Microsoft Research
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v5 · 7,126 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v5 · 7,126 words
Add missing citations, update stale details, or suggest a clearer explanation.
Microsoft Research (often abbreviated MSR) is the basic and applied research division of Microsoft Corporation, founded in 1991 in Redmond, Washington, and one of the largest computer science research organizations in the world, with more than 1,000 researchers across labs in the United States, United Kingdom, China, India, Canada, Japan, and Singapore[1][2]. MSR conducts open, publishable research spanning machine learning, systems, programming languages, human-computer interaction, computer vision, quantum computing, and theoretical computer science, and transfers that work into Microsoft products. Its researchers authored "Deep Residual Learning for Image Recognition" (ResNet), one of the most cited papers in the history of artificial intelligence, with well over 200,000 citations, and the 2023 study "Sparks of Artificial General Intelligence" on GPT-4[3][4][52]. MSR is organizationally distinct from the consumer-facing Microsoft AI division created in 2024, and its work is published largely without paywall on the MSR website[2].
| Microsoft Research | |
|---|---|
| Type | Corporate research laboratory |
| Founded | 1991 |
| Founders | Bill Gates, Nathan Myhrvold, Rick Rashid |
| Headquarters | Redmond, Washington, United States |
| Parent | Microsoft Corporation |
| Key people | Peter Lee (President of Microsoft Science), Eric Horvitz (Chief Scientific Officer), Jaime Teevan (Chief Scientist), Igor Tsyganskiy (Microsoft Research lead) |
| Annual Microsoft R&D spend (FY2025) | ~$32.5 billion (parent company total)[5][80] |
| Approximate staff | More than 1,000 researchers[1] |
| Website | microsoft.com/en-us/research |
Microsoft Research is Microsoft's in-house science organization. Unlike most corporate engineering groups, MSR operates with an academic publication culture: researchers identify their own long-horizon problems, publish openly at peer-reviewed venues, collaborate with universities, and contribute to Microsoft product groups when ideas mature. The lab measures its impact through both peer-reviewed publication and product transfer. Research from MSR underpins query optimization in SQL Server, ranking in Bing, grammar and proofing tools in Office, the computer vision pipeline of the Kinect sensor for Xbox, and features of Microsoft 365 Copilot[2][10][11].
Microsoft Research grew out of a 1991 push by Bill Gates and Microsoft's then chief technology officer Nathan Myhrvold to establish a basic research organization at the company. Myhrvold sent Gates a 21-page memo titled "Microsoft Research Plan" arguing that Microsoft needed a research arm modeled on the great industrial labs of the twentieth century, Bell Labs and Xerox PARC[6][7]. The timing was unusual. Through the late 1980s and early 1990s, large companies were trimming long-horizon research budgets, and Bell Labs in particular was being broken up. Gates and Myhrvold moved in the opposite direction, betting that a software company that aspired to do more than ride the personal-computer boom needed to invest in fundamental computer science.
In September 1991, Microsoft hired Rick Rashid to run the new lab. Rashid had been a professor of computer science at Carnegie Mellon University (CMU), where he led the development of the Mach kernel, an influential operating system project that would later underpin macOS and iOS. According to oral histories of MSR's founding, Rashid received a recruiting call from Gordon Bell while he was at home in Pittsburgh and initially thought the idea of Microsoft starting a research lab was a joke, given that Microsoft was still a comparatively small company at the time[7][8]. Rashid was the first head of Microsoft Research and stayed in the role for more than two decades, leading the lab through its expansion from a small Redmond team to a global organization with over 850 researchers across nearly a dozen labs by the time he stepped down in 2013[8][9].
Under Rashid's tenure, MSR adopted an academic style: papers were published in open conferences, researchers worked closely with universities, and the lab tracked its impact through both peer-reviewed publication and product transfer.
From the start, the lab was given an unusual amount of autonomy by Microsoft's standards. Researchers were expected to identify their own problems, publish openly, and contribute to product groups when ideas became practical. That model produced a steady flow of contributions to Microsoft products: query optimization in SQL Server, Bing's ranking algorithms, the early versions of Cortana, grammar and style features in Office, and the underlying computer vision pipeline of the Kinect sensor for Xbox[10][11].
The Bay Area lab, founded in 1995 by Gordon Bell and Jim Gray with a focus on servers and scalability, was an early example of MSR's hybrid research-and-advanced-engineering model. Gray, who joined Microsoft in 1995 and won the Turing Award in 1998 for work on transaction processing, set himself the goal of putting "all the world's scientific data online," a vision that anticipated cloud-scale data services by roughly a decade[12].
After Rashid stepped back in 2013, leadership of MSR rotated through a small group of senior figures. Jeannette Wing, the CMU professor who had coined the term "computational thinking," served as Corporate Vice President of MSR from 2013 to 2017 before moving to Columbia University[13]. Peter Lee, a former CMU department head and former director at DARPA, led Microsoft Research NExT (New Experiences and Technologies) from 2015 to 2020 and was promoted to lead the worldwide MSR organization, while overseeing Microsoft Health Futures and the lab's deep AI and biomedical engagements[14][15]. In November 2024, Microsoft created a new role for Lee, President of Microsoft Science, with a focus on AI-enabled virtual patients, populations, and labs, and on biomedical research. Igor Tsyganskiy took over operational leadership of the global Microsoft Research laboratory network as part of the same transition[15].
Eric Horvitz, who joined MSR in 1993 and previously directed MSR's Redmond lab, was named the company's first Chief Scientific Officer in March 2020, a role created to coordinate research direction across the company's product divisions[16]. Jaime Teevan, an MIT-trained information retrieval researcher who invented the first personalized search algorithm shipped in Bing, became Chief Scientist of Microsoft in 2021 and led the early integration of GPT-4 into Microsoft 365 Copilot starting in late 2022[17].
MSR runs research labs in multiple cities, with each lab having its own focus areas while collaborating across the network. Redmond remains the headquarters and the largest single lab. The network spans North America, Europe, and Asia, and continued to grow through 2024-2025 with new labs in Tokyo and Singapore.
| Lab | Year founded | Location | Notes |
|---|---|---|---|
| Microsoft Research Redmond | 1991 | Redmond, Washington, US | Headquarters; about 350 researchers; led by Donald Kossmann[1][18] |
| Microsoft Research San Francisco / Silicon Valley | 1995 | Mountain View, California, US | Founded by Gordon Bell and Jim Gray; closed in 2014 with most researchers moving to Redmond or NYC[12] |
| Microsoft Research Cambridge | July 1997 | Cambridge, United Kingdom | First MSR lab outside the US; founded by Roger Needham; over 100 researchers; long tied to the University of Cambridge[19][20] |
| Microsoft Research Asia (MSRA) | November 1998 | Beijing, China | Founded by Kai-Fu Lee; roughly 300 researchers and 300 visiting scientists/students; expansions to Shanghai, Vancouver, Tokyo, Singapore, Hong Kong[21][22][23] |
| Microsoft Research India | January 2005 | Bangalore, India | Founded by P. Anandan; initial focus on technology for emerging markets, multilingual systems, GIS, and sensor networks[24][25] |
| Microsoft Quantum (Station Q) | 2006 | Santa Barbara, California, US | Founded by Fields-medalist mathematician Michael Freedman; topological quantum computing on the UC Santa Barbara campus[26] |
| Microsoft Research New England | July 2008 | Cambridge, Massachusetts, US | Adjacent to MIT; established by Jennifer Chayes and Christian Borgs; theoretical machine learning, algorithmic economics, computational biology[27] |
| Gray Systems Lab | 2008 | Madison, Wisconsin, US | Named after database pioneer Jim Gray; partners with the University of Wisconsin-Madison on database systems[12] |
| Microsoft Research New York City | May 2012 | New York, New York, US | Founded with multiple researchers from the closed Yahoo Research East; computational social science, algorithmic economics, machine learning[28][29] |
| Microsoft Research Montreal | January 2017 | Montreal, Quebec, Canada | Came with the acquisition of Maluuba; deep learning and reinforcement learning for language understanding; Yoshua Bengio advisor[30][31] |
| MSR Asia Vancouver | 2023 | Vancouver, British Columbia, Canada | New lab aligned with MSR Asia, partly staffed by researchers relocated from Beijing and Shanghai during the "Vancouver Plan"[32][33] |
| Microsoft Research Asia Tokyo | November 18, 2024 | Tokyo, Japan | First MSR lab in Japan; embodied AI, well-being and neuroscience, societal AI; led by Yasuyuki Matsushita[34] |
| Microsoft Research Asia Singapore | July 24, 2025 | Singapore | First MSR lab in Southeast Asia; industrial AI for healthcare, finance, logistics, and embodied/spatial AI; talent development with NUS, NTU, and SMU[35][81] |
Each lab tends to have a distinctive culture. MSR Cambridge is known for programming languages and probabilistic programming, MSR Asia for computer vision and search, MSR New England for theoretical machine learning and economics, MSR New York for online experimentation and reinforcement learning, MSR India for systems and emerging-market computing, and MSR Redmond for systems and the broadest cross-section of disciplines[1][22].
When MSR opened the Singapore lab on July 24, 2025, its first in Southeast Asia, President of Microsoft Research Peter Lee framed the expansion this way: "Singapore plays a critical role in our global research strategy. Its leadership in science, engineering and digital governance, along with its strong culture of public-private-academic collaboration, makes it an ideal environment for advancing inclusive and impactful AI."[81]
MSR Asia in particular has been at the center of geopolitical attention. In June 2023, the Financial Times reported that Microsoft was working on a "Vancouver Plan" to relocate dozens of senior Chinese AI researchers from Beijing to Canada in response to U.S. export controls and concerns about technology transfer[32]. Microsoft publicly disputed some of the specifics, with MSR head Peter Lee stating that "there is no discussion or advocacy to relocate or close MSRA's locations in China"[33]. The company nonetheless asked a portion of the China-based AI staff (with subsequent reporting describing the affected group as hundreds rather than the originally reported dozens) to consider relocating to Canada, Australia, or Microsoft's Redmond headquarters, and stood up the new MSR Asia Vancouver lab in 2023 to receive some of them[36][37]. As of 2024-2025, MSRA still operates in Beijing while having expanded its presence in Vancouver, Shanghai, Hong Kong, Singapore, and Tokyo[22][34].
Beyond geopolitics, MSR Asia is widely regarded as the most influential industrial AI lab ever to operate in China. Founded in November 1998 with an initial five employees and a plan for ~100 researchers and ~$80 million invested over six years, the lab grew to several hundred researchers and an extended network of interns and visiting students[21][22]. Estimates put the number of MSRA "alumni" (former employees, interns, and visiting researchers) at 5,000 to 7,000 globally, with a heavy concentration in Chinese industry and academia[38]. Notable alumni include Harry Shum (later Microsoft Executive Vice President and chairman of MoonshotAI), Ya-Qin Zhang (later president of Baidu), Wang Jian (founder of Alibaba Cloud), Lin Bin (co-founder and president of Xiaomi), Zhang Hongjiang (former Kingsoft CEO), and computer vision pioneers Sun Jian (later chief scientist at MEGVII), Tang Xiaoou (founder of SenseTime), and Kaiming He (later FAIR, Google DeepMind, and MIT). Hsiao-Wuen Hon led MSRA as Managing Director from 2007 until the 2024 transitions and Corporate VP of the Asia-Pacific R&D Group[39].
MSR's research portfolio covers most of computer science. The largest groupings include:
MSR researchers regularly publish at NeurIPS, ICML, CVPR, SIGGRAPH, STOC, FOCS, SIGCOMM, OSDI, POPL, and PLDI. The lab is one of the most prolific industrial publishers in computer science, and its researchers have won numerous awards including the Turing Award (Tony Hoare and Butler Lampson, who joined MSR after their winning work, plus Jim Gray and Leslie Lamport, who received the Turing while at the lab), a Fields Medal (Michael Freedman), the Godel Prize, and ACM SIGCOMM Test of Time Awards[48].
MSR has produced or co-produced a number of widely cited results in artificial intelligence.
Kinect was the depth-sensing peripheral for Xbox 360, launched in November 2010. The body-tracking pipeline that turned raw depth data into a real-time skeleton was developed by MSR Cambridge researchers, including Jamie Shotton and colleagues. Their algorithm used randomized decision forests trained on over one million synthetic depth images. The paper "Real-Time Human Pose Recognition in Parts from Single Depth Images" won the Best Paper award at CVPR 2011 and is one of the most cited industrial computer vision papers of its era. A GPU implementation processed 30 frames per second using roughly 10 percent of the Xbox 360's hardware resources[45].
The learning-to-rank algorithms used in Bing for web ranking grew out of a long collaboration between MSR's Chris Burges and the Bing relevance team starting in 2004. RankNet, presented at ICML 2005 and later receiving the ICML Test of Time award, recast ranking as a gradient-descent learning problem. LambdaRank and LambdaMART followed, the latter winning Track 1 of the 2010 Yahoo Learning to Rank Challenge, and a LambdaMART variant remains a workhorse of search ranking inside Bing[49].
Project Adam was a distributed deep learning training system developed by Trishul Chilimbi, Yutaka Suzue, Johnson Apacible, and Karthik Kalyanaraman at MSR. Presented at OSDI 2014, Adam scaled stochastic gradient descent across thousands of cores and achieved roughly twice the accuracy of comparable systems on ImageNet at much lower cost[50].
Project Catapult, led by Doug Burger and James Larus at MSR Redmond with engineers including Andrew Putnam, introduced field-programmable gate array (FPGA) acceleration into Microsoft's data centers. FPGA-equipped servers were deployed at scale in Bing data centers in 2014, accelerating search ranking with a roughly 50 percent throughput improvement or 25 percent latency reduction[40][51]. Project Catapult subsequently expanded to Azure as part of the SmartNIC platform, and in 2017-2018 was repurposed to power Project Brainwave, a system for low-latency deep learning inference on FPGAs that targeted real-time AI workloads[40].
The single best-known paper to come out of MSR Asia is "Deep Residual Learning for Image Recognition" by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, posted to arXiv on December 10, 2015[3]. ResNet introduced residual connections, also called skip connections, which let very deep convolutional neural networks be trained without exploding or vanishing gradients. An ensemble of ResNet-152 models won the ILSVRC 2015 ImageNet classification challenge with a top-5 error of 3.57 percent, took first place across five separate ImageNet and COCO tracks (classification, detection, localization, plus COCO detection and segmentation), and went on to become one of the most cited papers in all of science, with well over 200,000 citations on Google Scholar[3][52]. The ResNet architecture is foundational to virtually every modern computer vision and Transformer-based model.
DeepSpeed is an open-source deep learning optimization library released by Microsoft in February 2020 and led by Yuxiong He's team at MSR. It includes ZeRO (Zero Redundancy Optimizer), introduced in the SC20 paper "ZeRO: Memory Optimizations Toward Training Trillion Parameter Models" by Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He, which removes redundant memory copies of optimizer state, gradients, and parameters across data-parallel processes[53][54]. Together with the ZeRO-Infinity and ZeRO-Offload extensions, DeepSpeed enabled training of language models with hundreds of billions of parameters on commodity GPU clusters and is widely used outside Microsoft, including in many open-source training pipelines.
Turing-NLG was a 17-billion-parameter Transformer language model produced by the Microsoft Project Turing team, trained with DeepSpeed and ZeRO. At the time it was the largest published Transformer language model, with 78 layers, hidden size 4256, and 28 attention heads. Turing-NLG improved state-of-the-art results on benchmarks such as LAMBADA and WikiText. Less than four months later OpenAI revealed GPT-3 at 175 billion parameters, an order of magnitude larger[55].
Project Florence is a vision foundation model effort at MSR. The original Florence model, described in a 2021 paper, expanded vision representations across coarse-to-fine and image-to-video tasks. Florence-2, released in 2024, is a unified prompt-based vision model that handles captioning, detection, segmentation, and visual question answering in a single architecture, trained on the FLD-5B dataset of 5.4 billion annotations across 126 million images[56][57].
The Phi project, led for several years by Sebastien Bubeck and colleagues, set out to test how much capability could be packed into very small models trained on carefully curated, often synthetic data. Phi-1 (1.3 billion parameters, June 2023) targeted code; Phi-1.5 and Phi-2 (2.7 billion parameters, late 2023) extended the approach to general language. Phi-3-mini (3.8 billion parameters) launched in April 2024 with 4K and 128K-token context variants, trained on 3.3 trillion tokens, and was positioned as an on-device alternative to much larger frontier models[58][59]. Phi-3.5 (August 2024) included three releases: Phi-3.5-mini-instruct (3.8B parameters), Phi-3.5-MoE-instruct (16x3.8B mixture-of-experts with 6.6B active parameters), and Phi-3.5-vision-instruct (4.2B), all with 128K-token contexts under MIT licenses[60]. Phi-4, released in December 2024 with 14 billion parameters and trained on roughly nine trillion tokens for 21 days, focused on advanced reasoning and relied heavily on synthetic data generated through multi-agent prompting and self-revision workflows[61].
In early 2025 Microsoft introduced Phi-4-mini (3.8B), Phi-4-multimodal (5.6B with vision and speech encoders), and the Phi-4-reasoning family (3.8B, 14B, and 14B-plus variants) trained in part on demonstrations from OpenAI's o3-mini[62][63]. Microsoft reported that the 14-billion-parameter Phi-4-reasoning and Phi-4-reasoning-plus "achieve performance comparable to the full DeepSeek-R1 model (with 671B parameters)" on AIME 2025, the American Invitational Mathematics Examination question set, and outperform OpenAI's o1-mini and DeepSeek-R1-Distill-Llama-70B on most benchmarks despite being a fraction of the size[63][83]. The Phi line is Microsoft's central demonstration that careful data curation can let small models rival much larger ones at reasoning.
In February 2024, researchers from Microsoft Research and the University of Chinese Academy of Sciences published "The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits"[64]. The paper introduces BitNet b1.58, a Transformer LLM whose weights are ternary {-1, 0, 1}, encoding log2(3) approximately 1.58 bits per parameter. At 3 billion parameters, BitNet b1.58 matched FP16 LLaMA in perplexity and zero-shot accuracy while using roughly 3.55x less GPU memory and running 2.71x faster. The work has since spawned bitnet.cpp, a runtime targeted at efficient inference on commodity CPUs[64].
"Sparks of Artificial General Intelligence: Early experiments with GPT-4," posted to arXiv on March 22, 2023, was authored by Sebastien Bubeck and 13 other MSR researchers including Eric Horvitz, Peter Lee, Ece Kamar, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Marco Tulio Ribeiro, and Yi Zhang[4]. The paper documented Microsoft's months of pre-release access to an early version of GPT-4 and concluded: "Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system."[4] The paper attracted both wide attention and methodological criticism, with Gary Marcus and others arguing that the AGI framing was untestable and that the authors' definition of AGI was idiosyncratic[65]. A separate critique titled "GPT-4 Can't Reason" by Konstantine Arkoudas pushed back specifically on the paper's claims about reasoning[66]. Despite the criticism, the paper shaped much of the public conversation about frontier model capabilities in 2023.
GraphRAG is an MSR-developed retrieval-augmented generation system that builds a knowledge graph from unstructured text using LLMs, applies community detection algorithms to cluster entities into thematic groups, generates LLM summaries of each community, and uses those summaries to answer both entity-level and corpus-level questions. Introduced in a Microsoft Research blog post in early 2024 and open-sourced on GitHub in July 2024, GraphRAG has been adopted as a reference architecture for combining LLMs with knowledge graphs and reports improved coverage and diversity on global queries over million-token corpora compared with conventional RAG[67].
AutoGen is an open-source Python framework for building multi-agent LLM applications, released by MSR in fall 2023 as a collaboration with researchers at Penn State and the University of Washington. By May 2024 the project had attracted more than 290 community contributors and 890,000 monthly downloads; a major v0.4 redesign in 2025 reorganized the framework around a layered, asynchronous, event-driven architecture for production-scale agentic AI[68].
Aurora is an MSR foundation model for the Earth system. The first version, described in a 2024 Nature submission, was trained on more than one million hours of geophysical data and outperformed operational forecasts in tasks including air quality, ocean waves, tropical cyclone tracks, and high-resolution weather forecasting; subsequent versions have been released openly for academic and operational use[69].
ONNX, the Open Neural Network Exchange, was announced jointly by Microsoft and Facebook in September 2017 as an open format for representing neural networks across frameworks. It became a de facto standard for model interoperability and is now stewarded under the Linux Foundation[70].
MSR's leadership has rotated through a small group of senior figures over the decades.
| Person | Role | Background |
|---|---|---|
| Bill Gates | Co-founder of MSR; Technology Adviser to Microsoft CEO since 2014 | Microsoft co-founder; stepped down from the board in 2020 but continues as adviser[71] |
| Nathan Myhrvold | Co-founder of MSR; former Chief Technology Officer | Wrote the 1991 memo proposing MSR; left Microsoft in 1999 to found Intellectual Ventures[7] |
| Rick Rashid | Founding head of MSR, 1991 to 2013 | Former CMU professor; Mach kernel architect[9] |
| Jeannette Wing | Corporate VP of MSR, 2013 to 2017 | CMU professor; coined "computational thinking"; later EVP for Research at Columbia[13] |
| Peter Lee | President of Microsoft Science (since 2024); previously President of Microsoft Research (2013-2024) | PhD Michigan; former CMU department head; former DARPA director; co-author of The AI Revolution in Medicine[14][15] |
| Igor Tsyganskiy | Microsoft Research operational lead (since November 2024) | Took on responsibility for the MSR global laboratory network during Peter Lee's transition to President of Microsoft Science[15] |
| Eric Horvitz | Chief Scientific Officer of Microsoft (since March 2020) | MD, PhD Stanford; previously Director of MSR Redmond; AAAI fellow; founder of the One Hundred Year Study on AI at Stanford[16] |
| Jaime Teevan | Chief Scientist of Microsoft (since 2021) | PhD MIT; over 280 publications and 66 patents; led the M365 Copilot effort from late 2022; ACM Fellow; Time 100 AI in 2023[17] |
| Christopher Bishop | Technical Fellow; Director of Microsoft Research AI for Science (since 2022) | Author of Pattern Recognition and Machine Learning; Lab Director of MSR Cambridge 2015-2022[47] |
| Donald Kossmann | Director of MSR Redmond | Distinguished Scientist; formerly ETH Zurich professor; ACM SIGMOD chair 2013-2017[18] |
| Kai-Fu Lee | Founding director of MSR Asia, 1998 to 2000 | PhD CMU; later President of Google China and founder of Sinovation Ventures[21] |
| Harry Shum | Managing Director of MSRA, 2004-2007; Executive VP at Microsoft | One of the founding members of MSR China; later led Bing product development[39] |
| Hsiao-Wuen Hon | Managing Director of MSRA, 2007 to 2024; Corporate VP of Asia-Pacific R&D | Joined Microsoft in 1995 and MSRA in 2004 as deputy MD[39] |
| Yasuyuki Matsushita | Head of Microsoft Research Asia Tokyo (since 2024) | Returned to Microsoft from Osaka University, where he had been a professor since 2015; earlier MSRA tenure 2003-2015[34] |
| Roger Needham | Founding director of MSR Cambridge, 1997 to 2003 | Cambridge computer lab head; pioneered Needham-Schroeder protocol; died in 2003[19] |
| Andrew Herbert | Managing Director of MSR Cambridge, 2003-2011 | Computer scientist who continued Needham's work expanding the Cambridge lab[20] |
| Jennifer Chayes | Founding Managing Director of MSR New England, 2008-2020 | Mathematician and physicist; co-founded MSR's Theory Group in 1997; later Dean of Computing at UC Berkeley[27] |
| P. Anandan | Founding Managing Director of MSR India, 2005-2014 | UMass Amherst PhD; led MSR India through its first decade; computer vision researcher[25] |
| Michael Freedman | Founding director of Station Q, 2006 to 2024 | Fields Medal winner; led topological quantum computing effort at UCSB[26] |
| Mustafa Suleyman | CEO of Microsoft AI (since March 2024) | Co-founder of DeepMind and Inflection AI; leads consumer AI products including Copilot, Bing, and Edge[72] |
| Karen Simonyan | Chief Scientist of Microsoft AI (since March 2024) | Co-founder of Inflection AI; led work on AlphaZero at DeepMind; later chief scientist of Microsoft's MAI Superintelligence team (announced November 2025)[72][73] |
| Sebastien Bubeck | Distinguished Scientist and VP of GenAI Research, MSR (2014 to October 2024) | Theoretical machine learning; lead author of "Sparks of AGI"; led Phi small-model effort; left for OpenAI in October 2024[74] |
The lab has also drawn well-known researchers across different periods: Susan Dumais (information retrieval, managing director of MSR labs in New England, NYC, and Montreal); Yoshua Bengio (advisor through the Maluuba acquisition); Don Syme (creator of F#); Anders Hejlsberg (creator of TypeScript and C#, working in adjacent Microsoft groups); and Tony Hoare and Butler Lampson, both Turing Award winners who later joined MSR Cambridge and MSR Silicon Valley/Redmond respectively[31][41][44][48].
MSR's role in Microsoft's overall AI strategy shifted dramatically with the rise of OpenAI. In July 2019, Microsoft announced a $1 billion investment in OpenAI and made Azure its exclusive cloud provider. In January 2023, Microsoft confirmed the third phase of the partnership, a multi-year, multi-billion-dollar investment widely reported at $10 billion, on top of additional rounds in 2021 and 2022. Total Microsoft commitments to OpenAI reached or exceeded $13 billion by most accounts[75][76]. The deal gave Microsoft access to OpenAI's frontier models and the right to integrate them into products such as Bing, GitHub Copilot, Microsoft 365 Copilot, and Windows.
For MSR, this raised a strategic question that the lab has had to navigate publicly: how should an in-house research organization position itself when the dominant AI partner sits outside the company. Peter Lee and others have argued that MSR retains an essential role in producing science that informs how Microsoft uses OpenAI models, in studying their behavior, and in pursuing directions OpenAI does not work on, including small models, safety research, scientific computing, and quantum.
In March 2024, Microsoft announced a separate Microsoft AI division, distinct from MSR, led by Mustafa Suleyman as Executive Vice President and CEO of Microsoft AI, reporting directly to Satya Nadella. Suleyman, a co-founder of DeepMind and of Inflection AI, brought along Inflection's co-founder Karen Simonyan as Chief Scientist of Microsoft AI, plus much of the Inflection technical team in what was widely described as a quasi-acquihire[72][77]. The new unit took over Copilot, Bing, Edge, and the consumer GenAI team. MSR remained organizationally separate, though the two divisions cooperate. The split increased pressure on MSR to focus on long-horizon research while Microsoft AI handled consumer-facing AI productization.
The transition came with high-profile departures. Sebastien Bubeck, who had led much of MSR's small-model and theoretical AI work, left for OpenAI in October 2024 to work on AGI research[74]. Most of his Phi team remained at Microsoft.
Microsoft AI subsequently began building its own foundation models. In August 2025 the company unveiled MAI-1-preview, a mixture-of-experts text model trained on roughly 15,000 NVIDIA H100 GPUs, and MAI-Voice-1, a high-throughput speech generation model that can produce a minute of audio in under one second on a single GPU[78][79]. MAI-Image-1, a text-to-image model, followed in late 2025 and was integrated into Bing Image Creator and Copilot[79]. In November 2025 Microsoft and OpenAI renegotiated their original 2019 partnership agreement, removing the contractual clause that had prevented Microsoft from independently pursuing artificial general intelligence. On the same day, November 6, 2025, Microsoft announced a new MAI Superintelligence Team led by Suleyman with Simonyan as chief scientist, oriented toward what Suleyman called "humanist superintelligence," which he described as "incredibly advanced AI capabilities that always work for, in service of, people and humanity more generally," adding: "We are not building an ill-defined and ethereal superintelligence; we are building a practical technology explicitly designed only to serve humanity."[73] The team's initial focus areas include medical diagnosis, digital companions, and renewable energy[73].
MSR has long had a publication culture closer to a university than a typical product organization. Researchers are encouraged to publish in top conferences and journals, and many work part-time as adjunct or visiting faculty at universities including Cambridge, MIT, Carnegie Mellon, and Tsinghua. The lab maintains a public list of publications, and most papers are available without paywall on the MSR site[2].
MSR has also been a steady contributor to open source. Notable open-source releases that grew out of MSR or related Microsoft research groups include:
Research from MSR has consistently found its way into Microsoft products. The SQL Server query optimizer relies on cost-based optimization techniques drawn from database research at MSR. Bing's ranking is driven by descendants of LambdaMART[49]. Office's grammar and proofing tools, the Visual Studio IntelliCode code completion service, the HoloLens spatial mapping pipeline, and many features of Microsoft 365 Copilot all have direct lineage in MSR work[10][17][49].