Full text 4,000 words | Reading time about 13 minutes
(Microsoft CEO Nadella discusses AI industrial revolution)
On November 12, 2025, in Atlanta, USA, a two-story data center was lit up.
Its name is Fairwater 2. On the surface, it's just another cloud computer room. The real specialty is underground: connected via high-speed fiber to the Fairwater facility in Wisconsin, 700 miles away, spanning 5 states.
Microsoft named this system not a campus or cluster, but: "Planet-scale AI Superfactory".
The biggest difference from traditional cloud data centers is what it does. Ordinary facilities serve thousands of applications, with each customer getting a small slice of resources; the AI Superfactory does one thing: coordinates GPUs scattered around like an assembly line to train and run next-generation AI large models.
In an interview the next day, Microsoft CEO Satya Nadella characterized it: this is an industrial revolution.
If lean production redefined manufacturing, then AI is redefining knowledge work.
And the starting point of this revolution is not releasing yet another killer app, but first building this generation's power plants and factories.
The Superfactory is the true starting point of AI.
Section 1 | Not Models, but Factories: Microsoft's AI Strategy Shifts Gears
While most companies are still competing over whose model is stronger, Nadella's focus in the interview is different:
We are truly focused on the underlying economic structure.
The so-called underlying layer is not the model capabilities themselves, but the foundational layer supporting the long-term operation of the entire AI system: power scheduling, GPU clusters, bandwidth networks, data center siting, inference architecture design. Microsoft no longer treats AI as a single product, but as a systems engineering project.
✅ How large is this factory?
(Fairwater 2 promo video: Microsoft is building the world's first AI Superfactory)
The Atlanta Fairwater 2 data center has 5 million network connections, with fiber optic cable volume equivalent to the total of all Microsoft Azure data centers two and a half years ago. Its training capacity is 10x what GPT-5 requires. Microsoft's goal is to increase training capacity 10x every 18 to 24 months.
More critically, the connection method. Through a 1 Petabit high-speed network, Fairwater 2 connects to the Milwaukee data center in Wisconsin. 700 miles apart, spanning 5 states, yet treated as one machine in the system for scheduling.
Standing in the noisy data center, Nadella joked: I run a software company, welcome to this software company.
Behind the joke is a pivot: Microsoft was once a typical software company, earning high profits from Windows and Office licenses. Now, they are building gigawatt-scale data centers, hundreds of thousands of GPUs, thousands of miles of high-speed fiber networks.
This is not just a change in investment scale. Nadella later said: Microsoft is now a capital-intensive business and a knowledge-intensive business.
✅ Not stacking GPUs, but building systems
But Microsoft is not becoming a hardware company; it's doing AI in a new way.
Nadella is clear: We can't build a moat by leading with one model; we need a system that continuously provides inference services to users.
In other words, models are just midstream processes in the AI economy; what truly determines long-term value is the generation, scheduling, and stable supply of tokens.
The key is not being locked into one generation of hardware.
To build Azure that excels at all AI stages, design an architecture flexible to hardware iterations. Quick deployment for GB200, not dragged by prior facilities for GB300, adaptable to Vera Rubin Ultra's different power density and cooling needs.
This is Microsoft's current thinking: not one powerful AI, but a sustainable, reusable, globally deliverable smart factory system.
✅ Architecture supporting factory operations
Microsoft internally calls this the AI Factory three-layer architecture:
• Training layer: GPU compute resources for GPT-5 and subsequent models
• Inference layer: Global response speed, real-time Copilot service
• Interface layer: Embed AI into development, office, search, and daily scenarios
At the Fairwater 2 construction site, Microsoft Cloud and AI EVP Scott Guthrie stated clearly: "The future is not one model winning and ending, but who can make token generation, inference, and delivery a closed-loop system."
This is their AI industrial revolution: not competing at the model layer, but rebuilding the entire underlying system from factories.
Section 2 | Data Centers: No Longer Cloud Warehouses, but AI Power Plants
In the past, data centers stored files and handled cloud tasks. To most enterprises, they were like warehouses: stable, scalable, cost-controlled.
But to Nadella, this definition is completely outdated.
Traditional data centers were designed for cloud; what we're doing now is rebuilding entire data centers for AI.
This isn't just adding servers; it's fundamentally changing function and structure. Scott Guthrie gave a precise positioning: We're turning data centers into AI power plants.
✅ Why power plants?
AI isn't just for training models; it must provide inference services daily at scale. Requirements for data centers have completely changed:
• Continuously output tokens, like power plants generating electricity
• Quick global response, like grid power scheduling
• Low latency, high throughput, precise scheduling
This requires Microsoft to rebuild the architecture: not cloud warehouses stacked with servers, but AI factories with supply capacity.
✅ Reconstructing core data center components
Guthrie noted Microsoft is reconstructing four core components for AI data centers:
1. Chip deployment logic - Previously storage-optimized, now for inference and training
2. Liquid cooling systems - Advanced cooling for lower energy and heat loads
3. Network connection structure - From API-facing to serving billions of global requests
4. Siting logic - From near customers to near clean energy and stable power
These designs must adapt to rapid hardware iterations. He quoted Nvidia CEO Jensen Huang: Execute at light speed.
What is light speed?
Atlanta Fairwater 2 went from acquisition to live workload in about 90 days. That's the execution speed Microsoft aims for each hardware generation.
✅ Complete ecosystem for AI workloads
These factories are deployed globally, not just one or two.
More importantly, Microsoft realizes: Each AI workload needs more than AI accelerators; much of profit will come from those other things.
What other things? Storage, databases, identity management, observability tools. AI inference is just the tip; full workloads need complete cloud support.
This explains why Microsoft considers data residency laws and EU boundaries. You can't just round-trip calls anywhere, even async. Need regional high-density facilities balancing power costs and regulations.
Nadella emphasizes: We're building a global AI grid supporting Copilot real-time across regions and time zones.
Microsoft reconstructs data centers not to launch models faster, but to build truly usable, controllable, profitable AI infrastructure.
But with power plants built, next is the grid.
Section 3 | AI-WAN: What Microsoft is Building is a Global Token Network
An invisible grid.
This grid has an internal Microsoft name: AI-WAN (AI Wide Area Network).
Unlike traditional cloud siloed by region, AI-WAN requires tighter inter-data-center linkage for intelligent scheduling. E.g., during Asia peak, pull idle capacity from US or South America, like cross-continent power dispatch.
The system's core goal: Every user instruction gets immediate AI compute response.
But why cross-data-center scheduling?
✅ Model parallelism + data parallelism
Nadella revealed a key design: You can see model and data parallelism. It's built for campus training tasks, super systems. Then via WAN, connect to Wisconsin data center, aggregate all resources for a training job.
What does this mean?
Fairwater 2 and 4 connected by 1 Petabit network can jointly do massive training, then switch to data generation or inference. Resources not forever tied to one workload.
Host asks: As AI tasks grow—30s for inference prompt, 30min for deep research, hours for software agents—why does data center location matter?
Nadella: "As model capabilities evolve and token usage changes, sync or async, you don't want disadvantage. That's why we think about Azure region layout and inter-region networks."
✅ Three-layer scheduling architecture
To realize AI-WAN, Microsoft built three-layer scheduling:
• Campus-level: Model parallelism for high-density training in single DC
• Regional-level: High-speed WAN for cross-state DC collaboration on large training
• Global-level: Dynamic inference allocation by workload type (sync/async) and data regs
Guthrie added: Databases and storage must be near compute. If Cosmos DB near Fairwater for session data or autonomous txns, it must be close.
This isn't simple networking; it's co-design of compute-storage-network architecture.
✅ From fixed workloads to fluid compute
E.g., Copilot email needs dozens to hundreds tokens. Unstable scheduling causes lag or failure. Microsoft solves every link from prompt to response.
Behind: Can latency be ms-level? Bandwidth crash at peak? Cache hit high enough to avoid recompute?
These details determine: Can AI supply stably like utilities?
Nadella direct: We're building a new supply network for inference capacity.
Tokens become new commodity, means of production. Microsoft masters global AI compute distribution.
When ubiquitous, users won't notice; sentence typed, result appears.
AI infra success: User imperceptible, system omnipresent.
Section 4 | Why Did Microsoft Hit the Brakes in 2023?
Grand AI-WAN blueprint, Fairwater 2 online smoothly—seems on track.
But Microsoft didn't charge ahead.
In H2 2023, amid fierce AI infra race, Microsoft surprised: paused batch of planned data center leases.
Why brake at peak competition?
✅ Not Be One Company's Hosting Provider
Nadella direct: We don't want to just be one company's hoster with massive business from one customer. That's not a business.
Points to Oracle model. Oracle grew from 1/5 Microsoft scale to possibly surpassing by 2027 end via large AI lab bare-metal. Though 35% margins, Nadella: Limited-contract hosting for one model co meaningless.
Any large-scale co eventually hyperscaler itself.
So Microsoft builds hyperscale network for long-tail customers, not bare-metal for few big ones.
✅ Software Optimization vs Hardware Costs
Microsoft capex tripled in 2 years. Others borrow for builds, FCF to zero.
Host: What's happening?
Nadella: We're capital and knowledge intensive. Must use knowledge to boost capex ROIC.
Meaning: For given GPT series, throughput (tokens per watt per dollar) grows massively quarterly/yearly via software: 5x, 10x, even 40x.
That's knowledge-driven capital efficiency.
Hardware vendors market Moore's Law; Microsoft fights costs with software. Classic hoster vs hyperscaler diff? Software.
✅ Market Share Drop Not Bad
Host notes: GitHub Copilot share from ~100% to <25%, chased by Cursor, Claude Code, Codex.
Nadella surprising: Shows market expanding fast.
Two reasons:
• First, Copilot still #1.
• Second, All listed cos born last 2-3 years.
To him, not share loss, market growth. Logic: Better 25% of big market than 100% of small. AI coding market vastly larger than past high-share biz.
This "market > share" logic permeates Microsoft decisions.
To Nadella, decisions clear industrially. Not chasing period margins, solving unique Microsoft problems.
Leads to decisions:
• Treat some spends as R&D, no short-term ROI force
• No blind overbuild, follow demand
• Flexible compute via lease, custom, GPUaaS etc.
• Welcome new cloud providers to Azure ecosystem
So 2023 pause not retreat, strategic adjust.
Microsoft seems slower, actually building 10-year growth system.
From DCs to AI-WAN, hardware iter to software opt, Microsoft doing bottom-up AI economy industrial revolution.
Revolution starts in these invisible infrastructures.
Conclusion | In This Industrial Revolution, Which Layer Are You On?
Microsoft's real investment logic?
Rebuild DCs not for storage, but power supply; AI-WAN not connection, but scheduling; Copilot not demo, but closed loop.
Core strategy: Not chase model power, master token gen, transmit, monetize.
From this, Microsoft not releasing AI products, quietly laying global smart grid.
So, in AI industrial revolution, which layer?
• App layer, eyeing stronger models hotter products;
• Model layer, competing params speed;
• Or infra layer, building DCs, power sched, nets?
Nadella's answer: Key not model strength, infra stability.
AI battlefield sunk to base layer.
Next opportunity under your feet.