More and more enterprises are integrating AI into their business models. However, many fall at a crucial hurdle: transitioning from proof of concept to production. After securing approval for AI workloads, you need to figure out where they will run.

CFOs have multiple options for AI data centres: colocation solutions, cloud providers, or on-prem builds. The choice you make has a huge impact on your bottom line and sustainability performance: here’s how these three options shape up.

Cloud services

Cost

62% of cloud enterprise customers went over budget in 2024, and one in four said their spending massively overran. This unpredictable spending is due to a couple of things. Firstly, knowing how much cloud you’ll need is hard to predict, leading enterprises to over-provision so they’re ready to run at maximum capacity. This means they end up paying for compute they don’t use, wasting their budget. Secondly, cloud providers charge egress fees when data moves out of the cloud. When you’re running AI models in the cloud, transferring data is part and parcel of your AI system, so these fees quickly add up. Because switching cloud providers is expensive and time-consuming, enterprises also often get locked into one vendor. When storage fees increase, they have to find room in their budget to match the new price.

Sustainability

Over-provisioning in the cloud doesn’t just impact your budget, but also your sustainability. Even when you’re not using your cloud storage, its GPUs are still running 24/7. The energy this requires contributes to your enterprise’s scope 2 and 3 emissions, harming your sustainability performance.

When it comes to reporting, cloud providers’ lack of transparency and reliance on offsets can also be a headache for enterprises. Without accurate data on your AI model’s sustainability, it becomes tricky to back up your ESG narrative. What’s more, cloud data centres (often situated far from heat off-takers in remote, industrial locations) aren’t suited to heat reuse. Data centres that don’t reuse heat are more inefficient and costly to run, leading to higher compute bills for enterprises.

Data sovereignty and performance

For British enterprises, choosing a US-based cloud provider for your compute comes with a couple of problems. Firstly, because your critical data isn’t housed in the UK, you lack the peace of mind that comes with data sovereignty. Secondly, running inference workloads at a distance leads to higher latency: millisecond gaps that mean your AI model’s performance suffers.

On-prem data centres

Cost

Enterprises looking to run AI models could also opt for on-prem: building their own data centre. As a CFO, you already know that this requires significant investment. Another factor to consider is that it’s hard for enterprises to keep their facilities up to date as compute tech advances. With GPUs doubling in performance every two years, legacy data centres soon become outdated as they don’t have the necessary cooling requirements for modern high-density racks. Over-provisioning is also an issue with on-prem: enterprises build AI data centres to meet their spikes in demand, but a lot of the time, their servers remain idle. This means compute bills are high even when AI usage isn’t. 

Sustainability

As for sustainability, over-provisioning also leads to a lot of wasted energy. Outdated cooling systems in legacy data centres require more energy to run, resulting in a higher PUE and a more inefficient facility. On-prem data centres are often built on remote sites far from heat off-takers, making it impossible to connect with district heating systems and donate excess heat. This is a missed opportunity for enterprises, which lose out on the emissions reduction and boosted sustainability performance that come with heat reuse programmes. 

On-prem becomes capital-intensive and operationally complex

In the words of Deep Green’s Chief of Staff, Olly Jones, “Building your own data centre is rarely a technology decision. It is a capital allocation decision. For most organisations, tying up millions in power, cooling, and real estate infrastructure generates no strategic moat and depresses return on invested capital. The asset is long-lived, capital-intensive, and operationally complex.”

Colocation solutions

Cost

Compared to the over-provisioning problems with cloud and on-prem, colocation pricing is usage-based. Enterprises can scale their compute servers when needed, but aren’t locked into the high additional costs of over-provisioning or the egress fees that come with the cloud.

Sustainability

Because colocation data centres are often situated in or near urban hubs, they’re suited to heat reuse set-ups. For example, facilities can donate their heat to nearby swimming pools or residential buildings, reducing community heating costs and energy usage. These count as avoided scope 4 emissions for enterprises that use these facilities to run their AI models. Benefitting the community and the environment through heat reuse also provides enterprises with a strong ESG narrative.

The sustainability pros of colocation also benefit your bottom line. When data centres donate heat, they receive cooled water in return from the heat off-takers. This reduces cooling and operational costs while improving the facility’s PUE. “This is why colocation data centres can run extremely efficiently with very low PUEs,” Olly explains. “It has a direct impact on the running costs that clients face.”

Deployment speed and location considerations

Compared to the legacy cooling systems often used by on-prem builds, colocation data centres are often purpose-built for high-density racks. This means enterprises can easily scale their compute needs even as GPU technology advances. Being able to rapidly scale reduces the costs associated with delays in deployment, too.

A UK colocation provider makes sense for British enterprises, firstly because your sensitive data is safer in UK servers. Secondly, choosing a colocation solution near your HQ supports the low latency needed for inference workloads. This means customers won’t experience lags in performance when using your AI model.

Cloud vs on-prem vs colocation

“Specialist colocation providers build at scale, optimise energy efficiency, and spread infrastructure risk across multiple customers,” Olly says. “In most cases, ownership of the infrastructure layer adds balance sheet drag without adding competitive advantage. The rational choice is to deploy capital into core products and outsource the rest.”

A final budgeting tip

CFOs, listen up: when your enterprise sidelines sustainability, this has a direct impact on your bottom line. Firstly, efficient compute is cheaper to run, especially when data centres reuse heat. Secondly, avoiding over-provisioning your compute means you’ll reduce emissions while saving costs. Finally, your enterprise’s sustainability performance matters for investors and customers alike. Improving your scope emissions and ESG narrative with sustainable AI models means your enterprise, and the planet, will benefit long-term.

H2: Run sustainable AI near Manchester with Deep Green’s latest colocation data centre 

For enterprises looking to run sustainable AI models near Manchester, Deep Green’s latest colocation data centre could be a perfect fit. A collaboration with Trafford Leisure that donates heat to the local community, it offers the predictable costs, scalable capacity, and low latency that enterprises need for their AI workloads.