Milan Radia: Liquid Cooling and the Future of AI Data Centres

If you still think data centers are just glorified real estate, the market has moved on. In this conversation from Web Summit, Milan Radia, Partner at Taranis Capital and CEO of Connected Compute, explains why data centers are now strategic infrastructure - and why AI is forcing a rethink of how they’re built, connected, and powered.The shift is bigger than more storage or more servers. You’re looking at a new era where 

compute, connectivity, power availability, and data sovereignty

all matter at once. If you work in fintech, infrastructure, cloud, or AI, this is the landscape you need to understand.

Why the Old Data Center Model No Longer Fits

For years, investors and operators could treat data centers as a simple play on capacity. Build the facility, provide the power, connect it to the network, and serve demand. That model worked when rack densities were low and most workloads were less demanding.But that era is over. Radia points out that the legacy view of data centers as “glorified real estate” made sense when power draw per rack was a fraction of what it is today. Back then, the industry was mostly focused on storage and basic compute. Now, the conversation has shifted to a much more complex stack of requirements: processing power, latency, inter-connectivity, and increasingly, specialized cooling.This matters because the value of a data center is no longer just in the building itself. It’s in what it can support. A facility that cannot handle modern workloads, or cannot be adapted quickly enough, may become obsolete far faster than investors expect.From storage to strategic infrastructure The cloud changed the game first. Instead of keeping applications and servers on-premise, companies moved workloads into remote environments where they could benefit from resilience, flexibility, and access anywhere. Hyperscalers like Amazon, Google, and Microsoft then expanded what was possible by offering capabilities many firms could not replicate internally.Now AI is pushing that evolution further.Instead of only hosting applications, data centers are increasingly doing heavy lifting for training models, running inference, and supporting real-time digital services. That changes what “good infrastructure” means. It’s not enough to have space and power - you need the right architecture for the workload. Why rack density is becoming a board-level issueOne of the clearest signs of change is rack density. Traditional facilities were often designed around 20-30 kilowatts per rack, which used to be considered strong performance.That assumption is breaking down. Radia notes that NVIDIA’s GB300 systems are pushing requirements toward 150 kilowatts per rack, with further movement toward even higher densities. Once you get into that range, air cooling starts to look inadequate, and direct-to-chip liquid cooling becomes far more relevant.For investors and operators, this creates a serious risk: 

technology obsolescence

. A facility that looked future-proof a few years ago may now need expensive retrofitting - or may not be suitable at all for the kind of demand coming next.

AI Is Reshaping What Data Centers Need to Do

The AI boom is often discussed in terms of software, models, and user experience. But underneath all that is a huge infrastructure story.Training large language models requires enormous amounts of compute. That part of the story is well known. What’s becoming more important now is 

inference

- the process of using AI models in real time to answer questions, generate content, and support decision-making.That shift changes the design requirements again.Inference is latency sensitive. If you ask a model for an answer, you expect a response immediately, not in 10 minutes. The same is true for image generation, video workflows, or any consumer-facing AI interaction where the conversation has to feel fluid. That means data centers can’t just be powerful - they need to be connected, distributed, and fast.Training is important, but usage is becoming the bigger story Radia makes a useful distinction here: a lot of the early AI conversation focused on model training, but the market is now moving toward usage at scale. In practice, that means more requests, more distributed compute, and more pressure on network performance.There’s also a reality check worth keeping in mind. As Radia references, one professor summed it up as: 

“AI minus BS equals software.”

In other words, much of what people are calling AI is still software - just software that is more intelligent, more responsive, and more capable of handling tasks on your behalf.That doesn’t make it less important. It makes infrastructure even more important, because software only works at scale if the underlying systems can support it.Why latency changes the geography of infrastructure A major implication of AI inference is that location matters more than people assumed.Historically, some data center projects were built in places where land and power were available, but connectivity wasn’t a major concern. That may have worked for certain workloads. It doesn’t work as well for real-time AI, online collaboration, or gaming, where speed and responsiveness are critical.Radia argues that the next wave of facilities will be more distributed, but not necessarily tiny. The idea isn’t that everything becomes a small edge node. Instead, expect larger facilities - perhaps 50 to 100 megawatts - placed closer to demand and tied into stronger network infrastructure.

The Middle East’s Advantages Go Beyond Power

One of the strongest themes in the discussion is why the Middle East, and especially the UAE and Saudi Arabia, is emerging as a major data center hub.It’s easy to assume the answer is simply abundant power. That’s part of it, but it’s not the whole story.According to Radia, several factors are converging:

• Open economies

  that are actively welcoming this kind of investment

• Regulatory structures

  that make it easier to set up and operate

• Reliable access to power

  with more certainty around delivery

• Strong political and financial support

  for AI and infrastructure initiatives

• Proximity to major subsea cable routes

  connecting Asia, the Middle East, and Europe

This combination is powerful. It means the region is not just consuming infrastructure - it’s becoming a strategic node in the global digital economy.Connectivity is now as important as capacity A key point Radia makes is that connectivity has become central to what makes a data center useful. It’s not enough for a facility to have power if it can’t move data quickly and reliably in and out.That’s why subsea cables matter so much. Around 98% of global IP traffic runs over subsea cables, making them one of the most strategic parts of the digital economy. In many ways, they are the internet’s highways.The analogy to roads is apt. A four-lane highway sounds sufficient - until traffic arrives and suddenly it’s not enough. The same thing is happening with digital infrastructure. Networks that once seemed generous are now hitting capacity constraints far sooner than expected. The region is building for the next wave, not the last one Because Asia-to-Europe traffic often passes through the Middle East, the region is benefiting from geography as well as policy. Add in sovereign wealth investment, state-backed AI programs, and access to GPUs through international deals, and the picture becomes clearer.The result is an ecosystem, not just a cluster of buildings. Data centers, subsea cables, compute supply, and government support are beginning to reinforce one another.That’s a much stronger position than simply having cheap land or abundant energy.

Data Sovereignty Is Forcing a New Build-Out

There’s another major force shaping data center demand: sovereignty.More governments and enterprises want sensitive data to stay within their own borders. That concern is growing for obvious reasons. If you put confidential information into open AI tools, you may be creating legal, operational, or reputational risk. Radia highlights concerns around public AI systems being used with sensitive government data. Even if a tool is not intentionally exposing that data, organizations do not want to rely on broad public systems for confidential workloads.That’s why proprietary AI models and private LLM instances are becoming more common.Why private AI needs private infrastructure When organizations build their own AI environments, they need infrastructure designed for controlled access, secure storage, and compliance with local regulations. That often means keeping the compute inside national borders or within tightly governed regions.This is especially important for governments and large enterprises, but it’s increasingly relevant for any business handling regulated data.As a result, sovereignty is not just a policy topic. It is becoming a demand driver for new data center capacity. If a country wants its data to remain onshore, it needs the infrastructure to support that requirement.That’s one reason the data center market is expanding in regions such as the UAE and Saudi Arabia. The need is not just for more capacity, but for capacity that aligns with local legal and strategic goals. The practical takeaway for investors and operators If you’re building or investing in data centers, sovereignty changes the brief.You are no longer just asking, “Can this facility handle demand?” You’re also asking:

• Can it support regulated workloads?
• Can it stay within local data rules?
• Can it run proprietary AI models securely?
• Can it adapt as governments tighten requirements?

These are not abstract questions. They directly affect leasing, design, location, and long-term asset value.

What Investors Should Worry About Now

The final and perhaps most important theme is risk.Radia is blunt that the risk of obsolescence is now higher than it has ever been. That’s because several shifts are happening at once: rack densities are rising, cooling requirements are changing, AI use cases are moving from training to inference, and customer expectations around latency are tightening.A facility built for yesterday’s requirements may not fit tomorrow’s market.That doesn’t mean existing assets are worthless. It means they need to be evaluated carefully. The more flexible the design, the better the network access, and the more future-ready the power architecture, the more defensible the investment.What “future-ready” actually meansA future-ready data center is not just one with more megawatts. It is one that can support changing workloads without requiring a complete rebuild.In practical terms, that means:

• Higher density support
• Liquid cooling readiness
• Strong fiber and cable connectivity
• Proximity to demand centers or network hubs
• Compliance with local sovereignty expectations
• Flexibility for both training and inference workloads

The challenge is that hyperscalers often want customized builds. They are not looking for generic shell space. They want facilities designed around their own specifications, which raises the bar for developers and operators.This is where the market gets sharper. The winners will not simply be the biggest builders. They will be the ones who understand the new requirements early and can deliver infrastructure that matches them.

Timestamps: 00:00 - Introduction to The FinTech Times News and Views. 00:41 - Milan Radia on his background in data centres and Taranis Capital. 01:32 - The transition from "glorified real estate" to high-complexity compute hubs. 05:15 - Subsea cables and the Middle East as a strategic information highway. 07:30 - Why the UAE and Saudi Arabia are winning the data centre race. 10:45 - The impact of Nvidia Blackwell architecture on data centre design. 13:20 - Moving from AI training to real-time inferencing and the latency challenge. 15:44 - Conclusion and closing remarks.