# Lightmatter

> Source: https://aiwiki.ai/wiki/lightmatter
> Updated: 2026-06-07
> Categories: AI Companies, AI Hardware
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

# Lightmatter

**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](/wiki/mit), 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]

## Infobox

| 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 |

## History

### MIT origins and founding (2014-2017)

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](/wiki/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](/wiki/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]

### Early funding and product development (2018-2020)

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]

### Scaling and Series B (2021)

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]

### Series C and unicorn status (2023)

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]

### Series D and post-money valuation of $4.4 billion (October 2024)

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](/wiki/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](/wiki/gpu) platforms in commercial training workloads.[^18]

### Manufacturing partnerships and Passage M1000 launch (2024-2025)

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]

### Nature paper (April 2025)

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]

## Technical details

### Photonic computing thesis

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](/wiki/ai_chip), 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]

### Mach-Zehnder interferometer mesh

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]

### Envise accelerator architecture

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](/wiki/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](/wiki/gpu) benchmarking via MLPerf was not yet available.[^12]

### Passage 3D photonic interposer

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 software stack

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]

## Funding history

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] |

## Variants and products

| 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 |

## Partnerships

- **GlobalFoundries**: A roughly seven-year manufacturing collaboration was expanded in November 2024 to mass-produce Passage on the foundry's Fotonix silicon-photonics platform, which integrates photonic and electronic components on the same CMOS wafer.[^19][^20]
- **Amkor Technology**: Advanced packaging partner for what the company described as the world's largest 3D photonic package, announced in late 2024.[^10]
- **Hewlett Packard Enterprise**: Strategic investor through both the Series B and the Hewlett Packard Pathfinder venture program in subsequent rounds.[^11][^15] HPE's interest is consistent with the data-center server market for which Passage is being positioned.
- **Lockheed Martin Ventures**: Strategic investor since the 2021 Series B; the relationship has been described by Lightmatter and by press coverage as reflecting U.S. defense interest in photonic computing for high-performance and energy-efficient applications.[^11][^13]
- **Alphawave Semi**: Chiplet supplier for the UCIe electrical interfaces stacked above the photonic interposer in Passage L200.[^25]

## Significance and applications

Lightmatter's products are positioned at two intersecting bottlenecks in modern AI infrastructure: the per-operation energy cost of matrix multiplication in [GPU](/wiki/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](/wiki/transformer) and [convolutional](/wiki/convolutional_neural_network) models, where the company has demonstrated end-to-end execution of [BERT](/wiki/bert), [ResNet](/wiki/resnet) and [reinforcement-learning](/wiki/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](/wiki/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]

## Limitations and challenges

Lightmatter and external analysts have publicly acknowledged several persistent challenges in photonic AI hardware:

- **Thermal sensitivity**: Mach-Zehnder interferometers rely on precise phase relationships between the two arms of each device. Local temperature variations on a silicon-photonics chip shift the refractive index of the waveguides and therefore the relative phases, which in turn changes the computed matrix coefficients.[^24] Industrial photonic processors must therefore include on-chip thermal management and ongoing phase calibration to maintain numerical accuracy.[^24][^9]
- **Calibration overhead**: Programming a large mesh of phase shifters to implement a target matrix requires either calibrating each element individually or running adaptive control loops. The IEEE Spectrum survey of optical-interconnect startups noted that Lightmatter's most aggressive interposer designs had been disclosed but, at the time of publication, not yet fully demonstrated in production hardware.[^25]
- **Process integration**: Silicon photonics requires fabrication steps (waveguide etching, modulator implants, photodetector integration) that are not standard parts of leading-edge logic CMOS processes. Lightmatter's reliance on GlobalFoundries' Fotonix platform is a strategic bet that monolithic electro-photonic integration on mature nodes is a more productive path than co-packaging discrete photonics with bleeding-edge logic dies.[^19][^20]
- **Software ecosystem**: Photonic accelerators must compete against the deeply entrenched [CUDA](/wiki/nvidia_cuda) software stack and the dominant [PyTorch](/wiki/pytorch) runtime path that targets [NVIDIA](/wiki/nvidia) hardware. Idiom is designed to abstract that gap by ingesting [PyTorch](/wiki/pytorch) and [TensorFlow](/wiki/tensorflow) graphs natively, but the absence of a published independent benchmark of Envise on standardized inference suites such as MLPerf has been a recurring observation in industry coverage.[^12][^5]
- **Commercialization risk**: A wider set of photonic AI startups, including some with substantial venture funding, has faced reorganization or strategic pivots. Luminous Computing, which raised significant capital in 2022, reportedly hit technical hurdles and pivoted away from pure photonic compute in 2023.[^27] Lightmatter's strategy of building a separate, near-term-shippable interconnect product (Passage) in parallel with its longer-horizon compute product (Envise) is partly a response to that commercialization risk.[^17]

## Competitive landscape

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](/wiki/nvidia)'s position in AI training and inference is well established, and incumbents [Cerebras Systems](/wiki/cerebras) (wafer-scale electronic engines), [Groq](/wiki/groq_hardware) (deterministic LPU), [SambaNova Systems](/wiki/sambanova), [Tenstorrent](/wiki/tenstorrent) and [Etched Sohu](/wiki/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]

## Comparison with electronic AI accelerators

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]

## See also

- [Massachusetts Institute of Technology](/wiki/mit)
- [AI accelerator](/wiki/ai_chip)
- [Graphics processing unit](/wiki/gpu)
- [NVIDIA](/wiki/nvidia)
- [NVIDIA A100](/wiki/nvidia_a100)
- [Cerebras Systems](/wiki/cerebras)
- [Groq](/wiki/groq_hardware)
- [SambaNova Systems](/wiki/sambanova)
- [Tenstorrent](/wiki/tenstorrent)
- [Etched Sohu](/wiki/etched_sohu)
- [Data Center](/wiki/data_center)
- [PyTorch](/wiki/pytorch)
- [TensorFlow](/wiki/tensorflow)
- [BERT](/wiki/bert)
- [ResNet](/wiki/resnet)
- [Transformer](/wiki/transformer)
- [DeepMind](/wiki/deepmind)
- [Deep Learning](/wiki/deep_learning)
- [Neural Network](/wiki/neural_network)
- [Reinforcement learning](/wiki/reinforcement_learning)
- [Convolutional Neural Network](/wiki/convolutional_neural_network)
- [CUDA](/wiki/nvidia_cuda)
- [Inference](/wiki/inference)

## References

[^1]: Lightmatter, "The photonic (super)computer company", lightmatter.co, 2025. https://lightmatter.co/. Accessed 2026-05-20.
[^2]: Lightmatter, "Vision", lightmatter.co, 2024. https://lightmatter.co/vision/. Accessed 2026-05-20.
[^3]: Zach Winn, "Startup accelerates progress toward light-speed computing", MIT News, 2024-03-01. https://news.mit.edu/2024/startup-lightmatter-accelerates-progress-toward-light-speed-computing-0301. Accessed 2026-05-20.
[^4]: Nicholas Harris, "Introducing Envise, Idiom and Passage: Next Generation AI Compute, Compile and Interconnect Platforms", Lightmatter (Medium), 2021-03-10. https://medium.com/lightmatter/introducing-envise-idiom-and-passage-next-generation-ai-compute-compile-and-interconnect-331878d6cea5. Accessed 2026-05-20.
[^5]: Lightmatter, "Idiom: Machine Learning and Neural Network Software", lightmatter.co, 2024. https://lightmatter.co/products/idiom/. Accessed 2026-05-20.
[^6]: Lightmatter, "Lightmatter Raises $400M Series D; Quadruples Valuation to $4.4B as Photonics Leader for Next-Gen AI Data Centers", Business Wire, 2024-10-16. https://www.businesswire.com/news/home/20241016498931/en/Lightmatter-Raises-$400M-Series-D-Quadruples-Valuation-to-$4.4B-as-Photonics-Leader-for-Next-Gen-AI-Data-Centers. Accessed 2026-05-20.
[^7]: Tracxn, "Lightmatter: Funding Rounds and List of Investors", Tracxn, 2024. https://tracxn.com/d/companies/lightmatter/__UzjACTIII619AffSL1FYPgASKG7oNo0LdOcRvQnFRl8/funding-and-investors. Accessed 2026-05-20.
[^8]: Lightmatter, "Lightmatter's Latest Research Published in Nature", lightmatter.co, 2025-04-30. https://lightmatter.co/blog/a-new-kind-of-computer/. Accessed 2026-05-20.
[^9]: Hamish Johnston, "Photonic computer chips perform as well as purely electronic counterparts, say researchers", Physics World, 2025-05-02. https://physicsworld.com/a/photonic-computer-chips-perform-as-well-as-purely-electronic-counterparts-say-researchers/. Accessed 2026-05-20.
[^10]: Contrary Research, "Lightmatter Business Breakdown and Founding Story", research.contrary.com, 2025. https://research.contrary.com/company/lightmatter. Accessed 2026-05-20.
[^11]: Edd Gent, "Corporates illuminate Lightmatter in $80m Series B", Global Venturing, 2021-05-10. https://globalventuring.com/blog/2021/05/10/corporates-illuminate-lightmatter/. Accessed 2026-05-20.
[^12]: Timothy Prickett Morgan, "Lightmatter Normalizing Silicon Photonics for AI", The Next Platform, 2021-03-17. https://www.nextplatform.com/2021/03/17/lightmatter-normalizing-silicon-photonics-for-ai/. Accessed 2026-05-20.
[^13]: FinSMEs, "Lightmatter Raises $80M in Series B Funding", FinSMEs, 2021-05-10. https://www.finsmes.com/2021/05/lightmatter-raises-80m-in-series-b-funding.html. Accessed 2026-05-20.
[^14]: Lightwave Editorial Staff, "Photonic processor startup Lightmatter lands $154 million in Series C funding", Lightwave, 2023-05-31. https://www.lightwaveonline.com/business/companies/article/14294632/photonic-processor-startup-lightmatter-lands-154-million-in-series-c-funding. Accessed 2026-05-20.
[^15]: Devin Coldewey, "Lightmatter's photonic AI hardware is ready to shine with $154M in new funding", TechCrunch, 2023-05-31. https://techcrunch.com/2023/05/31/lightmatters-photonic-ai-hardware-is-ready-to-shine-with-154m-in-new-funding/. Accessed 2026-05-20.
[^16]: Sacra, "Lightmatter valuation, funding and news", Sacra, 2024. https://sacra.com/c/lightmatter/. Accessed 2026-05-20.
[^17]: Chris Williams (citing Reuters), "Lightmatter: Photonic Supercomputing Company Raises $400 Million (Series D) At $4.4 Billion Valuation", Pulse 2.0, 2024-10-16. https://pulse2.com/lightmatter-photonic-supercomputing-company-raises-400-million-series-d-at-4-4-billion-valuation/. Accessed 2026-05-20.
[^18]: Bloomberg News, "Photonic Computing Startup Lightmatter Reaches $4.4 Billion Valuation", Bloomberg, 2024-10-16. https://www.bloomberg.com/news/articles/2024-10-16/photonic-computing-startup-lightmatter-reaches-4-4-billion-valuation. Accessed 2026-05-20.
[^19]: Sebastian Moss, "Lightmatter turns to GlobalFoundries to bring photonics interconnect to market in 2025", Data Center Dynamics, 2024-11-19. https://www.datacenterdynamics.com/en/news/lightmatter-turns-to-globalfoundries-to-bring-photonics-interconnect-to-market-in-2025/. Accessed 2026-05-20.
[^20]: Lightmatter, "Lightmatter and GlobalFoundries Partner to Mass Produce Passage Platform", lightmatter.co, 2024-11-19. https://lightmatter.co/blog/lightmatter-and-globalfoundries-partner-to-mass-produce-passage-platform/. Accessed 2026-05-20.
[^21]: Lightmatter, "Lightmatter Unveils Passage M1000 Photonic Superchip, World's Fastest AI Interconnect", lightmatter.co, 2024-12-09. https://lightmatter.co/press-release/lightmatter-unveils-passage-m1000-photonic-superchip-worlds-fastest-ai-interconnect/. Accessed 2026-05-20.
[^22]: Anton Shilov, "Lightmatter unveils high-performance photonic 'superchip', claims world's fastest AI interconnect", Tom's Hardware, 2024-12-09. https://www.tomshardware.com/tech-industry/lightmatter-unveils-high-performance-photonic-superchip-claims-worlds-fastest-ai-interconnect. Accessed 2026-05-20.
[^23]: Patrick Kennedy, "Lightmatter Passage M1000 at Hot Chips 2025", ServeTheHome, 2025-08-26. https://www.servethehome.com/lightmatter-passage-m1000-at-hot-chips-2025/. Accessed 2026-05-20.
[^24]: Daniel Brunner et al., "Photonic Computing Takes a Step Toward Fruition", Physics (APS), 2025-05-19. https://link.aps.org/doi/10.1103/Physics.18.84. Accessed 2026-05-20.
[^25]: Samuel K. Moore, "Start-ups Replace Copper with Optical Links for GPUs", IEEE Spectrum, 2025. https://spectrum.ieee.org/optics-gpu. Accessed 2026-05-20.
[^26]: Lightmatter, "Passage: 3D Photonics for AI Applications", lightmatter.co, 2024. https://lightmatter.co/products/passage/. Accessed 2026-05-20.
[^27]: ts2.tech, "Photonic AI Accelerators vs. GPUs: The Battle for AI's Future in Efficiency, Cost, and Scale", ts2.tech, 2025. https://ts2.tech/en/photonic-ai-accelerators-vs-gpus-the-battle-for-ais-future-in-efficiency-cost-and-scale/. Accessed 2026-05-20.