of each deployment. Second, EODHD's quote API can retrieve the June 4, 2026 closing price, but web quote sources may have cache delays, and market cap is roughly calculated from share capital and closing price. Third, public TAM forecasts for optical computing and CPO/NPO vary widely, and this report does not use exaggerated TAM as a primary valuation basis. Fourth, the company has not yet disclosed a full post-listing financial report after the IPO, so 2026 orders and cash flow still need to be verified by subsequent announcements. Fifth, this article does not reference any other blind-test variant outputs; all judgments come from public sources, repository templates, and this independent research.