Every CIO relies on their own or, more likely, someone elses datacentre capacity, but the nature of that reliance is increasingly unpredictable. Thats because of the surging demand for datacentre capacity in general, and the constraints on delivering on that demand. But its also because of the way artificial intelligence (AI) is changing how datacentres operate.Figures from real estate giant CBRE highlight the apparently inexhaustible demand for datacentre capacity. In the US, the average vacancy rate for primary markets hit a record low of 2.8% in 2024. Europe saw record new capacity come on stream in 2024, with take-up outstripping new supply the seventh consecutive year Europe has hit a record. Overall European vacancy rates are expected to hit 8.5% in 2025 bigger than the US, perhaps, but a record low for the region.This has further fuelled the datacentre building boom. But that boom is constrained by access to land, power and water even in the US.So, unsurprisingly, reservation signs are being slapped on capacity. Preleasing rates are expected to hit 90% or more in the US, CBRE predicts, with rental rates hitting the record highs last seen in 2011-12.So, how should CIOs understand this market when theyre thinking about planning their own datacentre needs?AI is driving the datacentre boom due to its relentless drive for graphics processing unit (GPU)-fuelled capacity. Thats certainly what CBRE expects. But AI introduces its own uncertainties.Certainly, it has highlighted the broader issues that are crimping datacentre buildouts. In the UK, the Labour government has pledged to overhaul planning and clear the way for datacentre building. How that might progress amidst a ramp-up in defence spending remains to be seen. In the US, the Trump administrations Stargate strategy has promised a $500bn public-private partnership to build out datacentre and related infrastructure to sharpen the countrys AI edge.Even if these government-backed initiatives are not thrown off course by events dear boy, events, they will still take years to come to fruition. How they will benefit businesses and other user organisations, rather than hyperscalers, is not immediately clear.But land, water and power apart, there are other factors at play.Dan Scarbrough, chief commercial officer of AI data mobility startup Stelia, says the current frenzy has undermined the traditional economics of the industry.One element is Nvidias rapid release cycle. Operators offering GPU-powered capacity find it becomes old hat very quickly.Datacentres historically have been built to last 15 years, and youre going from 40 kilowatts to forecast half a megawatt rack density over the course of a few iterations of the chip, he says.That rapid tech turnover means some customers are loathe to commit to long contracts. At the same time, keeping datacentre infrastructure up to date to cope with newer generations of higher-performing silicon is far more challenging. The datacentre has been pegged as a real estate asset, with relatively stable value for 15, 25 years. Its now becoming more like an iPhone Dan Scarbrough, SteliaThe datacentre has been pegged as a real estate asset, with relatively stable value for 15, 25 years, says Scarbrough. Its now becoming more like an iPhone.This has fuelled the rise of specialist operators, such as GPU-as-a-service firms.Josh Mesout, chief innovation officer at cloud services provider Civo, which both operates its own datacentre capacity and uses other providers, says GPUs raise their own issues. Access to chips is one thing, but its quite another to have the power, cooling or even management infrastructure that customers need. These are not easy things to use. They are very complex and require very deep system application, he says.Mesout suggests GPUs or at least Nvidias silicon are not necessarily going to be the only game in town. I think the interesting part were seeing is things like TPUs [tensor processing units] and NPUs [neural processing units] can generate the same thing for 10 times less power.Throw in concern over export restrictions, he says, and were almost building the perfect environment for someone to build an Nvidia competitor.More broadly, says Mesout, after years of the rush to cloud, enterprises are now more comfortable with multicloud and hybrid cloud. Rather than thinking about straightforward cloud migration, he says: Its now a full-fledged infrastructure digital transformation looking at things like, Should we buy a warehouse to put a datacentre in there? Ive got loads of spare real estate footprint. Ive got offices companies dont want. Could that be a datacentre?.Moreover, some companies will simply not be near a datacentre or cloud facility that can support their operations. This is particularly important for real-time operations, such as manufacturing, which cant tolerate latency.Penny Madsen, senior research director at IDC, cites the case of a large agriculture firm which uses AI to manage watering. It had to build an AI-capable datacentre as it was simply too far away from either a cloud or datacentre provider.Theres also those cases where you need to take the AI closer to the end user, she adds. Its not always the case that AI and cloud or general compute can co-exist, particularly in third-party datacentres, she says, as some tenants in a datacentre wont be comfortable having any water-cooled racks near their kit.Spencer Lamb, chief commercial officer of datacentre operator Kao Data, is more sceptical about the fate of on-premise. Beyond hardware that needs to be retained for regulatory or data sovereignty reasons, most businesses will be sourcing future data capacity from their current cloud providers, he predicts.Thats not necessarily the cheapest option for enterprises, however. It might be quite meaningful from a cost perspective, he says.Following the rush to the cloud, the cost implications should have prompted some companies to move back to on-premise, but it hasnt, according to Lamb. I thought it might happen with AI, because potentially the core per hour rate for AI is going to be far higher, but it hasnt.Lambs advice for CIOs is to be wary of being tied into particular providers or AI models, noting that Microsoft is creating models and not charging for them, knowing that companies will still be paying for the compute to use them.Lamb also says that, whether were talking on-premise, colocation or cloud, the potential for retrofitting existing capacity is limited, at least when it comes to capacity aimed at AI.After all, those GPUs often require liquid cooling to the chip. This changes the infrastructure equation, says Lamb, increasing the footprint for cooling infrastructure in comparison to compute. Quite apart from the real estate impact, this isnt something most enterprises will want to tackle.Also, cooling and power will only become more complicated. Andrew Bradner, Schnieder Electrics general manager for cooling, is confident that many sectors will continue to operate on-premise datacentre capacity life sciences, fintech and financial, for example. But he also expects power and cooling requirements to continue to rise, so they must be considered for new builds.If youre designing for todays chips, you also have to think about two or three generations that might come into the datacentre during that time, and how you design for that without a full retrofit of your infrastructure.Its not just on the cooling side, he adds. Its on the power side, because as you start to get over 200 kilowatts of rack, you have to fully redesign how you get power to those racks and servers. That includes heavier gauge cable, bigger breakers and ultimately a shift from AC power to DC power.How do you address these challenges? Because theyre real, theyre immediate, and were going to have to figure out how to do this as we look at that next generation of silicon thats going to be coming, he says.So, will there be net new datacentres? Undoubtedly. And CIOs will likely demand a mix of capacity cloud, colocation and possibly some on-premise depending on what their long-term strategy is.IDCs Madsen advises approaching datacentres from a data governance stance right from the outset.You need to start thinking about this at the very beginning of projects because the investment as you go through that project is really high, she says. So, consider what your return on investment is going to be, what your appetite for risk is with the data thats going to be used. And you need to have transparency on costs and pricing generative AI projects are being measured as business projects, not pure technology projects, she adds.And, given that cloud providers or datacentre providers are likely to be part of your strategy, Madsen advises having conversations around the innovation roadmap of those providers.Few will be able to go it entirely alone. I think youre going to see a lot of movement over the course of the next year, as things stabilise and people go, Actually, we do need a trusted partner to help with it.Read more articles about datacentre capacity planningFive datacentre trends to watch in 2025: Datacentre trends for 2025 focus on sustainability and AI, highlighting energy demand, hyperscale datacentres, innovative cooling methods, sustainable energy and edge computing.New componentry extends NetApp ASA and E-series block storage: One-time king of the filers adds anti-ransomware to its more recent block storage families, while also adding in an extra FAS array, all on the back of upgraded components.