The Campus Is the Product
Hyperscalers are committing $300–700B to AI infrastructure. Satya Nadella's framing reveals what's really being built: the campus itself is now the competitive moat.
Somewhere in the transition from “cloud computing” to “AI infrastructure,” we crossed a threshold I don’t think the industry has fully processed. The datacenter used to be a cost center — a necessary back-end that you kept off the product roadmap and managed for efficiency. What is being built right now is something different. The campus is the product.
Let me make that concrete with some numbers.
In 2024, the combined capital expenditure of Microsoft, Google, Amazon, and Meta crossed $200 billion for the year. In 2025, that roughly doubled. By 2026, Fortune is reporting the four hyperscalers are on track to spend approximately $700 billion combined on AI infrastructure in a single year — up 69% from 2025’s already-elevated baseline. Goldman Sachs projects cumulative capex from these four companies of $5.3 trillion from fiscal 2025 through 2030.
These are not software companies spending money on servers. This is infrastructure spending at a scale that rivals national energy grids.
What Satya Nadella Is Actually Saying
Satya Nadella has been one of the clearer voices on what this investment actually represents. At the Morgan Stanley TMT Conference in early 2026, he described the challenge as managing “a capital-intensive business” but doing so “using all of the levers that software gives us in managing TCO, managing utilization, optimizing the kernels by workload, ensuring that there’s a diverse class of customers.” His goal, stated directly: generating strong return on invested capital.
That framing matters because it reveals the mental model. Nadella is not thinking about AI infrastructure the way software companies thought about servers in 2005. He is thinking about it the way a capital-intensive manufacturer thinks about plant investment — long-duration assets, utilization rates, return on invested capital, TCO optimization across the full stack.
Microsoft spent the equivalent of building two gigawatts of datacenter capacity in fiscal year 2025. At SemiAnalysis’s estimate of roughly $50 billion per gigawatt of serious AI capacity, that represents $100 billion of infrastructure in a single fiscal year. Nadella called the Fairwater datacenter in Wisconsin — which connects hundreds of thousands of Blackwell B200 GPUs into a single seamless cluster — “the world’s most powerful AI datacenter.”
That language is not accidental. The datacenter itself is now a product announcement.
The Campus as a System
Here is what I find genuinely interesting from a systems design perspective. The challenge of building a serious AI campus at current scale is not about any single component — it is about integrating every component into a system that operates reliably at the limits of what each component can do.
A serious AI campus circa 2025-2026 has to solve:
Power at scale — hundreds of megawatts, increasingly with on-site generation because grid capacity in most locations cannot provision that much power on the timelines required. Aeroderivative gas turbines, fuel cells, and eventually small modular reactors are all being evaluated. This is an energy problem that requires energy expertise most tech companies did not have five years ago.
Cooling at density — liquid cooling is now standard in serious deployments. The old model of air-cooled server rows cannot handle the heat density of a rack of B200s. Direct liquid cooling, rear-door heat exchangers, immersion cooling experiments — the thermal engineering is a significant constraint on GPU density and therefore on the economics of the campus.
Networking at bandwidth — the AI campus runs on high-bandwidth interconnects between GPUs, between racks, between buildings. InfiniBand, NVIDIA NVLink at rack scale, emerging photonic interconnects. Bandwidth is the circulatory system of the AI factory. Starving it constrains every inference and training job.
Storage tiered appropriately — training requires enormous data ingestion at high throughput. Inference requires fast access to model weights, which can be 100-700GB for large models. The storage architecture of a training cluster is radically different from an inference cluster, and a campus often runs both.
Software and orchestration gluing it all together — scheduling workloads across thousands of GPUs, managing failures gracefully, optimizing for utilization while maintaining latency guarantees. This is why Google’s years of experience with Borg and Kubernetes gave them an operational advantage in early AI infrastructure; they already knew how to manage distributed systems at this scale.
The integration of all these layers, running reliably at production scale, with defined SLAs, is the actual product. Not the AI model. Not the cloud console. The campus.
The Moat That Is Being Built
I have a strong opinion here, and I want to be clear it is an opinion: what is being built right now is a form of infrastructure moat that will be very hard to replicate for the next decade.
Compute moats in software have historically been soft — if AWS builds something, Azure can build a comparable thing in 18 months. The cloud era democratized access to compute in a way that genuinely leveled the playing field for builders.
The AI campus moat has harder components. Land — secured in the right locations, with power rights negotiated, in jurisdictions that have approved permitting. Grid interconnection — with wait times averaging five years in many US jurisdictions, the companies that started securing interconnection rights in 2022 and 2023 have a multi-year head start. Utility relationships — the negotiated contracts for dedicated power substations and transmission capacity are not easily replicated. Liquid cooling expertise, embedded in facilities engineering teams. Custom silicon — as I described in an earlier post, the hyperscaler custom chips (TPU, Trainium, Maia) represent years of silicon engineering investment that a new entrant cannot accelerate.
The companies building this infrastructure are not just building for today’s workloads. The campus that takes five years to fully commission is being designed for model generations that do not exist yet. That is a significant bet on the direction of AI progress, and it requires a conviction about where things are going that extends well beyond the next product cycle.
Satya Nadella has been explicit about this: the infrastructure investment is predicated on a belief that AI demand will continue to compound, that the current generation of AI products will be replaced by more capable and more usage-intensive successors, and that the competitive position in AI infrastructure will prove durable in a way that software-only positions often don’t.
What It Means for Everyone Else
If you are not a hyperscaler — and nearly everyone reading this is not — this infrastructure buildout shapes your world in ways worth thinking about.
Access to frontier AI capability is going to be mediated through cloud APIs for the foreseeable future, because the campus required to run the largest models is not something you can build. This is not a criticism; it is a design reality. The question for builders is what to build on top of that infrastructure, and how to build it in a way that creates durable value.
The interesting flip side: all of this infrastructure is being built to be rented. Every gigawatt of AI capacity that gets built eventually comes to market as inference tokens, training compute, or cloud services. The supply is growing fast, which means inference costs will continue falling. A model that costs a dollar per million tokens today will cost ten cents in two years. That changes what is economically viable to build.
I have been watching Zero, the project I am building at zero.university, through this lens. The infrastructure investments being made now are not barriers for people who want to build on top of AI. They are the opposite — they are the foundation that makes building possible without owning the foundation layer. The question is not whether you have access to a campus. The question is what you build with the access you have.
The campus is the product. But the campus is also a platform. And platforms create ecosystems. The most interesting part of the next decade might not be the campus itself but what gets built on top of it.
