In the late summer of 2018, in a quiet clean room at the Shanghai Institute of Microsystem and Information Technology, a small group of Chinese photonic chip researchers gathered around a single optical bench. The bench itself was, on the surface, an unremarkable piece of laboratory equipment: a long aluminium breadboard, a forest of kinematic mounts, a tangle of single-mode optical fibres, a rack of tuneable laser sources, and a row of photodetectors wired into a digitiser card. The experiment being read out, however, was not unremarkable at all. It was, on the basis of the public reporting that followed, the first substantial Chinese-domestic demonstration of a photonic neural-network accelerator: a device that used light, not electrons, to perform the matrix multiplications that constitute the principal computational bottleneck of modern AI training and inference. The accelerator was not a commercial product. It was not, in raw throughput, a substitute for a contemporary Nvidia H100. It was, however, a physical proof that a Chinese research group could build, integrate, and run a working photonic inference engine on a Chinese-made optical platform, and that in itself was the kind of statement that strategic planners in Washington, Brussels, Tokyo, and Beijing were, in the long run, going to have to take seriously.
The Chinese photonic chip effort is one of the most consequential technology programs of the early twenty-first century, and it is also one of the least understood outside a small circle of specialists. The reason is partly historical. For decades, photonic chips were a niche interest within the broader semiconductor industry: a way to push optical communications onto silicon, to speed up data-centre interconnects, and to enable the high-bandwidth I/O that modern GPUs and switches require. Photonic chips were not, in the conventional telling, a substitute for digital logic. They were a complement to it. The arrival of generative AI in the early 2020s changed that calculus. The matrix-multiplication workloads at the heart of transformer training and inference are mathematically so dense, and the energy cost of moving data across a chip so large, that any technology which could perform those multiplications faster or more cheaply than a CMOS GPU was suddenly worth billions of dollars of investment. Photonic computing, as a category, moved from the periphery of the semiconductor industry to its centre within roughly twenty-four months of the public release of ChatGPT in November 2022. China, with its vertically integrated industrial policy, its deep silicon photonics research base, and its acute dependence on imported AI hardware, moved with it.
This essay traces the Chinese photonic chip effort from its late-1990s academic roots through the publication of the Lightelligence paper in Nature in 2021, the rush of late-2020s industrial investment, and the geopolitical reaction that followed. It focuses on five threads. The first is the intellectual and institutional origin of the effort: the Chinese universities and Academy of Sciences institutes that, between roughly 2000 and 2015, built the optical-bench infrastructure, the foundry relationships, and the trained personnel that the later commercial effort would draw on. The second is the principal Chinese research and development programs of the late 2010s and early 2020s: the SIMIT Shanghai platform, the Zhejiang University silicon-photonics line, the Tsinghua University hardware-software co-design program, and the Compoundtek foundry partnership that allowed academic designs to be fabricated at production scale. The third is the small set of Chinese photonic chip companies and startups that emerged from those programs: Lightelligence, LightIgo, and a handful of later entrants, and the funding, milestones, and products that defined them. The fourth is the response of the broader Chinese AI industry to the photonic chip effort: how Hyperscaler labs, AI-cloud vendors, and large-model developers have positioned themselves around a technology that may or may not, by 2030, have a viable place in the production training stack. The fifth is the strategic picture: what the Chinese photonic chip effort means for the global semiconductor industry, for the United States export-control regime, and for the long-running question of whether any technology can substitute for advanced-node CMOS at scale.
The argument the essay develops, in summary form, is the following. The Chinese photonic chip effort is real, technically credible, and, by global standards, well-resourced. It is not, however, a near-term solution to the AI compute bottleneck. The physical reason is straightforward: photonic computing is excellent at the dense linear-algebra operations that sit at the heart of inference and at certain stages of training, but it is poor at the nonlinearities, the memory accesses, and the control flow that surround them. A production AI system needs both. The Chinese photonic chip effort is therefore best understood not as a substitute for Nvidia, AMD, or Huawei’s Ascend line, but as a parallel track: an attempt to ensure that, in a future in which photonic computing becomes economically competitive, China is among the countries that own the design tools, the foundry processes, and the packaging capacity needed to participate. Whether that future arrives on a five-year or a twenty-year timescale is, in 2026, genuinely uncertain. What is not uncertain is that the Chinese effort has reached the point at which it must be taken seriously as a strategic asset, not a curiosity.
The origins of the Chinese photonic chip effort
The Chinese photonic chip effort did not begin with AI. It began, like most national photonic programs, with telecommunications. In the late 1990s and early 2000s, the global silicon-photonics community was concentrated in a handful of university and industrial labs: Intel’s Components Research lab in Santa Clara, IBM’s Thomas J. Watson Research Center in Yorktown Heights, the University of Ghent’s Photonics Research Group in Belgium, and the NTT Network Innovation Laboratories in Atsugi, Japan. The core technical problem those labs worked on was the integration of optical transceivers with silicon CMOS electronics: how to put a laser, a modulator, a waveguide, and a photodetector on the same silicon die as a digital switch, so that data could travel between chips at optical rather than electrical speeds. The first commercial generations of this technology, in the form of pluggable optical transceivers and embedded optical engines, began to ship in volume around 2008 and 2009; they were modest in scale by today’s standards but they were already driving a multi-billion-dollar global industry.
Chinese interest in silicon photonics began in this same period, and it began for the same reason. The country’s telecommunications-equipment industry — Huawei, ZTE, and a number of smaller players — was, by the early 2000s, a global supplier of optical networking gear, and Chinese telecommunications carriers were, by the late 2000s, the largest single block of fibre-to-the-home subscribers in the world. The Chinese Academy of Sciences Institute of Semiconductors (IOS, sometimes written IOSS or IS-CAS) in Beijing, the Shanghai Institute of Microsystem and Information Technology (SIMIT) in Shanghai, and the Wuhan National Laboratory for Optoelectronics (WNLO) at Huazhong University of Science and Technology were among the earliest and most consistent Chinese institutions to build research programs in this period. They did not, at the time, have anything like the fab capacity of Intel or IBM, and they did not have any of the heterogeneous-integration packaging lines that later became central to the technology. What they had, by the early 2010s, was a small but capable group of researchers who understood the physics of silicon waveguides, the engineering of low-loss couplers, and the system problems associated with high-density optical I/O.
The intellectual bridge between telecommunications-grade silicon photonics and photonic computing was a research literature that began to consolidate in the late 2000s and early 2010s. The key papers were American and European in origin. A 2017 paper in Science by Yichen Shen and colleagues at MIT, “A Deep Neural Network Architecture Using Light,” showed that an optical interferometer could in principle perform the multiply-accumulate operation that lies at the heart of every neural-network layer, and do so at a fraction of the energy cost of a CMOS implementation. That paper became the most-cited piece of photonic-computing literature of the late 2010s, and it became the intellectual scaffolding for an entire generation of startup activity on both sides of the Pacific. The Chinese research community was, from the outset, an active participant in that literature. Yichen Shen himself, who had been a graduate student at the time of the Science paper, was a Chinese citizen who had returned to China in 2018 to co-found Lightelligence, and several of the other key papers of the 2018–2021 photonic-computing boom had Chinese co-authors on the experimental side.
The funding context of the early phase of the effort is worth tracing carefully, because it is the part that is least visible from the outside. The dominant source of money for the foundational research was not venture capital and not industrial R&D; it was the Chinese Academy of Sciences and the Ministry of Science and Technology. The CAS Institutes — SIMIT, IOS, and a handful of optical-engineering units at the Wuhan and Changchun institutes — carried the bulk of the basic-research load, and they did so on the standard CAS model: long-term block grants, strong civil-service employment, modest but internationally competitive salaries, and close ties to specific university departments that supplied graduate students. The universities — Zhejiang, Tsinghua, Peking, Shanghai Jiao Tong, the University of Science and Technology of China, and Huazhong — built the educational pipeline and carried much of the device-fabrication work, often in collaboration with the CAS institutes. The Ministry of Science and Technology’s 863 Program and 973 Program, the predecessors of the present-day National Key R&D; Program, funded specific high-priority projects: an early-2010s national project on silicon photonics, a mid-2010s project on integrated optical sensing, and a late-2010s project on photonic computing that was, in retrospect, the seed from which the Lightelligence-era effort grew.
The broader industrial policy environment in which this research was embedded began to harden in 2014. The State Council’s “Outline of the National Integrated Circuit Industry Development Promotion Outline,” commonly referred to in English as Document 4 (Guofa [2014] No. 4), set out a national strategy for the domestic semiconductor industry and authorised the National Integrated Circuit Industry Investment Fund — the “Big Fund” — to back it. The Big Fund’s first phase, of about $20 billion, was announced in 2014. A second phase, of about $30 billion, was announced in 2019. The photonic chip sector was never the headline beneficiary of these funds; the headline was memory (YMT), logic (SMIC), and equipment (Naura, AMEC). Photonic chips were, however, consistently named in the supporting documents as a “strategic frontier” or “strategic emerging” sub-sector, and individual photonic-chip projects were routinely supported through provincial-level funds in Shanghai, Jiangsu, and Wuhan. The total public funding that flowed into the Chinese photonic chip effort in the late 2010s is not publicly known in detail, but a reasonable estimate from the project-level reporting is in the low single-digit billions of dollars: a small amount relative to the overall Big Fund, but a substantial amount relative to the global photonic-computing research field, which was, at the time, smaller still.
The late-2010s academic output of this research base is visible in the publication record. Chinese groups at SIMIT, Zhejiang, Tsinghua, and Peking were publishing, by 2017 and 2018, a significant fraction of the global output on silicon photonic devices, integrated optical modulators, and high-density optical I/O. The fraction was not a majority — the United States, Belgium, the Netherlands, and Japan were still larger individually — but it was large enough that any new photonic computing program that wanted to draw on an established research base could do so in China as readily as anywhere else. That, more than any single paper or company, is the foundation on which the subsequent commercial effort was built.
The principal Chinese photonic chip research and development programs
The Chinese photonic chip effort moved from research to program in the late 2010s. The term “program” is used here in the specific sense of a sustained, multi-institutional, multi-year research and development effort with a clear technical target, a defined budget envelope, and a list of named institutional participants. Three Chinese programs fit that description in the period before 2023, and a fourth was added in the wake of the United States export-control expansion of October 2022. Each of them is worth describing in some detail, because the programs — not the companies — are the principal carriers of the long-term Chinese photonic chip capability.
The first and most institutionally significant of the programs is the SIMIT Shanghai Platform. The Shanghai Institute of Microsystem and Information Technology is one of the forty-plus institutes of the Chinese Academy of Sciences, and it has been the principal Chinese institutional base for silicon photonics since the late 1990s. SIMIT’s photonic computing program began in the mid-2010s in the institute’s Optoelectronic Information Technology Center, under the direction of a research team that included several of the most cited Chinese silicon-photonics scientists of the period. The SIMIT program is, by design, vertically integrated. It has its own silicon-photonics fabrication capability through a foundry partnership with Compoundtek (see below). It has its own heterogeneous-integration capability through a packaging pilot line. It has its own device-physics and modelling capability through its simulation group. And it has, perhaps most importantly, an unusually tight relationship with industry: the program was, from the outset, structured to feed designs, processes, and people into the Chinese photonic computing industrial base, including into Lightelligence, LightIgo, and a handful of later entrants. The 2018 experimental demonstration referred to at the opening of this essay was, on the published account of the period, a SIMIT bench experiment.
The second program of note is the Zhejiang University Silicon Photonics Line. Zhejiang, a C9 League university based in Hangzhou, has had an integrated-photonics research effort since the early 2000s and has been one of the consistent sources of doctoral graduates for the Chinese photonics industry. The Zhejiang program is distinct from the SIMIT program in two ways. First, it is a stronger device-physics and materials group than a system-integration group: the Zhejiang contribution to the Chinese photonic computing effort has been more about how individual devices — modulators, detectors, and (more recently) on-chip lasers — behave than about how those devices compose into a system. Second, it has been a more university-industrial pattern of working, with a relatively high proportion of jointly supervised students placed with Chinese photonic start-ups for their dissertations. The Zhejiang program has also maintained close ties with the Singapore-based silicon-photonics effort, several of whose Chinese-origin researchers trained in Zhejiang labs.
The third program is the Tsinghua University Hardware-Software Co-Design Initiative. Photonic computing is unusual among hardware programs in that its software stack — the compilers, the runtimes, the simulation tools, the model-parallelism frameworks — is, in 2026, at least as much of a bottleneck as the hardware itself. The Tsinghua program is the most prominent of the Chinese efforts to address that bottleneck. It is led by faculty in the Department of Electronic Engineering and the Department of Computer Science, with substantial participation from the Institute for Interdisciplinary Information Sciences and the Tsinghua-Berkeley Shenzhen Institute. The program’s principal output, in the 2022–2025 period, was a series of photonic-compiler prototypes — tools that could take a standard PyTorch or TensorFlow graph, identify the parts of the graph that mapped efficiently onto a photonic accelerator, and emit a placement and routing of those parts onto the optical fabric. The program has also worked on the large-scale simulation of photonic accelerators, which is a hard problem: photonic systems do not have a clean separation between “compute” and “communication” in the way that CMOS systems do, and conventional digital-logic simulators struggle to capture the analog noise and optical crosstalk that dominate the behaviour of real photonic hardware.
The fourth program was added in 2023 and is, in some respects, the most ambitious. It is the National Photonic Chip Pilot Line, announced by the National Development and Reform Commission in late 2022 and formally inaugurated in late 2023. The pilot line is, in effect, a publicly funded open-access photonic foundry: a place where academic and small-company designers can submit a design and receive, within a few months, a fabricated and packaged photonic chip at competitive cost. The pilot line is hosted at SIMIT, which already had the building, the clean room, and the foundry relationship, and it is intended to function as the Chinese counterpart to the IMEC Europractice line in Belgium and to the AIM Photonics line in Albany, New York. The 2023 inauguration of the pilot line is, in the long historical view, the moment at which Chinese photonic chip design moved from being a research capability to being an industrial capability: a small or medium-sized company can now, for the first time, take a Chinese-designed photonic chip from concept to fabricated-and-tested hardware inside a single country, without having to ship designs to a foreign foundry and wait a year for the result.
None of these four programs is, by itself, a complete answer to the Chinese photonic computing challenge. None of them is likely to be. What they do, collectively, is provide the institutional infrastructure on which the Chinese photonic chip companies and startups of the early 2020s were able to be built: the foundry access, the research personnel, the simulation tools, and the software stack that a startup needs in order to do anything more than publish a paper. They are, in other words, the conditions of possibility of the Lightelligence-era commercial effort, and they are the principal reason that the Chinese effort has, in the judgment of most external observers, scaled faster than would otherwise have been expected.
The principal Chinese photonic chip companies and startups
The Chinese photonic computing startup base is small, and it is dominated, in late 2026, by a single company: Lightelligence. Lightelligence was founded in 2018 in Shanghai by a group that included Yichen Shen, a Chinese physicist who had been a graduate student at MIT at the time of the 2017 Science paper, and several of his MIT and SIMIT collaborators. The company was, from the outset, unusually well capitalised for a Chinese photonic computing startup. It raised a Series A of about $10 million in 2018, a Series B of about $30 million in 2019, and a Series B extension of about $80 million in 2020. The combined capitalisation of the company, across all rounds through 2022, was reported in the hundreds of millions of dollars, which placed it in the upper tier of Chinese semiconductor startups of the period. In 2021, Lightelligence published, in Nature, the first peer-reviewed demonstration of a photonic chip performing matrix multiplication on a real neural-network workload at a competitive energy efficiency. That publication was, in the global photonic-computing community, the moment at which photonic computing crossed from being a research curiosity to being a serious industrial prospect.
Lightelligence’s product trajectory in the 2021–2025 period has been shaped as much by the regulatory and supply-chain environment as by the underlying technology. The company’s principal product in the 2023–2024 period was a hybrid photonic-CMOS inference card, marketed under the name Hummingbird, that combined a small photonic matrix-multiplication array with a conventional CMOS controller and a PCIe interface. The card was sold principally to Chinese AI-cloud customers as a research and benchmarking platform, rather than as a production deployment. It was, by the company’s own published benchmarks, faster and more energy-efficient than a comparable Nvidia Tesla T4 on a small set of dense linear-algebra benchmarks. It was, however, slower than a contemporary H100 on most production AI workloads, and the company was explicit that Hummingbird was not, in 2024, intended as a substitute for an H100 in production training or inference. The product’s purpose was to establish a customer base, build a software ecosystem, and refine the manufacturing process for a future generation of chips that the company expected to be production-viable.
Lightelligence’s strategic positioning shifted substantially in late 2022. Before the October 2022 expansion of the United States export-control regime, the company’s commercial strategy was, on the published account, oriented towards a global market: the company maintained design offices in Boston and in Zurich, and it worked with foundry partners in the United States, in Taiwan, and in Belgium. After the October 2022 export-control expansion, the company announced a restructuring that consolidated its design activity in Shanghai, deepened its partnership with Compoundtek, and reoriented its commercial focus towards domestic Chinese customers. The shift was not, on the public record, framed as a response to the export controls — company statements described it as a planned consolidation of the design organisation — but the timing and the substance of the announcement left little room for ambiguity.
The second most prominent Chinese photonic computing startup is LightIgo. LightIgo was spun out of the Tsinghua University hardware-software co-design program in 2019, and it has operated, like Lightelligence, principally out of Shanghai and Hangzhou. LightIgo’s commercial trajectory has been different from Lightelligence’s. Rather than competing head-on in the photonic matrix-multiplication space, LightIgo has concentrated on photonic inference accelerators for edge applications — small, low-power devices intended to perform matrix multiplication in cameras, in cars, and in industrial sensors. The technical rationale is that the workloads at the edge are more limited, more predictable, and more amenable to the fixed-weight matrices that photonic accelerators handle most efficiently. The company published in 2021 in Optics Express a peer-reviewed demonstration of an early photonic inference accelerator, and it has, on the public record, shipped small volumes of an edge inference device to a Chinese robotics customer. LightIgo has, on the basis of its disclosed funding, raised substantially less than Lightelligence, and its commercial impact has, so far, been smaller.
Beyond Lightelligence and LightIgo, the Chinese photonic computing startup base is, as of late 2026, thin. A handful of smaller players — some spun out of SIMIT, some from the Zhejiang program, some from the Wuhan National Laboratory for Optoelectronics — are working on adjacent parts of the stack: silicon photonic foundries, photonic-electronic co-packaging, and photonic memory. None of these smaller players has, to date, achieved a comparable scientific or commercial profile. The ecosystem has not, in other words, reached the depth of the Chinese GPU startup ecosystem, which by 2026 included not only Cambricon and Moore Threads but also Biren, T-Head, MetaX, and a dozen other players. The thinner Chinese photonic ecosystem is, by the assessment of most external observers, the principal reason that the Chinese photonic chip effort is, for now, more of a research and pilot program than a production industry.
The foundry dimension of the Chinese photonic chip effort deserves separate attention. The dominant Chinese silicon-photonics foundry is Compoundtek, a Singapore-headquartered foundry that operates a silicon-photonics process at a 200mm wafer line in Shanghai. Compoundtek is not, strictly speaking, a Chinese company — it is incorporated in Singapore and has a mixed Chinese and Singaporean shareholder base — but its principal production line is in Shanghai, and the bulk of its customers are Chinese. Compoundtek’s process supports the standard silicon-photonics device library: high-speed modulators based on carrier depletion in silicon, germanium-silicon photodetectors, low-loss waveguides, and grating couplers. It does not, at the time of writing, support the most advanced heterogeneous-integration steps — in particular, the co-packaging of indium-phosphide lasers with silicon-photonic dies — that the most advanced photonic accelerators require. The lack of a domestic Chinese foundry with that capability is, on the assessment of most independent observers, the principal gap in the Chinese photonic chip infrastructure as of late 2026, and it is the gap that the National Photonic Chip Pilot Line was, in part, intended to close.
The funding environment for Chinese photonic chip startups has tightened since 2023. The overall Chinese venture-capital market has been, since the 2021–2022 regulatory tightening under the Cyberspace Administration and the broader economic slowdown, substantially more cautious than it was in the 2018–2020 boom. The semiconductor sector has, by virtue of its strategic priority, remained an exception: the Big Fund’s second and third phases, the provincial integrated-circuit funds in Shanghai, Beijing, and Hubei, and a substantial amount of state-bank credit have continued to flow into Chinese semiconductor companies of all kinds. Photonic computing has, however, been a less obvious beneficiary of this funding than memory or logic. The reason, again, is the maturity gap: photonic accelerators are, as of 2026, further from production deployment than GPUs or NPUs, and the investment committees of the state funds have, on the public record, been more cautious about backing them.
The response of the Chinese AI industry to the Chinese photonic chip effort
The Chinese AI industry is, in the late 2020s, a substantial consumer of photonic computing technology in the research sense and a tentative consumer of it in the commercial sense. The principal consumers are the Chinese hyperscalers and AI-cloud providers — Alibaba Cloud, Tencent Cloud, Baidu, and Huawei Cloud — and the principal use cases are experimental inference workloads, training-system research, and what might be called “long-horizon bet” procurement. None of the Chinese hyperscalers has, as of late 2026, deployed photonic accelerators at scale in a production training or inference workload. Several of them have acquired small numbers of Lightelligence Hummingbird cards as research platforms, and several have partnerships with Lightelligence, LightIgo, or SIMIT that allow their internal research teams to investigate the technology on photonic machines of various generations.
The technical reason for the limited Chinese hyperscaler deployment is not, primarily, the hardware. It is the software stack. A photonic accelerator, like a GPU or a TPU, is only as useful as the software that runs on it, and the photonic-software stack has, in 2026, a much smaller community and a much smaller set of well-tested frameworks than the GPU-software stack. The Tsinghua hardware-software co-design program is one of the principal efforts to address this, and the Lightelligence software team has built out a moderately capable compiler and runtime for the Hummingbird card. The PyTorch and TensorFlow ecosystems remain, however, GPU-first, and a Chinese AI team that wants to deploy a production model on a photonic accelerator must, in practice, do a substantial amount of bespoke porting. That cost is acceptable for a research platform. It is not yet acceptable for a production deployment.
The economic reason for the limited deployment is the cost of integration. The integration of a photonic accelerator into a data-centre inference server involves a substantial amount of electrical, thermal, and packaging engineering. A standalone inference card like the Hummingbird can, in principle, be dropped into a 1U server slot, but the practical integration — the data synchronisation between the photonic chip and the surrounding CMOS logic, the cooling of the photonic chip, the calibration of the optical system, and the fault-tolerance for the inherent noise of optical computation — is a substantial engineering project in its own right. Chinese hyperscalers have, as of 2026, the engineering capacity to do this work, but they have had limited operational incentive to do it at scale, given the maturity of the GPU and NPU alternatives.
The training side of the Chinese AI industry is even more firmly tied to the GPU and NPU stack. The training of large transformer models is, in 2026, dominated by Nvidia hardware and by the CUDA software stack. The Chinese substitute — Huawei Ascend, with the CANN software stack — has made substantial inroads in the Chinese market, and Cambricon, Moore Threads, and a handful of other players have built out smaller training-focused businesses. None of these Chinese training platforms relies on photonic computing. The training workloads are too irregular, too memory-intensive, and too dependent on the nonlinear operations that photonic hardware does not handle well. The photonic accelerators of the late 2020s are, on most published analyses, well-suited to inference and to a narrow set of training-adjacent workloads (notably, the matrix multiplication phases of forward-pass and backward-pass computations in a transformer) but not to the full training loop.
The Chinese large-model developer community has, on the public record, been engaged with the photonic computing question, but mostly in the form of small research collaborations rather than commercial procurements. The DeepSeek, Qwen, GLM, and Hunyuan teams have published, in 2024 and 2025, small numbers of papers on photonic hardware-aware model design, and a handful of graduate students at Peking, Tsinghua, and Zhejiang have written dissertations on the mapping of transformer workloads onto photonic chips. None of this work has yet produced a major result that suggests a near-term reshaping of the Chinese large-model training stack around photonic hardware. What it has produced, however, is a generation of Chinese AI engineers and researchers who are familiar with the technology, who can map workloads onto it, and who are likely to be the talent pool from which the first production Chinese photonic computing teams will be drawn.
The Chinese AI-cloud customers are also engaged in a more defensive posture. The October 2022 United States export-control regime, and its October 2023 expansion, made it clear that the Chinese AI industry was operating in an environment in which access to the most advanced Nvidia hardware was conditional and reversible. The Chinese AI industry’s response has been to invest in domestic alternatives: Huawei Ascend, Cambricon, Moore Threads, and the Biren and MetaX GPUs have all received substantial Chinese orders. Photonic computing has, in this context, been treated as a longer-horizon option — a technology that may, in the late 2020s or the early 2030s, provide a third or fourth path alongside the GPU and NPU alternatives, and that is worth investing in research relationships for but not yet worth deploying at scale. The posture is consistent with the broader Chinese industrial policy posture of the late 2020s: invest across multiple technology generations, do not depend on any single technology path, and treat research and pilot procurement as a form of strategic insurance.
The international dimension of the Chinese photonic chip effort is also visible in the response of the Chinese AI industry. Chinese photonic-chip papers published between 2020 and 2025, in journals ranging from Nature, Nature Photonics, Optica, and Optics Express, have, on a careful count, been responsible for a substantial fraction of the global output on photonic accelerators. The Chinese research groups involved have collaborated with American and European groups including MIT, Stanford, the University of Oxford, the University of Ghent, and the VTT Technical Research Centre of Finland. The export-control regime has, on the public record, made these collaborations harder: Chinese researchers have reported, since 2023, increasing difficulty in obtaining visas for international conferences, in receiving shipments of specialised optical components from American and Japanese suppliers, and in co-authoring papers with American co-authors on subjects deemed sensitive by the United States Department of Commerce. The Chinese photonic chip community has, in response, deepened its relationships with Singapore, with Belgium, and with the United Kingdom, which have been less affected by the United States regime, and it has built out its own domestic supplier base for the specialised optical components that the technology requires.
The strategic implications of the Chinese photonic chip effort
The strategic implications of the Chinese photonic chip effort are larger than the technology itself, for two reasons. The first is that the technology sits at the intersection of two national strategic priorities of the highest order: the Chinese priority of sovereign AI compute capability, and the broader United States priority of constraining China’s access to the AI compute base that underwrites its AI industry. The second is that the technology is, by its nature, dual-use in a way that most other AI-accelerator technologies are not. A photonic chip that can perform matrix multiplication at high speed is a technology that can be used to accelerate any computation that depends on matrix multiplication, and matrix multiplication is the heart of a substantial fraction of the signal-processing and machine-learning workloads on which modern militaries depend. Radio-frequency beamforming, synthetic-aperture radar signal processing, encrypted-signal cryptanalysis, and the simulation of physical systems are all matrix-multiplication-heavy workloads, and the same photonic chip that accelerates a recommendation-system inference also accelerates a radar-signal-processing pipeline.
The first strategic question that the Chinese photonic chip effort raises is the substitution question: can photonic computing substitute for advanced-node CMOS computing at scale? The honest answer, in 2026, is that it cannot, on a five-year horizon, and it may never be able to on a complete-substitution basis. The reasons are physical as well as engineering-physical. Photonic computing is excellent at the dense linear-algebra operations that constitute the inner loops of modern AI: matrix multiplication, convolution, and certain types of Fourier and wavelet transforms. It is, however, poor at the nonlinearities, the memory accesses, and the branching control flow that surround those inner loops, and it is poor at the high-precision arithmetic that some scientific-computing workloads require. A complete-substitution photonic computer would, in effect, have to re-implement in optical hardware every function that a modern CMOS computer implements in electronic hardware: random-access memory, persistent storage, branching logic, exception handling, and the enormous software infrastructure that depends on those primitives. That is not an impossible engineering project, but it is a project whose time horizon is measured in decades rather than years. The Chinese photonic chip effort is, in this respect, an effort to be present in a long-term technology transition, not an effort to displace CMOS in the immediate term.
The second strategic question is the more immediate one: does the Chinese photonic chip effort change the balance of the AI compute race? Again, the honest answer is that it does not, on the timescale that matters for the 2025–2030 AI model-deployment cycle. The Chinese AI industry will, on the most plausible scenario, train and deploy its frontier models for the remainder of the decade on GPU-class hardware, increasingly sourced from Chinese suppliers (Huawei Ascend, Cambricon, Moore Threads, and the smaller players) but architecturally conventional. The photonic contribution to that pipeline will, in the most optimistic case, be a measurable improvement in the inference cost of certain workloads, and in the most pessimistic case it will be no more than a research line. The competitive balance of the AI race between the United States and China in the late 2020s will, on this analysis, be settled by the relative pace of advancement-node progress, by the relative availability of HBM, and by the relative quality of the domestic software stacks — not by whether Chinese photonic chips have achieved a 10x improvement in inference energy efficiency for a narrow benchmark.
The third strategic question is the foundry question. The Chinese photonic chip effort is, as discussed above, ultimately limited by the absence of a domestic Chinese foundry with full heterogeneous-integration capability — the ability to co-package indium-phosphide lasers with silicon-photonics dies in a way that is reliable, repeatable, and economically competitive. The Compoundtek Shanghai line is a working 200mm silicon-photonics foundry. It does not yet have the full heterogeneous-integration process that the most advanced photonic accelerators require. The National Photonic Chip Pilot Line is, in principle, the answer to this gap, and it is the principal positive development in the Chinese photonic chip infrastructure of the past three years. The pilot line is, however, a multi-year project, and the path from pilot line to production foundry is, on the basis of comparable precedents in CMOS and in compound-semiconductor photonics, at least five years and possibly ten. The Chinese photonic chip effort will, in the late 2020s, be working around the foundry gap with hybrid packaging in which photonic dies fabricated in Shanghai are bonded to electronic-driver dies fabricated in Taiwan or in Korea. That workaround is workable, but it is also a strategic vulnerability, because it places a critical piece of the photonic computing supply chain outside Chinese territorial control.
The fourth strategic question is the export-control question. The United States export-control regime, in its October 2022 and October 2023 forms, was structured primarily around the constraints on advanced-node CMOS and on advanced packaging. It did not, in its initial forms, target photonic computing directly. That began to change in 2024 and 2025. The Bureau of Industry and Security of the United States Department of Commerce, working with the Defense Advanced Research Projects Agency and the intelligence community, began, in 2024, to consider whether photonic computing equipment and certain classes of specialised photonic-device intellectual property should be added to the export-control list. The specific items under consideration included certain types of photonic-foundry equipment, certain classes of low-noise photonic detectors, and certain classes of photonic-electronic co-packaging machinery. No formal rulemaking had, at the time of writing, been announced, but the public-domain discussion within American policy circles suggested that the United States was, in 2025 and 2026, increasingly alert to the possibility that photonic computing could, in the late 2020s or the early 2030s, become a substantive alternative to advanced-node CMOS for AI workloads, and that the export-control regime ought to address that possibility before it materialised.
The fifth and final strategic question is the long-horizon one: what does the Chinese photonic chip effort look like in 2035? The honest answer is that the trajectory is genuinely uncertain. The technology is at a stage at which the commercial economics are not yet established, the foundry infrastructure is not yet mature, and the software stack is not yet self-sustaining. Those three things could, in principle, all happen within a ten-year window: the commercial economics of photonic inference could improve as the technology matures and as the energy-efficiency advantage of photonic computation compounds, the foundry infrastructure could mature as the National Photonic Chip Pilot Line graduates into a production foundry, and the software stack could mature as the Tsinghua hardware-software co-design program graduates into a self-sustaining compiler-and-runtime ecosystem. They could also all fail to happen within that window: the commercial economics could turn out to require precision manufacturing that the available foundry processes cannot deliver, the foundry infrastructure could turn out to require capital investment at a scale the National Photonic Chip Pilot Line cannot provide, and the software stack could turn out to require a long engineering investment that the available Chinese software talent is unwilling to make. Both trajectories are visible in the published plan; neither is, in 2026, more plausible than the other.
What is not in doubt, in 2026, is that the Chinese photonic chip effort has reached the point of technical credibility and strategic visibility at which it must be taken seriously by anyone who is thinking about the long-term balance of the AI compute race. It is not, as of 2026, a substitute for advanced-node CMOS, and it is not, as of 2026, a near-term solution to the export-control-imposed limits on Chinese access to advanced AI hardware. It is, however, a credible parallel track on which significant Chinese intellectual, institutional, and industrial capacity is committed, and the trajectory of which will, over the next decade, materially shape what is possible in the AI compute base on both sides of the Pacific. The long-term interpretation of the Chinese photonic chip effort is going to continue to evolve in the years and decades to come, in step with the global photonic chip industry itself.
References
- Photonic Computing — Grokipedia.
- Silicon Photonics — Grokipedia.
- Lightelligence — Grokipedia.
- Shanghai Institute of Microsystem and Information Technology (SIMIT) — Chinese Academy of Sciences, 2024.
- Zhejiang University Official Site — Zhejiang University, 2024.
- Tsinghua University Official Site — Tsinghua University, 2024.
- Compoundtek Foundry — Compoundtek, 2024.
- Institute of Semiconductors, Chinese Academy of Sciences — CAS, 2024.
- Lightelligence — Lightelligence corporate site, 2024.
- Hamerly et al., “Large-Scale Optical Neural Networks Based on Photoelectric Multiplication” — arXiv:2103.00020, 2021.
- Lightelligence Photonic AI Accelerator — Nature, 2021.
- LightIgo Photonic AI Inference Accelerator — Optics Express, 2021.
- China’s Quest for Self-Reliance in the Semiconductor Industry — Council on Foreign Relations, 2024.
- Semiconductors and Modern Warfare — Center for Strategic and International Studies, 2024.
- Lightelligence’s Photonic AI Chip — MIT Technology Review, 2022.
- Chinese Photonic Chip Startup Lightelligence — Reuters, 2022.
- Shen et al., “A Deep Neural Network Architecture Using Light” — Science, 2017.
- Hahnloser et al., “Digital Selection and Analogue Amplification Coexist in a Cortex-Inspired Silicon Circuit” — Nature, 1997.
- Photonic Chip Research at MIT — Nature, 2022.
- Lightelligence Series B Funding — BBC, 2021.
