Satya Nadella – How Microsoft thinks about AGI

Microsoft CEO Satya Nadella and EVP Scott Guthrie give a tour of Fairwater 2, Microsoft’s newest and most powerful data center, located in Atlanta, and discuss Microsoft’s strategy across infrastructure, models, AI agents, and global expansion.

• Fairwater 2 represents roughly a 10x increase in training capacity over what was used for GPT-5, with network optics in the single building nearly equaling all of Azure’s global data center footprint from two and a half years ago, around five million network connections, and a design built to aggregate compute across multiple sites and regions for large training jobs.
• Microsoft is building a campus-scale topology: Fairwater 4 will be on the same one-petabit network, and an AI WAN will link the Fairwater campus to Milwaukee, allowing training jobs to run across all sites simultaneously, with model parallelism and data parallelism spanning the full campus.
• The design philosophy is to scale capacity roughly 10x every 18 to 24 months while maintaining fungibility across training, data generation, and inference workloads, and avoiding over-optimization for a single chip generation, since future chips like Vera Rubin Ultra will have very different power density and cooling requirements.

Nadella frames AI as a cognitive amplifier and guardian angel rather than a mystical force, and argues that even if this is the biggest technological shift since the Industrial Revolution, true economic growth requires diffusion into workflows, which takes time and change management.

• He references Raj Reddy’s metaphor that AI should function as either a guardian angel or a cognitive amplifier, and emphasizes that the real question is human utility.
• He notes that the Industrial Revolution took roughly 70 to 150 years to show up in economic growth statistics, and while he would love to compress that into 20 to 25 years, the work artifacts and workflows of corporations must genuinely change for productivity gains to materialize.
• On the question of whether AI tokens will replace human labor, he points out that even at cents per million tokens, the value of human output is enormous, and the real question is how much leverage humans gain from technology, not whether they are replaced outright.

On business models, Nadella argues that the traditional SaaS pricing levers (subscriptions, consumption, ads, transactions, device margins) will persist, but that AI massively expands the total addressable market, just as cloud computing expanded the server market.

• He draws a direct analogy to the server-to-cloud transition: moving to cloud introduced COGS and initially seemed like it would shrink margins, but the market expanded so dramatically that the business grew enormously, with examples like India going from minimal server purchases to widespread cloud adoption.
• He notes that low ARPU in cloud was more than offset by market expansion, and expects the same dynamic with AI, where the coding assistant category alone went from essentially zero to billions in revenue in about a year.

On GitHub Copilot and coding agents, Nadella acknowledges that Microsoft’s market share has dropped from dominant to below 25% in one year as Claude Code, Cursor, Cognition, Windsurf, Replit, and OpenAI Codex have all grown rapidly, but he sees this as evidence of massive market expansion rather than a loss.

• The overall AI coding agent market has grown from roughly $500 million run rate (mostly Copilot) to an estimated $5 to 6 billion run rate across all players in Q4 of this year, a 10x increase.
• GitHub itself is at an all-time high in repository creation and pull requests, with roughly one new developer joining GitHub per second, and about 80% of them falling into a Copilot workflow by default.
• Microsoft’s response is Agent HQ, announced at GitHub Universe, which includes Mission Control, a control plane where users can dispatch tasks to multiple agents (Codex, Claude, Cognition, Grok, etc.) working in independent branches, monitor their output, and triage results, all within a single subscription.
• Nadella argues that GitHub’s structural position as the repository layer gives Microsoft multiple shots on goal regardless of which individual coding agent wins, and that the real opportunity is in observability and management of agents at scale.

On whether value will migrate to model companies or to scaffolding and application layers, Nadella argues both layers will capture value, and that competition and open source models will prevent any single model company from capturing all the margin.

• He points out that Anthropic’s inference gross margins expanded from below 40% to above 60% this year despite growing competition from open source and other labs, suggesting the model layer can sustain strong margins.
• But he also argues that scaffolding matters: Microsoft’s Excel Agent is not a UI wrapper but a model embedded in the middle tier of Office that natively understands Excel’s formulas, business logic, and artifacts, giving it capabilities a pixel-level UI wrapper cannot match.
• He notes that model companies face a potential winner’s curse where their innovations are one copy away from being commoditized via open source checkpoints, and that the party with data liquidity, grounding, and context engineering can take those checkpoints and train on top of them.
• Microsoft’s strategy is to compete at every layer: hyperscale infrastructure supporting multiple models, access to OpenAI’s GPT family for at least seven more years, its own MAI models, and application-layer scaffolding in security, knowledge work, coding, and science.

On the future of Office and end-user tools, Nadella envisions a world where Microsoft’s tools business becomes an infrastructure business for AI agents, with agents provisioned their own computers, identities, security, and management layers.

• He describes two futures: one where humans steer everything with Copilot assistance, and another where the company provisions computing resources for fully autonomous agents that have embodied knowledge of analytical tools.
• He notes that there is already significant growth in demand for Windows 365 provisioning from companies running autonomous agents that need their own computers.
• The per-user business becomes a per-user-plus-per-agent business, with each agent needing a computer, identity, security, observability, and management, all of which are layers Microsoft already provides.
• He acknowledges that models will be used to migrate legacy systems (mainframes, Excel databases) to more modern infrastructure, but argues that Microsoft’s leadership in databases, storage, and structured data means the underlying infrastructure business will grow even as the surface-level tool usage shifts.

On Microsoft’s own models (MAI), Nadella says the strategy is to avoid duplicating OpenAI’s work and instead focus on product-optimized and research-oriented models, while building toward a world-class superintelligence team.

• MAI’s image model is ranked ninth in its arena and is used in Copilot and Bing for cost optimization; the text model debuted at 13th on LMArena trained on only around 15,000 H100s, proving the team’s capability.
• The next step is an omni-model combining text, image, and audio work, while the long-term goal is to build a first-class superintelligence team ready for the breakthroughs needed over the next five to eight years.
• Microsoft has hired key talent including Mustafa Suleyman, Karen Simmlan, Amar Subramanya (post-training lead for Gemini 2.5), and Nando de Freitas (multimedia lead from DeepMind), with Mustafa set to publish more details on the lab’s direction.
• Nadella argues that continuous learning and data liquidity will create network effects for broadly deployed models, but that the design space is large enough for multiple models to coexist across domains, geographies, and segments, just as multiple database types coexist today.

On the hyperscale business, Nadella explains that Microsoft paused some data center leasing and construction in the second half of last year to avoid overbuilding capacity optimized for a single chip generation or a single customer, and to rebalance toward global demand, fungibility, and workload diversity.

• The pause involved letting go of leased sites that were then picked up by Google, Meta, Amazon, and Oracle; Microsoft’s projected capacity by 2028 is roughly 3.5 gigawatts lower than pre-pause forecasts.
• Nadella says the decision was driven by a desire to avoid being a bare-metal hoster for one model company with a limited time horizon, and instead to build a fleet that supports multiple models, multiple customers, and multiple workloads across geographies including the EU, India, and the UAE.
• He emphasizes the importance of riding Moore’s law: Nvidia’s pace of chip migrations accelerated, and Microsoft does not want to be stuck with four or five years of depreciation on a single generation, citing Jensen Huang’s advice to execute at the speed of light.
• Microsoft is simultaneously securing capacity through neocloud deals with Iris Energy, Nebius, and Lambda Labs, and Nadella welcomes neoclouds joining the Azure marketplace so that customers using neocloud GPUs also buy Azure compute, storage, and databases.
• He distinguishes between the hyperscale business (long tail of AI workloads with higher margins from the full stack) and the bare-metal-as-a-service business (selling raw capacity to model companies), saying Microsoft participates in both but will not let the latter crowd out the former.

On in-house silicon, Nadella says Microsoft’s Maia 200 looks promising but that the bar for any new accelerator is the previous generation of Nvidia, and that the company will scale its own silicon in close loop with its MAI models.

• He notes that even Google and Amazon buy Nvidia chips because Nvidia is innovating and all models run on them, and that building your own silicon only makes sense if you have your own model to use it or can subsidize demand.
• Microsoft has full IP access to OpenAI’s silicon program (everything except consumer hardware), having built supercomputers together and contributed IP to bootstrap OpenAI’s efforts, and will instantiate what OpenAI builds before extending it.

On the new OpenAI agreement, Nadella clarifies that OpenAI’s API business (stateless API calls) is Azure-exclusive, while OpenAI’s SaaS business (ChatGPT) can run anywhere, and that even custom partnerships like OpenAI-Salesforce must run on Azure with limited exceptions such as US government deals.

• He frames the agreement as balancing Microsoft’s desire for partnership value with OpenAI’s need for flexibility as an independent company.

On the capex explosion, Nadella acknowledges that Microsoft has become both a capital-intensive and knowledge-intensive business, with capex roughly tripling over two years, and that the key to managing this is using software and systems know-how to increase ROIC on capital spend.

• He points out that for a given GPT family, software improvements have driven 5x to 40x improvements in tokens-per-dollar-per-watt quarter over quarter, and that the ability to schedule, evict, and shift workloads across the fleet is what distinguishes a hyperscaler from a classic hoster.
• He says Microsoft will allocate compute for research (treated as R&D expense, scaling perhaps 10x over a chosen period) and for demand-driven inference and training, and that the company’s cash flow allows both to be funded simultaneously.
• On talent, he says researcher-to-GPU ratios must be high and that spending on people is as important as spending on infrastructure.

On global trust and sovereign AI, Nadella argues that the most important feature for American tech is not model capability but trust, and that the US tech sector and government must collectively build trust in the American tech stack worldwide.

• He notes that the US has 4% of the world’s population, 25% of GDP, and 50% of global market cap, a ratio sustained by global trust in American institutions, capital markets, and technology stewardship.
• Microsoft is building sovereign clouds in France and Germany, offering Sovereign Services on Azure with key management and confidential computing (including confidential GPUs), and has made formal commitments to the EU on data residency and governance.
• He argues that every country will want agency and continuity, which means open source models and multiple model providers will always exist as a check against concentration risk, and that no single model will achieve runaway global deployment.
• On Chinese competition, he acknowledges China’s comparative advantage in industrial buildout speed and supply chain, but argues that long-term trustworthiness as a supplier may matter more than raw capability, and that Microsoft must respect sovereignty requirements as a first-class business requirement even as it operates globally.