Satya Nadella — How Microsoft is preparing for AGI
What this covers
Satya Nadella, Microsoft's chief executive, sits down with Dwarkesh Patel to explain how Microsoft is positioning itself for a world where artificial general intelligence may be imminent. Nadella and panelist Dylan Patel argue that the winner of the AI era will not be determined by model capability alone. Instead, value will distribute across three layers—infrastructure, models, and application scaffolding—and Microsoft intends to compete aggressively on all three while remaining agnostic about which model architecture or vendor dominates. The core strategic bet is fungibility and optionality: building a hyperscale fleet that can run many different models and workloads, rather than betting the company on a single chip generation, model family, or customer relationship.
Nadella diagnoses the risk of winner's curse for model companies—the paradox that whoever invests most in model innovation may see that innovation copied through open-source checkpoints and fine-tuning on others' data, leaving data liquidity and scaffolding as the defensible value layers. He positions tools as infrastructure—GitHub, coding agents, and workflow automation—as places where Microsoft has structural advantage regardless of which model wins. A second pillar is trust as a competitive moat: Nadella argues that global adoption of American AI will hinge not on raw capability but on whether the world trusts the U.S. tech stack and national institutions as reliable long-term suppliers, a trust he contrasts with Chinese competition. The conversation also covers technology diffusion lag, the historical parallel that AI's economic payoff requires real organizational change—not just fast algorithms—and why sovereign AI and supply-chain resilience will shape which providers win in each geography.
Nadella argues that value in the AI era will not accrue solely to whoever has the best model, but will be split across infrastructure, model, and application-scaffolding layers, so Microsoft should compete on every layer while keeping its fleet fungible and model-agnostic rather than betting everything on a single model or a single bare-metal customer.
- Models are at risk of commoditization because open-source checkpoints plus data liquidity and scaffolding let others replicate capability ('winner's curse'), so scaffolding and data ownership retain value.
- A fungible, multi-model hyperscale fleet avoids being stranded by a single chip generation or model architecture breakthrough.
- Trust in the American tech stack and respect for national sovereignty/resilience will shape which providers win globally, advantaging firms that build sovereign infrastructure.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
True economic growth from AI requires not just fast technology diffusion but a change in the work, the work artifact, and the workflow; just as the Industrial Revolution took ~70 years of diffusion before economic growth appeared, AI's payoff depends on corporations undertaking real change management, which shouldn't be discounted.
“for true economic growth to appear it has to diffuse to a point where the work, the work artifact, and the workflow has to change. So that's one place where I think the change management required for a corporation to truly change is something we shouldn't discount.”
Model companies may suffer a 'winner's curse': they do all the hard innovation work, but their capability is 'one copy away' from being commoditized by open-source checkpoints, so whoever holds the data liquidity, grounding, and context engineering can take a checkpoint, train on their data, and capture the value—meaning value need not migrate solely to the model.
“if you're a model company, you may have a winner's curse. You may have done all the hard work, done unbelievable innovation, except it's one copy away from that being commoditized.”
If models reach human level and can continuously learn on the job—with copies deployed across the economy amalgamating their learnings back into one model—this creates a continuous-learning exponential feedback loop resembling an intelligence explosion, meaning the leading model company could win game-set-match because one model would know how to do every job while others wouldn't.
“you have copies of one model broadly deployed through the economy learning how to do every single job. And unlike humans, they can amalgamate their learnings to that model. So there's this sort of continuous learning exponential feedback loop, which almost looks like a sort of intelligence explosion.”
Microsoft feared that moving Office customers from on-premise servers to the cloud would shrink margins due to COGS, but instead the cloud massively expanded the market by letting customers worldwide fractionally afford IT, eliminating costs like SharePoint storage servers (working capital outflows), and the same market expansion will happen with AI.
“what happened was the move to the cloud expanded the market like crazy. We sold a few servers in India, we didn't sell much. Whereas in the cloud suddenly everybody in India also could afford fractionally buying servers”
What ultimately matters for a country is the use of AI in its economy to create economic value—the 'diffusion theory'—which is not about owning the leading sector but about the ability to use leading technology to create one's own comparative advantage.
“what matters is the use of AI in their economy to create economic value. That's the diffusion theory, which ultimately, it's not the leading sector, but it's the ability to use the leading technology to create your own comparative advantage.”
Because the US is 4% of world population but 25% of GDP and 50% of market cap—a ratio that rests on the world's trust in US capital markets, technology, and stewardship—the key priority is building global trust in the American AI tech stack; if that trust breaks it harms the US, and trust ('can I trust you, your country, and its institutions to be a long-term supplier') may be the thing that wins the world against Chinese competition, more than raw model capability.
“It's 4% of the world's population, 25% of the GDP, and 50% of the market cap... That 50% happens because quite frankly the trust the world has in the United States, whether it's its capital markets or whether it's its technology”
Infrastructure should not be optimized for one model architecture, because an MoE-like breakthrough could change the network topology and strand capital optimized for a single architecture; you must build a fleet capable of supporting multiple model families and lineages, and to be a serious hyperscaler you must stay open and support an ISV ecosystem rather than owning every category.
“if you fall behind... all the infrastructure you built will be a waste. You kind of need to build an infrastructure that's capable of supporting multiple families and lineages of models. Otherwise the capital you put in, which is optimized for one model architecture, means you're one tweak away, some MoE-like breakthrough that happens, and your entire network topology goes out of the window.”
The single-dominant-model 'game set match' scenario does not match reality: like databases, multiple models get deployed for different use cases, and continual-learning network effects ('data liquidity') will not occur across all domains, geographies, segments, and categories simultaneously, so the design space is large enough for many winners and Microsoft should compete on the merits at each layer rather than assume vertical-stack dominance.
“There is not one model that is getting deployed broadly. There are multiple models that are getting deployed. It's like databases... There are multiple types of databases that are getting deployed for different use cases.”
Countries will demand continuity and avoid concentration risk in AI, so there will always be a check against any single model achieving runaway deployment: open-source models and multiple model options let a nation move its data and 'liquidity' to another model, giving it agency/sovereignty—so concentration risk and sovereignty (agency) are the two forces that will drive AI market structure.
“there's always going to be a check to 'Hey, can this one model have all the runaway deployment?' That's why open source is always going to be there... Concentration risk and sovereignty, which is really agency, those are the two things that will drive the market structure.”
Each major technological transition—railroads, the internet, replaceable parts, industrialization, the cloud—has gotten faster in the time from discovery to economy-wide pervasiveness, and AI's three-year ramp to hyperscalers spending $500 billion of capex next year is unmatched in speed relative to prior revolutions.
“each revolution has gotten much faster in the time it goes from technology discovered to ramp and pervasiveness through the economy”
After lessons like the pandemic, every nation state—including the US—will do what it takes to be more self-sufficient on critical supply chains, so a multinational must treat resilience/sovereignty as a first-class requirement; globalization can't simply be rewound at the same pace, but a credible plan toward resilience (e.g. TSMC Arizona, semiconductor plants) will be demanded and should be respected.
“Any nation state, including the United States, at this point will do what it takes to be more self-sufficient on some of these critical supply chains. So I, as a multinational company, have to think about that as a first-class requirement.”
Microsoft going from near-100% share in VS Code/repos to sub-25% share in coding agents in one year is acceptable because the new market is vastly larger; the existence proof is hyperscale, where Microsoft has much lower share than it had in client-server computing but the business is orders of magnitude bigger and supports multiple winners.
“You could say we had a high share in client-server server computing. We have much lower share than that in hyperscale. But is it a much bigger business? By orders of magnitude.”
Building your own vertical accelerator only makes sense if you have your own model to generate or subsidize demand for it, which is why even Google and Amazon still buy Nvidia (the general-purpose option all models run on); Microsoft's birthright to do its own silicon comes from creating a closed loop between its own MAI models and its silicon microarchitecture.
“if you build your own vertical thing, you better have your own model, which is either going to use it for training or inference, and you have to generate your own demand for it or subsidize the demand for it.”
The difference between a classic hoster and a hyperscaler is software: the systems know-how to optimize by workload and by fleet, including the ability to evict a workload and schedule another (fungibility), is what enables capital efficiency, with software improvements yielding 5x-40x tokens-per-dollar-per-watt gains for a given model family quarter-over-quarter.
“what is the difference between a classic old-time hoster and a hyperscaler? Software. Yes, it is capital intensive, but as long as you have systems know-how, software capability to optimize by workload, by fleet...”
Every AI workload requires not only the AI accelerator but a whole host of other services, and much of Microsoft's margin will come from those other things, so Azure should be built to be excellent for the long tail of AI workloads while remaining competitive in high-end bare-metal training—but bare-metal for a few customers can't be allowed to crowd out the broader hyperscale business.
“every AI workload does require not only the AI accelerator, but it requires a whole lot of other things. In fact, a lot of the margin structure for us will be in those other things.”
Despite AI's high COGS, the underlying business model levers remain similar—ads, transactions, device gross margin, consumer and enterprise subscriptions, and consumption—and a subscription is essentially an entitlement to consumption rights, making tiering a pricing decision; Microsoft's advantage is operating across all these meters at a portfolio level.
“there will be some ad unit, there will be some transaction, there will be some device gross margin for somebody who builds an AI device. There will be subscriptions, consumer and enterprise, and then there'll be consumption. So I still think those are all the meters.”
Microsoft's capacity 'pause' (dropping leasing sites taken up by Google, Meta, Amazon, Oracle) was a deliberate course-correction toward a fungible fleet usable across all AI stages (training, mid-training, data gen, inference) and able to serve models worldwide, rather than becoming a hoster for one company with one massive single-customer book of business—which Nadella argues 'is not a business.'
“We didn't want to just be a hoster for one company and have just a massive book of business with one customer. That's not a business, you should be vertically integrated with that company.”
Features like GitHub Copilot's 'auto' setting will optimize and arbitrage tokens across multiple models to complete a task, potentially fully autonomously, which turns the model into the commodity—especially with open-source checkpoints that can be fine-tuned on one's own data—while the scaffolding that handles model jaggedness becomes the value layer.
“one of my favorite settings in GitHub Copilot is called auto, which will just optimize... It could arbitrage the tokens available across multiple models to go get a task done. If you take that argument, the commodity there will be models.”
The category of AI coding—the 'software factory' category—is likely to be one of the biggest categories, possibly even bigger than knowledge work, and the fact that it scaled from nothing to billions in run-rate revenue in one year demonstrates cloud-like market expansion.
“Fundamentally, this category of coding and AI is probably going to be one of the biggest categories. It is the software factory category. In fact, it may be bigger than knowledge work.”
GitHub benefits from the entire coding-agent boom regardless of which agent wins, because the repos generated by all these agents go to GitHub, which is at all-time-high repo creation with a new developer joining roughly every second and 80% falling into a GitHub Copilot workflow; Microsoft will get 'many structural shots' via primitives like Git, issues, actions, and the new Agent HQ / Mission Control control plane.
“guess where all the repos of all these other guys who are generating lots and lots of code go? They go to GitHub. GitHub is at an all-time high in terms of repo creation, PRs, everything.”
The most immediately valuable AI use is migration—converting mainframe systems to cloud, Excel databases into real SQL databases, and Word/Excel artifacts into more programmatic, efficient forms—which could shrink the Office ecosystem's relative use, though even mainframes have continued growing for two decades despite such migrations.
“that is probably the most immediately valuable thing: converting mainframe-based systems to standard cloud systems, converting Excel databases into real databases with SQL”
Because new chips like Vera Rubin Ultra will have radically different power density and cooling requirements, datacenters should not be built entirely to one spec; you want to be scaling in time across generations rather than scaling once and being stuck with depreciating, single-generation infrastructure.
“you kind of don't want to just build all to one spec. That goes back a little bit to the dialogue we'll have, which is that you want to be scaling in time as opposed to scale once and then be stuck with it.”
Building a leading AI lab is a capital-intensive talent game—evidenced by Meta spending north of $20 billion on talent and Anthropic and Meta poaching reasoning/post-training teams from Google—so a company spending $100 billion on infrastructure should commensurately spend on the people who make breakthroughs, keeping researcher-to-GPU ratios high.
“You've seen Meta spend north of $20 billion on talent. You've seen Anthropic poach the entire Blueshift reasoning team from Google last year... These sorts of talent wars are very capital intensive.”
Even as AI task horizons lengthen (from 30-second prompts to multi-day agent jobs), datacenter location and regional networking topology still matter because of data residency laws (e.g. the EU Data Boundary prevents round-tripping calls to Texas even asynchronously) plus power costs, latency for session data, and proximity of storage like Cosmos DB.
“what are the data residency laws? There's the entire EU thing, where we literally had to create an EU Data Boundary. That basically meant that you can't just roundtrip a call to wherever, even if it's asynchronous.”
Microsoft has made a series of governance commitments to Europe on how it runs its hyperscale investment so the EU and European countries retain sovereignty, and is building sovereign clouds in France and Germany plus Sovereign Services on Azure offering key management and confidential computing (including in GPUs) developed with Nvidia.
“We made a series of commitments to Europe on how we will govern our hyperscale investment there such that the European Union and the European countries have sovereignty. We're also building sovereign clouds in France and in Germany.”
Microsoft's end-user tools business will become an infrastructure business in support of autonomous agents: a future autonomous AI agent gets provisioned a computer (e.g. Windows 365) plus embodied tools (because using tools is more token-efficient than raw computer use), plus storage, archival, discovery, identity, and observability—so per-user revenue evolves into per-agent revenue, growing faster than the number of users.
“our business, which today is an end-user tools business, will become essentially an infrastructure business in support of agents doing work.”
Microsoft has full access to OpenAI's silicon/chip program IP (the only excluded IP being consumer hardware), partly because Microsoft gave OpenAI IP to bootstrap them and built supercomputers together, so as OpenAI innovates at the system level Microsoft gets access and will first instantiate what OpenAI builds and then extend it alongside its own MAI silicon lineage.
“Dylan Patel: What level of access do you have to that? Satya Nadella: All of it. ... So the only IP you don't have is consumer hardware? ... That's it.”
Over the course of the year, Anthropic's gross margins on inference went from well below 40% to north of 60%, expanding significantly at the model layer despite intensifying competition from Chinese open-source models, OpenAI, Google, and X/Grok.
“at least at Anthropic, their gross margins on inference went from well below 40% to north of 60% by the end of the year. The margins are expanding there despite more Chinese open source models than ever.”
Microsoft's Fairwater campuses are built to aggregate flops across interconnected buildings and regions—via a one-petabit network on campus and an AI WAN to Wisconsin—so a single large training job can run with model parallelism and data parallelism distributed across sites, and the same fleet can later be repurposed for data gen and inference.
“The goal is to be able to aggregate these flops for a large training job and then put these things together across sites... you'll use it for training and then you'll use it for data gen, you'll use it for inference in all sorts of ways.”
SaaS companies have historically had very low incremental cost per user (high R&D and customer acquisition costs but cheap to serve), but the high COGS of AI completely breaks these business models, which is why SaaS companies (excluding Microsoft) have underperformed in the markets.
“Software-as-a-service has incredibly low incremental cost per user... This is sort of why, not Microsoft, but the SaaS companies have underperformed massively in the markets, because the COGS of AI is just so high, and that just completely breaks how these business models work.”
Following Jensen Huang's advice to achieve 'speed-of-light execution' (Atlanta's ~90-day turnaround from receipt to live workload), Microsoft avoids overbuilding one chip generation and instead scales each generation in a balanced flow, because Nvidia's accelerating migration pace means committing to four or five years of depreciation on a single generation is risky.
“I didn't want to go get stuck for four or five years of depreciation on one generation. In fact, Jensen's advice to me was two things. One is, get on the speed-of-light execution.”
Independent AI labs' projections of $100 billion revenue in 2027-28 growing 2-3x a year should be read in light of incentives: an independent lab raising money must put out big numbers to fund its compute bills, which is acceptable because they've shown real traction, but the numbers reflect fundraising incentives rather than certainty.
“What do you expect an independent lab that is sort of trying to raise money to do? They have to put some numbers out there such that they can actually go raise money so that they can pay their bills for compute”
The depreciating GPU asset is about 75% of the total cost of ownership of a datacenter over five-six years, and Nvidia takes roughly a 75% margin on that, which is why hyperscalers develop their own accelerators to reduce this cost and increase their margins.
“this depreciating asset, in five or six years, is 75% of the TCO of a data center. And Jensen is taking a 75% margin on that.”
Under the new OpenAI agreement, OpenAI's PaaS business (its API) is Azure-exclusive while its SaaS business (ChatGPT) can run anywhere; any partner wanting to use OpenAI's stateless API, or any custom co-developed/co-trained agreement, must come to Azure, with only a few exceptions like the US government.
“think of OpenAI having a PaaS business and a SaaS business. The SaaS business is ChatGPT. Their PaaS business is their API. That API is Azure-exclusive. The SaaS business, they can run it anywhere.”
Microsoft's Fairwater 2 datacenter represents a roughly 10x increase in training capacity over what GPT-5 was trained with, achieved by 10x-ing training capacity every 18-24 months, with network optics in a single building nearly equal to all of Azure two and a half years ago.
“We've tried to 10x the training capacity every 18 to 24 months. So this would effectively be a 10x increase from what GPT-5 was trained with.”
Even in a future where agents work with agents, core primitives—storage, e-discovery, observability, a unified identity system across multiple models, and databases—remain necessary; the agentic world will grow the underlying infrastructure business by enabling better joins between structured and unstructured data, with consumption driven by agents rather than users.
“Even when agents are working with agents, what are the primitives that are needed? Do you need a storage system? Does that storage system need to have e-discovery? Do you need to have observability? Do you need to have an identity system”
Given long AGI/ASI timelines, the right way to handle massive AI compute investment is to split it: allocate an order-of-magnitude-scaling 'research compute' budget accounted for like R&D expense (table stakes), while the remainder must be strictly demand-driven—you can build ahead of demand, but you'd better have a demand plan that doesn't go off kilter.
“There needs to be an allocation to, I'll call it, research compute. That needs to be done like you did R&D... The rest is all demand driven. Ultimately, you're allowed to build ahead of demand, but you better have a demand plan that doesn't go completely off kilter.”
Microsoft will use OpenAI models maximally across all products for the next seven years, adding value via RL fine-tuning and mid-training runs on the GPT family using unique data assets, while deliberately not duplicating those flops, and will simultaneously build its own MAI models for cost/latency-optimized or specialized capabilities.
“we are absolutely going to use the OpenAI models to the maximum across all of our products. That's the core thing that we're going to continue to do all the way for the next seven years, and not just use it but then add value to it.”
AI is best understood, following Raj Reddy's metaphor, as either a cognitive amplifier or a guardian angel—a tool whose human utility is to amplify human capability, rather than a mystical entity, since historically many things only humans did were eventually done by tools.
“He had this metaphor for AI, it should either be a guardian angel or a cognitive amplifier. I love that. It's a simple way to think about what this is.”
Microsoft's Excel Agent is not a UI-level wrapper but a model embedded in the middle tier of Office, taught the native artifacts and tools of Excel via markdown so it can natively understand formulas and fix its own reasoning mistakes—wrapping a cognitive layer around traditional business logic so Excel ships with an analyst bundled in.
“Excel Agent is not a UI-level wrapper. It's actually a model that is in the middle tier... teaching it all the tools of Excel. I'm giving it essentially a markdown to teach it the skills of what it means to be a sophisticated Excel user.”
Microsoft tracks competitors like AWS, Google, and Oracle but won't chase them merely for the gross margin a business represents in a period of time; the relevant question is what unique book of business Microsoft can clear that makes sense for it over the next 50 years, not the next five.
“The thing that you have to think through is not what you do in the next five years, but what you do for the next 50.”
Microsoft's MAI text model debuted around rank 13 on LMArena despite being a small model trained on only about 15,000 H100s, which Nadella presents as proof of core capability (instruction following) and, given scaling laws, of what Microsoft could achieve with more flops; the next step is an omni-model combining audio, image, and text.
“we started on the text one and it debuted at like 13. By the way, it was done only on around 15,000 H100s. It was a very small model.”
Rather than gobbling up all capacity itself, Microsoft welcomes neoclouds (Iris Energy, Nebius, Lambda Labs) into its marketplace and takes leases, build-to-suit, and GPUs-as-a-service where it lacks capacity, because a customer arriving through Azure will use the neocloud's compute plus Azure's storage, databases, and services—a win for both.
“I would even sort of welcome every neocloud to just be part of our marketplace. Because guess what? If they go bring their capacity into our marketplace, that customer who comes through Azure will use the neocloud... and will use compute, storage, databases, all the rest from Azure.”
In the most recent quarter Microsoft grew its Copilot subscriber base from 20 to 26 million subscribers.
“we did our quarterly announcement and I think we grew from 20 to 26 million subs.”
The fact that Microsoft's coding-agent competitors (Claude, Cursor, Codex) are all companies born in the last four or five years—rather than legacy rivals like Borland—is the best sign that Microsoft is heading in the right direction in a healthy, fast-growing category.
“all these companies that are listed here are all companies that have been born in the last four or five years. That to me is the best sign. You have new competitors, new existential problems... it's not Borland. Thank God.”