How AI will disrupt the data centre model

Summary

• AI equipment refresh cycles and capital costs are challenging traditional long-term data centre investment models.
• Colocation operators may find a different role hosting latency-sensitive inference near enterprise systems and users.
• Open-weight models, chip replacement costs, and uncertain revenues complicate the return on new AI infrastructure.

By Chetan Visrolia, Business Development Manager, Data Center at SHI

The data centre as we know it could soon become a thing of the past due to the disruptive influence of AI. While the focus to date has been on data centre build there are currently 568 facilities in the UK, up from 477 in August 2025, placing the UK in third place worldwide and ahead of China there’s been less attention given to how the business model itself might change.

Data centres require hefty upfront investment in physical infrastructure and hardware which the provider then recoups over a 20-30 year period. But AI changes that. It requires higher rack densities, burns through chips and networking equipment, and is energy intensive, so requires rapid cooling systems. In effect, that shortens the lifespan of the facility. Another equally worrying issue is being saddled with depreciating assets, and or assets that have a short shelf life as the typical GPU release cycle is down to once a year. Newer chips use less power and provide greater density and bring new capabilities and a different dynamic to the normal multi-year asset life cycle. But those that find it difficult to adapt and invest in AI are likely to become left behind and their facilities become all but redundant.

COLO conundrum

Recognising this risk, the likes of Equinix, an established data centre operator with colocation (COLO) facilities, have had to reposition themselves to capitalise on the AI boom. The company doesn’t have the scale and capacity that hyperscalers or neoclouds have which are able to offer large training environments so instead it has partnered with Nvidia to enable its 10,500 customers to run their AI models over the Together AI open source cloud platform. The company has set out a clear strategy that will see it focus on AI inference workloads, network connectivity, and strategic partnerships rather than competing directly with hyperscaler AI data centre builders.

The Equinix model lends itself to inference because they have built small capacity services next to large metropolitan environments, thus placing these data centres close to where customer interactions are at their most intense and where data is created to infer decisions and make money. But we’ll come back to this later.

It’s an interesting business model and one we can expect to see emulated elsewhere, with dedicated corporate data centres likely to elect to house large training clusters in COLO/HyperScaler facilities to provide the necessary GPU and associated power scalability. This will see the data centre become cloud-adjacent, with data connected to multiple services, providing a viable model for those that haven’t got the deep pockets of the hyperscalers. They’ll be able to remain close to the action while utilising a distributed architecture.

Inference as a key factor

These providers can concentrate on what matters most AI inference, the production element that follows the AI training phase. Inference requires low latency, local data access, and proximity to operational systems but has predictable costs. All of this means it makes sense for inference to happen closer to the customer, requiring integration with various customer facing platforms to ensure that the existing business processes stay intact.

According to McKinsey, inference will account for 50% of AI use between 30-40% of data centre traffic by 2030 and COLO providers are ideally placed to capitalise on that demand, as opposed to chasing hyperscaler spend who will require gigawatt processing capabilities.

At the other end of the spectrum, the message among those borrowing to build is go big or go home. Many are heavily leveraged, with recent reports suggesting AI hyperscalers will invest over a trillion dollars next year in data centres and a further four trillion in the three years following. Morgan Stanley has further stated that more than half of the $2.9 trillion expected to be invested between 2025-2028 stems from external capital. That investment will be further compounded by interest rates and the need to reinvest in the next generation of chips by switching from GPUs to custom made silicon further down the line.

A wise investment or fool’s gold?

The question then becomes how they will recoup that spend? That remains an unknown, particularly in the wake of China’s open-weight models. The Chinese are betting that these models, which are 60-90% cheaper than the closed models of Anthropic and OpenAI, will allow it to outpace the US and corner the market, at which point the revenue models or tokenomics which underpin those AI investments driving data centre builds begin to look a little shaky. There is a very real risk that the likes of the MAG7 (comprising Alphabet, Amazon, Apple, Meta, Nvidia and Tesla) could then walk away from some investments, leaving behind empty data centres.

It’s no exaggeration to say that we’re now at critical juncture in the global AI industry. Open-Weight is not too dissimilar to Open-Source in its versatility. The open-weight model can be downloaded and programmed with all the sensitive data that corporates are not willing to upload to those closed cloud-based services from the frontier providers. This ensures a clean line of delineation for when work must be done inside the corporation, within its walled garden of security, privacy, and regulatory bound processes.

However, when the project requires something brand new to a company i.e. it cannot be fed by existing data in the company because the scope is beyond the current strategy, the closed model in the cloud makes sense. The parameters are broader and the data the likes of OpenAI have is more conducive to the creation of new strategy. This means we can expect to see a tiering of models play out, where BAU will typically sit across a set of open models, and then anything new where there is not historic data in the organisation to infer on earns the right to use an expensive frontier service.

An important distinction between the US approach to AI versus the Chinese is that the Chinese have shaped their market around on-device models, with robots in factories, companion robots, cars and other devices to a vast extent, totalling approx. 5-700M model device based models. This practical application is a key differentiator but it nonetheless points to an ability to confidently invest in data centres and AI. Who will win the race and achieve global dominance? That remains to be seen but for now each is taking a very different approach and playing to its strengths.

---