Data Centres & Cloud Infrastructure
Layer 03 of seven · Data Centres & Cloud Infrastructure
- Why it matters
- This is where compute becomes a service someone can buy. The choice between renting from a hyperscaler, colocating your own fleet or committing to a specialised operator determines the cost structure of every AI product above it.
- The bottleneck
- Powered land with a secured grid connection, and the capital to fill it. Construction is rarely the slow part.
- Who captures value?
- Operators with long term contracted tenants, cheap capital and high utilisation. Empty capacity is expensive, and depreciating GPU fleets are unforgiving of idle time.
- What could change?
- A loosening of accelerator supply would compress specialised cloud pricing quickly, and a shift of inference toward smaller models or edge devices would change where capacity is needed.
The buildings, campuses and cloud platforms where AI actually runs. Hyperscale cloud, colocation, specialised AI and HPC facilities, and sovereign infrastructure are four different businesses.
If chips are the brains, this is the ground they stand on. Hyperscale cloud sells capacity as a service at enormous scale. Colocation sells space, power and the ability to interconnect with everyone else in the building. Specialised AI operators build for dense GPU clusters and little else. Sovereign infrastructure exists because some data is not permitted to leave a jurisdiction. The economics of each are quite different, even though they all look like large buildings full of computers from the outside.
Tap or hover a box to see what it does
Hyperscale cloud
The largest platforms sell compute, storage and managed AI services globally, organised into regions and availability zones so that a failure in one place does not take the service with it. Their own AI campuses are increasingly described as AI factories: purpose built sites optimised for training and serving models rather than for general computing.
Microsoft
Runs Azure, distributes AI through Microsoft 365 and Copilot, partners commercially with OpenAI, and designs its own Maia accelerators.
Few businesses touch as many parts of the ecosystem at once. It owns infrastructure, a commercial route to frontier models, and the enterprise distribution to sell the result.
Very heavy capital spending against uncertain payback, dependence on partner model roadmaps, and buyers questioning per-seat AI pricing.
Amazon
Runs AWS, designs Trainium and Inferentia accelerators, hosts third party models through Bedrock, and invests in Anthropic.
The largest cloud business is trying to own its own accelerator roadmap rather than rent it, which is the clearest example of the custom silicon threat to merchant GPUs.
Custom silicon needs software adoption to matter, and cloud growth is now compared against very large numbers.
Alphabet (Google)
Builds Gemini models through Google DeepMind, designs TPU accelerators, operates Google Cloud, and distributes AI through Search, Workspace and Android.
It is one of the few organisations that designs the chip, runs the data centre, trains the model and owns the consumer surface it ships on.
AI answers can cannibalise the advertising business that funds everything else, and antitrust proceedings could reshape distribution.
Oracle
Runs Oracle Cloud Infrastructure with a strong GPU cluster business, and sells the databases and applications that hold a great deal of enterprise data.
It sits in two layers at once. The database estate gives it a claim on enterprise data, and OCI has become a meaningful venue for large scale AI training capacity.
Building GPU capacity is capital hungry and the customer list is concentrated. Committed backlog is a promise, not yet revenue.
Colocation and interconnection
Customers place their own equipment in someone else's building and connect to networks, clouds and each other. Power density per rack is the number that decides whether a given facility can host modern AI hardware at all.
Equinix
Operates colocation data centres where customers place their own equipment and interconnect with each other.
Colocation is where networks meet. The value is less in the floor space and more in the density of parties you can connect to inside the building.
Older facilities were not designed for very high rack densities, and retrofitting power and cooling is expensive.
Digital Realty
Develops and leases large scale data centre capacity to hyperscalers and enterprises.
Land with a secured power connection has become the genuinely scarce asset. Owning it in the right markets is most of the business.
Development is capital intensive and rate sensitive, and a small number of tenants account for a large share of leasing.
Specialised AI and HPC clouds
Operators built specifically around dense GPU clusters, high speed interconnect and the workloads that need them. Sovereign AI infrastructure is a related category, driven by rules about where data and models may physically reside.
CoreWeave
Operates GPU cloud infrastructure purpose built for AI training and inference rather than general computing.
Specialised operators showed that a cloud tuned narrowly for AI clusters could win work from the incumbents, at least while capacity is scarce.
Debt funded fleets depreciate, revenue is concentrated in a few large contracts, and pricing softens if GPU supply loosens.
Nebius Group
Operates AI focused cloud infrastructure in Europe. It emerged from the old Yandex international business after the Russian operations were sold in 2024, with the Dutch parent keeping the global assets.
European customers with data residency requirements need somewhere local to train and serve. Sovereign preference is a real commercial tailwind.
Sub-scale relative to the hyperscalers, capital hungry, and still proving durable demand.
Repurposed and converted sites
Bitcoin mining, and historically Ethereum mining, required many of the same ingredients now sought by AI infrastructure developers: large power allocations, suitable sites and industrial cooling. Ethereum itself moved from proof of work to proof of stake in September 2022, so it is no longer mined.
Core Scientific, IREN, Applied Digital, Cipher Mining
Operators that built large powered sites for cryptocurrency mining and now convert or develop capacity for AI and high performance computing tenants.
They hold the two things that take longest to acquire: energised sites and power contracts. That is why AI landlords started paying attention to mining companies.
Converting a mining shed into an AI facility is a substantial rebuild, not a relabel. Financing, cooling and network requirements are all different.
How to think about the economics
- GrowthVery high
How quickly demand in this part of the ecosystem is expanding.
- Capital intensityVery high
How much money has to be spent up front before revenue arrives.
- Competitive moatModerate
How difficult it is for a credible new entrant to take the business.
- Customer concentrationHigh
How much revenue depends on a small number of buyers.
- Disruption riskModerate
How exposed the layer is to a technical or commercial shift.
This is a framework for thinking about the economics of a layer, not a recommendation. An important AI company, a strategically advantaged company, an investable security and an attractively valued security are four different things.
Where value may accrue
Capital intensity, power access, financing cost and utilisation. This layer rewards patient balance sheets rather than clever ones, and it punishes anyone who builds speculatively into a soft market.
Key terms in this layer
- Data centre
- A building full of racked computers with industrial power and cooling. AI data centres are built around dense GPU clusters rather than ordinary web servers.
- ExampleA single large AI campus can draw as much electricity as a small city.
- Related termsGPUHyperscalerThe gridNetwork switch
- Hyperscaler
- The handful of companies operating cloud infrastructure at global scale: Amazon, Microsoft and Alphabet. They own the largest GPU fleets and rent capacity to everyone else.
- ExampleAWS, Azure and Google Cloud.
- Related termsNeocloudData centre
- Neocloud
- A cloud provider built purely for AI workloads rather than general web hosting. Dense GPU clusters, rented by the hour.
- ExampleCoreWeave.
- Related termsHyperscalerGPUData centre
Related questions
Why are crypto miners relevant to AI?
Infrastructure overlap. Bitcoin mining, and historically Ethereum mining, required many of the same ingredients now sought by AI infrastructure developers: large power allocations, suitable sites and industrial cooling. Ethereum itself moved from proof of work to proof of stake in September 2022, so it is no longer mined. When mining economics weakened, several operators held energised sites that would take years to permit from scratch. Converting them for AI tenants is a substantial rebuild rather than a relabel, but the scarce ingredient, power, was already secured.
What is agentic AI?
Most AI use is still reactive: you ask, it answers. An agent is given a goal instead, and works out the intermediate steps, calling tools, reading data and looping until it decides the task is finished. That shift is why permissions, audit trails and identity suddenly matter so much. A chatbot that is wrong wastes your time. An agent that is wrong can take an action on your behalf, which is a different category of problem.
What is the difference between training and inference?
Training is not a single event. Pre-training is the long, expensive phase where a model learns general patterns from very large amounts of data. Post-training shapes that raw model into something useful and better behaved, often using human feedback and reinforcement learning. Fine-tuning adapts an existing model to a narrower task or domain, and can be done repeatedly and comparatively cheaply. Inference is the model actually answering, which happens billions of times a day. Much of the current demand for chips, power and data centres is driven by inference at scale rather than by headline training runs alone.
What is a foundation model?
A large, general purpose model trained on a very broad dataset, which can then be adapted to particular tasks. The GPT, Claude, Gemini and Llama families are all foundation models. They are called foundations because applications, tools and agents get built on top of them. Training one is an expensive bet that sitting at the base of the stack is where lasting value accrues, which is precisely the question the rest of this site is trying to help you think about.
Sources and further reading (4)+
- 1Investor relations and interconnection reporting · Equinix
- 2Investor relations disclosures · CoreWeave
- 3Energy and AI, data centre electricity demand · International Energy Agency
- 4The Merge: Ethereum's move to proof of stake · Ethereum Foundation
The AI ecosystem changes rapidly. Company positions, technologies and market data reflect information available at the date above.