Lightmatter: Ending AI's Energy Crisis with Photonic Computing

Whether it's ChatGPT, Claude, or DeepSeek, every conversation carries a heavy electricity bill behind it. SVTR's AI database shows global AI data center energy demand is set to double by 2026, reaching the level of an entire country's power consumption. Liquid cooling and other emerging technologies are cutting cooling-related power draw, but a fundamental break past the efficiency ceiling requires rethinking the computing architecture itself.

Lightmatter, a pioneer in photonic computing, is challenging the limits of traditional electronic chips. Its Passage photonic interconnect technology aims to transform data center scaling while moving data at the speed of light, cutting the energy losses that come with interconnects. If Lightmatter's vision plays out, it would mean not just a leap in compute, but a rebuilding of the energy foundation of AI computing itself.

As AI compute demand grows exponentially, is photonic computing hype or the future? Could it be the "Moore's Law 2.0" that resolves AI's energy crisis?

Global data center energy demand is projected to double by 2026, reaching a level equivalent to Japan's total annual electricity consumption. As of October 2024, data centers dedicated to AI workloads accounted for 70% of new energy demand. The growth is driven mainly by the high energy cost of training foundation AI models, plus the continuous power draw from inference on AI services like OpenAI's ChatGPT and Anthropic's Claude.

Take ChatGPT: as of August 2024, it had 200 million weekly active users and consumed roughly 564,000 kilowatt-hours of electricity per day. Six months later, its user base had doubled to 400 million, suggesting its energy consumption likely doubled as well, reflecting the sharp rise in AI energy demand. Even DeepSeek, known for its more efficient R1 model, likely consumes as much energy to run as other AI platforms, if not more.

Beyond compute itself, data center cooling is another major driver of rising overall energy demand. Traditional cooling systems struggle to manage the enormous heat generated by AI processors and require continuous power to prevent overheating. As a result, data centers are testing more efficient cooling technologies such as liquid cooling. By August 2024, 22% of data centers had adopted liquid cooling, a share expected to rise to 50% by 2029.

Against this backdrop, Lightmatter is working to cut the energy cost of AI computing through photonic computing. The company builds dedicated chips and photonic interconnects that use light for data processing and communication, boosting compute capability while reducing energy consumption, particularly for AI workloads and data-intensive applications. Lightmatter CEO Nicholas Harris has said: "In the coming years, every GPU on Earth used for AI training and inference will be built on Lightmatter's Passage." Passage is the company's photonic interconnect product, designed to optimize data center scaling and sharply cut the energy demands that traditional interconnects cannot overcome.

I. Founding Story

Lightmatter was founded in September 2017 by Nicholas Harris (CEO), Darius Bunandar (Chief Scientist), and Thomas Graham (former COO and CFO).

Before founding Lightmatter, Harris worked as an R&D engineer at Micron Technology starting in 2005, focusing on DRAM and NAND circuit and device physics. At Micron, he studied integrated chips in depth and came to believe that the limitations of traditional computing architectures could slow the progress of computing technology. To address this, Harris left Micron for MIT, where he pursued a PhD and postdoctoral research in electrical engineering and computer science.

At MIT, Harris began exploring quantum computing and again ran into fundamental technical bottlenecks. His doctoral dissertation, "Programmable Nanophotonics for Quantum Information Processing and Artificial Intelligence," proposed a nanophotonic architecture to break through computing bottlenecks. He also led MIT's Quantum Photonics Laboratory, developing silicon-based integrated photonic chips that let information be processed with light instead of electricity. As a researcher, Harris published 69 academic papers and holds 17 patents.

That research laid the groundwork for Lightmatter's initial concept. Harris noticed that "everyone knew deep learning was going to be a massive trend," so he dug into the field and realized that the system he had built for photonic quantum computing could also serve deep learning. That realization led him to abandon plans to become a professor after his PhD, and instead co-found Lightmatter with Bunandar (his labmate) and Graham (an MIT MBA student at the time). That same year, the startup won $100,000 at MIT's startup competition.

Before joining Lightmatter, Bunandar earned degrees in physics and mechanical engineering from UT Austin in 2013 and completed his physics PhD at MIT in 2018. Graham held various finance, operations, and business development roles at Morgan Stanley and Google before earning his MBA from MIT Sloan in 2018.

Lightmatter's leadership has changed several times since 2017. Graham served as COO until May 2023, then moved into the CFO role. In July 2024, Simona Jankowski, a former Goldman Sachs managing director and NVIDIA VP of investor relations and strategic finance, took over as CFO from Graham. As of February 2025, Graham remained a full-time employee at Lightmatter, with the title of "Founder."

II. Products

Over the past several decades, gains in compute power have relied mainly on two key trends: Moore's Law and Dennard Scaling. Moore's Law holds that the number of transistors on a chip roughly doubles every two years, while Dennard Scaling holds that as transistors shrink, their power density stays constant. Together these trends made computing cheaper, faster, smaller, and more energy-efficient. But since around 2005, these gains have begun to slow.

Lightmatter CEO and founder Nicholas Harris recognized the limits of traditional chips and proposed that photonic computing chips could keep these growth trends alive. As Harris explains:

"With photonics, computation can happen simultaneously, because data can travel on different colors of light. For instance, one color of light could carry a picture of a dog, another color a picture of a cat, another a tree. All that data can pass through the same optical compute unit, the matrix accelerator, at the same time. That approach increases compute density per unit area while making fuller use of the hardware, which improves energy efficiency."

To address the compute performance bottleneck, Lightmatter has launched two core products: Passage and Envise. Passage is an optical interconnect chip focused on efficient data transfer within data centers, while Envise is a photonic computing chip that uses optical computation to run deep learning tasks, cutting power consumption and boosting compute speed.

Passage: Optical Interconnect Chip

Passage is a silicon photonics interconnect system that uses light instead of electronic signals to move data between compute components, targeting the primary bottleneck in AI training and inference: data transfer.

In a compute cluster packed with GPUs, the main bottleneck isn't the AI model's matrix computation, it's reading, writing, and moving data. Traditional fiber-optic communication can achieve extremely high data transfer speeds, but it is typically used for long-distance communication, such as transcontinental undersea cables. Lightmatter's Passage miniaturizes fiber-optic communication down to the single-chip level, enabling efficient data transfer at that scale.

How it works:

Passage is essentially an interposer that bridges compute components. For example, it can sit between a processor and SRAM memory, letting data travel through Passage at the speed of light. Through Lightmatter's optical I/O technology, Passage supports both chip-to-chip and node-to-node high-speed communication.

Key features:

1. Compatible with existing hardware: integrates with standard commercial chips, no need to replace existing infrastructure.
2. Ultra-high-speed optical switching: delivers 100 Tbps of optical communication bandwidth, suited to photonic, CPU, or GPU compute.
3. High-density interconnect: supports up to 48 photonic, CPU, or GPU chips mounted directly on the interconnect system.
4. Efficient optical communication architecture: equipped with 40 photonic channels, with a total width equivalent to a single fiber, for efficient data transfer.

These features make Passage a strong fit for the AI data centers of the future. As AI workloads keep growing, Passage can efficiently link large numbers of compute processors, replacing traditional PCB (printed circuit board) communication and enabling greater scalability and energy efficiency.

Envise: Photonic Computing Chip

Envise, launched by Lightmatter in 2021, combines electronic and photonic computing to cut data center power consumption and operating costs. Billed as the world's first general-purpose photonic AI accelerator, Envise uses optical computation to accelerate AI tasks and can run a range of neural network workloads, from voice assistants to autonomous driving to AI models like ChatGPT.

As of 2024, a 16-chip Envise server blade draws just 3 kilowatts, compared with 6.5 kilowatts for an equivalent NVIDIA A100 system, a 54% reduction in power draw. A single Envise chip also offers:

1. 6.4 Tbps optical interconnect, scalable to multi-server systems
2. 1TB of DDR4 DRAM and 3TB of solid-state storage

Envise integrates directly with PyTorch, letting machine learning developers deploy neural networks without changing how they program. As the world's first commercial photonic computer, Envise can handle a wide range of neural network applications, from speech recognition to autonomous driving systems to large language models.

Idiom: Software Platform for Photonic Computing

Idiom is Lightmatter's companion software stack, helping users compile and run neural network models optimally on Envise. The platform offers a set of tools to optimize compute performance, including:

1. idCompile (graph compiler): optimizes deep learning workloads and improves Envise's efficiency across multi-server systems.
2. Automatic neural network partitioning, splitting large models into smaller pieces that run in parallel.
3. Data parallelism, running multiple Envise units simultaneously.
4. Mixed-precision computation, reducing compute load while preserving model accuracy.
5. idProfiler (performance analysis tool): monitors neural network execution on Envise devices, analyzing memory usage and performance data to help developers find bottlenecks and optimize models.
6. idBug (debugging tool): purpose-built for detecting errors in multi-chip parallel computation, speeding up issue diagnosis and resolution.
7. idML (machine learning optimization toolkit): integrates with PyTorch for model quantization and optimization, improving compute efficiency.
8. Support for knowledge distillation, letting smaller models learn from larger ones to improve compute efficiency.
9. Quantization-aware training, cutting compute cost without sacrificing accuracy.
10. Visual parameter analysis, helping developers track how optimizations affect model performance.

The core goal of the Idiom software stack is to let Envise run neural network tasks efficiently while minimizing wasted power and compute resources. Its compilation and optimization capabilities let Envise quickly adapt to a wide range of AI workloads, including autonomous driving, speech recognition, and large language models.

III. Market Landscape

Customers

Demand for efficient, energy-saving data processing draws three core customer groups to Lightmatter's Passage and Envise products:

1. Data centers — aiming to cut operating costs and boost compute capacity

Lightmatter's Passage optical interconnect chip improves the efficiency of data transfer between chips and servers, making it easier for data centers to scale generative AI and high-performance computing (HPC) applications. Lightmatter customizes its silicon photonic chips to the floor plan of a specific chip to optimize data transfer. As early as 2020, Lightmatter was already pitching its solution to large data center operators like Amazon and Google, aiming to help them escape the "giant spaghetti of cables" currently connecting thousands of servers.

2. AI technology companies — reliant on large-scale compute resources to develop and run AI models

Companies building specialized AI models, such as healthcare-tech firms DeepScribe and Freed, depend on large-scale compute resources to deliver AI-driven medical transcription services. But as AI workloads grow, simply adding more GPUs doesn't solve the data transfer bottleneck between compute components. For these large, industry-specific AI application companies, the interconnect layer becomes the key factor in system performance optimization.

3. Cloud service providers — needing high-bandwidth compute infrastructure to scale capacity

Cloud giants such as AWS, Google Cloud, and Microsoft Azure run massive compute clusters where traditional network architecture has become the bottleneck to scaling compute capacity. Lightmatter CEO Harris notes that cloud providers "have already recognized the limits of traditional networking solutions; efficient, high-bandwidth connectivity has become essential." Lightmatter's photonic interconnect technology can offer these companies a more scalable, energy-efficient solution.

Market Size

The global AI products and services market is projected to reach $780 billion to $990 billion by 2027, up sharply from $185 billion in 2023.

1. AI hardware market

The global AI hardware market was worth roughly $53.7 billion in 2023 and is projected to grow to $473.5 billion by 2033, a compound annual growth rate (CAGR) of 24.5%. This growth is driven mainly by the widespread adoption of AI technology, advances in semiconductor technology, and demand for edge AI solutions.

The AI hardware market covers specialized chips, processors, and compute systems, with the AI inference chip segment growing especially fast: it was worth about $15.8 billion in 2023 and is projected to reach $90.6 billion by 2030, a CAGR of 22.6%. AI inference compute accounts for 40% of the AI hardware market, and as more powerful compute chips are developed, inference costs may come to dominate overall AI costs.

2. Photonic integrated circuit market

The global photonic integrated circuit (PIC) market was worth $8.6 billion in 2023 and is projected to grow to $47.6 billion by 2032, a CAGR of 20.9%. The power demands of high-performance computing are the core driver of growth in photonic computing. Notably, while Lightmatter is a pioneer in photonic computing, the technology has yet to see wide adoption. Chip giants like NVIDIA and AMD dominate the market, and whether photonic computing can be seamlessly integrated into existing chip architectures will determine its commercial viability.

IV. Competition

Lightmatter competes across two product lines, Envise and Passage, in the high-performance AI compute and photonic interconnect markets.

Envise competes mainly with companies focused on high-performance AI compute hardware, such as NVIDIA's H100, Google's TPU, and Cerebras' WSE-3. These competitors all focus on optimizing compute speed and energy efficiency to meet demanding AI compute needs. Envise uses photonic computing rather than traditional electronic transistors, giving it stronger compute capability and lower heat output, a distinct advantage in the AI and high-performance compute markets.

Passage, as a photonics-based interconnect product, faces relatively fewer direct competitors. Its main rival is Ayar Labs, which also focuses on using photonic technology to boost interconnect capacity in high-bandwidth compute systems.

Key Competitors in Photonic Computing

Celestial AI

Celestial AI focuses on developing high-performance memory and interconnect solutions using silicon photonics. Its flagship product, Photonic Fabric, offers ultra-high bandwidth (14.4 Tbps) and low-latency memory interaction, combined with a chiplet architecture, addressing the limitations of traditional electronic interconnects and making AI hardware scaling more efficient and energy-saving.

Celestial AI competes directly with Lightmatter's Passage in the photonic interconnect market. Both aim to increase interconnect density and bandwidth for AI compute systems, but Celestial AI focuses more on memory expansion and inter-chip communication, while Passage aims to build a more comprehensive photonic interconnect platform, providing high-density optical interconnects for AI infrastructure while integrating lasers and dynamic network switching.

As of February 2025, Celestial AI has raised a cumulative $338.9 million, with investors including Samsung Catalyst Fund, Temasek, and the US Innovative Technology Fund, valuing the company at $1.2 billion.

Ayar Labs

Founded in 2015, Ayar Labs focuses on using optical technology to improve data transfer efficiency between server chips, reducing the power overhead of traditional electronic transmission. Its products compete directly with Lightmatter's Passage.

NVIDIA invested in Ayar Labs' $155 million Series C round in May 2023, aiming to integrate its optical interconnect technology into GPU racks to improve energy efficiency. Ayar Labs went on to complete a $155 million Series D round in December 2024, pushing its valuation past $1 billion, with investors including AMD Ventures, Intel Capital, and NVIDIA. As of February 2025, the company had raised a cumulative $374.7 million.

Competition from Established Chip Giants

NVIDIA

NVIDIA, the world's leading compute infrastructure company, was valued at $3.3 trillion as of February 2025 and holds a 70%-95% share of the AI chip market. Its flagship H100 sells for roughly $30,000, a high cost that may push some customers toward alternatives.

Lightmatter's Envise enters the AI compute market to challenge NVIDIA's dominance in electronic GPUs (such as the A100 and H100). Envise uses photonic computing focused on the matrix operations at the heart of AI compute, potentially offering a better solution on energy efficiency and scalability. Even though NVIDIA still held 88% of the AI electronic GPU market as of June 2024, Envise aims to win over data centers and large-scale AI compute applications on the strength of its energy efficiency.

Additionally, as AI model sizes keep growing, so does the memory they require. NVIDIA's GPUs, for instance, come with just 80GB of memory, while modern generative AI models (some with trillions of parameters) demand far more, forcing companies to deploy multiple GPUs and adding cost and management complexity. Envise, by contrast, is optimized for large-scale AI compute, offering a more energy-efficient, lower-cost alternative.

Notably, NVIDIA may itself be a potential Lightmatter customer. Passage could be used to replace or augment NVIDIA's NVLink (the high-speed interconnect between GPUs and CPUs), delivering higher bandwidth and lower power draw.

Cerebras

Cerebras Systems is known for building the "world's largest chip." Its flagship product, the Wafer-Scale Engine (WSE), which the company calls "the fastest AI processor on Earth," surpasses traditional processors in AI-optimized core count, memory speed, and on-chip interconnect bandwidth.

Its latest compute system, the CS-3, powered by the WSE-3 chip, reached speeds of 1,800 and 450 tokens/s when training smaller models like Llama 3.1 8B and 70B respectively. WSE-3 addresses the memory bandwidth bottleneck of traditional GPU systems; its memory bandwidth, 7,000 times that of NVIDIA's H100, eliminates the need for frequent data movement, making AI training more efficient.

Cerebras focuses on solving the memory challenges of GPU systems. Llama 70B, for instance, requires 140GB of memory, and Cerebras' CS-3 system uses multi-machine distributed computing to support trillion-parameter-scale models. While both Lightmatter and Cerebras are working to optimize AI compute performance, they take different technical paths: Cerebras relies on wafer-scale electronic computing, while Lightmatter relies on photonic computing.

V. Business Model

As of February 2025, Lightmatter's pricing model has not been publicly disclosed. However, by embedding optical interconnect technology directly into chip packaging, the company aims to substantially improve the performance and efficiency of data centers to meet growing AI compute demand. This technical breakthrough would let data centers scale better and handle increasingly complex AI workloads. Lightmatter's revenue model likely relies mainly on direct product sales, alongside long-term partnerships with enterprises that need cutting-edge compute solutions. CEO Harris has said the pricing of its Envise product is "quite competitive compared to standard silicon chips, though the highest-performance tiers may carry a premium."

Lightmatter may also offer technology licensing or integration services, particularly for large-scale infrastructure deployments where photonic technology can optimize data transfer and compute efficiency. As cloud computing and AI data centers continue to expand rapidly, this space could offer substantial commercialization opportunities for Lightmatter. At the same time, the industry's growing adoption of silicon photonics in next-generation AI chips and systems gives Lightmatter further room to grow, positioning it well as the market gradually shifts toward photonic architectures.

VI. Major Developments

Lightmatter is working with several semiconductor giants, including AMD, NVIDIA, Intel, and Qualcomm, providing optical interconnect substrates for their chips. The company's CEO has said Lightmatter's substrate is "the cornerstone of the future of computing technology. It will dramatically cut the energy consumption of compute clusters and let them scale to a million nodes or beyond, which is also key to achieving wafer-scale computing, building the world's largest chip, and reaching artificial general intelligence (AGI)." On common compute tasks, Lightmatter's technology is up to 10 times faster than NVIDIA's A100 chip, while its server blade delivers eight times the throughput of an NVIDIA server at one-fifth the power consumption.

As of February 2025, Lightmatter employed roughly 191 people across its Boston and Mountain View offices, up 46% from 2023.

In November 2024, Lightmatter partnered with Amkor Technology, a leading global provider of semiconductor packaging and testing services, to build the world's largest 3D packaged chip complex, using Lightmatter's Passage technology to optimize the package design. Combining Lightmatter's photonic technology with Amkor's multi-die packaging capabilities will help address the demanding interconnect scalability and power requirements of AI workloads.

Multi-die packaging breaks large semiconductor chips into smaller chip units and integrates multiple chips within a single package. Combined with Lightmatter's photonic interconnects, this can use optical communication to optimize data transfer within a chip, sharply cutting energy consumption. Ritesh Jain, Lightmatter's SVP of Engineering and Operations, called the partnership "a key step toward building a world-class ecosystem that helps customers achieve unprecedented bandwidth and efficiency, driving technological breakthroughs in AI and high-performance computing (HPC)."

VII. Funding & Valuation

In October 2024, Lightmatter closed a $400 million Series D round at a $4.4 billion valuation. The round was led by T. Rowe Price, with existing investors Google Ventures and Fidelity participating. Lightmatter has now raised a cumulative $850 million across six funding rounds, ranking second in SVTR's AI venture capital ranking for the photonic/quantum tech sector.

Previous investors include well-known names such as Matrix, Spark Capital, SIP Global Partners, Sequoia, Lockheed Martin Ventures, FAST (a GettyLab fund), Hewlett Packard Enterprise, Viking Global Investors, Trajectory Ventures, and Eastward Capital Partners.

Lightmatter CEO Harris has said the Series D is likely to be the company's last funding round. He also disclosed that as of October 2024, Lightmatter had already locked in some large-scale orders, and it was the strength of those orders that enabled the company to raise such a large round. That suggests Lightmatter has entered a strong commercialization phase, with the capital strength and market recognition to sustain continued growth.

VIII. Key Opportunities

Photonic Interconnect Technology

As Moore's Law slows and silicon chip manufacturing approaches its physical and economic limits, high-performance computing faces a major opportunity in photonic interconnect technology. Data centers today struggle to handle ever-growing AI workloads, while silicon-based chip scaling is nearing its ceiling. Lightmatter's photonic technology is positioned to solve exactly this bottleneck, with its high-speed data transfer and low energy consumption providing the infrastructure needed for large-scale chip interconnects in modern data centers.

Using photonic technology, Lightmatter can meet AI infrastructure's need for high-density interconnect environments. Its Passage technology is designed to integrate seamlessly with hyperscale chips, even near-wafer-scale chips, letting an entire data center operate like a single high-performance computer. This trend aligns with the push toward artificial superintelligence (ASI), and Lightmatter's photonic interconnect technology could become a key enabler of AI's push to higher performance.

Currently, NVIDIA's GPUs and Intel's Xeon CPUs are growing at compound annual rates of 13% and 18% respectively, and are expected to evolve toward wafer-scale chips by 2040. But as chip counts rise, interconnect demand grows fourfold, worsening problems around routing, bandwidth, and latency. Photonic technology is the best available answer to that problem.

New Chip Substrates

Beyond chip interconnects, Lightmatter is also developing new chip substrates to further extend chip interconnect capability using photonic technology. In October 2024, Lightmatter CEO Nicholas Harris predicted that wafer-scale chips, of the kind Cerebras pioneered, would become an industry standard within the next decade. However, he believes current Cerebras technology has not yet fully realized the potential of that breakthrough.

A chip substrate is the base material that supports and connects microelectronic devices, determining a chip's stability, electrical performance, and heat dissipation. Advanced substrate technology could become another major opportunity for Lightmatter to extend its photonic technology. By developing next-generation substrates, Lightmatter could integrate larger, more complex photonic components (such as wafer-scale chips), improving performance, reducing latency, and meeting the hyperscale interconnect demands of AI and ASI data centers.

Laser Technology

Lightmatter is also exploring chip-scale erbium-doped waveguide lasers, a technology that could further strengthen its photonic computing and interconnect solutions. Erbium-doped lasers have long been used in fiber-optic communication, relying on erbium ions as a gain medium to amplify optical signals at specific wavelengths, offering high coherence and stability well suited to long-distance, high-precision optical applications.

Applying this technology to Lightmatter's photonic products could bring several benefits:

1. Improved precision and energy efficiency: stable erbium-doped lasers can improve the precision and efficiency of data transfer, meeting high-bandwidth AI compute needs.
2. Better scalability: miniaturized lasers can integrate into more compact photonic chips, improving a product's ability to scale.
3. Lower manufacturing cost: this technology could improve compatibility between photonic chips and existing semiconductor manufacturing processes, lowering production costs.

Beyond data centers and AI compute, this technology could also open new markets for Lightmatter in telecommunications, LiDAR, and high-precision sensing. As chip-scale laser technology advances, Lightmatter could use this innovation to gain an edge across multiple high-growth sectors.

IX. Key Risks

Market Acceptance of Photonic Computing

Integrating photonic computing into existing electronic systems presents significant technical challenges. Given the inherent complexity of manufacturing and designing photonic technology, as of 2024 there remain real barriers to adoption of this emerging technology. While Lightmatter claims its technology is compatible with existing electronic infrastructure, it has not yet announced any formal customers, suggesting the market may still be taking a wait-and-see approach to this innovation.

For data centers with mature, established electronic systems, adopting photonic computing carries the risk of operational disruption. How to lower switching costs, improve compatibility, and deliver a compelling performance gain will be key to Lightmatter winning market acceptance.

Limitations in Logic Operations

Photonic computing faces fundamental limits when it comes to logic operations, especially for general-purpose computing tasks that require control flow and conditional logic. The industry has tried to solve this before, such as Bell Labs' optical transistor research in the 1980s, but the complexity of nonlinear light-beam interactions ultimately prevented commercialization.

To work around this limitation, Lightmatter has chosen to focus on linear algebra computation, which aligns well with light's electromagnetic properties. But that focus also means photonic computing's range of applications is limited, suited mainly to fields like machine learning, aerospace engineering, chemical simulation, and quantum computing. While these fields have enormous demand for efficient computation, photonic computing will still need to overcome its inherent logic-operation limitations before it can play a role in broader computing tasks.

X. Summary

As AI compute demand surges, traditional silicon-based computing is approaching its physical and economic limits. The slowdown of Moore's Law has left data centers and AI companies scrambling for new approaches that can handle growing workloads while cutting energy consumption and operating costs. Lightmatter is addressing this challenge with photonic computing hardware, using light instead of electrical signals to compute, breaking through the performance ceiling of traditional chips.

Lightmatter's core product, Passage, is a photonic interconnect solution that integrates seamlessly into existing data center infrastructure, delivering higher-bandwidth data transfer while cutting chip power consumption. Its Envise product is a photonic accelerator purpose-built for AI compute, capable of running AI tasks more energy-efficiently and opening new possibilities for the future of high-performance computing.