Equinix, Inc. (EQIX) Earnings Call Transcript & Summary

It could serve as an accelerant to the growth as we talked about, but our long-term guide is 8% to 10% as we gave at Analyst Day.

Right. So we've kind of sized up a few of these things kind of talked a little bit about the basics. So this is kind of where the rubber meets the road, I think, a little bit for the market where Justin -- the market were -- I guess, we're getting ready for a wave of kind of AI training engine development, which is going to require a lot of new investment, likely for dedicated new facilities, which are optimized to address some of these power density requirements that the training engines require. So, to what extent -- if you agree with the argument that the first instance of kind of AI development in the data center is the training part before we get to the next part, which is the inference and applications adoption part. Would you -- first of all, how will Equinix participate in this? And would you agree with the assertion that given that super dense data center development isn't necessarily wheelhouse, you're not necessarily going to be the first data center guy to benefit?

Great question. Well, as I mentioned earlier, AI is a pretty broad umbrella. And then with AI and machine learning use cases, when we look at deep training, yes, training and inference are specific stages of deep training, but across the Board, there are many more ML/AI use cases that have a very wide range of requirements in terms of power density, in terms of compute, interconnect and storage capabilities. As I mentioned, I think while AI plays an increasingly critical role for businesses to stay competitive, I believe AI will be one of the many pieces that will enable businesses to succeed today in the digital age. I personally believe that AI will deliver an optimal value when seamlessly integrated with the rest of the end-to-end digital infrastructure architecture. And while some AI training use cases, specifically large or very large models can introduce new infrastructure requirements, building dedicated data centers for AI with brand-new designs might not necessarily be an optimal approach. And today, we are closely investigating the emerging AI-driven requirements in terms of power density, cooling, and more performance interconnect and storage capabilities, and we continue to evaluate next-gen technologies in these areas, both internally as well as with our partners to stay ahead of the curve. And it is also important to note that not AI workloads are equal. And as such, there won't be a one-size-fits-all approach to meet the wide range of requirements. In our view, the model size, the training data set size, the cleanliness of the data set, the requirements around retraining frequency and specific requirements around compliance and security altogether will determine the right training solution for a particular use case. For example, our xScale portfolio today could well be suitable for large-scale training requirements, such as Gen AI or large language models, and our retail portfolio and the digital edge can offer an ideal home today for use cases where the AI models need to be frequently retrained and where strong security and compliance requirements need to be met, for example. And demand forecasting is a good use case. Risk analytics is another good use case. And there are many other enterprise use cases that we see across the board. So I hope that answers your question, but I would just summarize it by saying that AI workloads come in many different shapes and forms, and there is a very wide range of infrastructure requirements, both across training and across inference. And again, AI in order to offer the best value to the business, it has to be part of an overall digital infrastructure architecture. It has to sit close to where sources of data and the users of data are, and again, depending on the use case, whether you need to retrain or whether there are latency requirements, the optimal architecture can be quite different.

Okay. I think we want to -- this might dovetail another question. For those on the webcast, if you have questions that are coming up, there's a little bar, I think, underneath our presentation here that you can type in some questions, and they'll pop up here then I can try to feather them in, and they're coming in already. So you guys already figured it out. Okay. And so we'll get to some of those. So the thing that I think people think, and Justin, maybe you just talked a little bit about it, is that there's a monolithic sense of the training engine being this hyper-dense 40kW plus per rack, dense, liquid cooled data center occupant and that your facilities are simply -- were never simply designed to allow for that. And while the occasional hotspot might be able to be accommodated, you can't really scale it inside your existing portfolio and maybe not even exist on your existing build-out road map. But your xScale program, as you say, is much more of a kind of raw material kind of -- it represents an opportunity to kind of do greenfield things that are more designed to address those types of builds, that may be right or wrong. But you pointed out and I think many people believe that the inference element, which is typically going to be likely multiples of existing traditional CPU compute, but likely fractions of training density in terms of power consumption per rack, these are the things that could more easily live at some meaningful scale inside the Equinix facilities as designed and then obviously inform future development at the margin. So do you agree with the idea that Equinix is more of a business model because of its latency interconnection centricity, density of customers that it's more of an inference play than a training play?

I would say it's an end and we can come back to that. I would say, in addition to the elements of training that we talked about and many training use cases that we can support today and will be supported as part of our technology road map. Let's talk about inference for a second. So as it retains to a retail portfolio, I think as we briefly talked about it, we think that our retail portfolio today will be an ideal home for inference, particularly for use cases where the AI models are highly dynamic and where the insights generated based on these models are more real-time and mission-critical in nature, because these really drive very stringent performance and latency requirements. Overall, even though we are in early days of AI adoption within the enterprise, we are seeing wins. For example, as we discussed in our Q2 Earnings Call in early August, we have seen specific instances of interconnect to support AI. In fact, we had a pretty significant win in Q2 in the AI space with an AI as-a-service provider and that put their network nodes with us to really drive interconnection to the multi-cloud connectivity and to support the inference and interconnection to the cloud. Now from a customer lens, as we talk with customers and many AI practitioners, some customers are saying that they have private data that for a number of reasons, they want to retain control of, but also want to be able to connect to AI services. Customers can use interconnection today at Equinix to be a direct connect -- I'm sorry, direct cross-connect their way to the hyperscaler of their choice, leveraging our 40% market share of cloud on-ramp in the markets where we operate today globally. On the other side of the equation, you're seeing some newcomers enter the market to deploy GPUs as a service. And these players might deploy their training in a low-cost power market. But the fact is that once their model is trained and they think about inference, which with scale, it usually requires a much more distributed footprint. They look at Equinix, which with our broad geographic footprint of data centers across the world means that 80% of the population in North America, Western Europe and many large Asian metros today lives within the 10 millisecond round-trip network latency of our facilities. So with that kind of geographic reach and latency and proximity to the population, Equinix is the logical place to put inference nodes. And we believe that cloud service providers are likely to play a significant role in AI, and we already have a deep relationship with them as well as with enterprises that want to consume those cloud services. We also believe that new ecosystems could be created by the demand for AI, and we are working to actively seek those ecosystems in our facilities. And finally, our view is that with scale and adoption, inference will be increasingly in edge play and infrastructure providers with densely interconnected ecosystems and distributed footprint will be best positioned to serve those types of workloads going forward.

So we've got a number of questions that have come in. Before we kind of start hitting those and again, if you guys want to just type in something, I'll try to get it as we kind of move into the final phases here. But I guess this is a big question, Justin, for us, which is, are data centers like Equinix, given that they've been designed to 110, 120 watts per foot types of density. That's like what exists today. You guys have expressed an inclination towards a little bit more density, but not dramatically larger. Does the data center industry represent an enabler or an obstacle to AI adoption. And does Power represent a limiting factor in the pace at which it's possible to adopt it?

Great questions. So Equinix is an enabler for AI for sure and will continue to be an enabler for AI. And as we have shared, at the crux of AI, it's really about data and the ability to optimally process the right data sets and share the insights from that data processing. And data is mostly generated and consumed at the edge. And today, Equinix is well positioned to support the building blocks for AI, again, both for training that requires proximity to data and the distribution of the insights in terms of inference, which will increasingly be an edge play as we talked about. So at Equinix today, we provide the most efficient ways for enterprises to effectively move their data from the private infrastructure to end users as well as from their private infrastructure into the multi-cloud environment and to the SaaS providers. These are all capabilities which have been core to our [ Equinix ] history. We don't believe that there will be a significant need to retrofit facilities. We certainly continue to invest in emerging technologies and architectures for our next-gen design to ensure it's evolving and keeping pace with the market and latest innovations. One of the strengths of our retail business, for example, is that when you serve a very broad range of customers with deferring density requirements, you're actually able to sort of dense up and extract more from the system over time. On the xScale side of the equation, it's a little bit more challenging as you typically allocate power for the building to one or two customers. And within our co-innovation facility in Ashburn, we are actively investigating and testing various technologies, including liquid cooling to further optimize our current designs and be able to implement it as a more standard capabilities -- capability in our go-forward builds. Now in terms of power availability if we have time, with power being a critical element of what Equinix delivers to customers, we take a long-term lens on our power procurement planning by securing power ahead of our breaking ground on our developments. In some cases, we are working with municipalities or local utilities that are facing resource constraints around power to secure our build capacities, for example, in Northern Virginia or in Singapore. We also partner with utilities to share a roadmap of our own future power requirements, because our fill rates are pretty predictable, that can be done pretty well, and we are leaning into our sustainability agenda as well to demonstrate to municipalities and governments that Equinix is and will be a responsible user of their resources in the community. In contrast to wholesale and hyperscale players are incremental power demands in any given year are smaller in scale, but have actually consistently grown over time. In certain markets, such as in Silicon Valley or in Dublin, we have also implemented self-generation via the use of, for example, Bloom fuel cells as a primary power source. And because of our retail focus model, we have multiple options around power availability as we continue to expand our footprint and bring up new locations in the coming years. Now in addition to power distribution availability, sustainability is obviously top of mind in order for us to measure our power consumption and the right type of renewable energy coverage. In discussions with AI practitioners, as I mentioned before, really, the key decision drivers for them today are location of their data sets, and data gravity and a distributed footprint, which all play well into Equinix's retail colocation sweet spot. I think we are still in early days for AI deployment for the enterprise. And as innovation continues to develop in this space, I think we will see a pervasiveness of AI use cases and workloads across many new verticals and industries. And as AI infrastructure continues to get more distributed and move toward the edge, we actually see a wide adoption of hybrid multi-cloud architectures for AI use cases, which again lands itself very well to our strengths. We very much look forward to continuing the AI journey with customers. And if you have any additional questions at this point, I'm happy to answer them. Thanks very much, Dave.

Okay, tons of questions. So one question, and I don't -- actually the beginning of the answer doesn't even -- I don't really get. So Cisco was earlier talking about the evolution from InfiniBand to Ethernet. And does that create any obsolescence risk for your interconnection business as you talk about the necessity of low latency interconnect with respect to AI?

My short answer is I don't see that as an impediment or any issue to make our interconnection capabilities obsolete. Yes, I will be closely and actively investigating and monitoring these technology trends. Interestingly, there's a lot of innovation in the networking space these days because networking, if not done right, could become a bottleneck in the AI training and inference and as you know, interconnection and highly performing interconnect capabilities have been core to our business model for many years. So the short answer is no, just the evolution and the trend making Ethernet one of the viable technologies in the interconnect space is a natural thing that we see as many other developments in the industry. So that is in no way a showstopper or any impediment for our product portfolio.

So another question is with respect to kind of the rising density of power consumption in the data center, obviously, power exists, right? But in order for it to exist in a data center, in order to increase the power density of a data center, it requires an increase in cooling capability, maybe a diversity of cooling technologies to require an increase in N+1 generation capability, diesel backup power, battery backup power and so could you talk about the step function increase in capital expenditure intensity for building an ideal Equinix hybrid AI regular data center versus what maybe Equinix thought these data centers would look like from a capital intensity standpoint 5 years ago?

Yes. I can talk about several elements. The short answer is that I don't think there is a one single formula that works for all use cases, as I think the keyword that you mentioned is that it's going to be a hybrid mixed environment. Different use cases will require different types of energy and power requirements. And I think as the CTO, I spend a lot of time with my team and with my peers and looking at how we can actually leverage software and AI to create a much more dynamic architecture such that we can measure the power consumption and distribution at a very granular level and effective data feedback loop where we can actually act upon those insights and control how power is distributed and how it's consumed. And based on different use cases and different requirements, even use different sources of energy for different use cases. And it's a very fascinating space that is fast developing. It's really the application of AI in the physical infrastructure. So we are investing R&D efforts in that space. But to answer your question, because of our neutrality, kind of a tenet as part of our business model, we are really trying to enable as many use cases as possible on platform Equinix. And different use cases have different requirements in terms of power density, flexibility between different sources of power, the granularity at which power needs to be measured and modified or configured differently.

And Justin, I would just add on, Dave, as you think about it, it's something we've done over the course of our history is managing across a number of different customer workloads. When you go into our facilities, it's not a homogeneous type of workloads. You have a number of different customers with a number of different types of workloads and jar caps and things like that. And so today, we may have in a facility, it's always kind of like playing tetris as we lease up a facility where we may have what we call today a very dense deployment of, call it, 10 kW per cabinet in a certain spot of the data center for. Next that we'll put a networking deployment where we have 2 kWs' per cabinet to kind of balance that out, but it's something we've always managed through across our portfolio.

So I have lots of follow-ups on that, but I want to keep asking the customer questions. One question, Justin, you kind of mentioned this idea of data gravity. And one, client kind of notes, I'll throw out an example that Meta has been reported to be investigating a gigawatt deployment in Wisconsin because the land is cheap, the power is available and there seems to be an indifference, if you will, to the notion of data gravity because meta can create its own data gravity. So are you at all concerned that as large language models built in huge data centers, move data center gravity away from city centers by necessity that somehow your advantage as dilutes down through time?

The short answer is, I don't believe so. I think there is -- the pie is big enough for a lot of use cases. There are definitely use cases that can benefit from a large training kind of deployment at a place where maybe power is more accessible or that use case does not necessarily require a lot of interconnection to other data sets. And typically, that's the case with applications like maybe ChatGPT or public data, I would say, as opposed to enterprise private data. And in many discussions with customers, their data is actually their most valuable asset, and they want to keep it where it is, and they want to continue to update it and they want to continue to augment that with other data sets. And moving that data out of where the compute is to a remote place and then do all that computation and then sending the inference workloads back to the edge, I don't think that's the most natural process. And again, AI practitioners today say, how can I leverage my data set today that is sitting at Equinix or in the cloud? And how can I just augment that and derive new insights from it and make sure that it stays private, it stays secure, and that I don't have to do a lot of extra work to do AI on it. And again, in the example that you mentioned, that use case might benefit from that particular deployment type. But across the board, we are seeing that a lot of AI practitioners are saying how can you help me leverage my data? How can you leverage your compute storage and interconnect capabilities for me to do AI training and inference where I am today, any close proximity to cloud providers within single digits and latency.

So Justin and Katie, thank you for being with us again today. As we wrap up, what I wanted to do is just kind of really tackle like the number -- and you've outlined all the opportunities and all these things, and you've kind of produced a lot of numbers, and we appreciate that. I want to tackle the one, I think, lingering criticism. It's that AI has changed everything that everything that we used to think could be done in a data center can no longer be done there. It requires far more density of power and far more ingenuity and cooling, and no data center that you know exist today could possibly handle it, your roadmap couldn't possibly handle it. All of this has to be done by brand-new people coming in with brand-new tools and that Equinix is a dinosaur and it can't really participate. Justin take us home.

I have a slightly different view. As we mentioned, we talked about the fact that AI is not necessarily a new thing. And we have seen a continuous progression of Moore's Law that has really enabled so many different use cases. And I continue to believe that as you look at the end-to-end AI/ML workflow, there are so many ways to optimize the output that just throwing power density to the problem might not be the single kind of bullet that solves all the problems. And sometimes, it might actually make things less optimal. So again, across the board, we are seeing that today, there is a mix of hardware, technologies and compute requirements that can successfully meet the wide range of enterprise AI use cases. And the fundamental building blocks are continue to be around compute, networking and storage and when these things come together at the right place close to the data that matters close to the ecosystem participants that matter, all of that together actually define the success or the failure of AI kind of deployment model. And again, I'd be happy to continue the conversation and share the experience and the learnings that we see across the board with the 10,000 customers that we have on our platform, every one of them is looking at ways to leverage AI and again, there are recurring themes that we are seeing and we are able to share with the community. And again, AI is about where data is. AI is about bringing compute to the right data sets and doing that privately, especially in the case of enterprise and being able to elastically increase your inference capabilities and have inference where it's needed. Obviously, we are not saying that there won't be any new power density requirements. And we are working on these technologies, and we are enabling them at our data centers as well as our co-innovation facility in Ashburn. I hope that answers your question somehow.

That's a great place to leave it. Justin and Katie, thank you so much for joining us. We really appreciate you being part of our first ever global 2023 AI conference. It was really, really fun to talk to you guys looking forward to staying in touch.

Thank you very much.

Thank you.