AI Needs Data Centres. But Who Should Pay for Them?
Keep up with IT Club
Add IT Club as a preferred source in Google Search.
AI depends on data centres, electricity, cooling, water, land and network infrastructure. This balanced IT Club analysis, based on a Kara Swisher podcast discussion, examines who should pay for the resources AI requires and what responsible data-centre development should look like for communities and businesses.
Artificial intelligence feels like software. Ask ChatGPT a question, use Microsoft Copilot or run an AI agent and the result appears on a screen.
But none of it is weightless. Every AI service ultimately depends on physical computing infrastructure somewhere.
- Data-centre buildings
- Servers and specialist processors
- Electricity generation and grid capacity
- Cooling and water systems
- Land and construction
- Network infrastructure
- Backup power and resilience systems
As investment in AI accelerates, the scale of that infrastructure is becoming a political, economic and environmental issue.
A recent episode of Kara Swisher's podcast brought together environmental campaigners, an energy researcher and a city mayor to ask a useful question: can data centres actually be built the right way?
The answer is more nuanced than either “data centres are bad” or “AI is inevitable, so just build them”.
The Quick Answer
AI needs data centres. Pretending otherwise is pointless. But accepting every proposed facility on whatever terms a technology company requests is equally simplistic.
The useful debate is not whether society should have data centres. It is how they should be built, where they should go, who should pay for the electricity and resource demands they create, and what affected communities should receive in return.
A responsible project should be transparent about energy, water, land, backup generation and grid requirements. It should pay an appropriate share of the costs created by its demand, use resources efficiently, choose a suitable location and remain accountable after construction.
The cloud still lives somewhere.
AI Data Centres Are Getting Much Bigger
Data centres are not new. Cloud computing, online services, financial systems, streaming, storage and enterprise applications have depended on them for years.
The change is the scale and density of some newer developments. The podcast discussion contrasts older data centres that might operate around 10–50 megawatts with newer AI-related developments reaching hundreds of megawatts, alongside proposals that could eventually approach gigawatt scale. These are rough orders of magnitude rather than a universal classification for every facility.
AI infrastructure can include large training facilities, where models are built or updated, and inference facilities, where live requests are processed. Some sites may support both. The practical point is straightforward: more computing demand translates into more infrastructure requirements.
| Workload | What it needs |
|---|---|
| Training | Large clusters of specialist processors working together for intensive, time-limited runs. High-speed networking, electricity and cooling are especially important. |
| Inference | Computing that serves live questions, generated content, searches, agents and business applications. Demand may be more continuous and closely linked to user activity. |
The exact boundary between a traditional cloud facility and an AI facility is not always clear. The important distinction for communities and businesses is the physical demand created by the particular equipment and workload, not the label attached to the building.
AI Has a Physical Footprint
Software can hide the infrastructure that supports it. A user sees an answer, image or summary; the user does not see the processors, power connections, cooling equipment and fibre routes involved in producing it.
The physical footprint can include:
- Electricity generation and grid connections
- Substations, transformers and transmission infrastructure
- Cooling equipment and water systems
- Land, construction materials and access roads
- Backup batteries and generators
- Fibre and network connectivity
- Noise, emissions and other local environmental effects
AI may be delivered through software, but the infrastructure behind it is physical.
This does not make AI inherently harmful. Hospitals, manufacturers, universities, public services and ordinary businesses all depend on data-centre infrastructure. It does mean that a sensible assessment needs to include the whole system rather than treating a data-centre proposal as an invisible extension of the internet.
The Water Question Is More Complicated Than It Looks
Water is one of the most visible concerns around data-centre development, but “data centres consume huge amounts of water” is too broad to be a useful conclusion on its own.
The podcast discussion distinguishes between water used directly at the data centre and water used indirectly by the electricity generation supplying it. Those are different parts of the wider resource footprint.
| Question | Why it matters |
|---|---|
| Direct water use | Water may be used for evaporative cooling, maintenance or other on-site processes. The amount depends on the design, climate, workload and operating conditions. |
| Indirect water use | Power stations and fuel supply chains can have their own water requirements. Moving electricity demand does not necessarily remove every wider resource impact. |
| Water availability | A volume that is manageable in one region may be significant in another, particularly where communities, agriculture or ecosystems already face pressure. |
Closed-loop cooling systems may reduce direct water consumption in some designs. But the podcast also raises the possibility that electrically intensive cooling can shift some water demand into power generation rather than eliminating the wider resource footprint.
Look at the whole system, not just the water meter attached to the building.
A credible proposal should explain its cooling approach, direct water demand, wastewater arrangements, seasonal assumptions and the relevant indirect resource questions. It should avoid presenting one attractive metric as if it describes the entire environmental impact.
Who Pays for the Electricity Infrastructure?
This is one of the strongest questions in the debate. A very large new electricity consumer may need new generation, grid reinforcement, transmission upgrades, connection capacity and local infrastructure.
Those requirements cost money. Depending on the market and regulatory structure, the bill may be divided between the data-centre developer, a utility or grid operator, an energy producer, other customers or taxpayers.
If somebody else pays for everything, there is less reason to choose the efficient option.
The podcast's core analogy is simple: when the user of a resource pays the bill, efficiency suddenly matters. The same principle can apply to large computing demand. If companies creating extraordinary new demand are required to pay appropriately for the infrastructure they need, they have a stronger incentive to improve efficiency, manage demand, invest in generation and use existing infrastructure intelligently.
That does not mean every cost should be assigned to one company or that wider infrastructure investment has no public value. Grid upgrades can support homes, manufacturers, transport, public services and future businesses as well. The important question is whether the allocation is transparent and proportionate, rather than quietly transferring a private growth cost to people who did not create it.
Could Data Centres Reduce Some Energy Costs?
The balanced answer is potentially, under the right regulatory structure.
Large new electricity consumers could help fund grid improvements, support advanced transmission, finance new generation, increase utilisation of infrastructure and potentially spread some fixed grid costs across a larger base.
Done well, large data-centre investment could potentially support wider grid improvements rather than simply adding costs. But that outcome is not automatic. It depends on where the facility is built, what capacity already exists, what the developer pays, how flexible the demand is, how the project is regulated and whether the infrastructure benefits extend beyond the site.
The claim should therefore be treated as a condition to be demonstrated, not a benefit to be assumed. A planning application should show what will be funded, who will benefit and how those commitments will be measured.
Demand Flexibility Matters
Electricity networks need enough capacity for the moments when demand is highest. Some extreme peaks last for relatively limited periods, but the system still needs to be built and managed around them.
A data centre that can reduce, delay or move some computing during peak periods may create less strain than one requiring maximum power continuously regardless of grid conditions.
A Plain-English Test
When everybody needs electricity at once, can some computing workloads wait?
The answer may be yes for some batch training, scheduled analysis, backups or non-urgent processing. It may be no for some live inference, safety-related services, customer interactions or workloads with strict latency requirements.
Demand flexibility can help, but it does not mean every AI workload can simply be switched off.
A responsible proposal should describe which loads are flexible, what notice is required, how performance is protected and whether the flexibility is a real operational capability or only a promise made during planning.
Data Centre Location Matters
The podcast uses Lansing as an example of why not all data centres are equivalent. The proposed facility discussed there was approximately 24 MW, intended for cloud infrastructure rather than AI training, proposed on existing urban land and near existing infrastructure. Local conditions around water, electricity and community contribution still mattered, and public opposition prevented the project progressing.
The example makes two points. First, a relatively modest cloud facility and a hyperscale AI development should not be treated as the same planning question. Second, a poorly handled proposal can damage public trust in better-designed projects that might have a more reasonable resource profile.
Potentially appropriate locations may include existing industrial land, brownfield sites, areas with suitable grid capacity or places where water and other resource pressures are manageable. None of those is a universal planning formula. A site still needs a local assessment of transport, noise, emissions, land use, resilience and community impact.
Transparency Is Becoming Part of the Problem
The podcast discussion repeatedly criticises non-disclosure agreements, secret negotiations, limited public information, communities finding out late and unclear resource requirements.
Large infrastructure projects create less resistance when people understand what is being built, how much energy it requires, how water will be managed, what upgrades are needed, who pays, what local impact is expected and what the community receives in return.
Trust is difficult to build after a project has already been agreed behind closed doors.
This is not an argument for publishing commercially sensitive security details or disclosing information that would create a genuine safety risk. It is an argument for making the public case with enough information to allow an informed discussion before commitments become irreversible.
What Does a “Good” Data Centre Look Like?
There is no single design that is right for every country, grid, climate, workload or community. The following framework is a practical way to test whether a proposal is being treated as a serious piece of infrastructure rather than a collection of attractive promises.
1. Transparent planning
Publish realistic information about energy demand, water demand, backup generation, land use, grid upgrades, construction impacts and environmental effects. Explain the assumptions behind forecasts and identify what may change if the workload grows.
2. Pay for the infrastructure required
Large technology companies should bear an appropriate share of the costs created directly by their demand. Avoid blanket household or taxpayer subsidies for infrastructure whose primary purpose is supporting private computing growth, while recognising that some network investment may deliver wider public value.
3. Use energy efficiently
- Efficient processors and hardware
- Efficient cooling
- Optimised software and model selection
- Better workload management
- Demand flexibility during grid stress
- Clear measurement of actual rather than theoretical consumption
4. Use appropriate sites
Prefer locations where suitable infrastructure already exists or where the development creates less environmental and community disruption. A site with spare grid capacity is not automatically suitable if its water, transport or local environmental conditions are poor.
5. Build cleaner generation where practical
The podcast discusses solar, batteries, geothermal, nuclear and natural gas. No single generation type is the universal answer. Each has different costs, build times, emissions and reliability characteristics. The sensible question is what mix is appropriate for the location, the workload and the timescale.
6. Understand the full water footprint
- Direct cooling requirements
- Water associated with electricity generation
- Local water availability and seasonal pressure
- Wastewater and discharge arrangements
- Maintenance and flushing requirements where relevant
- The difference between water withdrawal and water consumption
7. Create genuine local benefit
Ask what the community receives in return for hosting major infrastructure. Possible benefits include infrastructure investment, local tax contribution, energy programmes, redevelopment or a defined community-benefit arrangement. Do not oversell permanent jobs as the main benefit without evidence: construction employment and permanent operational employment are different things.
8. Measure performance after construction
Planning promises should not be the end of scrutiny. Where appropriate, monitor energy use, water use, emissions, backup generation, infrastructure commitments and community commitments. A project should be able to explain what happens when actual performance differs materially from the original forecast.
Why Should an SME Care?
Most IT Club readers will never build a data centre. They should still care because AI infrastructure costs flow through the economy and through the services businesses already use.
- Microsoft 365 and Copilot subscriptions
- Cloud computing and storage costs
- AI software pricing
- Electricity prices and network investment
- Taxes and infrastructure funding
- Availability of grid capacity for other businesses
- The resilience and location of important suppliers
Small businesses are increasingly dependent on Microsoft, Amazon, Google, specialist AI providers and cloud applications. The physical infrastructure supporting those services affects price, reliability, sustainability, availability and future competition.
AI infrastructure is not somebody else’s problem if your business depends on AI and cloud services.
The practical questions for an SME are not usually about planning permission for a hyperscale campus. They are about supplier concentration, service resilience, pricing changes, regional availability, data location and whether an AI feature is important enough to need an alternative if the provider changes its terms.
Efficiency May Be the Most Important Long-Term Answer
The podcast raises a historical parallel. Earlier predictions suggested that internet growth would cause huge permanent increases in electricity consumption. Efficiency improvements helped limit some of that growth.
AI may follow a similar path. Companies have strong incentives to develop more efficient models, better chips, improved cooling, smarter scheduling and smaller models for suitable tasks. As computing becomes expensive, those improvements can make a meaningful difference.
That does not mean efficiency will definitely cancel out future demand. If AI becomes cheaper and easier to use, people and businesses may use it more. Efficiency could materially change today’s infrastructure forecasts, which is another reason not to assume every current hyperscale projection will unfold exactly as predicted.
The IT Club View
AI needs data centres. Pretending otherwise is pointless.
But accepting every proposed data centre on whatever terms technology companies request is equally simplistic. The sensible question is not “data centres: yes or no?”
The Better Question
Who benefits, who pays, what resources are required and what happens if today’s assumptions prove wrong?
A well-designed project should be able to demonstrate transparent resource requirements, appropriate infrastructure funding, efficient energy use, responsible water use, a sensible location, grid resilience, community benefit and ongoing accountability.
If the AI industry genuinely expects to create enormous economic value, it should also be capable of paying appropriately for the physical infrastructure that makes that value possible. That is not anti-AI. It is a requirement for treating AI as an infrastructure industry rather than as weightless software.
The cloud is not weightless. AI may be digital, but the bill for running it is very physical.
Related IT Club Intelligence
America Loves AI — So Why Does It Hate the Data Centres That Power It? →
AI Is Starting to Outbid Bitcoin for Data-Centre Power →
Why Is AI Computing Power Becoming a Tradable Commodity? →
What Happens When an AI Agent Acts Beyond Its Authority? →
Source and Further Reading
Primary source: On with Kara Swisher — “Can Data Centres Be Done Right?” This article was based primarily on the supplied podcast transcript and paraphrases the discussion rather than reproducing long quotations. A verified public episode URL was not provided, so no external link has been invented.
Questions About AI Infrastructure?
Ask the IT Club Advisor. One straightforward IT question. Independent guidance without the sales pitch.
Plain-English Takeaway
AI needs data centres, but accepting every proposed facility on any terms is as simplistic as pretending AI can run without physical infrastructure. Responsible development should explain its resource requirements, pay an appropriate share of the costs it creates, use energy and water efficiently, choose sites carefully, provide genuine local benefit and remain accountable after construction.
Frequently asked questions
Why does AI need data centres?
AI services need physical computing hardware for both training models and serving live requests. Data centres provide the servers, specialist processors, electricity, cooling, network connections, security and resilience needed to operate those services at scale.
How much electricity does an AI data centre use?
There is no single figure. The podcast discussion contrasts older facilities that may operate around 10–50 MW with newer AI-related proposals reaching hundreds of megawatts or, in some cases, a potential gigawatt scale. Actual demand depends on the facility, workload, design and operating conditions.
Do data centres use a lot of water?
It depends on the cooling design, climate and location. Some water is used directly at the facility, while additional water may be associated indirectly with the electricity generation supplying it. Direct and indirect water use should be assessed separately rather than reduced to one universal number.
Who should pay for new grid infrastructure for data centres?
The answer depends on the regulatory and electricity-market structure, but the companies creating extraordinary new demand should bear an appropriate share of the infrastructure costs directly attributable to that demand. If costs are hidden or broadly socialised, the incentive to use power efficiently is weaker.
Can data centres help improve the electricity grid?
Potentially. A large, well-regulated consumer could help fund generation, transmission, storage or other grid improvements and increase utilisation of infrastructure. That is not guaranteed: the outcome depends on the contract, location, timing, regulation, flexibility and who actually pays.
Can data-centre computing be reduced during peak electricity demand?
Some workloads may be delayed, moved or throttled during grid-stress periods, while latency-sensitive inference and other critical workloads may not be flexible in the same way. Demand flexibility can help, but it does not mean every AI workload can simply be switched off.
Why should an SME care about data-centre infrastructure?
Businesses experience the infrastructure economics through cloud and AI pricing, availability, resilience, supplier concentration, electricity costs, taxes and access to future grid capacity. If your organisation depends on Microsoft, Amazon, Google or another AI platform, the physical infrastructure behind that service is part of your technology dependency.
Related Articles
Someone Can Fake an Intimate Image of You. Who Should Stop Them?
AI-generated intimate images can turn an ordinary photograph into a tool for harassment and coercion. This serious, UK-focused guide examines consent, safeguarding and the technology ecosystem that can enable abuse.
Read articleClaude Is Watermarking AI Content — But What Does That Actually Prove?
Anthropic is adding machine-readable marks to supported Claude-generated text and files. Here is what a watermark can show, what it cannot prove and what UK businesses should do with that information.
Read articleDo You Need to Be an AI Expert to Lead AI?
You do not need to be an AI engineer to lead AI adoption. You do need enough AI literacy, business authority and access to technical expertise to make good decisions.
Read article