Submind YouTube summaries
Thumbnail for Distributed Inference: How to Fix Agentic AI at Scale | Jon Alexander, Akamai

Distributed Inference: How to Fix Agentic AI at Scale | Jon Alexander, Akamai

Watch on YouTube

Video summary

The video addresses a critical shift in the architectural approach required for artificial intelligence as the industry moves from model training to inference and the rise of agentic AI. While the centralized cloud model has been highly successful for training large foundational models by connecting GPUs into high-speed clusters to optimize memory usage, this same model is proving inadequate for modern agent applications. Early inference workloads like simple chatbots could be efficiently handled in a single location due to economies of scale, but the new wave of agents represents a fundamentally different class of software. These agents are no longer one-dimensional; they are multi-stage systems that execute complex loops, call various tools, and interact with external APIs, making physical geography a significant constraint rather than an afterthought. The core problem with forcing these distributed agent workflows into legacy centralized infrastructure is the compounding latency introduced by the distance between components. In a real-world scenario, data, users, devices, and the APIs agents need to call are inherently distributed across different locations. An agent typically runs on CPU infrastructure in one place, calls GPU-based models that may be hosted elsewhere, and interacts with third-party systems like OpenAI or Anthropic that are physically distant from the user. When an agent loops multiple times to gather context and execute tasks, every interaction adds network delay. For instance, adding just 100 milliseconds of latency per loop can result in a total wait time of ten seconds if the system loops one hundred times, which is acceptable for a chatbot but disastrous for real-time physical AI systems where such delays render the technology unusable. To fix these issues at scale, the industry must adopt a distributed inference architecture that respects the physical reality of where data and users reside. Instead of trying to force all compute into a single massive data center, solutions need to be designed so that agents can interact with the real world without incurring prohibitive network delays. This involves recognizing that the power of an agent lies in its ability to access file systems, execute code, and call external tools, all of which happen outside the confines of a centralized cluster. By acknowledging that the real world is not centralized, organizations can build systems that minimize latency by keeping inference closer to the data sources and users, thereby enabling agents to function effectively in time-sensitive applications without suffering from runaway token usage or performance degradation.
Read the full video transcript
We have spent decades moving everything to the cloud and it's solved a lot of problems, but when it comes to AI inferencing specifically, it demands a fundamentally different architectural approach. Why is the centralized cloud model not working for agent AI? >> It's a good point, and I mean the the the cloud has been massively transformational for for for many enterprises. So, the ability to rent infrastructure versus own infrastructure, the ability to consume infrastructure as a as a service versus as as as infrastructure. Like these have been transformational. Um but we think about AI we we're going through a big transformation in terms of the types of applications that we're seeing running um on top of the the infrastructure. So, a couple of years ago, um a lot of focus was on um training models. And so, a a lot of the um uh discussion and and architecture focus was on how do we support very large clusters of coherent GPUs so we can uh connect these uh GPUs together with very high-speed interconnects to create this large uh coherent cluster with high-speed interconnects between each of the GPUs so they can share the memory uh that that's available to them. And so, memory was really the constraint that we're optimizing for until we create these very closely connected um clusters to achieve um that high-performance computing infrastructure that allows us to create these amazing foundational models um that we're all familiar with today. That's all going to shift as we uh put more of our focus into inference. So, running workloads on top of those those models. Early inference workloads were were were chatbots. They weren't particularly uh complex. Um and so, those could be deployed into centralized infrastructure. And again, having um having those run on centralized infrastructure has got some some value. So, if you've got massive compute all in one location, you can often drive um economies of scale there. You you you can drive efficiency in terms of placing workloads and optimizing usage across many many users. We're at the cusp now of the the next wave of adoption where we're moving into the rise of agents. So, Open Claw, I'm trying to remember exactly how long ago it was, 6 months ago I think it kind of exploded onto the scene, really kind of started the um uh main main wave of adoption around kind of agents. And what we're seeing from our customers now is as they think about agents, these aren't sort of one-dimensional chat type applications, they're multi-stage, multi-turn applications that are calling tools, they're they're looping multiple times around these kind of inference loops. They're pulling in huge amounts of context to give the best answers possible. And physical geography is becoming a really important constraint in these architectures. >> When enterprise try to force these new agenting workflow through that same legacy centralized infrastructure, what exactly breaks? Because often they don't realize it, all they see is performance hit, higher bill, or runaway token usage. But what is actually breaking underneath? >> I think that the kind of simple answer is like, hey, the the real world isn't centralized. Data, users, the devices, the APIs, and all the decisions you're making are distributed. Like the when we're delivering a a service in production, it doesn't all sit in one data center and like not everyone is physically close to that data center. And so as you think about it, like what what an agent really looks like, you've got a combination of the the actual agent which runs on top of CPU infrastructure, it calls out to models that are running on GPUs. Those can often be multiple models. So, you can have different types of models that's calling. Maybe it's calling Open AI, maybe it's calling Anthropic, maybe it's calling open source models that are hosted on on different GPU infrastructure elsewhere. So, that even the models aren't all co-located. But, increasingly what we're seeing is the agents are dominated by the tools that they're calling. Um so, the power of agents um is around the memory system that it's got, the access to to the file system, the ability to execute code um in some kind of a sandbox, generate um output from the code that it's running, uh to call out to uh to third parties. So, to call a a APIs, to call MCPs, to call um um out to these external systems so it can interact with with the real world. So, this is the real power of an agent. And as it's interacting with the real world, that's not all in one location. And so, this is the problem that we're seeing is agents have to interact with the real world. That real world is distributed, forcing all of that to run in one location means that you're compounding latency. And these agentic systems, they run in a loop, they run multiple interaction on multiple times around the loop. And you're compounding that latency on every loop. And so, even adding 100 milliseconds of latency, if you're looping 100 times, that's 10 seconds of latency that you've added, which might be okay for a chatbot. Maybe a human's willing to wait 10 10 seconds for an answer to come back. But, for physical AI, for um for any type of real-time system, 10 seconds is way too long. It's it's a lifetime.