Satya Nadella — Microsoft’s AGI plan & quantum breakthrough

Summary

• Hyperscalers are the two-confident-calls winner: “if intelligence is log of compute, whoever can do lots of compute is a big winner,” and agents exponentially multiply demand — “one human invoking programs that invoke lots more programs.” Models, by contrast, won’t be winner-take-all: enterprise buyers “will not tolerate” it, open source is a structural check, and models still need hyperscale compute, storage and state — “that nexus… is going to be there forever.”
• Capex has a governor: Microsoft reported $13B in AI revenue; Nadella points to real revenue as a governor because “you can go off the rails completely when you are hyping yourself with the supply-side.” Nadella calls AGI self-claims “nonsensical benchmark hacking” — his real benchmark is Industrial-Revolution-type, inflation-adjusted growth in the developed world, not self-claimed AGI milestones. Dwarkesh’s pushback stands: if 10% growth on a $100T economy is at stake, why $80B and not $800B?
• Overbuild is coming and Nadella wants to rent it: “I’m so excited to be a leaser… I’m going to be leasing a lot of capacity in ‘27, ‘28… the only thing that’s going to happen with all the compute builds is the prices are going to come down.” Companies AND countries will race to deploy capital.
• Majorana One is pitched as quantum’s transistor moment: a topological-qubit chip that Nadella says is the “first chip that will be capable of a million qubits, physical,” with a fault-tolerant machine “maybe ‘27, ‘28, ‘29” — and quantum + AI as a hoped-for way to compress “250 years of chemistry” into 25.
• SaaS is structurally at risk: “the business logic will go more into this agentic tier,” CRUD apps will be fundamentally changed, while Office is repositioned as “the UI layer for knowledge work” and Copilot as “the UI for AI.” Muse, the new world human action model, makes gaming data “to Microsoft what YouTube is perhaps to Google.”
• The rate limiters aren’t technical: no deployment without a human indemnifying the AI — “before it is a real problem, the real problem will be in the courts” — plus return on labor (“you can’t just have a return on capital and no return on labor”) and enterprise change management, which he calls “Lean for knowledge work.”

Deep dive

1. This wave is more full-stack than anything since the 90s

• Nadella’s opening frame: AI echoes the early-90s RISC-vs-CISC and Windows NT era — “the entire thing is being litigated,” from silicon to OS to app tier — and “this does feel a little more… full-stack than even the past.”
• The formative miss he keeps returning to: Microsoft embraced the browser well, but “missed what turned out to be the biggest business model on the web” — nobody expected search to organize a distributed web; “Google saw it and executed super well.” The lesson: “you have to not only get the tech trend right, you also have to get where the value is going to be created” — business-model shifts are “probably tougher than even the tech trend changes.”

2. Where value accrues: hyperscalers yes, single-model dominance no

• His two confident calls: hyperscalers win — “if intelligence is log of compute, whoever can do lots of compute is a big winner” — and agents will “exponentially increase compute usage” since “you’re not even bound by just one human invoking a program.” He thinks of his fleet as “a ratio of the AI accelerator to storage, to compute,” not just GPUs.
• Above the infrastructure layer “it becomes a little fuzzy,” and Nadella’s claimed core skill is knowing “what are winner-take-all markets and what are not — in some sense, everything.” Enterprise buyers “will not tolerate winner-take-all” — the same read that carried him past investors telling him Azure vs. Amazon was “game over.”
• On models: “there will definitely be an open-source alternative” checking closed-source — the Windows lesson — and “the state is not going to sit around and wait for private companies.” Consumer is the exception: ChatGPT “has already got real escape velocity,” always in the App Store top five.
• Dwarkesh’s pushback — worth keeping: cloud was “just a chip and a box” and margins emerged anyway; if AGI helps build better AIs, the lead compounds. Nadella’s partial concession: “At scale, nothing is commodity” — Azure’s 60-plus-region know-how “is just a tough thing to duplicate” — but models still need hyperscale compute, state, storage: “that nexus, I feel, is going to be there forever.”

3. Fleet philosophy: ride Moore’s law, lease into the overbuild

• The build-out logic: refresh the fleet yearly, depreciate, and get “very very good at placement” for high utilization across big training jobs, test-time compute, and RL distillation (“kind of like more training flops”). Pre-training “at some point has to cross data center boundaries” anyway — and “speed of light is speed of light,” so serving requires an inference fleet everywhere, with storage and compute co-located.
• The overbuild call, verbatim: “the memo has gone out” on energy and compute, countries will deploy capital alongside companies, and “I’m so excited to be a leaser… I’m going to be leasing a lot of capacity in ‘27, ‘28… the only thing that’s going to happen with all the compute builds is the prices are going to come down.”

4. The real AGI benchmark is 10% world growth — and $13B revenue is the governor

• Against the $13B AI revenue (Dwarkesh extrapolates 10x in four years): before Microsoft’s revenue, “the first thing we have to observe is GDP growth.” Developed world runs 2% — “if you adjust for inflation it’s zero.” If this is the Industrial Revolution, prove it: “10%, 7%, developed world, inflation-adjusted, growing at 5%. That’s the real marker.”
• The signature line: “Us self-claiming some AGI milestone, that’s just nonsensical benchmark hacking to me.” And “the big winners here are not going to be tech companies” — it’s the broader industry consuming abundant intelligence.
• Dwarkesh presses the contradiction: Microsoft bet on OpenAI in 2019 pre-revenue, and the railway build-outs lost money while building the future — so why not hundreds of billions now? Nadella: supply and demand “have to map”; you need “existence proof that you are able to parlay yesterday’s capital into today’s demand” before investing “maybe exponentially even.” It’s “not a race to just building a model, it’s a race to creating a commodity that is getting used.”

5. Jevons, DeepSeek, and “Lean for knowledge work”

• Dwarkesh’s challenge on Jevons: intelligence is already two cents per million tokens — he’s bottlenecked on smarts, not price. Nadella: both must happen — DeepSeek-style breakthroughs bend the performance-per-token curve and expand demand, exactly as cloud did (“the cloud in India is so much bigger than anything we did in the server era”). Cheap healthcare tokens in the Global South “would be the biggest change ever.”
• The real deployment challenge is “change management or process change.” His analogy, as told: pre-PC forecasting ran on faxes and interoffice memos until “somebody said, hey, I’m just going to take an Excel spreadsheet, put it in email, send it around” — the artifact and workflow changed together. AI needs the same reinvention: “This is like Lean for knowledge work” — reduce waste, increase value — “and that’s going to take its time.”
• His own practice: at least ten Copilot agents, podcast prep done by Copilot (it found both Nature papers; “go give it to me in a podcast format”). The coming scaffolding is an agent manager — “a new inbox… my millions of agents will have to invoke some exceptions to me” — with Copilot as “the UI for AI”: agents do the knowledge work, but “you still have a knowledge worker.”

6. Majorana One: the “transistor moment” — with a hedged timeline

• A 30-year bet (“I’m the third CEO of Microsoft who’s been excited about quantum”): the thesis was always that “you need a physics breakthrough to build a utility-scale quantum computer.” The Nature result is existence proof of Majorana zero modes — theorized in the 1930s — in a new topological phase of matter that lets them “reliably hide the quantum information, measure it, and fabricate it.”
• The claim: Majorana One is “the first chip that will be capable of a million qubits, physical,” thousands of error-corrected logical qubits, on a chip this size — without which “you’ll never be able to build a utility-scale computer.” Timeline, exactly as hedged: “maybe ‘27, ‘28, ‘29, we will be able to actually build this” fault-tolerant machine. Separately, 24 logical qubits already announced via neutral-atom/ion-trap partners — software stack deliberately decoupled from hardware.
• The usage model: “AI is like an emulator of the simulator. Quantum is like a simulator of nature.” Quantum suits data-light, exponential-state exploration — chemistry, physics, biology — generating synthetic data that trains better AI models; he hopes to swap quantum in for pieces of today’s HPC stack, en route to compressing “250 years of chemistry” into 25 for the move off carbon.
• On funding 20-year bets: MSR’s budget goes in first, “knowing that most of these bets are not going to pay off in any finite time frame. Maybe the sixth CEO of Microsoft will benefit.” The hard part is capitalizing — “I can tell you the thousand projects from MSR that we should have probably led with, but we didn’t.” Related: “longevity is not a goal; relevance is,” and for “mere mortal CEOs, it’s more like refounder mode.”

7. Muse: gaming data is Microsoft’s YouTube

• Muse, the world human action model (the second Nature paper), is trained on gameplay data to generate games that are consistent, diverse, and “persistent to user mods.” Phil Spencer’s demo — an Xbox controller driving model-generated output — was for Nadella “kind of like the first time we saw ChatGPT complete sentences.”
• The strategic read: “gaming data is to Microsoft what YouTube is perhaps to Google” — though he insists gaming must be an end in itself (“Flight Simulator was a Microsoft product long before Windows”), not a means to models.
• The three bets set five-to-seven years ago — AI, quantum, mixed reality — mapped as business-logic breakthrough, systems breakthrough, and UI (presence). An honest change of mind on MR: “I thought it was going to be more solvable. It’s tougher, perhaps, just because of the social side of it” — continuing via Anduril/Palmer on IVAS.

8. The rate limiters: courts, labor, and alignment engineering

• On his 2017 “new species” line: trust comes first, and “the one biggest rate limiter will be how does our legal infrastructure evolve” — a world built on humans owning property and being liable. Categorical: “you cannot deploy these intelligences unless and until there’s someone indemnifying it as a human.” On takeoff risk: “before it is a real problem, the real problem will be in the courts.”
• Dwarkesh’s counter: some other society’s legal system might be more amenable — takeoff needn’t happen in America. Nadella: rogue actors and rogue states exist, but “the world doesn’t sit around and say we’ll tolerate that.”
• Second limiter: “you can’t just have a return on capital and no return on labor” — that 60% of labor has to be revalued for democracies to stay stable; “what is today considered high-value human labor may be a commodity.”
• From Sydney Bing to agent swarms: allocate compute explicitly for alignment, build observability like cyber (“we don’t just write software and then let it go”), sandbox runtimes, control whether agent-written code is ephemeral, treat physical embodiment as a threshold — and take “our own liability,” because otherwise “the social permission for that is not going to be there.”

9. SaaS shifts toward the agentic tier; “knowledge work ≠ knowledge worker”

• Could LLMs commodify Office? His answer via the tumor-board doctor demo: Copilot reasons over SharePoint cases to build the agenda, transcribes the Teams meeting into a recallable database entry, then turns it into a teaching deck. “Office is not just about the office of today; it’s the UI layer for knowledge work.”
• The sharper call: today’s SaaS “CRUD applications are going to fundamentally be changed because the business logic will go more into this agentic tier” — but with the world’s IT backlog, code-gen plus agents means “the greatest explosion of apps, they’ll be called agents.” Incumbents “can’t stay still.”
• On whether AGI is even a thing — the episode’s deepest disagreement: “Don’t conflate knowledge worker with knowledge work.” Today’s work gets automated (“who said my life’s goal is to triage my email?”), new cognitive labor emerges. Dwarkesh’s horses analogy gets this reply: “it’s only 200 years of history where we have valued some narrow sort of things called ‘cognitive labor.’” And yes, he’d welcome an AI facilitator in board meetings — humans are Herbert Simon’s “bounded rationality,” and “a cognitive amplifier outside, that’s great.”