Data, Scale, and the Future of Inference at AI Infrastructure Field Day 4 – Tech Field Day Takeaways

Hi, I'm Alistair Cook. I'm an event lead here at Tech Field Day. And here are my takeaways from AI Infrastructure Field.

Day four, this edition of AI Infrastructure Field Day felt like there was a lot of networking. We had multiple vendors talking about networking, but one of the things that I saw in it was that that networking has more compute actually inside the networking. There's more thinking going on inside the networking than we previously seen.

That's an interesting trend to see a different requirement from the network in order to support ai. My first takeaway is that AI infrastructure must be designed to handle very large data volumes. Now, we've seen this before when we're talking about the, the training phases where there are these massive hundreds of gigabit links run and multi terabit switches being connected to the GPUs in order to feed them with data and, and move the data between them.

But that's the training data. Uh, and while that's still vital, and of course we saw the network that deliver this, the storage network that would deliver, uh, information into the same training servers, but also we're seeing inference, you know, the place where we could get business value out of AI driving networking too. And so the ability to handle large volumes of high quality data has been really much a, a focus in this, this particular event and this of switching and delivery across the entire, uh, event.

My second takeaway is reluctant to the first in, in that data always matters. Now, this is really not something AI infrastructure specific, but I think there was some discussion early on that AI might be able to deal with the fact that we have low quality data and to actually refine that data. For us, what we've seen is that having a good data pipeline into your AI pipeline is absolutely vital and having good data from the start.

Now we understand good data being fed into training, but also good data being fed in for inference. Uh, we really need to be thinking about the data pipeline for inference to make sure we're getting fast, efficient and accurate results outta our relatively expensive large language models that we're using. So having good quality data fed in means that we're not using excessive wealth tokens, which is a proxy for cost these days.

Data still matters. It always, well, My third takeaway from our infrastructure field day four is that AI will be different everywhere in terms of there isn't just a single workload that is ai. We started with the awareness that the training phase was a pipeline and there were different workloads feeding into that pipeline.

Well, fine tuning like training, uh, that will often happen in large data centers and the cloud and, uh, be running in a, a very dense location. But we're increasingly seeing the inference phase again, the time when you get business value out of having a night. This doesn't always happen in the cloud or in a data center.

We're seeing increasing amounts of AI inference happening where data is being generated, whether it's on physical devices or embedded devices. We're seeing entire new classes of compute being available for these embedded devices for these smaller devices. So where you might have previously needed a, a big server with, uh, big GPUs attached to it out as your edge compute to handle data streaming from multiple cameras, you may now start to see a more dedicated device.

And you may well start seeing AI models that are dedicated to specific functions rather than just using the largest large language model that you can fit out on top of whatever hardware you've got. As we're gonna inference, gonna business value and seeing this being an ongoing workload that's gonna be running for days, weeks, months, years, we need to have specialized hardware and specialized models that fit together to efficiently deliver that business value. I, I really enjoy all tick Field Day events.

AI infrastructure Field Day is certainly one that I, I find really enlightening because there's a lot of new things going on. We particularly saw Excite Labs showing us brand new hardware that they had designed that was specialized to some of these new functions, particularly for scalable AI for training, but very much for AI inference going on somewhere, which is not the century of our data center. The other new sponsor, new presenting company that we had at AI Infrastructure Field Day, uh, this time around was fabrics ai.

And again, as we saw at the previous edition of AI Infrastructure Field Day, moving from a few test cases to lots of production deployment is really hard. And that's where Fabrics AI comes in. Uh, they will build agents using an agent framework to allow you to relatively easy scale from a couple of agents to potentially hundreds and thousands of agents running.

Another interesting insight that we got at, uh, the, the event was Brian Martin from Signal six Five Labs. Now Brian's been a, uh, recurrent delegate because he has some great insights, but he shared with us some of the amazing capabilities that the six five, the Signal six five team have to do actual physical lab work around AI building out through large infrastructure. Uh, do check out that video and check out Signal six five as well.

They produce a lot of really good performance and comparison oriented, uh, reports. Thanks for watching this episode of the Tech Field Day takeaway series on the Tech Field Day plus YouTube channel. If you enjoyed it, be sure to like, subscribe and share your thoughts in the comments.

Follow Tech Field Day on X, Twitter, blue Sky, mastodon, any other social media that happens to appear and look for our updates and check out our presentation videos on the Tech Field Day website and the Tech Field Day YouTube channel. Our next event, my next trip to Silicon Valley will be for Cloud Field Day 25, March 11th and 12th, 2026. Tune in live on our website, on our LinkedIn live page, uh, also on text, on tv, as well as the on the Main Tech Fields Day YouTube channel.

Thanks for watching and we'll see you next time.