AI Model Routing Explained: Save Costs and Prevent Outages | Utilizing AI Podcast Episode 32

One of the challenges of building AI applications is the constant churn of new models and services, and this is where model routers come in. This episode of "Utilizing AI" features a discussion of AI model routing with Jay Cutrell of Nexus Tech, Dave Graham of ML Commons, and myself, Stephen Foskett. Welcome to "Utilizing AI," the podcast focused on practical applications of artificial intelligence from the Futurum Group.

Every Wednesday, we explore news and use cases of the ways in which AI is transforming enterprise IT and the industries it serves. I'm your host, Stephen Foskett, president of the Tech Field Day Business Unit here at the Futurum Group. Before we dive into this discussion, let's meet who's on the panel today.

Hi, I'm Jay Cutrell. I'm the chief product officer at Nexus Tech. org.

It's had many different names. You might know of "Fudge Factor," "What the Fudge" is the podcast. There's even "Fudge Sunday" for the Sunday compendiums I was putting together.

org and get a good scoop and mm-mm, good. That is just the pinnacle of excellence, and I'm in marketing. Hi, my name is Dave Graham.

I'm the director of marketing for ML Commons Association, a technical standards organization that exists to serve, benchmark, characterize, and quantify the AI industry at large. So, you can find me on LinkedIn. Unfortunately, I do not have any nom de guerres that Jay over here has, or any of that fancy stuff, or domain names for that matter.

com, which has nothing to do with anything that we're talking about today. So, there you go. Well, so the three of us were at AI Field Day together, or have been at various Field Day events in the past.

We've known each other for a long time. I think you can tell that we're a little bit comfortable with each other here. " This week, we are going to choose our own adventure.

We are going to talk about model routers. Now, this is one of those things that has risen in popularity, and anyone who's tried to build any kind of AI application, AI-powered application, has encountered a situation where maybe the model you're trying to hit is unavailable right now. Maybe you ran out of tokens.

Or maybe, and this is my dreaded, hated use case, that model's no longer available because we introduced a new one three weeks ago, and you didn't hear about it, and you didn't change your thing. And anyway, what if there was a system that you could just point your application at and say, "Hey, give me a fast model. Give me a good model.

" Well, that's what model routers do. So, Jay, I'm going to throw this to you first since you were the one who suggested this topic. Great topic, by the way.

" Talk to us a little bit about the reality of this world. I think of it as a little bit of a throwback to the past in some ways. I think we've got enough things in the AI world that are taking place where someone might say, "Boy, that's a really nice, cool story about that toy you have.

" And so I think there's going to be this first grasp of, "Okay, great. I'll figure out how to talk to this one. If I can't get to that one, I'll put a conditional in.

" It's almost like you're writing software, more software, which means you have more problems. And so I think routing will come into this realm of possibility. And, one thing that I tried a couple of weekends ago was Elmo.

If you're not familiar with the Elmo project, it lets you just download, do some node stuff. It scaffolds. You put in your Open Router API, you put in your other API keys, and it just helps you figure out how to monitor your marketing brand online.

It's pretty clever. And what I realized was, wow, what a great one-to-one experience in making it super simple and easy to put in your Open Router key. Great.

What's Open Router? Well, if you didn't know, Open Router is one of these amazing new unicorns with, I think they're over a billion-dollar valuation now. That's the definition of a unicorn, which also sounds kind of quaint because we have trillion-dollar and quotably larger companies through.

But Open Router is probably processed across multiple different models and providers. I think they're over a quadrillion tokens that they've processed, which again, depending on when you listen to this, that might sound quaint, too. But they've done a quadrillion of these tokens in the past year or so.

And so I think of it as there's more things that'll probably be flowing through Open Routers. Maybe there's going to be other Open Routers that challenge Open Router. Maybe there's a more open Open Router, it just hasn't been created yet.

But I think that that's going to be an essential building block for the next set of AI-enabled or infused powered things that people believe are essential for day-to-day business. Yeah. It's a menu.

It's basically a fancy menu. It's, I want to do X, Y, and Z. Pardon the Canadian expression there.

To pick things, it gives you choice. When the alternative is buying down a fixed line of Anthropic or OpenAI, or God forbid, Grok, sorry, xAI, it gives you the choice to pick and choose the latest Kimi models or Moonshot or any of these things, right? And I think, as you mentioned, Jay, it's the model obsolescence these days is going so quickly, right?

6 is deprecated. It's now legacy And these things go away so quickly because we want to be on the bleeding edge. We want those capabilities, we want those concepts to be able to be played back in real time or to be integrated as quickly as possible.

I would argue that the best thing a model router has ever done, and OpenRouter does this, is model provider/model:latest, right? And then all of a sudden, every single routing request goes to that latest and greatest model and whatnot. So yeah, it gives you a lot of flexibility, it gives you a lot of choice.

It can be part of a tool chain, like Go, you can prioritize your tiers. OpenRouter has Owl, which is their text-only kind of Haiku alternative, for example. Which is great.

It's free, it's cheap, as in I get zero, zero, right? Zero input, zero output kind of thing, and it's great for just doing initial. So then you can kind of do comparisons, and you can look at different capabilities between that.

So there's a lot of good work that's been done by these model routers, specifically OpenRouter, that we're talking about right now. Yeah, and this is, I think the thing that it may not be obvious if you haven't built applications like this that have lasted more than a chat session, but what you're describing is reality. 0 Flash," and suddenly that one's no longer available.

6 thing. It's like, "What are you talking about? I just set something up using this one.

That one's new. " But it happens a lot, and it can cause problems. Is that really what it's for, though, is basically for availability, or are there other angles here that these routers or other possibilities that they serve?

I think availability is certainly one of the bigger kind of classic use cases for it. You're calling to a single API, which then ladders down to multiple providers in the backend. So if one provider disappears, you still have that model call that's going to proceed through there.

ai will help you with. You can see the pathing, and you can understand a little bit more context there. But that's one of the first defensible points.

It provides that kind of robustness of, I call once to an API, it ladders and makes sure that that pathing is done correctly. I think the second thing is, again, with the principle of single API call equals, yes, it gives you the flexibility of now plugging different types of things. Not every agent needs tool calling, not every agent needs reasoning.

And so the ability to still maintain from a single call to intentionality or purpose or agency, to be really punny about it, to maintain that kind of stream through there. And then I think one of the third things that we're noticing these routers do is cost control. So the ability to task a specific API key to a specific agent type, which will give you a certain cap.

" So it's cost optimization. It's kind of the economy or token economics, if you will, but in a different type of spin. 5 Pro Max whatever thing you got, $15 per every input token kind of thing.

So I don't know, Jay, your lived experience is probably valuable here, too. Yeah. At the risk of letting people know exactly how old I am, I remember back in the old days when internet access actually came in two flavors.

One, high-performance business class internet access at your office, and, mm, cheap packets. It was delivered sometimes in the same order that you asked. And so you would have these, sometimes you'd have circuits, multiple circuits to the internet, and you'd have a router that would load balance for you effectively.

And so in the old days, you'd set up something, there's a term called BGP. So if you go to these network field days, of course, people have a lot of opinions about things like that. But there was also the internal version, that IBGP, for within your border.

And you'd figure out, is this the business class, high-performance circuit, and is this the cheap and deep version of a circuit? And then how do I decide at what time and what person is using the service? Do they go out the cheap and deep, or they go out the business class, and who gets what and why?

And so that unit economic story, I think that I'm talking about things that are 20-plus years ago, but I think that it's relevant now because, as Steven said, these applications aren't just little small mayflies that have a brief but beautiful life and then go away. Some of this stuff is going to get baked into the institution. People are going to come to rely on that particular agent.

And so, Dave, your point about resilience, resilience, resilience, economics and resilience paired together, that might be the new Oreo cookie flavor for 2027. But that's what I think about this right now. Yeah.

The economic aspect, I think, is really important as well. Like you said, well, first off, I should point out that there is still business class and non-business class internet, and most of it now is about oversubscription rather than about actual throughput. But that being as it may, I think that's a good metaphor for what's happening here.

On the economic factor, I think that that's really true. It's very difficult To keep up with the models, because the pricing changes constantly, the token usage changes constantly. As you said, you want to make sure if you exhaust your tokens, you want to make sure your application doesn't fail.

It occurs to me that the flip side of that is also true, right? Many of these applications or subscriptions include an allocation of tokens that you want to make sure you're using. I actually got an email the other day from OpenAI saying, "Hey, buddy, we haven't heard from you lately.

What happened to your token use? " Well, routers can help you with that aspect, too. Essentially not just keeping you from using tokens, but making you use tokens, making you use the tokens that you've bought, right?

That's an important point, because if you're, for example, a customer of Microsoft, you may have what is known as a commitment to consume or an MACC in the parlance. And so you making the best use of what you've already decided you're going to spend with that company is an important point, and that's a function of the procurement side of an organization ensuring that, "Hey, we went to the, was it the mat? " And some of that could be consuming tokens.

And to your earlier point, what kind of token and what is the periodicity we're seeing in the token usage? Because we're already starting to see the first stories of token maxing, I think is what the kids are calling that these days. But there's already people taking the token maxing awards away because it was really like most things, experiments.

We get data, now we move on to the next experiment. But how you're token maxing or thinking more carefully about what you're doing with your tokens is probably another important element of this that most people don't build that telemetry in day one. They only find out about the bill later.

Shocker. Cloud bills, meter is running. Lather, rinse, repeat.

It reminds me of the subscription sprawl that we have for streaming services, right? I subscribed to, I don't know, Peacock. Do I still watch Peacock?

Is there anything good on Netflix? Could you imagine if... I mean, wouldn't it be cool if your Apple TV or Roku or whatever was like, "Hey, bro, you haven't watched anything on Peacock in the last month.

" It's kind of similar, because a lot of us have, I don't know about you, but I definitely have Google, Anthropic, OpenAI subscriptions that I would like to make use of, but when I get frustrated and want to strangle one of them, I switch to the other one, and then I am not making use of what I'm paying for, just like I'm actually not watching anything on Peacock. Yeah. I think, yeah, that's part of the entirety of this thing.

So again, I'm probably the classic Anthropic user. I have my 20X account with them or whatever it is this month that gives me the features and functionality. I'm designing a slide deck in the background as we're talking here, using Claude Designer.

I'm consuming those tokens, but Designer sits there mostly idle during the week. So a function I don't use. When we start to look at things like Open Design, which is a Claude Designer alternative that's in the open sphere, starts to pull in multimodal models, it starts to pull in different features and functionalities.

The premise of this routing concept becomes even more brilliant because, again, what are we looking for on our menu? We're looking at the appetizers, we're looking at the mains, we're looking at the desserts, we're looking at the drinks. And all of these things have different flavors.

They lean into different sensibilities during the time that we're interested in those things, right? And so if we're looking at from marketing context to where I sit, it's not a Mac by any stretch of the imagination, Jay, but I have a commitment I have to... I have a business need that I have to commit to.

If I'm paying all this money, again, to the overall arguable point here, I also have a need to ensure that the outputs of those things mirror the reality of what I want to do. And so being able to plug in a VLM, a visual language model, into Open Design or utilize that, but also figure out which one gives me the best return on my investment, and I'm doing it in pennies, I'm not doing it on a subscriber basis. Right?

I'm willing to spend 15 cents on a series of token calls and whatever, because again, I can insulate and isolate and understand through my routing surface what I'm trying to use. Or if I'm doing enterprise-level embeddings for a RAG-type engagement, I can go and choose the most current embedding model, but I can pick the one that I know based on testing and trying. So I think a lot of this is playground to production type thing that you made the allusion to the SaaS, that everything's a subscription.

Yeah, everything is a subscription. Even our hardware is getting to the point of it's just all subscription, right? We buy it.

But this becomes very powerful within here, especially if you're buying into an ecosystem of development and buying into these things. Open routers, one DigitalOcean just announced their routing interface. We've seen Ollama with Ollama Cloud is another router, even though it's not called a router on the surface, but Ollama's always been a router.

It's always been the playground of choice for a lot of self-development folks like myself, like LM Client, or any of these that are Misty, and all these kind of IDs that kind of integrate all these things together. And so what you're seeing is this absolute plethora of choice that we're going into, again, allows us to say, if I'm not watching Peacock, which I don't watch TV anymore because it's absolutely asinine stuff, and if I'm not watching Peacock, then I don't have to worry about it. But I can still get the on-demand flavor of the things like if CSI.

I don't even know if CSI is on Peacock, but if I wanted to watch that one CSI episode, I can go in there and get it, but I only have to worry about that perimeter, and it's only going to cost me 30 cents. 99 a month. The danger in that is that the economics of that at scale can suddenly exceed the bounds of that.

So if you're bought into an ecosystem, an Anthropic, for example, where I am going to literally use the cost basis of $200 a month or whatever it is, then it becomes more economical to just consume it there. It's predictable latency, it's predictable cost. It's I'm not going to exceed that.

The danger that we've seen, Jay, your token maxi comment is apropos to this. What we're seeing is people flowing by that boundary willy-nilly and not remembering or not putting the stopgaps in place in order to make use of that. And that's where the economics of just a couple of pennies more per API call versus first-party services becomes hugely interesting, right?

Again, that cost control and constraint idea. Do you think that proxies are just going to multiply like rabbits? Are we going to have our choice of different proxies to choose from, just like in the networking world before we had our choice of this appliance, that appliance, this thing, it all reduces down to it's an appliance experience, but do you think proxies are the next kind of wave of what we're going to see?

That's a Neocloud. A Neocloud is a proxy. There you go.

Well, so yes, I think this is important. I think the differentiation, this is going to be my cynical take on this. So cynical/optimistic.

Can you be both? Yes. It's a little bit bipolar.

Silicon diversity right now is almost nonexistent anymore, right? We have known quantities of stuff until you get into especially ASIC stuff, the Cerberuses of the world or those burned weight ASICs that are out there that are coming that are pretty incredible, right? But silicon is, for all intents and purposes, it is a fixed commodity.

We know who's going to be putting V200s, V300s. We know who's going to be having instincts. We know who's going to have...

It is. What really comes down to it is that proxy service or that routing service, proxy router is almost interchangeable, I guess, in a sense, and how you differentiate your things. So how does a CoreWeave differentiate themselves from a Vulter or a Nebulous or, it's not necessarily in the silicon offerings anymore.

It's maybe you get them first, but this is a war of attrition, right? Everybody's going to have that silicon at some point along the way. It's what you're able to do to reduce the amount of friction between user and the outputs that they need.

So what is that agentic interface that you're doing? How are you routing? How are you giving them the models that are the current hot thing, or how do you allow your users to token max the most effectively?

Hardware is certainly part of the equation, and I don't want to disclaim that, but a lot of that is reducing the friction. This is why DigitalOcean, for example, I keep coming back to them. They offer hosting, GPU droplets.

Now they have a model router that enables you to just stitch together at the same point. They've reduced the amount of friction it takes. You don't have to go outside the house anymore.

You go inside a house. You can go from everything from, I think you can do bare metal. I don't use bare metal, but you go everything from VPCs to GPU-enabled stuff, and you're set.

I don't have to go outside the walls anymore. That's a value chain that has just reduced the amount of sprawl and cost containment and all the things that matter to me as an enterprise reuser. And they've made it addressable literally through my cloud desktop now.

So I just call DO. DO provision. And I now have an entire production chain that's ready to go.

So anyway. Yeah. And to your point, guys, one of the things that occurs to me is this is classic aggregation theory.

So those of us who read Stratechery and Ben Thompson or Clayton Christensen are aware of his whole theory of how aggregators have emerged as the real powerhouse on the internet. And it occurs to me that a model router is an incredibly powerful aggregator in the age of AI. That being said, I am not sure that OpenRouter has yet embraced the classic aggregator model because I'm not sure that they, I don't want to say they don't understand it, but maybe they didn't see it in that way.

But I think that that's what's going on here. So essentially, according to sort of Thompson's writing for literally a decade on this topic, the idea is that if you look at the successful business models on the internet, most of them are aggregators, not providers. In other words, they are companies like Google who collect everybody's work and then aggregate it and then present it to users, or Uber, who collect the work of people with cars and then present it to users as a unified service, or Amazon, or all these other companies out there that are doing this.

A model router is really an ideal aggregator in the age of AI, and it gives them incredible power. So that's the aspect that I think it's important. So the aggregator is where the relationship with the users is, not the provider of the service.

The aggregator should have a very flexible service and a very friendly service that encourages a user to just go to one-stop shopping rather than dealing with all these other things. So when you order from Amazon and it comes from a marketplace seller, you don't know, and hopefully you don't care. However, unfortunately, for the case of Amazon Marketplace, you absolutely do care because half the time they suck.

But anyway, point is you don't know, and you don't care who's actually servicing that order, who's the one who provided that. You just click the button, and the thing comes to your house, right? It should be that way with AI models, but my experience with OpenRouter has not been quite that friendly.

I mean, specifically, it's actually fairly difficult to attach OpenRouter to your OpenAI or Anthropic or Google account ... basically you give them the key, and then you can manually mess around with it and tell them limits and so on. But what I would love to see is a true AI model aggregator that would basically say, "Hi there.

Tell me about which services you use. Okay, I found those accounts, I found your limits, I have discovered everything, and we're good now. " Because I think that would be a much friendlier thing.

And then, essentially, that aggregator would have incredible power over this entire industry. Because could you imagine that once that took off, that aggregator, and OpenRouter can certainly do this, could say, "Hey, Anthropic, we want a better deal. " So I want you guys, because Jay, I know I've talked to you about aggregation theory in the past.

I know you know about this. Jay, react to this. I am holding in my hand an AI factory.

I am pulling the pin on the AI factory. I am throwing the AI factory into our conversation because I do believe that this notion of a supervisor, a scheduler, an economic engine will probably sit atop of something inside of an AI factory that's deployed into your favorite good old-fashioned data center that was never, ever imagined or designed for supporting the silicon that Dave was talking about. But someone's going to buy one, someone's going to put one in their data center, and that will be the appliance, going back to that appliance story again, that'll be the appliance that does that beneficial look ahead that you're talking about, Stephen.

And I believe that someone's going to wait for the AI factory to ship that does that. I think there'll be software players that will want to partner with the larger OEMs, the Dells, the HPEs, the Lenovos, et ceteras of the world that are going to ship AI factory in a can. But I do believe that'll be a central component piece.

Let us be your arbiter of where you go inside of the thing that we are helping you buy more of, or go outside, to your point, and make it simple, easy, elegant, beautiful, all the other phrases that would make you want to buy more of it. But that was me throwing the AI factory grenade. Well, listen, GSIs have, the Deloittes, KPMGs, Bains, whatever of the world, McKinseys, you name it, have built their business on being the router, if you will, or the proxy for all these type of things.

And they have a line of business around AI factories. We know this, right? It's stuff that I think even in early days, Jay, going back to your VC days, when we-- The inception of this, thankfully, I had left EMC at that point, but I still maintained my hooks into that space.

But even the perspective of the GSIs was, yeah, we want to give you advice on what to choose, when to choose it, and how to implement it. Right? And we'll go to market with a Dell, we'll go to market with an HP, a Lenovo, whatever.

Supermicro will build this stuff. They'll build the factory that produces the bits and pieces, and go in here, and we'll charge you $5,000 an hour to tell you what to do. Right?

And so the idea and the concept is sound, and I think this is great. Stephen, going back to your point about aggregation theory, one of the benefits of OpenRouter was that it collected and cascaded everything together into one piece. And then OpenRouter got the brilliant idea, which remains to be vetted out in the longer term, of introducing their own models.

So their collection, so we have Owl, which is their textual-based models. And so they're starting to offer a little bit more of this first-party type thing. But I think one of the challenges/opportunities that I've seen from OpenRouter's perspective, and again, remember, I'm coming from the characterization space.

This is what ML Commons does. Part of one of the beautiful things that OpenRouter did early on was that they integrated artificial analysis in the back end and said, "Hey, listen, you now can get cards. You can look at it.

It's pretty pictures. It will give you capabilities. " It still relies on the grunt work of the human doing the discovery work.

Key point is the human side. One of the things that becomes interesting that you've represented is this idea of an agent autonomously picking the best choice for the tool at hand. Now, that requires a lot of meta intelligence.

Again, it's not just colon latest of a model variation. It's how does DeepSeek take JSON and formulate it? How do you trigger an individual model's capabilities without-- Because they differ, and they don't always differ in the same way.

I've had problems with flipping between DeepSeek and Claude before, and the difference is slight, but it's enough that it breaks the tool chain. So an agent needs to be able to dip the proverbial toe in the water and say, "Oh, that's cold. That's my reaction.

I need to adjust my temperature. " Temperature is a thing in whatever. The analogies fall apart at some point.

But that idea of a kind of autonomizing. Autonomizing? No, it's not a word.

Being able to autonomously go in there, sense what's working, what's not. It's that feedback loop that determines, hey, I can do these things. That's where we need to get to.

OpenRouter uses an OpenAI conforming call, OpenAPI conforming call, which is great. So there is a standardization that's happening regardless of whether or not we want it. But the next step is then, all right, now give me determinism in the same way that Claude Desktop gives me, asks me choices when I'm doing a plan.

Like, what do you need? What are you trying to accomplish, right? And then is able to pick and choose based on whether I set a certain type of budget or capability or those types of things around it.

So anyway. Just to reflect the reality of it, these are the rooms for improvement that would make it even more first party and more valuable, I think, to your point, Stephen, about becoming the assembler, and, Jay, becoming the factory of factories, right? I think it's also interesting when you talk about the progression, and I'm thinking about the coffee I'm going to go get next because I see you sipping.

I'm having one of those moments where I need to go grab my coffee now because I'm watching you sip. I know that the caffeine- Air pressure ... it's coursing through your veins now.

But people feel different ways about Starbucks and different coffee places. But even in the early days, Starbucks, when it was just getting started, the myth or the origin story was, I think it was actually a distributor of Peet's, which is also great coffee. Peet's is great coffee.

I like Peet's Coffee. I like coffee. Coffee is good.

But it started off as a distributor, but today, Starbucks is a fully vertically integrated edutainment, infotainment experience that happens to also have coffee on the menu. But you could eat at a Starbucks now. You could drink all kinds of things at a Starbucks now.

I don't pretend to even imagine that I understand the entirety of what Starbucks actually is as a company. But they started as distributing, and just like we talked about with OpenRouter, they're currently distributing, but now you're talking about, like, "Ow. " And I have to imagine that the Cambrian explosion, to use that phrase I always use, the Cambrian explosion of startups and starting a company, how fast could you start a company now?

Going back to the OpenAI claim that eventually there'll be that one person that gets to a billion-dollar valuation, a unicorn, which, again, will seem quaint if you're looking at this 20 years from now. But one person, and now they're a unicorn. And so I think that you've got to start somewhere, but I think longer term, they will probably come into a lot more than just what we think of them today in the year 2026.

And yeah, I think that we're really in the early stages of here. Obviously, we're in the early stages of AI overall, but I do think that the idea of an AI access aggregator, like OpenRouter or whatever comes next, is incredibly powerful. And I think that if a company like Google, Amazon, Microsoft, or even a company like OpenAI or Anthropic decided that they wanted to provide that sort of flexibility and reliability and choice and cost control and so on to customers, I think that would be a big hit.

The other thing I think is in many cases, this sort of aggregation aspect or routing aspect is going to be taken by operating system device type people. So already Apple supports multiple, or at least the rumors are that they're going to have multiple different AI model choices available, sort of like the way that you can set a different search engine or different mail application or whatever. I think that that would be an interesting aspect as well if Windows, macOS, iOS, Android, et cetera, allowed you to just sort of flip the switch or even make it more dynamic like we're talking about.

So I guess as we wrap up here, what's your vision for this? Where is this going? What's the future of model routing?

Dave? The future of model routing is yes. Its existence, I think it's provided a great sanity check for a lot of folks in terms of their development processes, right?

It's allowed people to play with new experiences, new models, new capabilities in ways that a walled garden like an AWS Bedrock or an Azure service or a Google Vertex or any of these things have required more-- They're more cumbersome, right? The ability to go to one place and get a look and get a glance and understand, at least visually, and then as you integrate to do that. I think the futures of routers in general, if I look at it, is similar to what I see DigitalOcean in.

And again, DigitalOcean as a prototype or an archetype of an emerging neo cloud is that friction reduction. It's the, we'll give you every model that you could possibly dream of. We'll provide some intelligence behind it.

Silicon is table stakes. Software is not. So we'll give you the playground, we'll give you the ability to do that at a low cost or low friction, right, in order to enable a kind of constant velocity, if it's the appropriate word.

So I think that's the reason why, going back to even the original thesis, OpenRouter being a billion-dollar unicorn, that's the reason why the thesis works, was the reduction of friction, the increased surface area of models to applications and developers. The next step beyond that to turn them into a $50 billion company is to say, all right, now how do we provide the telemetry? How do we provide the accountability?

How do we provide the reliability that a lot of these over-indexed model providers, to be fair, are not giving you, right? How we can contain the storm and channel that into your usefulness, right? Whether you're a GSI, like a Deloitte or someone like me that's sitting at home writing a Chimera agent to do things.

I don't know. That's kind of my prognostication. I think you're going to see a couple of other players we haven't mentioned.

I think you're going to see the CloudFlares of the world. You're going to see the Fastlys of the world, the Akamis, the OGs of the old content delivery network making their play. I also think you're going to see different silicon that's not necessarily always going to be NVIDIA.

I think you're going to see AWS Outposts that are micro-sized and just enough to put Trainiums and Inferentias and Root Gravitons and Nitros into your back office. I think you're going to see Maias and Cobalts coming from Microsoft eventually, and I think you're going to see, maybe not Torus, 3D connected, eighth generation, post-Ironwood TPUs from Google, but I think Google's going to ship hardware in places we would've never imagined possible. Yes, there's NVIDIA Jetsons, and they're very cute, and you can build little robots with them.

But I do believe there's going to be more than just NVIDIA Universe and more than just the DigitalOcean example. I think that that pattern, that motif, will be far wider than we can possibly imagine now in 2026. So are we going to start calling them MDNs, model delivery networks?

Sure. There you went. You coined it.

You heard it here first. Who is the Postmates of that delivery service? That's the other question, right?

Yeah. Truly. All right.

Well, thank you guys so much for joining us. As always, it's great to have you be part of the Field Day community. com to see the events that these folks have been to, and check out their sites as well.

Why don't you give us a little shout-out? Where can we continue this conversation with you all? org.

com. And you can find me on LinkedIn, sadly. LinkedIn.

com. Yes, all one word. org.

That's where I spend most of my time. Join one of our working groups. It's free, and we'd love your participation in building what comes next.

Excellent. Well, thank you very much for joining us, everyone, and thank you for listening to the podcast today. If you enjoyed "Utilizing AI," please do subscribe on YouTube or in your favorite podcast application, and consider giving us a rating and a review.

This podcast was brought to you by the analysts and experts from the Futurum Group, where insights meet AI, as well as our Tech Field Day delegate community. ai, the "Utilizing AI" YouTube channel, or Techstrong's TV app. Thanks for listening, and we'll catch you next week.

And now here's my