Microsoft Build 2026 - Keynote
Satya Nadella
Where do you want to go today? What do you want to build today? Satya Nadella welcomed those to Build talking about the opportunity for developers as well as the broader world, with one key take away, how do we build a frontier intelligence ecosystem about the value developers can build on top of a platform as frontier developers. Looking at the stack which starts with infrastructure with the edge a nd cloud, then there is the emerging layer of models, tools and context with runtime where you deploy the agents and the best tooling to do all of this with the security and governance.
Start at the edge with Windows, the amount of compute at the edge is astounding, if you aggregate this that is a lot of compute power. If can deliver unmetered intelligence to every desk and home, can you do that in this era of AI, such as Outlook Summarise which is using onboard AI locally or Teams Super Resolution or Adobe AfterEffects and Premiere using Windows ML and local processing. Tap into this power to expand scope of Windows ML and Windows AI and expanding Windows AI APIS to more Pcs. Microsoft are announcing two new models on box with a new SLM (small language model) of Aion and a local reasoning and too calling for agentic apps with Aion Plan.
Microsoft are thrilled to see the innovation in hardware with Qualcomm with Snapdragon Elite and this brings things to nVidia RTX Spark with their single SoC for AI capabilities coming together with one of the first devices using this with the Surface Ultra with 128GB of unified memory with an all-day battery life coming later in the Autumn. There are others taking advantage of this platform and brining out exciting machines, but what can be done next and push the architecture for developers to max compute and memory and build a Surface RTX Spark DevBox that is built for more, it is truly a dream machine with a petaflop of AI computer and twenty cores and can join the wait list at microsoft.com/devbox. Why stop there and do one more thing with Windows coming to a PGX workstation running a 1 trillion parameter AI model which was used to build GPT 2.5, with a datacentre on your desktop.
Extending developer endpoints to cloud with Windows 365 now dev maxed for a consistent developer optimised experience with developer optimised Windows with updates including a distraction free dev environment and intelligent terminal with built in GitHub Copilot and Linux love with Coreutils along with seventy plus others. You will be able to personalise your terminal and announcing WSL containers for switching environments and first-class support for containers.
Kayla Cinnamon
Kayla Cinnamon gave a walk though of the tools they have been working on to make Windows a great place for development on Windows with the Surface RTX Spark Dev Box where can be free of distractions. It is possible now to move the Start Menu to the left, top or right and Surface RTX Spark has many dev tools already installed and to get same experience with the Windows Developer Config with WinGet configure. Now you can enable End Task to close applications without needing to open Task Manager and File Explorer is now Git aware with the details of the repository and branch name. Intelligent Terminal has AI agent support and can apply fixes such as RegExes or work with OpenClaw with WSL containers and there is a WSL profile for those who feel comfortable with tools from Linux along with utilities. Surface RTX Spark is designed for running large language models locally with millions of tokens used locally, you can run agents and agents with sub agent tasks to make better use of resources locally, there are also new command line utilities to use for those who love staying in the command line. It was possible to use three models unmetered locally while going through regular developer flows.
Satya Nadella
Satya Nadella talked about this showing the idea of unmetered intelligence and as you move to the cloud where driving equation is tokens per dollar per watt and how to think about systems optimisation end to end with design of data centre with CPUs, storage, GPUs, networking and memory along with offload to local compute. They have a community first AI infrastructure plan to ensure data centres don't increase electricity prices, replenish water, create jobs locally, add to tax base and invest in training and only if they live up to these do they gain the permission to innovate.
Azure has over five hundred datacentres with more expansive out there and have added more in last eighteen months than the first decade of Azure and have build a datacentre for heterogenous workloads. When think about three dominate workloads which is training, inference and agent usage. They can pack hardware densely with more effective usage and how to deliver hundreds of watts of power and minimise losses with new approaches and have a cooling loop with zero water consumption where water usage over a year is what a single restaurant would use.
They have first party and partner silicon including nVidia, AMD and Intel with improvements in better use of tokens per dollar, when it comes to running agents is no longer about AI accelerator and CPU it is important with CPU and GPU for cloud native and agentic workloads with Cobalt delivering 50% better performance with benchmarking with GitHub Copilot traces to show better throughput. They can also look at how data goes around data centres and have to have a coherent system and keep scaling the network architecture and connect to a continent spanning WAN for a fungible AI stack.
Jensen Huang & Satya Nadella
Jensen Huang founder and CEO joined from Taipei who had delivered a keynote over the weekend where Satya mentioned that RTX Spark has delivered a revolution in AI. Jensen mentioned speaking to Satya three years ago about designing a capability and software spec to deliver all the things they are doing with AI, they have build an incredible AI chip which is supported by the software in Windows to have an autonomous agent running on the PC, have went from developing DirectX together to build a computer with autonomous features that can be used by an AI assistant where it can get some coding done or something for an idea they had and it would iterate with them when away from the PC, this idea the PC evolved from a personal computer to personal AI and see this come to life is great.
Jensen mentioned it is possible with RTX Spark to fit a couple of hundred billion parameter model on your computer locally. Talking about the datacentre side and cloud side on how they are pushing this where they build the first AI datacentre together with Microsoft and focus has been on pre-training and then post-training with reasoning models, but these need a large capability from one node to one rack today. They are able to increase token generation rate and reduce token generation cost. They have platforms designed to run agents with the same system as RTX Spark but much larger and be able to process many of these simultaneously where data and memory is encrypted in transit and in use for greater security.
Jensen mentioned that past CPUs were designed for humans, and new ones are designed for Agents, we were more patient, but agents need low latency. Satya mentioned all of this is to power the ecosystem around us and to provide the opportunity to build on the work we do. Broader vision as an opportunity for people to create value, what happened in last few months and have worked together for a decade and a half, but with the convergence of AI models, the commits into GitHub has gone parabolic and it is clear agentic systems are being used and compute demand has gone through the roof and are making sure that tools that agents use are fully accelerated such as Fabric and will make sure that all tools on Azure are fully GPU accelerated as the faster they can iterate and generate tokens to be highly efficient and profitable for developers.
Satya Nadella
Satya mentioned talking about edge and cloud but current form factors, same form factors can offer new functionality with onboard AI functionality with laptop and desktop, but sets up next question if have capability to put this into new form factors, can you purpose build a new platform even for agent era and this is motivation about Project Solara, a new platform for a new agent first world for agent first devices which is ready when you are when you speak it understand you and everything else falls way and when you want it to keep going, it keeps going with a whole ecosystem build to clear the way wherever you are, whenever it matters.
Steven Bathiche
Steven Bathiche started with a why, back at Build 2023 they talked about the outside AI application structure to work across and coordinate across entire workflows, devices and timescales. What if there was a compute for that specific type of workflow and transformational interaction technology with Project Solara, what is the next form factor? It is not just one but creating a system for a constellation of devices working together as one system, whenever and wherever you need them. Two challenges are specific platforms exist which are difficult and expensive to deploy, build and maintain and organisations are already building agents which are constrained on where they exist so Project Solara extends agents onto easy to use form factors to extend where computers don't exist or where they are not optimal enabled by three things which is enterprise readiness with an agent interactive model with a just in time UI and where you can bring your own AI and brought together with AI.
Steven mentioned there are two categories for devices with a device that belongs on your desk from Mediatek to help with the information worker and surface what matters next and help plan or act by delegating tasks and can support hand off between devices including existing Windows PC or connect with your PC or Windows 365. The second device is portable which is an access pass from Qualcomm which is designed for agent interactions and adaptable for on-the-go workflows in a pass form factor, the agent can perform multiple tasks based on interaction. It is all agent-driven with so many opportunities where you can have the right agent for the role or workflow with enterprise secure access where can have documentation or dictation and can use a purpose built wearable to access, gather and act on information.
Steven mentioned that agent first devices are designed to be highly flexible with different input methods for the same software to be adapted to different workflows and verticals with flexible workflows to bring agent workflows where they didn't naturally fit before, think about where agents live and what work they can do and can find out more at aka.ms/ProjectSolara.
Cristiano Amon & Satya Nadella
Cristiano Amon from Qualcomm spoke to Satya about how do you make this an open platform, the job of devices today is to extend the job of the smart phone and make sense to have a vertical platform for mobile where how apps were built but with agents it is an open and horizontal platform to interact with best possible device to build applications, not to be bound to current devices that can carry intelligence wherever you need it.
Satya Nadella
Satya mentioned it is a new set of platform rules that don't pen in the form factor and where your agent lives and can rewrite the rules where new platforms operator and have agents be ubiquitous. Microsoft are building a new intelligence layer starting with model choice from over 11,000 Foundry Models with largest catalogue out there, and considerations about models when building agentic system to have context is becoming super important at data tier and are changing from user facing to agent facing where agents are continuously storing, retrieving and analysing data in a continuous loop where have unified data platform with Fabric or a variety of different storage options including the new Postgres managed service of Horizon DB which is zone redundant with 128 terabytes of storage per cluster and are seeing 3x throughput over self-managed Postgres setup.
Data warehouse is becoming mission critical and bringing CPU acceleration to fabric is delivering increased performance and data is the IQ layer that brings together model capabilities along with the data to deliver the right context to deliver the data, for token efficiency is where if you structure context right and optimise models then you can improve this and web grounding is important, Web IQ is build on their structure built for AI workflow which is model agnostic and MCP native to ground responses on fresh reliable content from the web which is best quality, fastest and most efficient to help build out agentic systems. Developers want to ground data across the enterprise and bringing a unified IQ layer for an understanding of your organisation.
Elijah Straight
Elijah Straight talked about unifying intelligence for every organisation with Microsoft IQ, with agents in Microsoft Foundry you can build and deploy agents in Microsoft 365 that can be grounded with knowledge for the latest information using WebIQ for the web which indexes fresh content from the web grounded in reality which can be embedded in custom applications where can combine external knowledge using Foundry IQ with industry context which can use Fabric IQ which takes Power BI and can include enterprise intelligence with a single living model of the business that an agent can use. Agents can access the outside world and own systems and have a final layer of Microsoft IQ with Work IQ which can be the latest documents in Microsoft 365 that are up-to-date and is your knowledge no matter the agent which is reasoning over them and then leverage functionality from Microsoft 365 with Microsoft IQ that is grounded in world, people and information.
Satya Nadella
Satya talked about deploying agents and the runtime, when building a first-class agentic system need a first-class runtime which will be shipped on Windows and Azure, Windows is the best place to run agents. They are a new paradigm, they can take actions across files and devices and there is a lot of power in this, which can create new risks so are introducing Microsoft Execution Containers which is a new policy layer for isolation to have containment built into the operating system with a Micro-VM and beyond to pick the right containment option for the workload and Microsoft Execution Containers will enforce it and to build real developer workflows, where OpenClaw runs on Windows leveraging MXC.
Samantha Song & Scott Hansleman
talked about OpenClaw which has been out since November where Scott has used it to help manage his blood sugar to managing GiutHub Issues and Samantha is using it to help with triathlon training. Microsoft has been collaborating in the open on a Windows companion app and sandbox the OpenClaw tool calls for Windows and WSL. The OpenClaw Windows Companion app looks like a native app as written in WinUI 3 and can access chat, canvas and main dashboard there is a sandbox using MXC with process isolation but in the future there will be more options and settings, you will have granular security options for internet access and specific access to folders which expands on the rich safety layer of OpenClaw. If these layers are turned off, then the additional safety layer can stop actions you don't want to perform and the OpenClaw Companion app being a great way to show off OpenClaw on Windows.
Peter Steinberger
Peter Steinberger, the creator of OpenClaw, spoke about it being built to have access to everything and is very powerful but makes companies nervous and have been spending time to add observability with granular permission like readonly folders and you can now run OpenClaw inside your company and you can bring your own rules and models and have OpenClaw run on top of this. It has been exciting to see OpenClaw grow into something much bigger and it will stay open and neutral for any model and operating system as are entering a new era, with capabilities for those who don't code and more for those who do.
Satya Nadella
Satya talked about how exciting it is to see OpenClaw come to Windows and have agents and unmetered intelligence come together. They are building Microsoft Foundry into the system with Agent Service, Models, IQ and Tools, with Hosted Agents where can scale up and down with all the IQ layers, tools along with durability, memory and state with a super-fast sandbox where you can have all the safety and guardrails around an agentic system. Also excited to announce partnership with Fireworks AI on Foundry to bring more choices to developers to build next generation application.
Satya then spoke about the tools with GitHub at the heart of this beyond the code and becoming a control plane, pull requests, commits and new repositories per month are growing exponentially powered by agents, seeing tremendous growth in CLI form factor and are becoming the thing everyone goes to, but when you have hundreds of CLIs it becomes hard and need something new and led to a tool that has the speed and flexibility of CLI and capability of IDE with the GitHub Copilot App where you can start your GitHub Agents to work with them and even run parallel sessions.
Cassidy Williams
Cassidy Williams spoke about the GitHub Copilot App which is the homebase for development where you can kick off an agentic coding session, which can kick off separate sessions for different issues and can use Git Worktrees for sessions where agents can work together without stepping on each other, so don't need to stash anything. Then can have agent merge where can deal with any issues or merge conflicts. Cn also see from work a focused view of activity and in automations can see actions that can be run locally and on the cloud. You can start a session by pulling from a local or GitHub Repository.
Cassidy mentioned there are features such as Pick and Polish where you can focus on elements to be amended and can access features from any subscription model you are using and choose the right model for the task or request a rubber duck review from other models, so this helps catch model blind spots. Working with AI should be more than chat with concept of a canvas where agent can build a custom UI to communicate with you, you could approve a PR with a thumbs up from the camera and you can do many tasks where the GitHub Copilot app can help you start or finish what you are working on.
Amanda Foster
Amanda Foster spoke about integrating agents into business and scaling them at scale where can build agents and have them consumed by MCP and can also block PII by applying rules to protect agents and can help get agents ready for Foundry by including the necessary code. You can also interact with Foundry with a micro-VM, and own persistent file system and Foundry now has server-side tracing and evaluations to see what is happening on every run. You can also see governance outcome from production traces performed by Foundry, There is also an agent optimiser to compare candidates such as prompt changes, model changes and more and allow the best candidate to be deployed for an agent and every evaluation helps and they get better the more they are used.
Amanda mentioned can have agents with own identity that can work on its own behalf for a agent that keeps track of features, but to get something like this deployed it needs to be enterprise ready such as who has the ability to talk to an agent with admin mode who can decide access or block agents at any time. Foundry accelerates development from local to enterprise ready, you build a agent and we handle the rest.
Satya Nadella
Satya mentioned are building security for AI with MDash but how do you defend yourself against attacks that may be using AI, are bringing across agents and custom models to find exploitable bugs to prepare, scan, validate, de-duplicate and prove security issues.
Sarah Young
Sarah Young talked about the security scans that can be done which can be seen in GitHub App which can include usual issues with coding along with AI vulnerabilities with scans to look for these which produces a report and then can dig into these vulnerabilities and see where it is in the code along with severity and then can use the defender fix command to create fixes that are needed and can see what has been done and still have a human in the loop check, this is all done locally but can also do a PR locally and then push this and use a tool such as GitHub Advanced security to manage this. Bugs can have parts of it distributed across the codebase but can have teams of agents that can argue and find an issue as well as prove an issue and will be coming soon to the CLI and Microsoft Defender Portal.
Alex Pall, Drew Taggart & Satya Nadella
Alex Pall and Drew Taggart from The Chainsmokers joined Satya, who are venture capitalists who spoke about their experience, they have been doing music for fourteen years and have met people from startups and found out about early stage investments and found a lot in common about their business and how they started and have got to participate in deals and in 2020 started their own fund. How they see the opportunity going forward is to have authenticity and for investment is from producing outputs to producing actions and think about instead of humans but machines producing outputs. When they look at founder what is the advice they give them and what are they looking for in founders? It is really difficult to find what is authentically do, you have to be connected to product you are created, have to iterate and have consistency to then build something special and unique.
Satya Nadella
Satya talked about the developer stack so what is the opportunity for every company, how to be an AI native company from Saas to enterprise so when building AI apps and agents have to make sure these are discovered and customers are building line of business applications discoverable in Copilot Studio and tools have become destination for human agent interaction and what to be able to find agents and super charging this. Copilot started with Chat, then Cowork to generate stunning artifacts and solve multi-step problems, code with GitHub and now bringing coding to all knowledge work with one Copilot super app, with Chat, Cowork and Code all in Copilot but today are bringing Autopilots, which are enterprise claw that are autonomous agents that can have a name, personality, custom connectors that can help with toil and get you into what you love.
Satya talked about the first Autopilot is Scout that can monitor things, powerful out of the box and personal by design which works where you work including Microsoft Teams, it can also handle threads In Outlook and is your always-on personal agent for work and you can build more of these autopilots and is the future of the Copilot ecosystem itself. Agents can be discovered in Copilot, Teams and Windows but what makes organisations unique is its tacit knowledge so how do you preserve and compound this knowledge when models can learn anything so to do this, they believe that any organisation will need to build their own hill climbing machine which continuously improves against your own information. Working with frontier tuning to supercharge this capability with innovation.
Mustafa Suleyman
Mustafa Suleyman talked about the compute to train frontier models has increased on trillion fold in fifteen years so an increase in computation results in an increase of capability and an increase of compute is needed where intelligence is now a function of compute, and building a humanist super intelligence to serve people rathe than replacement that prioritises human wellbeing and purpose and is behind Microsoft's super intelligence plan. Microsoft are announcing seven new models cross image, transcription and voice to make practical and efficient tools with MAI-Image 2.5 and Flash with a score higher than Nano banana 2 on image editing for professional grade performance available in PowerPoint today, MAI-Transcribe which is the best transcription model in the world and optimised for real-world use and is five times faster than all other models and is available in GitHub App and is most efficient most effective transcriber model. MAI-Voice 2 has voice capabilities and find-grained voice control and with best value and speed along. MAI-Thinking-1 is a reasoning model and is a 35 million parameter model with 256K token window for a medium sized model and achieves 97% in general purpose reasoning abilities and what is remarkable is it has climbed from the bottom with zero distillation with an enterprise cleared and enterprise data lineage that can be deployed in complete confidence. MAI-Code-1-Flash is for Visual Studio Code and is a five billion parameter model for coding processes.
Mustafa mentioned models are watermarked from scratch and have a full transparent report on how the models have be put together and co-designing models for the hardware from Microsoft where model and silicon co-design helps develop the most efficient and cost-effective agents out there. This is what owning the full stack looks like where you can customise models and the disciplined and relentless engineering in these models can be used to create custom agents you can control. Reinforcement Learning Environments help find best agentic use cases as well as help improve efficacy and have better fine tuning, you keep the benefits of your hard-earned knowledge, and you get to control your models and marks a new era in AI. One final announcement for taking co-creation of models where are partnering with Mayo Clinic to create a new frontier model for health and was joined by Gianrico Farrugia talked about most people won't have access to Mayo Clinic where they have build a platform that created the largest healthcare dataset so now have the opportunity to create a frontier model for healthcare to get clinical and logistical insights or find out what happens next or prevent harm to improve patient safety and deliver better healthcare, models have the knowledge and textbooks but are lacking clinical practice and expertise. They will build a trusted scalable solution to build this frontier model to offer safe, trustworthy solutions from Mayo Clinic in partnership with Microsoft. We are taking exciting steps with models and working with people to co-create agents with a new agent of AI that you control on your terms we can build together.
Tanaya Yadav
Tanaya Yadav talked about frontier tuning which is taking models tuned on your workflow to build your own hill climbing machine, where you can deploy a model as-is and then you can use the fine tune model to add your own dataset and grader and then can see how the model is learning to start the hill climb but you can also have more control with a low-level training API, where you can configure the lora strategy to see how your training is going to work and define the tools the AI model will work with. You can create a place where your agents can continuously learn how to work with higher accuracy, you can have evaluation criteria or you can spend a lot of time doing things in Microsoft 365 that can define how you work with tools and can simulate the tools without impacting live state of the business and you can generalise these learnings into the main model and are able to hill climb for tasks that require higher accuracies and can even help models become ten times more efficient that the base model based on effective fine tuning. With frontier turning your agent continuously improves on the way you work and can't wait to see the environments you build.
Satya Nadella
Satya spoke about how have seen a pretty significant shift from moving from using a frontier model to fully utilising these with own enterprise models and create that differentiator. There is a new operating point at frontier where you can use a very efficient model where you can get the frontier performance to enrich models to hill climb and is a game changer for how people operate at the frontier. New frontiers that push the ecosystem and what is the next big thing, building that scientific discovery loop can have that societal impact and make it more programmable and bring this together in Microsoft Discovery with scientific reasoning, hypothesis reasoning and more.
David Carmona
David Carmona spoke about advancing research with Microsoft Discovery, which has many parallels with agentic software development, you can ask it to write a hypotheses, discover what is needed and create a ranked set of candidates, much like planning, building / testing and deployment in software but for products. Microsoft Discovery includes products and tools across many domains, but science is not sequential and can perform long-running simulations dynamically that can be used to create a paper where it uses a knowledge graph of scientific knowledge and internal knowledge. You can then find what is needed and Microsoft Discovery will use the right agent for the task, and if not, it will create one for the task, then to determine a solution can apply different variations that can be created such as the DNA sequences to create proteins needed. These can then be sent to a lab to be produced and evaluated with many steps that can be automated with just human supervision to unify the digital and physical world to embrace a new era in scientific discovery.
Satya Nadella
Satya mentioned about making progress with scalable quantum computer and take a radially different approach to barriers about building a quantum computer for reliability, and size alogn with working with partners to build quantum computers and compressing years of research, Majorana 2 combines next generation of Qubits that can have a 20 second mean lifetime or a thousand times higher than Majorana 1 and key operations are one microsecond for complex Quantum operations where it would be possible to fit a million of these chips in something smaller than a credit card and with Majorana 2 are building out the engineering scale after proving the concept with Majorana 1. It is not about building the next platform but is how do we build this frontier ecosystem together, technology concentrates power or we use this next wave to unlock opportunities in every community, and our job is to make that second story true and let's all build together.