Microsoft Research

Microsoft Research (MSR) is Microsoft's industrial research organization. Established in 1991, it conducts fundamental and applied research in computing and other sciences through regional laboratories and mission-oriented groups.^[1][2] Its stated model combines academic publication and external collaboration with work alongside Microsoft product teams.^[1][3]

MSR is not the same thing as all research and development conducted by Microsoft, and Microsoft's company-wide research spending is not an MSR budget. It is also distinct from Microsoft AI, the consumer product and research organization created in 2024 around Copilot, Bing, Edge, and other consumer AI work.^[15] No current organization-wide staff total or separately reported budget is established by the sources used here.

Scroll sideways for more →

Purpose and boundaries

Microsoft describes MSR as a network of regional labs and mission-oriented organizations. Its current public statement emphasizes fundamental science and technology, collaboration with academia, government, industry, and Microsoft product teams, and publication or release of much of its work.^[1] Those statements describe the organization's operating model, not independent proof that every project is open, long-term, or transferred into a product.

The distinction from product organizations matters because Microsoft conducts research and engineering in many parts of the company. In March 2024, Microsoft announced a separate organization named Microsoft AI, led by Mustafa Suleyman and focused on Copilot and other consumer AI products and research. The Copilot, Bing, Edge, and GenAI teams named in that announcement moved into the new organization.^[15] A project appearing in a Microsoft product, a corporate research-spending figure, or employment by Microsoft therefore does not by itself establish that the work belongs to MSR.

MSR's own founding accounts describe two linked goals: advancing the state of the art through publication and open interaction with academic researchers, and moving suitable technologies into products.^[3] The original planning document was also explicitly applied. It proposed research on unsolved problems important to Microsoft's strategic needs and anticipated incorporation of some work into products within two to five years.^[5] MSR is consequently better described as an industrial lab balancing research autonomy and product relevance than as either a university-style institute or a product-development division.

Founding and expansion

The organization began in stages during 1991. Three researchers, George Heidorn, Karen Jensen, and Steve Richardson, started on June 10 within Microsoft's Advanced Technology and Business Development group. Gordon Bell advised Nathan Myhrvold on forming the lab and recruiting a director. Rick Rashid, then a computer-science professor at Carnegie Mellon University, decided to accept the role on August 31 and moved to Microsoft later that year.^[3] Microsoft's current profile identifies Rashid as the founder and records that he directed the worldwide organization from 1991 through 2013.^[4]

These records support calling Rashid the founding director. They do not support a simple three-person "founders" list made up of Bill Gates, Myhrvold, and Rashid. Gates and Myhrvold sponsored the initiative, Bell helped recruit its leader, the first researchers arrived before Rashid, and Rashid built and led the research organization.^[3][4][5]

Early organizational choices connected scholarly work with possible product transfer. The Redmond researchers were placed near product groups rather than on a remote campus, and the lab included software engineers who could turn research ideas into prototypes.^[3] At the same time, managers described publication and academic interaction as measures of research progress. This arrangement created a route into products without making every research project a product commitment.

MSR expanded outside Redmond during the 1990s. Its oral history places the Cambridge lab after discussions with Roger Needham and describes the move into China the following year.^[3] Microsoft announced its first Asian research facility in Beijing on November 4, 1998. The announcement recorded five initial employees and planned growth, figures that describe the launch rather than current staffing.^[6] Microsoft Research India began operations in Bengaluru in January 2005 with a mission covering long-term basic and applied work and collaboration with Indian universities and research institutions.^[7] Microsoft announced the New England lab in February 2008 and opened it in Cambridge, Massachusetts, later that year.^[8]

Rashid's profile says that by the end of his 1991-2013 tenure the organization had grown to more than 850 researchers across nearly a dozen labs.^[4] This is a dated historical figure, not a basis for a present headcount. The network has also contracted. In September 2014, Reuters reported that Microsoft closed its Mountain View research lab, affecting 50 positions, and consolidated its United States research work in Redmond, New York, and the Boston area.^[16]

Organization as of July 2026

At the article cutoff, Microsoft's directory separated the network into research missions, regional labs, and "other labs."^[2] The structure was:

Scroll sideways for more →

The four United States and Canadian labs in Redmond, New England, New York City, and Montreal were grouped under Microsoft Research Americas, led by Lili Cheng.^[10] The Cambridge page named Aditya Nori as lab director, while the India page named Venkat Padmanabhan as managing director.^[11][12] The Africa lab's public page described a multidisciplinary program in cloud and AI technologies grounded in community-based and human-centered methods.^[13] These are dated descriptions of public organization pages, not permanent titles or exhaustive accounts of each lab's work.

Microsoft Research Asia's own groups page listed Shanghai, Singapore, Tokyo, and Vancouver beneath the Asia organization.^[9] They should not be counted again as peer members of the eight-lab global directory. Microsoft launched the Singapore group in July 2025 as MSR Asia's first lab in Southeast Asia and identified Lidong Zhou as managing director of Microsoft Research Asia.^[26]

Leadership labels also require dates. The Singapore announcement called Peter Lee President of Microsoft Research in July 2025.^[26] By the July 2026 cutoff, Lee's current profile identified him as President, Microsoft Science and said that he had previously led the worldwide MSR laboratories.^[14] The later profile does not support continuing to list Lee as the current president of MSR.

Research model and product transfer

MSR's research record covers many fields, but examples should be attributed to the specific papers and teams that produced them. The cases below illustrate different paths: creation of a language or tool within a lab, joint research with a product group, publication of an architecture, and evaluation of research systems in production or scientific settings.

Programming languages and verification

Don Syme's history of F# traces the language to Microsoft Research Cambridge. Work connecting functional programming with the .NET framework began in the late 1990s; F# was initiated in 2002, its core early features were developed from 2004 to 2007, and Microsoft productized it during 2007-2010. The language later developed as an open-source, cross-platform project.^[18] This is a documented MSR-to-product path, not evidence that every language project at Microsoft began in MSR.

Leonardo de Moura and Nikolaj Bjorner introduced Z3 as an MSR satisfiability modulo theories solver for software verification and analysis. Their 2008 paper describes support for arithmetic, bit vectors, arrays, quantifiers, and other theories, and records an external prototype release in September 2007.^[19] The paper establishes Z3's origin, technical purpose, and early evaluation; later adoption and awards require separate, dated evidence.

Computer vision

The Kinect pose-recognition paper was a collaboration between Microsoft Research Cambridge and Xbox Incubation. It transformed a depth image into per-pixel body-part classifications and then proposed three-dimensional joint locations. The paper reported 200 frames per second on its consumer-hardware test setup and evaluations on synthetic and real test sets.^[20] The result identifies a component and measured performance under the study's conditions. It does not justify attributing the whole Kinect system to MSR.

The 2016 ResNet paper came from Microsoft Research authors Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. It introduced residual learning to train substantially deeper image-recognition networks. The reported 152-layer model and ensemble achieved a 3.57 percent top-five error on the ImageNet test set and won the ILSVRC 2015 classification task.^[21] These competition results are well defined; broader influence or citation-rank claims depend on a database and date and are not necessary to describe the contribution.

Datacenter and model-training systems

The Project Catapult paper reported an FPGA fabric deployed across 1,632 servers and evaluated on part of Bing's ranking stack. Under the paper's measured conditions, it improved ranking throughput per server by 95 percent at a fixed latency distribution, or reduced tail latency by 29 percent at equivalent throughput.^[22] The evidence concerns one production experiment and one part of the ranking system. It does not establish that Catapult accelerated every Microsoft datacenter service.

The 2020 ZeRO paper addressed duplicated memory use in distributed model training by partitioning model states across devices. Its implementation trained models above 100 billion parameters on a 400-GPU cluster and reported about 15 petaflops of aggregate throughput.^[23] The paper's statement that the method could scale beyond one trillion parameters was an analysis of potential capacity, not a reported trillion-parameter training run.

Graph-based retrieval and Earth-system modeling

The February 2025 GraphRAG preprint proposed constructing an entity graph and hierarchical community summaries for broad questions over large text collections. On two corpora of approximately one million tokens each, the authors reported more comprehensive and diverse answers than a vector-retrieval baseline.^[24] The evaluation used language models to generate questions and judge many answer comparisons. The authors called for tests on more domains and direct study of fabrication rates, so the experiment does not establish general superiority for all retrieval tasks.

Aurora, reported in Nature in 2025, was pretrained on more than one million hours of geophysical data and fine-tuned for air quality, ocean waves, tropical-cyclone tracks, and high-resolution weather. The authors reported performance exceeding several specialized operational systems at lower computational cost in their evaluations.^[25] They also stated that Aurora still depended on initial conditions from traditional data-assimilation systems and that the accelerated development drew on decades of data produced by numerical forecasting research. The paper therefore supports specific benchmark results, not the claim that numerical forecasting has been replaced.

Geographic and corporate constraints

The public directory still listed the Asia lab in Beijing at the cutoff and placed Shanghai, Singapore, Tokyo, and Vancouver within MSR Asia.^[2][9] Reporting about employee relocation needs a narrower description than earlier versions of this article provided. In July 2024, Rest of World reported that Microsoft had asked as many as 800 China-based employees working across artificial intelligence and cloud computing to consider moving to Canada, Australia, or Redmond.^[17] That reported group was broader than MSR Asia, so it cannot be converted into a count of MSR researchers or proof that the Beijing lab relocated.

Together with the 2014 Mountain View closure, this episode shows that industrial research labs are affected by company reorganizations and geopolitical policy as well as scholarly priorities.^[16][17] A current lab directory establishes that a unit is publicly listed at a given date. It does not establish its staffing, budget, permanence, or influence.

Interpreting the record

Several evidence boundaries recur across MSR's history:

• Corporate scope: Microsoft-wide spending, employees, products, and AI initiatives are not automatically MSR spending, staff, output, or management.^[15]
• Dated organization: lab lists and leadership titles change. The global directory, regional pages, and Peter Lee's profile were reconciled to the July 28, 2026 cutoff.^{[2][10][11][12][14]}
• Authorship and collaboration: Kinect joined MSR Cambridge with Xbox Incubation, while Catapult's author list and deployment involved multiple Microsoft teams. Product transfer should be credited to the named collaborators.^[20][22]
• Experimental scope: ResNet's 3.57 percent result belongs to a specified competition; ZeRO's trillion-parameter statement was projected capacity; GraphRAG's study covered two corpora; and Aurora's four-domain results retained conventional data-assimilation inputs.^{[21][23][24][25]}
• Source type: Microsoft histories and launch announcements establish what the organization said and did at a date. Peer-reviewed papers support technical methods and measured results. Independent reporting is needed for closures, workforce changes, and contested corporate conditions.^{[3][6][7][8][16][17][26]}

These distinctions make it possible to describe an industrial research organization without turning corporate claims, citation counts, proposed projects, or product marketing into measures of scientific impact.

See also

• Microsoft
• Artificial intelligence
• Machine learning
• Computer vision
• Kinect
• ResNet
• DeepSpeed
• GraphRAG
• Bing
• Carnegie Mellon University

References

1. ^Microsoft Research. "About Microsoft Research." microsoft.com/...about-microsoft-research
2. ^Microsoft Research. "Labs & locations." microsoft.com/...labs-locations
3. ^Rob Knies. "The Meteoric Rise of Microsoft Research: An Oral History." Microsoft Research, September 26, 2006. microsoft.com/...e-microsoft-research-oral-history
4. ^Microsoft Research. "Rick Rashid." microsoft.com/...rashid
5. ^Nathan Myhrvold. "Microsoft Research Plan." Microsoft, 1991. microsoft.com/...Microsoft-Research-Plan.pdf
6. ^Microsoft. "Microsoft Announces Beijing as Site for Its First Asian Research Lab." November 4, 1998. news.microsoft.com/...its-first-asian-research-lab
7. ^Microsoft. "Microsoft Research Launches Operations in India." January 12, 2005. news.microsoft.com/...launches-operations-in-india
8. ^Microsoft. "Microsoft Research to Open New Doors for Scientists in Cambridge, Mass." February 4, 2008. news.microsoft.com/...scientists-in-cambridge-mass
9. ^Microsoft Research. "Microsoft Research Lab - Asia: Groups." microsoft.com/...groups
10. ^Microsoft Research. "Lili Cheng, Corporate Vice President and Managing Director, Microsoft Research Americas." microsoft.com/...lilich
11. ^Microsoft Research. "Microsoft Research Cambridge." microsoft.com/...microsoft-research-cambridge
12. ^Microsoft Research. "Microsoft Research India: Overview." microsoft.com/...overview
13. ^Microsoft Research. "Microsoft Research Lab - Africa, Nairobi." microsoft.com/...osoft-research-lab-africa-nairobi
14. ^Microsoft Research. "Peter Lee." microsoft.com/...petelee
15. ^Satya Nadella. "Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot." Official Microsoft Blog, March 19, 2024. blogs.microsoft.com/...s-microsoft-to-lead-copilot
16. ^Reuters. "Microsoft lays off 2,100." *South China Morning Post*, September 19, 2014. scmp.com/...microsoft-lays-2100
17. ^Viola Zhou. "Pressured to relocate, Microsoft's AI engineers in China must choose between homeland and career." *Rest of World*, July 23, 2024. restofworld.org/...ft-china-ai-engineer-relocation
18. ^Don Syme. "The Early History of F#." *Proceedings of the ACM on Programming Languages* 4, HOPL, article 75, 2020. fsharp.org/...hopl-fsharp.pdf
19. ^Leonardo de Moura and Nikolaj Bjorner. "Z3: An Efficient SMT Solver." *Tools and Algorithms for the Construction and Analysis of Systems*, 2008, pp. 337-340. link.springer.com/...07%2F978-3-540-78800-3_24.pdf
20. ^Jamie Shotton et al. "Real-Time Human Pose Recognition in Parts from Single Depth Images." *2011 IEEE Conference on Computer Vision and Pattern Recognition*, 2011, pp. 1297-1304. microsoft.com/...BodyPartRecognition.pdf
21. ^Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep Residual Learning for Image Recognition." *2016 IEEE Conference on Computer Vision and Pattern Recognition*, 2016, pp. 770-778. openaccess.thecvf.com/...rning_CVPR_2016_paper.pdf
22. ^Andrew Putnam et al. "A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services." *2014 ACM/IEEE 41st International Symposium on Computer Architecture*, 2014, pp. 13-24. microsoft.com/...Catapult_ISCA_2014.pdf
23. ^Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. "ZeRO: Memory Optimizations Toward Training Trillion Parameter Models." arXiv:1910.02054, version 3, May 13, 2020. arxiv.org/...1910.02054
24. ^Darren Edge et al. "From Local to Global: A GraphRAG Approach to Query-Focused Summarization." arXiv:2404.16130, version 2, February 19, 2025. arxiv.org/...2404.16130
25. ^Cristian Bodnar et al. "A foundation model for the Earth system." *Nature* 641 (2025): 1180-1187. nature.com/...s41586-025-09005-y.pdf
26. ^Microsoft. "Microsoft Research Asia Launches Singapore Lab to Drive AI Innovation, Industrial Transformation, and Talent Development." July 24, 2025. news.microsoft.com/...ation-and-talent-development