Navigating the AI infrastructure mismatch: Equinix CEO Adaire Fox-Martin

As AI moves from model training toward widespread inference, the infrastructure underpinning it will need to evolve as well. In a wide-ranging conversation, Equinix CEO and President Adaire Fox-Martin talks to McKinsey Senior Partner Pankaj Sachdeva about the constraints shaping data center development, the growing importance of connectivity and sovereignty, and how she sees the next generation of digital infrastructure taking shape. This interview has been edited for length and clarity.

From a training to an inference-based AI economy

Pankaj Sachdeva: You’ve spoken about an infrastructure mismatch as more AI demand shifts from training toward inference. What do you mean by that?

Adaire Fox-Martin: We’ve seen an enormous amount of capital deployed against infrastructure that lends itself to a training environment. Those training environments tend to be situated in more remote locations and are often singular facilities for a singular tenant.

As the world moves toward inference, a different infrastructure is required. It’s much more distributed and more local to where humans and devices live and operate. That is the mismatch the industry will have to navigate as we move from primarily a training economy into an inference-based economy, where we get to put these models to work, embed them inside our business processes, and drive business outcomes with them.

Pankaj Sachdeva: What happens to tokenomics and the unit cost of inference in that environment, and how should enterprises plan around it?

Adaire Fox-Martin: We’re still at a very early stage, but what we can see is that the overall cost of AI queries is coming down. As inference becomes cheaper, you get higher volume, and that can become a self-fulfilling prophecy of economic outcome.

But for any business, if the ROI of an application isn’t relevant to the business outcomes you’re driving, it’s not necessarily a project worth escalating to the top of your priorities. Our enterprise customers are watching this extremely closely to determine the type of return they are getting on the projects where they are deploying AI capabilities.

Pankaj Sachdeva: That uncertainty around ROI also creates uncertainty around the pace of AI adoption and infrastructure buildout. How is Equinix navigating that?

Adaire Fox-Martin: It’s a combination of uncertainty and absolute excitement about where the company can grow and the kinds of services we can offer our customers.

We’re navigating two speeds: the speed of capacity development and infrastructure build, and the speed at which the technology itself is evolving. They are very different speeds, and we’re very conscious of that.

What gives us confidence is that our demand base is broad and durable. We’re not focused on a handful of customers. We have an exceptionally broad base of enterprise customers, and we run standard workloads as well as AI workloads for them.

We also have very intimate relationships with many of those customers because we’re helping to build the foundations that allow their business processes to execute efficiently, effectively, and within the regulatory frameworks in which they operate. That requires engagement, and quite often that engagement is two or three years out. It gives us a long-term view of our customers, their growth profiles, and the work they are doing with us to secure their energy and compute futures. That’s why we have a strong sense of the durability of the demand, particularly for what Equinix offers.

Managing the constraints on infrastructure

Pankaj Sachdeva: This is an unprecedented infrastructure buildout, but supply chains, labor, energy, permitting, and other resources are constrained. How do you navigate those constraints while moving at the speed customers require?

Adaire Fox-Martin: We’re in a time of tremendous opportunity, where capacity enables our customers’ business outcomes. At the same time, it is much more difficult to deliver that capacity than ever before.

Whether you’re looking at supply chains, energy, the energization of land, permitting, or societal feedback on data centers, you need a clear strategy for each constraint.

Take supply chain. We’re fortunate to be a significant player in the industry and therefore have relevance to the suppliers we work with. But we also practice open planning with those suppliers so that they understand what our footprint looks like and we understand their risks and constraints in delivering against it.

We’ve also had to prepurchase some of the major equipment that goes into our data centers. If we tell a customer that June is the date, then June needs to be the date when that data center is ready for service. It’s a combination of relationships, understanding where prepurchasing can ensure delivery of the necessary components, and investing in the long-term partnerships that matter at times like this.

We’re also elevating our supply chain–planning capabilities and applying technology to make the process more effective. Because we use standard formats across different geographies, we have some fungibility. If one build accelerates and another slows, for example, we can move equipment between them and continue to meet our overall capacity goals.

At the same time, while we try to standardize our formats, because we build in metro locations, there is often something unique about the local environment that must be taken into consideration. You can’t always have 100 percent conformity to a single design specification. We use AI as we review our design specifications, and applying technology in that way has produced significant benefits for us.

Pankaj Sachdeva: Power and time to power have become major bottlenecks in this AI supercycle. What will it take to overcome that constraint?

Adaire Fox-Martin: It’s a multifaceted issue, and the procurement and securing of energized land is almost job number one for our design, construction, and real estate teams.

There are several dimensions to solving it. First is efficiency: The more efficiently you use energy, the less energy you need, and that’s better for everyone. We continue to work on reducing the energy usage inside our data centers.

Then there is innovation. We’re looking at alternative sources of energy and at opportunities for self-generation. We’ve made investments in companies exploring technologies such as small modular reactors, should they become mainstream. We’ve used gas fuel cells in our data centers for about a decade, and we’re exploring opportunities around hydrogen and other sources.

The third dimension is partnership. That includes power purchase agreements that support investment in renewable energy, as well as our relationships with utilities and the communities in which we operate.

With utilities, our objective is to help manage grid infrastructure alongside them, to be transparent about the loads we require so they can plan effectively. In some cases, we will build infrastructure to the utilities’ standards to support the delivery of power to one of our facilities, which can reduce the overall cost for those utilities.

And we feel exceptionally strongly that ratepayers should not bear the cost of a large energy user such as Equinix operating in their jurisdiction. Large energy users have a genuine responsibility to make sure the costs associated with their requirements are not passed on to the community.

How customer requirements are changing

Pankaj Sachdeva: Customers have traditionally valued Equinix for service, resilience, and reliability. As AI workloads grow, are their requirements changing?

Adaire Fox-Martin: Some aspects of what customers expect have not changed, and I don’t believe they ever will. Resilience is one of them. We run significant workloads in sectors such as financial services and healthcare, and we value the trust our customers place in us. We design our data centers to meet that requirement for resilience.

What has changed is density. Over roughly the past five years, our customers’ energy requirements have changed significantly as workload density has increased, and we’re now designing for that.

One thing I’m exceptionally proud of is the way our teams have been able to engineer the latest technology into our existing data center footprint. We have 281 facilities globally, and they have been built across many different generations of technology over the past 28 years. In some of our existing facilities, we’ve been able to introduce liquid cooling so customers can run highly dense workloads requiring GPU capability.

Pankaj Sachdeva: Does that complexity mean customers need more help deciding where and how to design and deploy workloads?

Adaire Fox-Martin: It’s extremely important that we understand the workloads our customers are deploying with us. Understanding the workload helps us architect the best place to land it. Our mantra has always been putting the right application in the right location for all the right reasons.

That requires industry domain knowledge—understanding the language, semantics, and business of the customer. At a strategic level, we’ve organized our sales team along industry lines so that domain knowledge builds over time. We can then work with customers in the context of their industry, including the regulatory requirements that need to be considered when designing infrastructure.

The importance of connectivity and sovereignty

Pankaj Sachdeva: You’ve argued that interconnection, rather than simply land, power, and space, could be the real unlock for scaling AI. Why?

Adaire Fox-Martin: I believe the true bottleneck as we move to large scale will be connectivity, not compute.

Over the long term, it’s possible that compute becomes a commodity differentiated primarily by price. But even if you have the strongest, most resilient GPU cluster, the moment you leave it to access data—from different sources, from the cloud, or from a private environment—the latency and capabilities of the network can become the bottleneck.

This type of connectivity will differentiate those workloads that can deliver the ROI we talked about earlier from those that cannot. Latency and sensitivity to speed can become detrimental to the customer experience if the infrastructure is not structured correctly.

Pankaj Sachdeva: How important is sovereignty in the AI era?

Adaire Fox-Martin: Sovereignty is a very important topic, although the degree of emphasis can vary geographically. In Europe, for example, it is very much front and center.

We are naturally sovereign by design in the sense that we exist physically in the countries where we operate. But sovereignty isn’t only about where data sits on a machine or in a facility. Managing sovereignty as data moves is just as important.

That is part of what our Equinix Fabric capability addresses: issuing instructions to workloads so that data moves only within the jurisdictional boundaries a sovereignty program defines. If a transfer would break those guidelines, the network doesn’t route around them. The transaction simply doesn’t proceed.

The next generation of digital infrastructure

Pankaj Sachdeva: Looking ahead five to ten years, what major changes do you expect in the enterprise technology ecosystem?

Adaire Fox-Martin: We’ve been through a period of tremendous change, but that creates tremendous opportunity for companies to understand how technology can be deployed in service of their business models.

I think quantum will be a significant driver of opportunity over the next five years. Perhaps we’ll see a time when customers navigate from CPUs to GPUs to quantum capability depending on the requirements of the workload. There is a lot to consider in bringing quantum into the mainstream, including how you design facilities such as ours, and those are things we’re already thinking about.

I also think we’ll see agentic networking. Rather than networks being defined primarily by humans, as agents become more prevalent, you could see real-time propagation of network capability to deliver against the requirements of an agentic platform or workload. I expect the network to evolve substantially over the next five years and become much more agentic in nature.

Pankaj Sachdeva: And how will data centers themselves change?

Adaire Fox-Martin: They will continue to evolve. We’ve already seen that in our design specifications, and we’ve been fortunate that the way we designed our facilities has allowed us to evolve them rather than face obsolescence. The ability to retrofit liquid cooling into existing facilities is one example.

Design specifications will continue to change. I think we’ll also see more construction done offsite and brought onsite later in the process, which can accelerate overall construction.

AI itself will be an accelerant to data center design and construction and to project execution for example, helping optimize processes when schedules change.

Technology already plays an important role in data center operations, but there is much more opportunity to embrace it. I see data centers becoming more automated while remaining highly observable environments.

Making Equinix ‘customer zero’ for AI

Pankaj Sachdeva: You’ve talked about “AI for AI.” How do you see AI changing Equinix itself?

Adaire Fox-Martin: Like all companies, we’re exploring opportunities to deploy AI within our own business, focusing on where we believe those deployments will generate ROI.

At the same time, we’re imagining Equinix as “customer zero” for what we do. We want to be the first to utilize these capabilities and, when we see returns and benefits from the programs and solutions we develop, take them to our customers.

One example is what we call Equinix Brain, which brings together asset, people, and customer data and gives us a powerful view of how our business is connected across the assets we operate and the customers we serve. From there, many applications can sit on top of that data and use it in differentiated ways.

Building the AI ecosystem

Pankaj Sachdeva: You came to Equinix after leadership roles at SAP and Google Cloud. How has that unique experience of being adjacent to the industry but not directly in it shaped the way you lead the company?

Adaire Fox-Martin: I’ve always approached business from an outside-in perspective: Understand the customer and the customer requirements and make sure that what we do internally has a genuine customer-first mentality.

If something isn’t going to enhance the customer experience, engage customers in a positive way, or improve the outcomes of the workloads they run at Equinix, then you need to question why you would do it. That outside-in lens helps you focus on what is most important: the customers and partners that use Equinix to run a significant proportion of their businesses.

Pankaj Sachdeva: The technology ecosystem is also changing quickly, with AI-native companies and new kinds of infrastructure providers emerging. How does Equinix’s ecosystem evolve for the AI era?

Adaire Fox-Martin: I believe we’re leading in our thinking about the role of ecosystem here. It’s a continuation of a role Equinix has played through previous technology eras.

As the internet became pervasive, Equinix served as a neutral platform for carriers and network service providers. The same was true as cloud grew into the technological backbone for many of our customers. The connectivity and native cloud on ramps in our ecosystem helped customers navigate architectures in which their data might be present in more than one cloud.

As we move into the era of agents and inference, we believe we can play a similar role by providing an ecosystem and becoming a trusted exchange for these kinds of transactions. We’re looking at who the participants in that AI ecosystem will be: model companies, frontier-model providers, neoclouds offering capacity to enterprise customers, and service providers offering capabilities such as security—all housed on a neutral platform that allows customers to connect, distribute, and operate AI-based workloads.

Pankaj Sachdeva: Why does neutrality become especially important in this environment?

Adaire Fox-Martin: A neutral platform provides an environment that does not require lock-in to a single vendor. That is particularly important when technology is evolving so rapidly. Some companies that are large-scale participants in the AI economy today may not even have existed a few years ago.

You have to be open-minded about where this is going. An open, neutral platform allows a company to experiment—to access the best models, inference technologies, and service providers for whatever workload or process it is executing—while being confident that they are not locked in for the long term.

Explore a career with us