Satya Nadella — How Microsoft is Preparing for AGI

In this wide-ranging interview, Microsoft CEO Satya Nadella walks Dwarkesh Patel and SemiAnalysis founder Dylan Patel through the company's newly built Fairwater 2 data center — the most powerful AI facility in the world — and lays out his vision for how Microsoft is positioning itself across every layer of the AI stack. The conversation is grounded and pragmatic, with Nadella pushing back against AGI hype while simultaneously committing hundreds of billions in capital to what he calls "the biggest thing since the Industrial Revolution." The stakes are enormous: Microsoft is transforming from a software company into an industrial-scale infrastructure operator, and Nadella must navigate competition from model companies, hyperscaler rivals, sovereign AI demands, and the question of whether the scaffolding layer or the model layer will capture most of the economic value.

0:00Fairwater 2: The Scale of the Bet

The episode opens inside Microsoft's Fairwater 2 data center in Atlanta, which Nadella describes as the current most powerful AI facility in the world. The building represents a 10x increase in training capacity from what GPT-5 was trained on, and the numbers are staggering: the number of network optics in this single building is almost as much as all of Azure across all data centers two and a half years ago, representing roughly 5 million network connections. Microsoft is building multiple Fairwater facilities, each with hundreds of thousands of GB200s and GB300s, and linking them together with a one-petabit network so that training jobs can span across sites and even across regions — including a connection to Milwaukee where additional Fairwater buildings are under construction.

Nadella emphasizes that the design space is enormous and that the coupling between model architecture and physical infrastructure is both an opportunity and a risk. New chips like Vera Rubin Ultra will have dramatically different power density and cooling requirements, meaning Microsoft cannot afford to build everything to one spec. "You kind of don't want to just build all to one spec," he says. "You want to be scaling in time as opposed to scale once, and then be stuck with it." This philosophy of fungibility — building infrastructure that can support multiple model families and generations — becomes a recurring theme throughout the conversation.

3:20Business Models for AGI: The Long View

Nadella grounds the conversation in historical perspective, noting that past technological transitions — railroads, the internet, cloud computing — have each gotten faster in their diffusion through the economy. While many AI optimists believe this is the "final technological revolution," Nadella takes a more measured view. He cites Raj Reddy's metaphor that AI should be either a "guardian angel" or a "cognitive amplifier" — a tool for human utility rather than something mystical. "I start with the excitement that after the Industrial Revolution, this is the biggest thing," he says. "But at the same time, I'm a little grounded in the fact that this is still early innings."

When pressed on whether the economic value will accrue to the model layer or the scaffolding layer, Nadella argues that the answer depends on the incentive structure. He points out that even if models become incredibly capable, there will always be open-source checkpoints available, and the company that controls the scaffolding — the data, the context engineering, the observability — can take those checkpoints and fine-tune them with proprietary data. "I can make the argument that if you're a model company, you may have a winner's curse," he says. "You've done all the hard work, done unbelievable innovation, except it's kind of like one copy away from being commoditized."

The key insight Nadella offers is that the market expansion from AI will dwarf any single company's market share. Just as moving Office from on-premise licenses to the cloud expanded the market massively — suddenly everyone in India could afford fractional server capacity — AI will create entirely new categories of consumption. "If you take coding, what we built with GitHub and VS Code over whatever decades, suddenly the coding assistant is that big in one year," he notes.

12:48Copilot and the Competitive Landscape

The conversation turns to GitHub Copilot's competitive position, where Dylan Patel presents a striking chart: GitHub Copilot went from roughly $500 million in revenue early in the year with no close competitors, to a market where Claude Code, Cursor, and Copilot each have around $1 billion in revenue, with Codex catching up at $700-800 million. Nadella's reaction is surprisingly positive: "I love this chart for so many reasons. One is we're still on the top. Second is 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."

Nadella argues that Microsoft's share going from near 100% to sub-25% in a single year is actually a sign of market expansion, not failure. "There's no birthright here that we should have any confidence other than to say, hey, we should go innovate," he says. He points out that GitHub itself is at an all-time high in repo creation and pull requests, and that 80% of new developers joining GitHub naturally fall into some Copilot workflow. Microsoft's strategy is to turn GitHub into an "Agent HQ" — a cable TV-like marketplace where users can subscribe to multiple coding agents (Claude, Cognition, Grok, etc.) and manage them all from a single mission control interface with observability and steering capabilities.

The deeper question is whether this scaffolding layer will remain relevant as models become more capable. If future AIs can use computers as well as humans, why would they need deep integration with GitHub or Office? Nadella's answer is that the entire substrate underneath human tools — storage, identity, security, observability, management — becomes the bootstrap for AI agents as well. "The per-user business is not just per user, it's per agent," he says. "And if you sort of say it's per user and per agent, the key is what's the stuff to provision for every agent? A computer, a set of security things, an identity around it, and all those things."

20:02Whose Margins Will Expand Most?

The hosts press Nadella on a fundamental question: as AI models become more capable and autonomous, why wouldn't the model companies — which are already seeing their gross margins expand from below 40% to north of 60% — capture all the value? Nadella pushes back by arguing that the model layer faces structural commoditization pressure from open-source checkpoints and multiple competing frontier labs. He points to the database industry as an analogy: no single database dominates everywhere, and different use cases demand different solutions.

Nadella introduces a concrete example of how Microsoft is building AI into its products at a deeper level than simple wrappers. The Excel Agent, he explains, is not a UI-level wrapper but a model embedded in the middle tier that has native understanding of Excel's artifacts — formulas, cells, data structures. "It's not just, hey, I just have a pixel level understanding," he says. "I have a full understanding of all the native artifacts of Excel." This means Microsoft is effectively bundling an analyst with the tool, using the model to teach the tool's business logic back to itself.

The conversation then explores two possible futures: one where humans remain in the loop with copilots and agents, and another where companies provision computing resources for fully autonomous AI agents. Nadella argues that even in the fully autonomous world, Microsoft's infrastructure becomes more valuable, not less. "Our business, which today is an end user tools business, will become essentially an infrastructure business in support of agents," he says. He reveals that Microsoft is already seeing significant growth in Windows 365 provisioning for autonomous agents — companies want to provision actual computers for their AI agents to use.

36:17MAI and Microsoft's Own Model Strategy

Dylan Patel raises a pointed question about Microsoft's own AI models: the most recent MAI model ranked 36th on Chatbot Arena, despite Microsoft having IP rights to OpenAI's models and the ability to fork their monorepo. Why is Microsoft behind, and how can investors have confidence that they'll catch up?

Nadella's response is strategic and nuanced. He explains that Microsoft is taking a dual approach: fully exploiting the OpenAI partnership for the next seven years while building its own world-class superintelligence team. "The last thing I don't want to do is use my flops in a way that is just duplicative and doesn't add much value," he says. Instead, Microsoft is using its compute for targeted optimizations — an image model that ranked 9th in the image arena, an audio model optimized for Copilot, and a text model that debuted at 13th on only 15,000 H100s. The next step is an omni model combining text, image, and audio capabilities.

Nadella reveals that Microsoft is assembling a world-class team under Mustafa Suleyman, with key hires including Karen, Amar Subramaniam (who led post-training at Gemini), Tufai, and Nando (who worked on multimedia at DeepMind). He promises that Mustafa will soon publish more clarity on the lab's direction. But the core strategic insight is that Microsoft doesn't need to be the single best model company — it needs to be the best hyperscaler supporting multiple models, while having enough model capability to compete effectively.

When pressed on the "intelligence explosion" scenario — where one model with continuous learning deployed across the entire economy creates an insurmountable lead — Nadella pushes back on the premise. "Even today, for all the dominance of any one model, it's not the case," he says. "Take any corporate coding, there's multiple models. In fact, every day it's less the case where there is not one model that is getting deployed broadly." He argues that network effects of continual learning will be domain-specific, geography-specific, and segment-specific, leaving room for multiple winners.

47:47The Hyperscale Business: Pause, Pivot, and Strategy

Dylan Patel reveals that Microsoft recently paused or walked away from several data center leasing deals, which were then picked up by Google, Meta, Amazon, and Oracle. This caused Microsoft's projected 2028 capacity to drop from 12-13 gigawatts to about 9.5 gigawatts, while Oracle is projected to grow from one-fifth Microsoft's size to larger than Microsoft by 2027. Why did Microsoft pull back?

Nadella's explanation is revealing. He says the fundamental issue was that Microsoft realized it didn't want to be a "hoster for one model company with limited time horizon." The hyperscale business, he argues, is about serving a long tail of diverse AI workloads — not just five contracts with five frontier labs. "That's not a Microsoft business," he says flatly. "That may be a business for someone else."

The other critical factor is technology generation risk. With Nvidia's pace of innovation accelerating — GB200s now, GB300s coming, then Vera Rubin Ultra — Nadella didn't want to be stuck with five years of depreciation on a single generation of hardware. He cites Jensen Huang's advice: "Get on the speed of light execution." The Fairwater 2 data center was built in 90 days from receiving hardware to handoff, and Nadella wants to maintain that cadence of continuous scaling rather than building massive capacity in one shot and then pausing.

Nadella also emphasizes that location matters enormously for inference workloads. Even if inference becomes asynchronous, data residency laws in Europe mean you can't round-trip a call to Texas. Microsoft needs capacity distributed globally, with the right balance of training, inference, and research compute. He reveals that Microsoft is now buying capacity from neo-clouds like Iris Energy, Nebius, and Lambda Labs, and welcomes them into the Azure marketplace — because customers who come through Azure will still use Microsoft's compute, storage, and databases.

1:02:44In-House Chips and the OpenAI Partnership

When asked about Microsoft's internal chip efforts, Nadella acknowledges that the company has been behind Google and Amazon in deploying custom accelerators at scale. Google is making 5-7 million TPUs, Amazon 3-5 million, while Microsoft's orders are "way below that number." But Nadella argues that the bar for a new accelerator is the total cost of ownership (TCO) of the previous Nvidia generation, and that Microsoft has a proven track record of managing heterogeneous fleets — Intel, AMD, and Cobalt (ARM-based) chips all running in Azure.

The key insight Nadella offers is that custom silicon makes most sense when there's a closed loop between the model and the chip. "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," he says. Microsoft's MAI models will drive demand for its own silicon, while the OpenAI partnership provides access to OpenAI's chip program. "We gave them a bunch of IP as well to bootstrap them," Nadella reveals. "And now as they innovate, even at the system level, we get access to all of it."

In a surprising disclosure, Nadella clarifies the terms of the new OpenAI agreement. OpenAI's API business (the PaaS layer) is Azure-exclusive. Their SaaS business (ChatGPT) can run anywhere. But crucially, if any partner wants to build a product with OpenAI that involves state — memory, databases, storage — they must come to Azure. "If Salesforce wants to integrate OpenAI, it's not through an API," Nadella explains. "They actually work together, train a model together, deploy it on, let's say Amazon. Now is that allowed? They will have to come to Azure." The only exceptions are for US government workloads.

1:09:35The CAPEX Explosion and Industrial Transformation

Nadella acknowledges that Microsoft is undergoing a fundamental structural transformation from a software company to a capital-intensive industrial business. Capex has tripled in two years, and other hyperscalers are taking massive loans — Meta did a $20 billion loan in Louisiana. But Nadella argues that the key to making this work is software-driven capital efficiency. "The hardware guys have done a great job of marketing Moore's law," he says. "But the improvement in tokens per dollar per watt that we are able to get quarter over quarter, year over year is massive — 5x, 10x, maybe 40x in some cases."

He frames the challenge as one of knowledge intensity driving capital efficiency. The ability to schedule workloads, evict jobs, manage fungibility across a heterogeneous fleet — these are software problems that determine whether the massive capital investment generates adequate returns. "That's the type of stuff that we have to be world class at," he says.

When asked how to think about investment in a world where AGI might arrive in 3 years (as Sam Altman anticipates) versus 20 years, Nadella offers a practical framework: treat research compute as R&D expense, not capital investment. "You should think of it as just R&D expense and you should say, hey, what's the research compute and how do you want to scale it?" The rest of the investment should be demand-driven, with the understanding that you can build ahead of demand but need a credible demand plan. He notes that labs projecting $100 billion in revenue by 2027-2028 have strong incentives to be optimistic, but their demonstrated traction makes the risk reasonable.

1:15:07Will the World Trust US Companies to Lead AI?

The final section tackles geopolitics and sovereignty. Dylan Patel points out that the US has dominated every previous tech stack — Windows, Office, semiconductors — and deployed them globally. But AI feels different because every government cares about it, and the world is becoming bipolar between the US and China. How does Microsoft navigate sovereign AI demands?

Nadella's answer is striking in its emphasis on trust. "The key, key priority for the US tech sector and the US government is to ensure that we not only do leading innovative work, but we also collectively build trust around the world on our tech stack," he says. He notes that the US has 4% of the world's population, 25% of GDP, and 50% of market cap — and that 50% exists because the world trusts US capital markets and technology stewardship. "If that is broken, then that's not a good day for the United States."

He argues that American companies should take credit for their foreign direct investment — AI factories being built all over the world by American companies is the best marketing the US could have. Microsoft is making concrete commitments to Europe on data sovereignty, building sovereign clouds in France and Germany, and offering confidential computing with GPU support. "I feel very, very good about being able to build, both technically and through policy, this trust in the American tech stack," he says.

When pressed on whether countries will demand locally trained models or just locally hosted weights, Nadella argues that the key is continuity and resilience. Open-source models provide a check against concentration risk, and countries want the ability to move their data and liquidity to another model if needed. He draws a parallel to semiconductors: everyone knows TSMC is better, but the pandemic taught the world that resilience matters. "Globalization was fantastic, it helped supply chains be globalized and be super efficient. But there's such a thing called resilience."

Conclusion

What stays with the listener is Nadella's remarkable ability to hold two seemingly contradictory positions simultaneously: he is investing hundreds of billions in AI infrastructure while arguing that the technology is still in "early innings"; he is building world-class models while insisting that the scaffolding layer may capture more value; he is deepening the OpenAI partnership while preparing for a world where Microsoft competes with its own models. The episode matters because it offers a rare window into how the CEO of the world's most valuable company is thinking about the most consequential technology transition in decades — not with hype, but with the grounded pragmatism of someone who has to make capital allocation decisions that will play out over 50 years. The central tension is whether Nadella's bet on fungibility, openness, and infrastructure will prove more durable than the model companies' bet on intelligence as the ultimate moat.

Key takeaways

• Microsoft's Fairwater 2 data center represents a 10x increase in training capacity from GPT-5, with 5 million network connections and the ability to link multiple buildings across regions for single training jobs.
• Nadella believes the scaffolding layer (data, context engineering, observability) may capture more value than the model layer, because open-source checkpoints will always be available for fine-tuning with proprietary data.
• GitHub Copilot's market share dropped from near 100% to sub-25% in one year as competitors emerged, but Nadella sees this as evidence of massive market expansion, not failure.
• Microsoft is taking a dual model strategy: fully exploiting the OpenAI partnership for seven years while building its own world-class superintelligence team under Mustafa Suleyman.
• The company paused data center expansion not because of demand concerns, but to avoid being stuck with a single generation of hardware and to focus on building fungible infrastructure that supports multiple model families.
• Microsoft's custom silicon strategy is tied directly to its MAI models, with the closed loop between model and chip providing the "birthright" for custom accelerators.
• Under the new OpenAI agreement, the API business is Azure-exclusive, and any partner building stateful products with OpenAI must use Azure infrastructure.
• Nadella frames sovereign AI demands as a legitimate business requirement, arguing that American tech companies must build trust globally through data residency, confidential computing, and local investment.