Lightmatter
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,191 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 20, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,191 words
Add missing citations, update stale details, or suggest a clearer explanation.
Lightmatter is a U.S. silicon-photonics company that designs and manufactures optical computing hardware and photonic interconnects for artificial intelligence and high-performance computing.[^1][^2] Founded in 2017 as a spinout from the Massachusetts Institute of Technology, the company is headquartered in Boston (Mountain View, California secondary office) and develops three principal product lines: Envise, a photonic inference accelerator that performs matrix multiplications optically using arrays of Mach-Zehnder interferometers; Passage, a 3D photonic interposer that provides chip-to-chip optical I/O; and Idiom, a compiler and runtime stack that targets photonic hardware from PyTorch and TensorFlow models.[^3][^4][^5] In October 2024 the company closed a $400 million Series D round led by T. Rowe Price Associates that valued it at $4.4 billion and raised cumulative funding past the $850 million mark.[^6][^7] In April 2025, Lightmatter and academic collaborators published a paper in Nature describing a multi-chip photonic processor that executed ResNet, BERT and DeepMind-style Atari reinforcement-learning models at accuracies comparable to electronic baselines.[^8][^9]
| Field | Value |
|---|---|
| Type | Private company |
| Industry | Silicon photonics, AI accelerators |
| Founded | September 2017 |
| Founders | Nicholas Harris, Darius Bunandar, Thomas Graham |
| Headquarters | Boston, Massachusetts |
| Secondary office | Mountain View, California |
| Products | Envise, Passage (L20, L200, M1000), Idiom |
| Funding raised | approximately $850 million (through Series D, October 2024) |
| Last reported valuation | $4.4 billion (October 2024) |
| Lead Series D investor | T. Rowe Price Associates |
| Manufacturing partner | GlobalFoundries (Fotonix platform) |
| CEO | Nicholas Harris |
Lightmatter's technical foundation traces to Nicholas Harris's doctoral research at MIT, where he worked in the quantum photonics group led by Dirk Englund in the Department of Electrical Engineering and Computer Science.[^3] Harris had previously been a research and development engineer at Micron Technology, where he worked on DRAM and NAND circuits, and that industry exposure shaped his view that gains from transistor scaling were reaching a plateau.[^10] At MIT he initially pursued photonic quantum information processing, and his thesis "Programmable Nanophotonics for Quantum Information Processing and Artificial Intelligence" treated arrays of programmable interferometers as both a quantum-computing substrate and a substrate for analog linear algebra.[^10] As the modern wave of deep learning expanded in the mid-2010s, Harris and his collaborators observed that the same meshes of integrated optical components used in photonic quantum experiments could be applied directly to the matrix multiplications that dominate neural network inference.[^3]
The company was founded in September 2017 by Harris together with Darius Bunandar, who earned his PhD in physics from MIT in 2018 and had worked in Englund's group, and Thomas Graham, then an MBA student at MIT Sloan with prior finance and operations experience at Morgan Stanley and Google.[^10] The team's commercial trajectory began with the 2017 MIT $100K Entrepreneurship Competition, where their photonic computing pitch was awarded a top prize and Harris and Bunandar received initial seed grants from Harvard Innovation Labs and the MIT delta-v accelerator.[^7][^10] Harris served as chief executive officer, Bunandar as chief scientist, and Graham initially as chief operating officer before later transitioning into a chief financial officer role.[^10]
Lightmatter's first institutional capital arrived in early 2018 in the form of an $11 million Series A tranche co-led by Matrix Partners and Spark Capital.[^7] In early 2019 GV (formerly Google Ventures) led a $22 million Series A extension that brought GV onto the cap table alongside Matrix and Spark.[^7][^11] During this period the company concentrated on internal prototypes of programmable photonic meshes and on adapting Mach-Zehnder interferometer arrays, which were originally explored for quantum experiments, for use as the linear-algebra engines of an AI inference accelerator.[^12]
In March 2021 Lightmatter publicly disclosed a three-part product strategy comprising Envise, Idiom and Passage. The original announcement described Envise as "the world's first general-purpose photonic artificial intelligence accelerator," packaged as a 4U server blade containing sixteen Envise chips, 1 TB of DDR4 DRAM, 3 TB of solid-state storage and 6.4 Tbps of optical interconnect bandwidth for multi-blade scale-out.[^4] Idiom was introduced as a compiler and runtime that ingests PyTorch and TensorFlow models and partitions them across multiple Envise blades, with companion debugging and profiling tooling.[^4][^5] Passage was first described as an 8 inch by 8 inch wafer-scale programmable photonic interconnect integrating transistors with lasers, modulators and photodetectors and providing roughly 100 Tbps of chip-to-chip optical fabric.[^4]
In May 2021 Lightmatter announced an $80 million Series B led by Viking Global Investors, with participation from existing investors GV, Matrix Partners and Spark Capital and new strategic investors Hewlett Packard Enterprise and Lockheed Martin Ventures.[^11][^13] The round expanded the engineering organization and was used to harden the Envise hardware and to advance the Passage interconnect.[^13] Lockheed Martin's participation, in particular, reflected the U.S. defense industry's interest in photonic computing for high-performance and edge applications where energy-per-operation matters.[^11]
In May 2023, Lightmatter closed a $154 million Series C round co-led by SIP Global Partners and Aliya Capital Partners, with participation from Fidelity Management & Research Company, Viking Global Investors, GV and the Hewlett Packard Pathfinder venture program.[^14][^15] The press coverage at the time placed the round at the inflection point between research-grade prototypes and a productization push, with the company stating that pilots were underway and that volume manufacturing was being targeted for 2024.[^15] An extension of the Series C in late 2023 lifted total cumulative capital above the $300 million threshold and brought the company's valuation to roughly $1.2 billion, granting it unicorn status.[^16]
On October 16, 2024 Lightmatter disclosed a $400 million Series D round led by new investor T. Rowe Price Associates, with participation from existing backers including Fidelity Management & Research Company, GV, Viking Global Investors and the Hewlett Packard Pathfinder program.[^6][^17] The post-money valuation was reported at $4.4 billion, approximately four times the level set by the December 2023 Series C extension, and cumulative funding crossed roughly $850 million.[^6][^7][^17] In a Reuters interview accompanying the announcement, Harris said "this is probably our last private funding round," signaling an expected path toward an eventual initial public offering.[^17] Bloomberg framed the round as evidence that hyperscale data center capital expenditure on AI infrastructure was creating a viable opening for photonic-interconnect startups, despite the fact that no photonic accelerator had yet displaced electronic graphics processing unit platforms in commercial training workloads.[^18]
In late 2024 Lightmatter announced an expanded multi-generation manufacturing partnership with GlobalFoundries, building on a seven-year-long collaboration that spanned several silicon process generations.[^19] Production of the Passage platform was assigned to GlobalFoundries' Fotonix process, which integrates electronic transistors and silicon-photonic components on a single CMOS wafer.[^19][^20] Lightmatter also disclosed advanced packaging work with Amkor Technology to support a large 3D photonic interposer.[^10]
In December 2024 the company unveiled the Passage M1000, marketed as the world's fastest photonic AI interconnect.[^21] The M1000 was described as a 7,735 mm² multi-reticle active photonic interposer hosting up to 4,000 mm² of stacked die complex consisting of as many as 34 chiplets, 1,024 serial data channels at 56 Gbps each, 256 external fiber-optic lines (each carrying eight wavelengths) and up to 114 Tbps of aggregate off-package optical bandwidth.[^22][^21] In 2025 the M1000 reference platform was presented in technical detail at the Hot Chips conference, and the L200 and L20 co-packaged optics products were positioned as commercial follow-ons.[^23][^21]
In April 2025, Nature published a peer-reviewed paper authored by Lightmatter researchers and collaborators describing a multi-chip photonic processor that ran modern AI workloads end-to-end.[^8][^9] The reported system vertically combined six chips in a single package: four 128 by 128 photonic tensor cores together with two 12 nm digital control interfaces.[^9] The processor reported 65.5 trillion adaptive block floating-point 16-bit operations per second on 78 W of electrical power and 1.6 W of optical power, and executed ResNet, BERT and a DeepMind-style Atari reinforcement-learning model at accuracies the authors described as comparable to standard 32-bit floating-point electronic baselines without quantization-aware retraining.[^9] Physics World and other outlets characterized the result as the first photonic processor to run state-of-the-art neural networks unmodified at competitive accuracies.[^9]
Lightmatter's core technical thesis is that the matrix multiplications dominating modern neural network workloads can be performed more efficiently in the optical domain than in CMOS digital arithmetic.[^12][^24] In an electronic AI accelerator, each multiply-accumulate operation consumes energy in transistor switching, in moving data over short metal interconnects and in driving register files; in a coherent photonic processor, the same multiplication can be implemented by setting the transmission coefficients of optical components and letting light propagate through the mesh in essentially constant time, with energy cost dominated by laser sourcing and electro-optic configuration rather than per-operation switching.[^24][^12]
The basic building block of Lightmatter's photonic tensor core is the Mach-Zehnder interferometer, an integrated silicon-photonic device in which an input optical waveguide is split into two arms, each arm experiences a programmable phase shift, and the two arms are recombined.[^12][^24] By cascading large two-dimensional grids of these interferometers, Lightmatter's chips realize unitary linear transformations on a vector of optical amplitudes; combining two such unitary meshes with intermediate amplitude attenuators implements arbitrary real-valued matrix-vector products in the optical domain, a construction based on singular value decomposition.[^12] At the output, photodetectors convert the optical result back into the electronic domain for accumulation, nonlinearity application and digital control.[^12]
The Envise accelerator integrates large numbers of these nanometer-scale interferometers on a silicon-photonics die. Different optical wavelengths can be multiplexed through the same physical mesh to perform multiple matrix multiplications in parallel, a wavelength-division-multiplexing approach that Lightmatter has repeatedly cited as a key reason that photonic throughput can scale beyond what a comparably sized electronic accelerator can deliver.[^3][^15]
The first-generation Envise chip integrated a photonic tensor core with a substantial on-chip electronic substrate. Disclosed specifications include 500 MB of on-chip memory for activations and weights, 256 RISC cores per processor for handling offload control, support for 8-bit and 16-bit integer plus bfloat16 precision and dynamic scaling for precision adjustment.[^12] Each Envise chip provides 400 Gbps of Lightmatter photonic interconnect bandwidth, and two-chip Envise cards reach 6.4 Tbps of aggregate optical interconnect.[^12] A 4U blade carries sixteen Envise chips, 3 TB of NVMe SSD storage and two AMD Epyc 7002 host processors at roughly 3 kW of power, which Lightmatter contrasted at announcement time against an equivalent eight-way NVIDIA A100 reference configuration drawing roughly 6.5 kW.[^12] Harris claimed at the time that each Envise chip was faster than a single A100 on a broad set of inference benchmarks, though independent GPU benchmarking via MLPerf was not yet available.[^12]
Passage is the company's solution to the inter-package bandwidth wall: as die sizes and chiplet counts grow, the rate at which signals can escape the package edge becomes the binding throughput constraint.[^22][^25] Conventional 2.5D interposers route electrical wires laterally across a passive silicon substrate, and the bandwidth per millimeter of beachfront is fundamentally limited by transmission-line losses, SerDes power and pitch.[^22] Passage replaces or augments that lateral electrical fabric with a 3D photonic interposer: an active silicon-photonics die on which stacked compute and memory chiplets sit, with electro-optical I/O placed anywhere across the surface rather than only at the edges.[^22][^25]
The Passage M1000 reference platform exemplifies this design. Eight active photonic tiles tile a 7,735 mm² package, accommodating a die complex of up to 4,000 mm². Stacked chiplets connect downward to the photonic interposer via short UCIe electrical links, which the interposer then converts to optical signals carried by 256 external optical fibers (1,024 serial data channels at 56 Gbps each, plus eight-wavelength wavelength-division multiplexing for an aggregate of 114 Tbps).[^21][^22] The interposer delivers up to 1.5 kW of power to the stacked compute die complex and supports built-in solid-state optical circuit switching, allowing on-package optical paths between chiplets to be reconfigured under software control.[^21][^22]
Two co-packaged-optics products complement the M-series. Passage L200 is described as offering 32 to 64 Tbps of aggregate bandwidth using 112G PAM4 signaling and is intended for frontier-scale AI training fabrics.[^26] Passage L20 provides 12.8 Tbps of aggregate bandwidth with four times the pluggable density of incumbent transceivers and is intended for higher-density deployments where short-reach optical links are required.[^26] Both products use chiplets supplied by external partners, including Alphawave Semi's UCIe interface, stacked on top of Lightmatter's photonic circuitry using standard chip-on-wafer packaging.[^25]
Idiom is the compiler and runtime that links existing deep-learning frameworks to Envise hardware. The stack ingests PyTorch and TensorFlow model graphs, applies optimization passes specific to photonic tensor cores (including precision lowering and operator fusion), and emits an execution plan that can target single Envise chips or large clusters of blades.[^5][^15] Lightmatter identifies three subcomponents publicly: idCompile, the graph compiler; idProfiler, a runtime profiler; and idBug, a debugger.[^10][^15] Idiom can automatically detect Envise topology and partition large models across blades, an important property given the multi-chip, multi-blade scale at which photonic systems are intended to operate.[^5][^15]
The following table summarizes Lightmatter's disclosed primary funding rounds. Cumulative capital is reported by multiple sources as approximately $850 million following the October 2024 Series D.[^7][^17][^6]
| Round | Date | Amount | Lead investor(s) | Notes |
|---|---|---|---|---|
| MIT $100K + grants | 2017 | ~$0.1M | MIT $100K, Harvard Innovation Labs | Founding seed capital[^7] |
| Seed | Early 2018 | ~$11M | Matrix Partners, Spark Capital | First institutional round[^7] |
| Series A extension | Early 2019 | ~$22M | GV (Google Ventures) | Matrix, Spark also participating[^11] |
| Series B | May 2021 | $80M | Viking Global Investors | HPE, Lockheed Martin Ventures, GV, Matrix, Spark[^11][^13] |
| Series C | May 2023 | $154M | SIP Global Partners, Aliya Capital | Fidelity, Viking, GV, HPE Pathfinder[^14][^15] |
| Series C extension | December 2023 | ~$155M | Existing investors | Lifted valuation to ~$1.2B (unicorn status)[^16] |
| Series D | October 2024 | $400M | T. Rowe Price Associates | Fidelity, GV, Viking, HPE Pathfinder; valuation $4.4B[^6][^17] |
| Product | Function | Disclosed key specifications | First disclosed |
|---|---|---|---|
| Envise (blade) | Photonic AI inference accelerator | 16 chips per 4U blade, 6.4 Tbps optical interconnect, 1 TB DDR4, 3 TB SSD, ~3 kW[^4][^12] | March 2021 |
| Envise (chip) | Single accelerator | 500 MB on-chip memory, 256 RISC cores, Int-8/Int-16/bfloat16, 400 Gbps interconnect[^12] | 2021 |
| Idiom | Compiler and runtime | PyTorch/TensorFlow ingestion, idCompile, idProfiler, idBug[^5][^10] | March 2021 |
| Passage (wafer-scale) | Original photonic interconnect | 8 in by 8 in wafer-scale chip, ~100 Tbps chip-to-chip, integrated lasers and photodetectors[^4] | March 2021 |
| Passage L20 | Pluggable co-packaged optics | 12.8 Tbps aggregate, 4x pluggable density[^26] | 2024 |
| Passage L200 | Frontier-scale co-packaged optics | 32-64 Tbps aggregate, 112G PAM4[^26] | 2024 |
| Passage M1000 | 3D photonic interposer | 7,735 mm², up to 4,000 mm² die complex, 1,024 channels at 56 Gbps, 114 Tbps total, 256 fibers, ~1.5 kW power delivery[^21][^22] | December 2024 |
Lightmatter's products are positioned at two intersecting bottlenecks in modern AI infrastructure: the per-operation energy cost of matrix multiplication in GPU inference and the bandwidth wall imposed by electrical chip-to-chip I/O.[^18][^22] The Envise accelerator is targeted at neural-network inference workloads, including transformer and convolutional models, where the company has demonstrated end-to-end execution of BERT, ResNet and reinforcement-learning benchmarks in its Nature paper.[^9] The Passage product line is aimed at the much larger problem of optically interconnecting thousands to millions of accelerator dies within hyperscale data center AI training clusters.[^6][^22]
The broader significance of Lightmatter, as repeatedly framed by Harris in press interviews and on the company's own materials, is the contention that continued scaling of large neural network training and inference will require co-design of compute and interconnect substrates rather than improvements to electronics alone, and that silicon photonics is the most production-ready substrate to fill that role.[^17][^18][^3] Passage in particular is described by Lightmatter as the world's first 3D-stacked photonics engine capable of connecting many processors at the speed of light at extreme scale.[^6]
Lightmatter and external analysts have publicly acknowledged several persistent challenges in photonic AI hardware:
Lightmatter operates in two adjacent markets: photonic AI compute and photonic chip-to-chip interconnect. The following table summarizes principal competitors as identified in industry surveys.[^27][^25]
| Company | Focus | Status (as of 2025-2026) |
|---|---|---|
| Lightmatter | Photonic compute (Envise), photonic interconnect (Passage), software (Idiom) | $4.4B valuation; Passage on GlobalFoundries; M1000 shipping reference platform[^6][^21] |
| Ayar Labs | Photonic I/O chiplets (TeraPHY) with UCIe interface; SuperNova multi-wavelength laser | Productized chiplets; backed by NVIDIA, AMD, Intel[^25] |
| Celestial AI | Photonic Fabric decoupling memory from compute | Reported acquired by Marvell in December 2025[^27] |
| Luminous Computing | Full photonic AI supercomputer architecture | Reportedly reorganized in 2023 with a pivot toward optical networking[^27] |
| Optalysys | Free-space and integrated photonics for fully-homomorphic-encryption and AI primitives | Earlier-stage commercial deployments[^27] |
| Avicena | MicroLED-based optical interconnects (LightBundle) | Targets dense GPU-to-switch fabrics[^25] |
| Xscape Photonics | On-chip frequency-comb laser sources for wavelength-division multiplexed interconnects | Earlier-stage[^25] |
Electronic AI-accelerator vendors are also implicit competitors at the system level. NVIDIA's position in AI training and inference is well established, and incumbents Cerebras Systems (wafer-scale electronic engines), Groq (deterministic LPU), SambaNova Systems, Tenstorrent and Etched Sohu all pursue alternative-architecture strategies, though none use silicon photonics as their primary compute substrate.[^27] At the interconnect layer Lightmatter must also compete against the trajectory of NVIDIA's own optical-interconnect roadmap and against established switch silicon vendors such as Broadcom and Marvell, both of which have been integrating co-packaged optics offerings of their own.[^27]
The qualitative case for photonic AI hardware versus mainstream electronic accelerators can be summarized along several axes. The table below collects publicly reported claims rather than independently audited benchmarks.
| Axis | Envise (photonic) | NVIDIA A100 (electronic, reference) |
|---|---|---|
| Substrate | Silicon photonics with Mach-Zehnder mesh | 7 nm CMOS with Tensor Cores |
| Energy/operation (claimed) | Lower for fixed matrix multiplication; ~7x efficiency claim for specific inference workloads[^12] | Reference baseline |
| Precision | Int-8, Int-16, bfloat16, with dynamic scaling[^12] | FP16, BF16, TF32, INT8 (Tensor Cores) |
| On-chip memory | 500 MB per Envise chip[^12] | 40 GB HBM2 (A100 40GB) |
| Optical interconnect | 400 Gbps per chip; 6.4 Tbps per 2-chip card[^12] | NVLink + electrical SerDes |
| Independent benchmarks | Limited public MLPerf data[^12] | Extensive MLPerf history |
Lightmatter's Nature paper has been read by external commentators as the first peer-reviewed evidence that a photonic processor can match electronic accuracy on production-scale neural networks without bespoke retraining, a claim the company itself has subsequently echoed in marketing materials.[^9][^8]