Microsoft AI: MAI Models, Copilot and Azure Infrastructure

Microsoft develops MAI models and Maia chips alongside Azure and Copilot. MAI-Base-2’s October 2026 pre-training report was not a public weight release.

kindCompany
foundedApr 4, 1975

Latest

Updated

A model developer as well as a cloud provider

Microsoft builds developer tools, enterprise software and Azure, and distributes AI through Copilot and Foundry. Its own anniversary account dates the company's start to April 4, 1975. The same product portfolio gives it places to train, serve and use models, alongside its investments in other AI companies.

Microsoft AI's October 15, 2026 technical account described completed pre-training of MAI-Base-2, with 2.8 trillion total parameters and 80 billion active. Microsoft also identified MAI-Thinking-1 as a model post-trained on its earlier MAI-Base-1. These are its own model-development efforts, not merely the resale of OpenAI access.

MAI-Base-2 was still undergoing post-training in that account. A proposed route to license weights to trusted Foundry partners was future work, rather than an unrestricted weight release. Microsoft's held-out comparisons and efficiency estimates were company-run evaluations.

Infrastructure and outside models

Microsoft’s July 2019 partnership with OpenAI paired a $1 billion investment with work on Azure supercomputing and an exclusive cloud arrangement. It was an important earlier infrastructure and financing contribution, separate from Microsoft’s later MAI model development and from the partnership terms subsequently revised.

Microsoft's Maia program connects those models to custom inference hardware. Its October 2026 announcement said Maia 200 was serving production workloads, while Maia 300 and multi-gigawatt expansion were planned for 2027. The distinction separates deployed systems from the announced next generation.

Microsoft also continues to supply outside models. Its April 2026 OpenAI agreement retained a non-exclusive license through 2032 and a primary-cloud relationship while allowing broader cloud distribution. The separate Anthropic partnership brought Claude to Azure. Investment, hosting and licensing establish commercial relationships; they do not make those developers Microsoft subsidiaries.

The earlier Tay chatbot incident remains part of this history. Microsoft's March 2016 account describes the abusive interactions that caused it to take Tay offline and the changes it intended to make, an example of deployment behavior that model capability alone did not settle.

Sources

  1. [1]
    Celebrating Microsoft’s 50 years

    Microsoft · Apr 4, 2025

  2. [2]
    Frontier Pre-training at Microsoft AI

    Microsoft AI · Oct 15, 2026

  3. [3]
  4. [4]
  5. [5]
  6. [6]
  7. [7]
    Learning from Tay’s introduction

    Microsoft · Mar 25, 2016

Microsoft history

2026

│
Forever publishes EMBRACE findings on embedded AI browsers

With Forever Security, Anysphere

Forever Security published EMBRACE research on embedded browsers in AI assistants, covering prompt injection, address spoofing and local-file exposure. It reported fixes for Cursor and Copilot, with Codex and VS Code findings unresolved at publication.

Forever Security
Research
│
Microsoft announces Maia 300, its third-generation inference chip

Due in production deployments in 2027, with compute chiplets on TSMC's N2P process and eight HBM4 stacks, and pitched at 40 percent more tokens per megawatt than unnamed leading GPU systems. Microsoft said its Maia family would scale to multiple gigawatts and become the main inference system for its own products.

Microsoft
Hardware
│
More than a hundred companies sign a call for collective action on cyber defense

With OpenAI, Anthropic, Google, Amazon, Cloudflare, Hugging Face, Oracle, Perplexity

OpenAI hosted the letter and Anthropic, Google, Microsoft, AWS, Cloudflare, CrowdStrike, Visa and Mastercard signed it, the day after the Hugging Face postmortem. The premise is a window: models now help defenders faster than attackers, and the signatories expect it to close within months.

OpenAI
Culture
│
Microsoft previews Frontier Tuning for enterprise AI agents

A private preview shown at Build 2026: reinforcement learning applied inside a customer's own compliance boundary, on their own data and workflows, to produce tuned models and a runtime harness. Microsoft's bet is that the most valuable AI in an organization is the one that already knows how it works.

Microsoft
Product
│
Google, Microsoft and xAI hand models to the government before release

With Center for AI Standards and Innovation, Google DeepMind, xAI

The Center for AI Standards and Innovation signed pre-deployment evaluation agreements with three more labs, extending the arrangement OpenAI and Anthropic made in 2024. Five weeks later the same government stopped a model rollout outright.

CNBC
Policy
■
OpenAI's models reach AWS as Microsoft exclusivity ends

With OpenAI, Amazon

The cloud exclusivity clause is gone, replaced by an Azure-first commitment, while Microsoft keeps a non-exclusive license through 2032 and its revenue share until 2030 under a cap. OpenAI had outgrown a single cloud.

CNBC
Deal
│
GitHub Copilot moves to usage-based billing

With GitHub

From June 1, premium requests give way to a monthly allowance of credits metered on tokens, with subscription prices unchanged. GitHub said it had been absorbing the cost of agents that run for hours on a single question.

GitHub
Product
■
Microsoft ships Copilot Cowork and puts Claude in mainline Copilot

With Anthropic

Anthropic's agent technology arrived inside Microsoft 365 as Copilot Cowork, for work that runs over hours rather than a single turn, and Claude appeared in ordinary Copilot chat alongside OpenAI's models. The company that bankrolled OpenAI now picks a model per task.

Microsoft
Product
■
Microsoft unveils its second-generation Maia AI chip

Maia 200 packs more than 100 billion transistors and roughly 10 petaflops of 4-bit compute, with Microsoft claiming three times the performance of Amazon's Trainium on some benchmarks. The chip is already handling GPT-5.2 inference and Microsoft 365 Copilot workloads in US data centers.

Microsoft
Hardware

2025

●
OpenAI's for-profit becomes a public benefit corporation

With OpenAI

The recapitalization left the nonprofit foundation holding about 26 percent of OpenAI Group PBC and Microsoft about 27 percent, after a year of negotiation with two state attorneys general. The structure that produced the 2023 board crisis was finally rewritten.

OpenAI
Company
│
Nscale announces an expanded Microsoft GPU contract

With Nscale, Aker ASA, Dell Technologies

Nscale announced a signed rollout covering approximately 200,000 GB300 GPUs across four countries, delivered through its operations and Aker joint venture with Dell collaboration. Capacity would arrive in geographically phased deployments.

Nscale
Deal
■
Microsoft adds Anthropic's Claude models to Copilot

With Anthropic

Copilot business users can now choose Claude Opus 4.1 or Claude Sonnet 4 alongside OpenAI's models for research and building custom agents. It follows an earlier deal bringing Anthropic's models into Office, another step in Microsoft's move beyond OpenAI exclusivity.

TechCrunch
Deal
■
OpenAI and Microsoft agree on terms for a for-profit restructuring

With OpenAI

A nonbinding memorandum sets out OpenAI's for-profit arm becoming a public benefit corporation, with the nonprofit holding a stake later valued above $100 billion and Microsoft keeping preferred access to its models and cloud. A definitive agreement was still being finalized.

TechCrunch
Deal
│
Microsoft and OpenAI investigate DeepSeek over possible distillation

With OpenAI, DeepSeek

Microsoft's security team flagged accounts linked to DeepSeek pulling unusually large volumes of data through OpenAI's API in late 2024, and the companies opened a joint investigation into whether the outputs were used to train a rival model. OpenAI's terms bar using its outputs to build competing systems.

Tom's Hardware
Culture

2024

│
Germany rules Microsoft's Inflection AI hire-out counts as a merger

With Inflection AI

The Bundeskartellamt found that Microsoft's move to hire almost all of Inflection AI's staff and license its technology met the definition of a merger under German law, then closed its review anyway because Inflection had no substantial German operations. The decision put AI acqui-hires on notice that they cannot automatically dodge merger review.

Bundeskartellamt
Policy
│
Frontier labs sign the Seoul safety commitments

With OpenAI, Anthropic, Google, Meta, Amazon, Mistral AI, xAI, Cohere, G42, IBM, Inflection AI, Naver, Samsung Electronics, Technology Innovation Institute, Zhipu AI

At the UK and South Korea's Seoul summit, sixteen frontier labs including OpenAI, Anthropic, Google, Meta and Zhipu AI committed to publish safety frameworks with risk thresholds and to hold back models that breach them. Voluntary, and the signatory list kept growing afterwards.

Department for Science, Innovation and Technology
Policy
│
Microsoft introduces Copilot+ PCs, with Recall

A new class of Windows machines built around a 40 TOPS neural processor requirement, Snapdragon chips first. The flagship feature, Recall, photographed everything on screen into a searchable memory and drew a privacy backlash that delayed it for months.

Microsoft
Product
│
Microsoft commits $3.3 billion to a Wisconsin AI datacenter

A Mount Pleasant campus on the site Foxconn once promised to fill, announced with President Biden on stage, plus a manufacturing AI lab at UW-Milwaukee and a pledge to train 100,000 residents. One of the first datacenter announcements framed as regional economic policy rather than cloud capacity.

Microsoft
Hardware
│
Eight Alden Global Capital newspapers sue OpenAI and Microsoft

With Alden Global Capital, OpenAI

The Chicago Tribune, New York Daily News, Denver Post and five sister papers filed in the Southern District of New York, the same court as the Times case, alleging millions of articles were copied to train ChatGPT and Copilot. They asked for damages, an injunction and the destruction of any model built on their work.

Chicago Sun-Times
Policy

2023

●
Llama 2 ships with a commercial license

With Meta

Downloadable models for research and commercial use under a community license, with acceptable-use rules and additional terms for very large services. Access to weights had become a distribution strategy, with conditions attached.

Meta AI
Model

2021

■
LoRA makes fine-tuning a 35MB file instead of a second copy of the model

Freeze the pretrained weights, train a thin pair of matrices beside them, and a task-specific GPT-3 checkpoint drops from 350GB to 35MB. Microsoft's method became the default way to adapt an open-weight model, and once quantization was stacked on top, the reason one consumer GPU could do it.

arXiv
Research

2019

■
Microsoft invests $1 billion in OpenAI

With OpenAI

A billion dollars, an exclusive Azure partnership, and a capped-profit structure OpenAI had invented four months earlier. The compute-for-equity template every frontier lab now runs on starts here.

OpenAI
Funding

2017

│
Microsoft and Facebook announce ONNX model interchange

With ONNX, Meta

Microsoft and Facebook announced the Open Neural Network Exchange, a shared model representation intended to improve interoperability across AI frameworks. This initial ONNX announcement preceded the separate version-one release in December.

Microsoft
Open source

2016

│
Technology companies establish Partnership on AI

With Partnership on AI, Amazon, Meta, Google, DeepMind, IBM

Amazon, Facebook, Google, DeepMind, IBM and Microsoft established a nonprofit forum for AI research and practices. The Partnership on AI was publicly announced in New York on September 28.

DeepMind
Company
■
Microsoft's Tay lasts sixteen hours

Microsoft released a chatbot that learned from whoever talked to it, and Twitter taught it to praise Hitler before the day was out. It is still the standard citation for shipping a model without thinking about who is on the other end.

The Guardian
Culture