The Data Centre Becomes a Factory
Clarity in a Changing World.
For informational purposes only. Not investment advice.
Introduction
For much of the internet era, data centres were treated as supporting infrastructure. They stored information, hosted websites and provided computing capacity. Economically, they were often understood through the characteristics of real estate: land, buildings, leases, power and occupancy.
Artificial intelligence is changing that model. An AI data centre does more than store or transmit information. It combines processors, memory, networking, cooling and electricity to produce computational output at industrial scale.
It is therefore increasingly useful to think of the modern AI data centre not simply as a building filled with servers, but as a factory for computation. Once viewed in this way, capacity, utilisation, efficiency and return on capital become as relevant as the technology inside the facility.
From Housing Computers to Producing Compute
Traditional data centres primarily supported relatively predictable workloads. Enterprise software, cloud storage, websites and databases required substantial computing resources, but those resources could often be distributed across large numbers of conventional servers.
AI changes the nature of the workload. Training and running advanced models require large numbers of accelerators to operate together, often at high utilisation for extended periods. The objective is therefore no longer simply to provide access to computing infrastructure; the infrastructure itself produces a measurable economic output: compute.
This creates an industrial relationship familiar from manufacturing. A semiconductor fabrication plant converts materials, equipment, electricity and engineering expertise into chips. A chemical plant converts feedstocks and energy into chemicals. An AI data centre combines electricity, processors, networking, cooling and software to produce computational capacity. The products are different, but the economics increasingly share important similarities.
Capital Intensity Changes the Economics
AI infrastructure is extraordinarily capital-intensive. The cost of an AI facility extends far beyond the building: advanced accelerators are expensive, high-speed networks must connect them, electrical systems must support unusually dense loads, and cooling infrastructure must remove the resulting heat. Reliable grid connections and backup power add another layer of investment.
As computing density rises, the infrastructure surrounding the processors becomes more demanding. Each additional unit of AI capacity therefore requires substantial investment not only in computing hardware, but also in the physical systems that allow it to operate.
This differs from much of the software era. Successful software companies could often expand revenue with relatively limited incremental physical investment. Distribution costs were low, and additional customers could frequently be served using existing infrastructure.
AI introduces a more capital-intensive model. Before additional computational capacity can be sold or deployed internally, physical assets must first be financed, constructed and commissioned. The boundary between technology and industry is beginning to narrow.
Utilisation Becomes Critical
Factories are valuable only when they are productive. The same principle applies to AI data centres.
A facility containing expensive processors creates limited economic value if those processors spend substantial periods waiting for data, power or other components. Utilisation therefore becomes a central variable in the economics of AI infrastructure.
Consider two data centres containing similar numbers of advanced accelerators. One has efficient networking, adequate cooling and reliable power delivery, allowing its processors to operate close to intended capacity. The other experiences network congestion, thermal constraints or interruptions in power availability. Both may possess similar nominal computing capacity, but their effective productive capacity—and therefore their economics—can be very different.
As capital expenditure rises, relatively small improvements in utilisation can materially affect the economics of these assets.
Competitive advantage may therefore come not only from owning more processors, but from operating them more efficiently.
Power Density Is Redesigning the Building
Traditional data centres were designed around relatively modest rack densities. AI clusters concentrate substantially more computing power into the same physical space. More computation produces more heat, and that relationship has consequences throughout the facility.
Electrical distribution systems must handle greater loads. Cooling systems must remove more heat. Backup power requirements increase, floor layouts change, and water availability may become relevant for some cooling configurations. Even the location of a facility can be affected by access to electricity, grid capacity and environmental conditions.
The building itself is becoming part of the computing architecture.
A poorly designed facility can reduce the effective output of the processors inside it. A well-designed one can support greater computing density, improve utilisation and generate more computational output from the same physical footprint.
Location Is Becoming a Strategic Decision
During the cloud era, data centres were often located according to a combination of land prices, network connectivity, tax incentives and proximity to customers. Those factors remain relevant, but AI adds another constraint: access to large quantities of reliable power.
A site may have abundant land and excellent connectivity but still be unsuitable if the local grid cannot support several hundred megawatts of incremental demand. Power availability is therefore beginning to reshape the geography of computing.
Regions with sufficient generation, transmission capacity and relatively predictable permitting processes may be better positioned to accommodate new AI infrastructure. At the same time, hyperscale operators are becoming more involved in energy procurement, grid planning and long-term power arrangements.
Data-centre strategy is consequently becoming intertwined with energy strategy. The location of future computing capacity may depend as much on electrical infrastructure as on proximity to traditional technology centres.
Capacity Is Not the Same as Demand
Capacity is not the same as demand.
Industrial history is full of investment cycles in which strong demand encouraged rapid capacity expansion, only for utilisation and economic returns to decline once supply caught up. AI infrastructure is not necessarily different.
The current scale of investment assumes that demand for computational capacity will continue to grow rapidly. That assumption may prove correct, but building more data centres does not automatically create economic value.
The relevant questions are whether the new capacity can be utilised, whether customers can generate sufficient value from AI to justify the cost of computing, and whether operators can generate sufficient economic returns after accounting for electricity, depreciation and continuing technology upgrades.
The number of data centres under construction tells us how much capital is being deployed. It does not tell us whether that capital is being deployed productively.
The Depreciation Problem
AI infrastructure also introduces a characteristic familiar to industrial businesses: depreciation.
The buildings may remain useful for decades, but the computing equipment inside them may not. Rapid improvements in processor performance mean that today’s advanced accelerators can become economically less competitive long before the surrounding facility reaches the end of its useful life.
Different components therefore have very different economic lives. Land and electrical infrastructure may remain productive for many years. Cooling systems and networking equipment may require periodic upgrades. Processors may face much shorter replacement cycles.
The economics of an AI data centre depend partly on whether productivity improvements and revenue growth can compensate for technological obsolescence. Capital expenditure alone therefore tells us little about whether AI infrastructure is economically productive. The assets must ultimately generate sufficient output to justify both the initial investment and the continuing cost of upgrading them.
A New Industrial Ecosystem
Once data centres are understood as factories for computation, the breadth of the AI investment cycle becomes easier to see. A modern AI facility depends on far more than semiconductor manufacturers.
It requires electrical distribution and power-management equipment, cooling and thermal-management systems, networking and optical infrastructure, engineering and construction capacity, grid connections, power generation, specialised real estate and software capable of managing increasingly complex computing clusters.
AI is therefore creating demand across an industrial ecosystem that extends far beyond the traditional technology sector. This helps explain why the current investment cycle is spreading from semiconductors into electricity, networking, cooling and industrial equipment.
But greater spending does not mean that economic value will be distributed evenly across the supply chain. Capacity expansion can attract competitors, technological standards can change, and some components may eventually become commoditised.
The relevant question is not simply where spending is increasing, but where scarcity, technical complexity or operational importance allow economic value to be captured.
What Should We Watch?
The scale of data-centre construction is useful evidence, but it should not be confused with economic success. As the investment cycle develops, several indicators can help distinguish capacity expansion from productive deployment.
Utilisation shows whether installed computing capacity is actually being used. Capital expenditure indicates the scale of investment, but its relationship with revenue and cash generation becomes more informative as the cycle matures. Power availability matters because delays in grid connections can postpone projects and leave committed capital underused.
Computing density provides another useful measure. Higher rack densities can increase output from a given site, but they also place greater demands on electrical and cooling infrastructure. Depreciation and upgrade cycles matter because rapid hardware replacement can materially change the economics of an AI facility.
Ultimately, these indicators lead back to one question:
How much productive economic output is being generated for each dollar of capital invested?
This provides a more useful test of the AI investment cycle as the focus shifts from capacity expansion to the productivity of the capital deployed.
What Could Change This Outlook?
The factory analogy has limits. Unlike conventional manufacturing, computational output is highly flexible, and software improvements can increase the productivity of existing hardware without requiring equivalent increases in physical infrastructure.
Model efficiency may also improve substantially. If future AI systems require less computation for equivalent performance, some projections of infrastructure demand could prove excessive. Conversely, lower computing costs could stimulate much greater usage, offsetting some or all of those efficiency gains.
There is also a conventional investment-cycle risk: overcapacity. If infrastructure construction grows faster than economically useful AI demand, utilisation and economic returns could weaken even while overall AI adoption continues to expand.
The long-term development of the data-centre cycle should therefore not be assessed simply by megawatts built or GPUs installed. The more meaningful test is whether the capital deployed produces sustained economic output.
Conclusion
AI is changing the economics of the data centre. As computing becomes more capital-intensive, the central question is shifting from how much capacity can be built to how productively that capacity can be used.
More GPUs, more megawatts and more data centres measure the scale of investment. They do not measure its productivity. Capacity creates value only when utilisation, efficiency and economic demand justify the capital required to build it.
The data centre is therefore becoming productive industrial capacity, and its economics increasingly depend on output relative to capital invested.
That shift also creates a new physical constraint. As more computing is concentrated into each facility, more heat must be removed.
In the next article, Cooling: The Invisible Infrastructure, we examine why rising computing density is making thermal management a critical constraint on AI infrastructure.