{"id":1142640,"date":"2025-11-26T06:21:41","date_gmt":"2025-11-26T14:21:41","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/"},"modified":"2026-01-05T14:22:20","modified_gmt":"2026-01-05T22:22:20","slug":"future-ai-infrastructure","status":"publish","type":"msr-group","link":"https:\/\/www.microsoft.com\/en-us\/research\/theme\/future-ai-infrastructure\/","title":{"rendered":"Future AI Infrastructure"},"content":{"rendered":"
\n\t
\n\t\t
\n\t\t\t\"Colour\t\t<\/div>\n\t\t\n\t\t
\n\t\t\t\n\t\t\t
\n\t\t\t\t\n\t\t\t\t
\n\t\t\t\t\t\n\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\t\t<\/span>\n\t\t\t\t\t\t\t\t\tMicrosoft Research Cambridge\t\t\t\t\t\t\t\t<\/a>\n\t\t\t\t\t\t\t\n\t\t\t\t\t\t\t\n\n

Future AI Infrastructure<\/h1>\n\n\n\n

Innovative hardware and systems for sustainable and efficient AI infrastructure<\/p>\n\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n\n\n\n\n\n\n\n\n\n\n\n

Our mission in Future AI Infrastructure is to co-design the next generation of hardware and systems for Microsoft\u2019s AI and Cloud datacenters\u2014spanning storage, networking and compute, and other foundational technologies for our AI stack. We work across multiple disciplines including computer science, physics, materials science, machine learning, engineering, electronics, and robotics to ideate and incubate full-stack innovations. We collaborate closely with academic colleagues and, through strong ties to product teams and industry partners, ensure we pursue keystone research problems that are deeply aligned with the needs of the company, the industry, and society.<\/p>\n\n\n\n

Our workstreams<\/h2>\n\n\n\n

Networking:<\/strong> together with Azure Core and Azure Hardware, we are developing innovative optical transceiver and switching technologies of the future, not just improving cost-effectiveness but also opening new opportunities for fundamentally reimagined AI infrastructure \u2013 from accelerators and memory subsystems to rack and datacenter designs.<\/p>\n\n\n\n

Green technology:<\/strong> we are developing scalable, low-cost sensors and AI-powered systems to enhance greenhouse gas monitoring accuracy and drive actionable climate solutions worldwide.<\/p>\n\n\n\n

Computing:<\/strong> we are building the world\u2019s first unconventional computing system capable of accelerating real-world AI inference and optimization workloads using analog computing, which has the promise of being far more energy efficient, and fast for these important problem type.<\/p>\n\n\n\n

Storage, memory and materials<\/strong>: we aim to discover and create innovative materials to increase performance per $ and enhance sustainability of our Cloud infrastructure. For example, we can use AI to evaluate new materials for data storage. We are also actively pursuing new avenues in memory technologies.<\/p>\n\n\n\n

Robotics:<\/strong> we are developing cutting-edge robotics to transform datacenter operations, starting with repair tasks in the networking infrastructure. Our long-term vision is to enable self-maintaining datacenters\u2014from initial rack assembly and deployment to ongoing maintenance, repairs, decommissioning, and reconfiguration\u2014aiming to reduce downtime, enhance reliability, and support the growing demand for scalable, sustainable cloud infrastructure.<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Networking<\/h2>\n\n\n\n

Historically, significant advancements in datacenter design have relied on disruptive network innovations that redefine the balance among bandwidth, latency, power consumption, and cost. For example, the introduction of high-bisection bandwidth networks in the late 2000s enabled cloud datacenters to separate compute from storage, greatly enhancing scalability and efficiency.<\/p>\n\n\n\n

Today, the rapid growth of AI workloads is once again stretching current network technologies to their limits, creating performance bottlenecks, driving up costs, and increasing power usage. In response, Microsoft Research is closely collaborating with Azure Core, Azure Hardware, and Microsoft 365 to overcome these challenges.<\/p>\n\n\n\n

Together, we are developing innovative optical transceiver and switching technologies designed to deliver significantly higher bandwidth and reliability while reducing power consumption and costs. These advancements not only substantially improve performance-per-dollar but also open new opportunities for fundamentally reimaging AI infrastructure, from accelerators and memory subsystems designs to rack and datacenter architectures, ensuring Microsoft remain at the forefront of datacenter innovation and shaping the next generation of infrastructure optimized for the AI era.<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Green technologies<\/h2>\n\n\n\n

Greenhouse gas (GHG) concentrations continue to rise, intensifying the global climate crisis. Yet, accurately measuring emissions\u2014especially at a local level\u2014remains a major challenge due to the lack of fine-grained, scalable data.<\/p>\n\n\n\n

We are advancing technologies to transform how emissions are monitored globally. Our work focuses on developing low-cost, miniaturized GHG sensors\u2014particularly for methane. By leveraging datacenter and Cloud technologies, along with advanced AI algorithms, we aim to enhance emission detection, improve measurement accuracy, and enable real-time, large-scale monitoring. In collaboration with external partners, we are also quantifying the value of hyperlocal emissions data through a distributed network of affordable sensors with a goal of empowering communities, researchers, and policymakers with actionable insights to drive climate solutions.<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Computing<\/h2>\n\n\n\n
\"Schematic
Schematic of the analog optical computer. In the foreground is the vector-by-matrix multiplication unit. This consists of a 1D array of micro-LEDs, a 2D modulator array (using display projectors), and a 1D array of Silicon sensors.<\/figcaption><\/figure>\n\n\n\n
<\/div>\n\n\n\n

As industries increasingly rely on AI models and complex optimization, computing demands are soaring\u2014just as digital hardware reaches its limits. To meet this challenge, Microsoft Research has developed the Analog Optical Computer (AOC):<\/strong> the world\u2019s first unconventional computing system capable of accelerating real-world AI inference and optimization workloads.<\/p>\n\n\n\n

At scale the AOC has the potential to solve problems 100x faster or more energy efficiently<\/strong>, than digital systems. This can be achieved by leveraging the parallelism of light and physical processes to perform computations, and avoiding the separation of compute from memory, operating on both continuous and binary data and adopting asynchronous operation.<\/p>\n\n\n\n

Built from low-cost, scalable, and high-volume optical and analog electronics, AOC operates at room temperature. A key innovation of AOC lies in the co-design of hardware and applications, reminiscent of the co-evolution between GPUs and deep learning models.<\/p>\n\n\n\n

To realize its potential, close collaboration on real industry applications is key.<\/p>\n\n\n\n

In partnership with Barclays, we solved a scaled-down version of a high-value financial optimization problem on AOC hardware. Similarly, with the Microsoft Health Futures team, we demonstrated the reconstruction of representative small-scale MRI data, pointing to better patient experiences through faster imaging.<\/p>\n\n\n\n

AOC also runs neural models for image recognition (e.g. MNIST and Fashion MNIST) and nonlinear curve fitting. Beyond what is running on hardware today, we trained a billion-parameter language model on GPUs that applies test-time compute compatible with AOC\u2019s capabilities.<\/p>\n\n\n\n

AOC was featured both at Microsoft Build (opens in new tab)<\/span><\/a> \u2014 our segment begins at 57:47 \u2014 and at Microsoft Ignite (opens in new tab)<\/span><\/a> \u2014 our segment begins at 41:20 \u2014 on Inside Azure Innovations with Mark Russinovich.<\/p>\n\n\n\n

We are partnering with M365 Research on this research.<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

<\/p>\n\n\n\n

Analog optical computer for AI inference and combinatorial optimization<\/a>Kirill P. Kalinin<\/a>, Jannes Gladrow, Jiaqi Chu<\/a>, James H. Clegg<\/a>, Daniel Cletheroe<\/a>, Douglas J. Kelly<\/a>, Babak Rahmani, Grace Brennan<\/a>, Burcu Canakci<\/a>, Fabian Falck<\/a>, Michael Hansen<\/a>, Jim Kleewein (opens in new tab)<\/span><\/a>, Heiner Kremer<\/a>, Greg O\u2019Shea, Lucinda Pickup, Saravan Rajmohan<\/a>, Ant Rowstron, Victor Ruehle<\/a>, Lee Braine, Shrirang Khedekar, Natalia G. Berloff, Christos Gkantsidis, Francesca Parmigiani<\/a>, Hitesh Ballani<\/a>, Nature<\/em> 645<\/strong>, 354\u2013361<\/em>.  Publication (opens in new tab)<\/span><\/a> | September 2025<\/p>\n\n\n\n

Research talk: Computing at the speed of light
<\/a>Video | October 2022<\/p>\n\n\n\n

Analog optical computing for sustainable AI and beyond
<\/a>Video | September 2024<\/p>\n\n\n\n

Implicit Language Models are RNNs: Balancing Parallelization and Expressivity<\/a>
Publication | June 2025<\/p>\n\n\n\n

Analog Iterative Machine (AIM): using light to solve quadratic optimization problems with mixed variables<\/a>
Publication | June 2023<\/p>\n\n\n\n

Roadmap on Neuromorphic Photonics<\/a>
Publication | January 2025<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Storage and materials<\/h2>\n\n\n\n

We aim to discover<\/strong> and create<\/strong> innovative materials to increase performance per $<\/strong> and enhance sustainability<\/strong> of our cloud infrastructure. For example, we are using advanced machine learning to speed up the search and discovery of new materials for data storage. We are also actively pursuing new avenues in memory technologies. We work as a tight cross-disciplinary team covering computer science, physics, and materials science.<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

Robotics<\/h2>\n\n\n\n

Modern AI infrastructure faces the challenge of balancing hardware reliability with cost management. As the infrastructure expands, hardware failures become common, resulting in costly downtime and frequent repairs. The traditional approach requires numerous technicians and can take many hours to days to service, hindering efficiency as demand for cloud-based services grows. Datacenter robotics and automation offer a transformative solution to this challenge.<\/p>\n\n\n\n

At MSR, we are developing cutting-edge robotics to revolutionize datacenter operations. By co-designing hardware and software<\/strong>, we ensure seamless integration of our modular robots. Starting with networking, our dexterous robots autonomously manipulate optical transceivers and cables and perform complex tasks such as optical fiber cleaning and installation. These robots navigate densely packed environments filled with delicate, intersecting cables, which are typically challenging due to occlusions and the complexity of deformable objects.<\/p>\n\n\n\n

Our goal is to enable self-maintaining datacenters<\/strong> by leveraging dexterous modular robotics, AI, and automation. Our robotic systems are engineered for dexterous manipulation of hardware throughout its lifecycle\u2014from initial rack assembly and deployment to ongoing maintenance, repairs, decommissioning, and reconfiguration. By implementing these autonomous solutions, datacenters can significantly reduce service window durations and enhance reliability, paving the way for a more resilient future cloud. The future of cloud infrastructure is adaptive, efficient, and resilient, and robotics will play a pivotal role in achieving this vision!<\/p>\n\n\n\n

Learn more:<\/h3>\n\n\n\n

Self-maintaining [networked] systems: The rise of datacenter robotics!<\/a>
Publication | November 2024<\/p>\n\n\n\n

Robust Optical Transceiver Manipulation in Cluttered Cable Environments Using 3D Scene Understanding and Planning<\/a>
Publication | May 2025<\/p>\n\n\n\n

<\/div>\n\n\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

Innovative hardware and systems for sustainable and efficient AI infrastructure Our mission in Future AI Infrastructure is to co-design the next generation of hardware and systems for Microsoft\u2019s AI and Cloud datacenters\u2014spanning storage, networking and compute, and other foundational technologies for our AI stack. We work across multiple disciplines including computer science, physics, materials science, […]<\/p>\n","protected":false},"featured_media":1142843,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr_group_start":"","footnotes":""},"research-area":[13556,13562,13552,13547],"msr-group-type":[243688],"msr-locale":[268875],"msr-impact-theme":[],"class_list":["post-1142640","msr-group","type-msr-group","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-research-area-computer-vision","msr-research-area-hardware-devices","msr-research-area-systems-and-networking","msr-group-type-theme","msr-locale-en_us"],"msr_group_start":"","msr_detailed_description":"","msr_further_details":"","msr_hero_images":[],"msr_research_lab":[199560,199561,1012650],"related-researchers":[{"type":"user_nicename","display_name":"Hitesh Ballani","user_id":32008,"people_section":"Section name 0","alias":"hiballan"},{"type":"guest","display_name":"Natalia Berloff","user_id":855339,"people_section":"Section name 0","alias":""},{"type":"user_nicename","display_name":"Richard Black","user_id":33417,"people_section":"Section name 0","alias":"rjblack"},{"type":"user_nicename","display_name":"Burcu Canakci","user_id":41910,"people_section":"Section name 0","alias":"burcucanakci"},{"type":"user_nicename","display_name":"Andromachi Chatzieleftheriou","user_id":37833,"people_section":"Section name 0","alias":"anchatzi"},{"type":"user_nicename","display_name":"Jiaqi Chu","user_id":39147,"people_section":"Section name 0","alias":"jiaqchu"},{"type":"user_nicename","display_name":"James Clegg","user_id":37794,"people_section":"Section name 0","alias":"jaclegg"},{"type":"user_nicename","display_name":"Daniel Cletheroe","user_id":31505,"people_section":"Section name 0","alias":"daclethe"},{"type":"user_nicename","display_name":"Paolo Costa","user_id":33218,"people_section":"Section name 0","alias":"pcosta"},{"type":"user_nicename","display_name":"Fabian Falck","user_id":44067,"people_section":"Section name 0","alias":"fabianfalck"},{"type":"user_nicename","display_name":"Kirill Kalinin","user_id":43191,"people_section":"Section name 0","alias":"kkalinin"},{"type":"user_nicename","display_name":"Thomas Karagiannis","user_id":34031,"people_section":"Section name 0","alias":"thomkar"},{"type":"user_nicename","display_name":"Doug Kelly","user_id":40711,"people_section":"Section name 0","alias":"dougkelly"},{"type":"user_nicename","display_name":"Vassily Lyutsarev","user_id":34498,"people_section":"Section name 0","alias":"vassilyl"},{"type":"user_nicename","display_name":"Francesca Parmigiani","user_id":37727,"people_section":"Section name 0","alias":"frparmig"},{"type":"user_nicename","display_name":"Bozidar Radunovic","user_id":31286,"people_section":"Section name 0","alias":"bozidar"},{"type":"user_nicename","display_name":"David Sweeney","user_id":31553,"people_section":"Section name 0","alias":"dasweene"},{"type":"user_nicename","display_name":"Hugh Williams","user_id":32055,"people_section":"Section name 0","alias":"hughwi"},{"type":"user_nicename","display_name":"Jiaolong Yang","user_id":36125,"people_section":"Section name 0","alias":"jiaoyan"},{"type":"user_nicename","display_name":"Yizhong Zhang","user_id":41200,"people_section":"Section name 0","alias":"yizzhan"}],"related-publications":[940299,1090110,1097973,1135077,1140743,1142749,1147643,1147670,1149294],"related-downloads":[],"related-videos":[888825,1080693],"related-projects":[1064820,955254],"related-events":[],"related-opportunities":[1162002],"related-posts":[],"tab-content":[],"msr_impact_theme":[],"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142640","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-group"}],"version-history":[{"count":37,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142640\/revisions"}],"predecessor-version":[{"id":1159630,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group\/1142640\/revisions\/1159630"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1142843"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1142640"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1142640"},{"taxonomy":"msr-group-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-group-type?post=1142640"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1142640"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1142640"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}