FOURWEEKMBA EXPLAINER · FROM THE NEWSROOM · UPDATED 16 NOVEMBER 2026
The pattern in one paragraph
A research group is organised around questions: what works, what is true, what is next. A model unit is organised around a release: one model family, one compute budget, one ship date. Between April 2023 and September 2026, Google, Meta and Microsoft each moved a named research group from the first kind of organisation into the second. None of the companies called this closing research. Each described the new unit by the models and products it builds.
Three cases, in the companies’ own words
- Google, 20 April 2023. Google announced that “Google DeepMind, will bring together two leading research groups in the AI field: the Brain team from Google Research, and DeepMind.” Demis Hassabis would lead “research that will help power the next generation of our products and services”. Jeff Dean became Google’s Chief Scientist. Google Research kept its fundamental work in areas such as algorithms, privacy, quantum computing and health. On 6 December 2023, Google DeepMind introduced Gemini 1.0, “the first realization of the vision we had when we formed Google DeepMind earlier this year”.
- Microsoft, 19 March 2024, then September 2026. Satya Nadella announced a new organisation, Microsoft AI, “focused on advancing Copilot and our other consumer AI products and research”, led by Mustafa Suleyman, with the Copilot, Bing and Edge teams reporting to it. On 28 August 2025, Microsoft AI previewed MAI 1, “MAI’s first foundation model trained end-to-end”, trained on about 15,000 NVIDIA H100 GPUs. In early September 2026, according to John Langford, Microsoft Research’s AI Frontiers group was “reorged into Microsoft AI”.
- Meta, 30 June 2025. Mark Zuckerberg’s staff memo, published by CNBC, created Meta Superintelligence Labs: “This includes all of our foundations, product, and FAIR teams, as well as a new lab focused on developing the next generation of our models.” On 8 April 2026, Meta announced Muse Spark, “the first in a new series of large language models built by Meta Superintelligence Labs”, saying the unit had “rebuilt our AI stack from the ground up” over nine months.
Why research groups get pulled in
The companies’ statements and Langford’s post give four reasons. Each is stated by one of the sources below; the grouping is ours.
Compute goes to one team
Google said that combining the talent “into one focused team, backed by the computational resources of Google, will significantly accelerate our progress in AI.” Langford called compute “an often-severe bottleneck” inside AI Frontiers.
Scale is easier to budget
Langford described Microsoft AI’s core philosophy as “scaling, and aggressively not research”, and granted its logic: “returns to scale are far more predictable than returns to research”.
Products need their own models
“MAI requires purpose-built models,” Microsoft AI wrote when it previewed MAI 1. Nadella framed the 2024 unit as “science, engineering, product, and design coming together”.
Duplicate pipelines cost money
Langford wrote he knew of “at least 4 efforts” inside Microsoft Research trying to build model pipelines, and that putting two AI organisations together “creates opportunities for better concentration of effort”.
What research loses, in one researcher’s account
Only one of the people involved has written publicly about what the move costs. Langford, who helped organise AI Frontiers in 2023, wrote that the merger “could make sense” on the surface, but that the philosophy of Microsoft AI’s leadership was “incompatible” with research. His cost case: ByteDance experiments suggest a Full Bandwidth Transformer variant “might be comparable to a model of twice the size”, which would cut training and inference costs. That is his reading of another lab’s results; we have not checked it. He left, and wrote that he would not go back to Microsoft Research either:
“I am just not comfortable being at a company where the research I do has no path to real use.”
Langford also wrote that “there will not be internships with the folks there in the coming year”. Microsoft has not responded to his post in any statement we found.
What this means for companies building AI
- Follow the compute. In all three cases the research group moved to the unit that trains the company’s main models. Where the compute budget sits is a good guide to where research will report.
- Expect a model clock. The three units announced their first flagship model 7.6, 9.3 and 17.3 months after they were formed. A research group inside such a unit works to that kind of timetable.
- Decide where long bets live. Google kept a separate Google Research for fundamental work. Microsoft’s case shows the risk when a research group is folded into a unit built for scale.
- Price the people. Langford wrote that the other founding AI Frontiers leaders “seem to” have reached the same conclusion he did. Talent is the asset research groups hold.
What is not established
Three cases are not a law. We count only research groups named in primary statements; other reorganisations may exist. The companies did not describe these moves as ending research: Google kept Google Research running, and Meta’s memo describes FAIR as part of the new unit. Langford’s account is one founder’s view, and the “early September” date is his. The months between dates are our arithmetic. “First flagship model” follows each company’s own description; Google DeepMind and Microsoft AI released other models and tools too. This page makes no forecast.
FAQ
Why do AI research labs end up inside product or model units?
Because the companies put compute, staff and deadlines in the unit that builds their main models. Google, Meta and Microsoft each moved a named research group into such a unit between 2023 and 2026.
What happened to Google Brain?
Google combined the Brain team from Google Research with DeepMind to form Google DeepMind, announced on 20 April 2023. Google DeepMind introduced Gemini 1.0 on 6 December 2023.
Is Meta’s FAIR still a separate lab?
Zuckerberg’s 30 June 2025 memo placed “foundations, product, and FAIR teams” inside Meta Superintelligence Labs, so FAIR is part of that unit.
What was Microsoft’s AI Frontiers?
A group of about 50 people inside Microsoft Research, organised in 2023 after the ChatGPT release, according to John Langford. It was moved into Microsoft AI in early September 2026, and Langford wrote on 6 November 2026 that “AI Frontiers is over”.
Does this mean Big Tech is giving up on research?
No company said so. Google kept Google Research for fundamental work. The shift is in who controls the AI research closest to the main models: the unit that ships them.
Related
Sources: Google, “Google DeepMind: Bringing together two world-class AI teams”, 20 April 2023; Google, “Introducing Gemini: Google’s most capable AI model yet”, Demis Hassabis, 6 December 2023; Satya Nadella, Official Microsoft Blog, 19 March 2024; Microsoft AI, “Two in-house models in support of our mission”, 28 August 2025; Mark Zuckerberg staff memo, 30 June 2025, as published by CNBC; Meta Newsroom, “Introducing Muse Spark”, 8 April 2026; John Langford, “The End of AI Frontiers”, hunch.net, 6 November 2026. Read 16 November 2026. Quotes are verbatim. Month counts and the grouping of reasons are ours.









