エージェント型AIのレイテンシーに対応する分散コンピューティング

In the age of compute, you see that people essentially are okay with their compute being somewhere else, with their intelligence, their data being somewhere else, and just counting on the internet to be able to deliver it to them when they need it. For many, the edge is the physical edge, the device edge. That, when we look at things like physical AI and robotics, is where when we talk about the edge, we're actually talking about the robot, we're talking about the assembly line. When we talk about the internet edge, which is a much more commonplace understanding of edge across the industry, that really is getting as close to that device, but not onto the device as possible. So AI to the edge will have nuances and different meanings across a variety of industries.

You cannot wait for large data sets to be sent to some faraway location where they will be processed and you will get a response. You need to be able to bring that intelligence to the edge where the data is being created so that you can act upon it instantly and be able to provide truly scalable performance. That's really where we think the edge matters.

Moving forward, getting AI to the edge is really going to be all about how do I route to the right hardware, to the right destination based on what my use case demands? And so it's not one-size-fits-all AI, it's AI tailored to the use case and leveraging the right provider, the right models and the right infrastructure depending on the outcomes that I'm looking to achieve.

One thing that will be pervasive is that every industry and every application of AI from core to edge will benefit from having distributed compute capabilities and dense computing at that internet edge so that decisions and data, one, could overcome the challenges of the speed of light. Two, we can apply security closer to where an attack might originate and further from where I might poison or otherwise compromise the models that I've spent millions or billions of dollars building and training. And ultimately it will provide us with a scalable ecosystem of partnerships where every enterprise will be able to maximize their version of AI investment with a consistent internet edge layer that will be able to manage their compute-intensive requirements and then make intelligent decisions, from caching to model routing to the leveraging of AI factories, depending on that use case in real time.

What we've seen in the last few years is that more information is generated at the edge today than at the core. And what we were able to realize is that the infrastructure that we built to disseminate information outwards can also help in addressing the information coming in, the many-fold information coming in. What we can do now is utilize that infrastructure, actually build intelligence on the edge that allows us to take this massive ocean of data coming in and process it at the edge where it matters. So that it does not need to flow all the way to the core before it can become useful.

The nature of the clouds is going to have to change. These are experiences that are going to be more about latency. They're going to be becoming product latency sensitive. Especially when you think about these agents that oftentimes are going to have, in some cases, hundreds of interactions to deliver the output for which you declare your intent, latency becomes one of the key constraints in how fast that experience is delivered to the end users.

And what you look at today, when you look at the internet, if you think about where the users are and where the big core data centers are, the latency may be about 100 milliseconds. And the model itself may take another 100 milliseconds to respond. When the numbers are both so large, it doesn't really matter that much. But as the models get better and the models start to respond within five milliseconds, to wait for 100 milliseconds of network latency just wouldn't make sense, especially when it's devices that need that inference to perform actions.

Inference allows us to action that intelligence and to tailor it for every individual user engagement. The thing that Akamai is looking to achieve with the Akamai Inference Cloud is the distributed dense computing capability so that you can apply those millisecond decisions and apply context in a cohesive and coherent way, applying security as you go, giving telemetry and observability data back to the enterprise to understand those customer engagements and where they need to tweak and tune them.

But the best thing about AI is that it's evolving for us. And if we build it correctly, if we have the right guardrails in place, and we follow the best practices that we've built up for the last several decades as an industry, then I think that AI and AI inference at the edge looks to unlock the promise that AI factories have been delivering to us so far. And that's going to carry the industry into the next five to 10 years of growth.

The edge is going to be transformational in the quality of all these experiences that get delivered. As these agentic experiences become more sophisticated and there's all kinds of interactions going back and forth, latency becomes one of the key multipliers, one of the key parameters that determine the quality of the experience, right? And then it all becomes about proximity, right? How close can you get the token factory or the thing that is ultimately doing the work close to the end user, right? And this is where Akamai, I mean, that's our bread and butter, right? That's what we know how to do best. And that is also why we built this massive network of servers all around the world.

We absolutely see customers zeroing in on a couple of key use cases. If I think about, for example, in the media and gaming space, where we need to be consuming very high bitrate video, so 8K video, immersive experiences that would turn into VR or open up new markets for video generation and video delivery in real time, there are a lot of excellent use cases, and the Blackwell RTX Pro 6000 Server Edition architecture is really designed to maximize that use case. Akamai has also been adding custom-built GPU capabilities to our platform to handle things like recommendation engines. Moving forward, we'll be looking at things like fraud detection, which we're already working on with certain Lighthouse customers.

And so when we look at what are the sorts of use cases that feel appropriate, it's where something needs to engage with the user, where that user has already high expectations for not only immediate delivery of whatever the insight or the intelligence is, but flawless quality, where they don't have to consider the security of engaging with whatever that experience is as they do so. And that's commerce, that's gaming, WebRTC communications evolving into more compute-oriented, integrated media experiences, moving into how we consume sports and all of the different markets that sporting is opening up for us, whether that be investing, understanding players, the way that we might cheer and review statistics for our favorite teams.

And if we go into commerce and retail or even travel and hospitality, it's making the engagements of those brands feel like they're finally doing something with all of that information and the loyalty that we've been paying and providing them, by really tailoring the experiences to us. So not making me work so hard when I want to buy whatever the latest good is from my favorite retailer. Or when I show up to the hotel of my choice, the ambiance in the room, what's playing on the television, my checkout experience, all of that is tailored for me. And all of those little micro decisions that I typically make, that frankly can be exhausting throughout the day, turn into a delightful experience for me because they're all happening based on my real-time interactions.

We have the data on our devices. We have the applications on our smartphones. Now we're adding generative AI and AI inference capabilities to those to provide us with that benefit of the real-time decisioning that many people have been forced to do on their own, that we're now able to not just automate, but make truly intelligent, and generate real-time responses and real-time adjustments based on what we do without having to think and interact directly with the code or with the application, but merely just allowing inference to carry us forward into that truly engaging future that we all imagine is right around the corner.

I have a hard time imagining that you're going to be able to do anything on the internet in the next three, five years, and you're not going to be doing something that is agentic or AI in nature. I think the transformation is happening incredibly quickly. Now, if we think about how businesses are adapting: right now, AI is happening. Whether you accept it or not, businesses are currently taking the steps that they need to be aware of where AI is impacting them and then to develop strategies for how they are going to mitigate or manage AI in that sense. So as enterprises understand how they're going to be evolving their AI strategies, Akamai wants to make sure that we can continue to partner with every business to provide them with a platform for their application and security needs today, but laying the foundation and the groundwork for the applications that they will evolve into tomorrow.