Running AI Inference at the Edge – The Last Great Cloud Transformation EP2

Hey everyone, I'm Alan Shimel of techstrong, and welcome to the last great Cloud transformation. This is, uh, episode two in our ongoing series of the last great cloud transformation, and it's a joint production between us here at Textron Group and our friends and sponsors of of, of this, uh, show, CloudFlare, and so many, many thanks to CloudFlare for, um, sponsoring and co-producing with us. We, and, uh, we're excited to be doing this.

If you miss the first episode, it is available on Techstrong tv, and I highly, highly encourage you to go back and take a look at it. But for today's show, new show via, as I said, episode two. Uh, the title for today's show is Running AI Inference at the Edge.

We're peeling down a little bit right on this I idea of a connectivity cloud and this last great cloud transformation. Uh, before we jump into the subject matter though, I want to introduce you to our panel for today. In addition to myself, I'm happy to be joined by Rainy Habi.

Is it, if I got it wrong, Habi is how I pronounce Rainy. Pronounce it for us so we get it right. Uh, hi, I'm Rani Habi, um, the CTO of Networking and Edge at the Williams Foundation, where I work with our open source communities, helping them kind of, uh, shape their strategic direction around open source networking and open source edge, uh, finding new technologies that are relevant, finding a lot of new synergies between our projects and between our project and external, uh, projects and organizations.

So I'm kind of trying to keep my hand on the pulse of what's the latest on networking and Edge, and obviously, uh, ai, uh, and network connectivity for ai and doing all that at the edge is of course a hot topic these days. Um, so I'm seeing a lot of activities in our community around that. Excellent.

Thank you. Radiant, thank you. And the Linux Foundation for participating today.

Our next panel member is Rita Lov Reader is the VP product management. CloudFlare. Rita, welcome.

And maybe give a little bit of your background. Sure. Thank you for having me.

I'm Rita. I have been a CloudFlare for the past eight years, and for the majority of that time building, building out Cloudflare's developer platform workers, which includes many services, um, some of which I, I think we're gonna talk about today. Uh, and before that started, you know, my career in software engineering and I have been deep down a big nerd and programmer since then.

Excellent. Love it. Then our third member of our panel today is, is my, uh, partner and co-host Mitchell Ashley.

He's the CTO here at Techstar Group and a CTA at futur. Hey Mitchell, welcome and thanks for coming on. Yeah, good to be here.

Boy, both listening the both of the backgrounds. I was like, oh man, I can't wait to talk about this. Hope this goes for four hours.

Is that how long we're gonna go? Not Quite, not quite. I don't know if, I don't know if the folks out there wanna sit in and listen once for four hours.

Maybe We'll have it after the show Conversation. Maybe we could chop it into 10 parts. But, um, anyway, I, but I agree, we've got some great panel members who are bringing some amazing focus on, on this issue.

So we're gonna talk about running AI inference at the edge, but before I jump into that, a a quick worried about AI at the edge. Both, both, all three of our panel members are, are talking about this in that, hey, it's amazing, right? We two years in generative AI is cool, it's cool, we love it, but we're recognizing that we can't run, we can't do all the great things we want to do with AI going, you know, backhoeing, so to speak, back to the, the, the, you know, the main cloud center, these hyperscale centers, we need to be closer to where the action is.

On the other hand, most of our endpoint devices, unless you listen to Apple, about Apple six, the Apple 16 phone right, are not super computers and they can't run everything. We, you know, larger models of AI on that device. It's, they just don't have the footprint.

So we need something in between. And the edge is that place where we can get closer to where the action is, but still have the footprint to do the heavy computational work that's necessary. And that's driving this hole.

There's a whole thing going on. Let's, I don't wanna call it a next generation of AI or heaven forbid, gen AI two oh or something like that, but that's what's driving it. And so we wanna get smart about running AI at the edge.

What, how do we run it? What parts of it do we run all of it? When should we go back when?

How did they all work together? Um, Randy, I get the impression this is something you are, you know, putting a lot of thought into. What do you think?

Yeah, so what, what we're seeing is indeed this desire to run, as you said, the inference mostly close to the edge. And when I say close, it means two things. One is close in terms of latency when decisions need to be made fast and, and provided back to the end users.

And that's one type of closeness. But the other kind of closeness is where maybe you don't want the da your data to travel all the way to the cloud. Maybe there are security issues, sovereignty, privacy, so on and so forth.

And this is when we see that in use cases, people want, wanna keep the data close to where it's generated, but still do all those fancy things of, uh, training on this data insurance based on this data and providing insights, uh, based on this data. So that kind of puts edge computing at the sweet spot of being able to do all that and still comply with these, uh, strict requirements for latencies regulations and, and privacy requirement. Fair, fair.

Mitchell, uh, Rita, thoughts on on what Randy said, or was I, I had some ideas on wanted to take this. Said, Well, I, I love the, um, developer perspective on this, right? Because the people that are creating applications, they're trying to, they're wrestling with where do we put things, right?

Do, what do we put on the device? What do we, is there someplace close to the edge or at the edge of what we're connecting to what resources? It's like, it's all one kind of big map of resources you, you have available to your application.

It's just, where is it? How do we get to it, and what's the optimal way to leverage that? And that may be different two years from now, right?

We may do more heavier things on the device then than we do today. So it's kind of a moving target. Um, I, I know one of the things that I really, and by the way, just full disclosure, we're a, we're a CloudFlare customer also, so we're, we're very familiar with using their services.

I've been really, I don't, this is not just being nice, I'm really impressed with the progress you've made on the developer portal, on the APIs, and really kind of programmatic programmability of the edge. And so I, I'd love to hear, Rita, your thoughts on, so how did you decide that the approach that you took, how do you help people do? Maybe they don't know what they're doing, what a one they wanna do yet, you're kind of helping them create the future.

Yeah, I, I think your point about this being interesting from the developer standpoint is really good. And that was the way that we approached it, where, from our point of view, we've been watching, you know, developers building applications for as long as LER has existed, right? So for the past 14 or so years.

And so we've seen developers struggling with all sorts of challenges from, you know, whether you're developing on-prem or in the cloud, uh, especially, uh, you, you still have to do a lot of work around provisioning infrastructure, right? Where, uh, when you're in the seat of the developer, actually your, your task for the day, uh, d doesn't end with, oh, great, I provisioned a server. Um, the actual ticket item that you're trying to check off, off is, you know, I ship this feature and now it's in customer's hands.

And we've been helping solve that challenge of, you know, how do you unblock developers to focus on just that for, uh, as long as the developer platform has existed, by allowing developers to deploy code directly to our edge. And what's been really interesting about, uh, you know, this rise of AI over the past two or so years is we've seen a lot of these problems come back, but in actually a kind of even hairier way where when you're deploying AI applications, um, first of all, there are a lot of different things that you need. Um, running the model itself is just one part of it, right?

And here you're trying to focus on bringing the best experience possible to the user, but you, you have to, um, you, you have to do all of this provisioning upfront, right? So, um, you, you need all of these, uh, different tools. You need the model, you need a vector database, um, and, you know, we, we can talk more about that later.

Um, but you also need to make sure that your application to Rainey's point is really, really performant. Um, you need to make sure that it scales. And I think that that's been one of the really hard challenges is when you're launching these new AI products and features, you actually have no idea is it gonna be successful on day one?

Maybe. Yeah. Or maybe not, Maybe not, right?

Um, and there's actually a real cost to getting that question wrong. Uh, there's a penalty in either direction, right? So either, um, you know, you're like, okay, I'm gonna have a bunch of users on the first day, I'm gonna go ahead and provision all of these resources, and then no one shows up.

And then you get your cloud bill at the end of the month and you're like, great, I had a bunch of boxes on, you know, standby, uh, or inversely, uh, you underprovision, and then all of these users show up, and all of your hardware kind of goes to nothing because the traffic falls on its floors and there's no way to swallow it. So the, the way that we approached it was really from that perspective of, okay, how do we like step one, let's enable people to build stuff, and we're good at infrastructure and we will handle it on our backend. And we really wanted to take this kind of serverless first approach where you focus on building in the application, we'll take care of provisioning the infrastructure and, you know, let, let's see what direction the industry evolves in and take it from there.

Agreed. Um, you know, what I find important? So look, the whole reason about being at the edge is to be close to your end points, or I've waited 40 years as a yes fan, Tom, say this, to be close to the edge.

Um, you gotta be my age to, to appreciate it. But anyway, the fact of the matter is, the edge is doesn't, it's not the edge in the edge alone, though. It, this is, we gotta think about this as an interconnected web of things here, right?

You can't do everything on the edge. The edge is, you know, one area, but we still need that core, if we could call it that core data center, right? The hyperscale data center, right?

We need, we need the horsepower that only those places bring to bear on, on some things that we need the edge and its unique capability of being proximity wise close, but still have, you know, moderate to, to heavy horsepower. And then we still can do some things on the endpoint, especially all kidding aside, as these next generation of endpoints Incorporate faster, better processing power and more, you know, more capability. And, and quite frankly, as we get more efficient in running AI applications and AI assisted applications, so now you got this picture right of this interconnected endpoint data center edge.

Well, you know what, emphasis on the word connected. We, we've gotta connect them. And that's part of this last great cloud migration.

The connectivity between them has to be, you know, 'cause latency counts here, right? Every millisecond counts. So we need providers, we need solutions that have that kind of connectivity.

And anytime you're doing that kind, any kind of connectivity, you need security. Right? And I mean, Rita, obviously these are two things that Cloudflare's known for, but Randy, you look at it with open source tools and there's open source edge going, a ton going on at lf.

Let, let's talk about what are the kind of unique connectivity needs to make this system work? Yes. I think you kind of alluded to that when you said that it's a very complex mesh of things on the edge and the cloud.

And on the other hand, we have developers, I think we touched upon it, saying that developers need to focus on, and this is what they know how to do on, on developing the application. They don't know much about cloud infrastructure. They don't know much about connectivity and, and VPCs and, and firewall rules and so on and so forth.

So what we see the open source community do, and then the service providers like Cloudflare's and others actually build their solutions using is open source projects that are Dell are, are dealing with, um, kind of abstracting all this complexity from the developers and doing all the nasty and and complicated stuff as much as possible behind the scenes for, for the end users and presenting the end users or developers with very simple to consume APIs in which they can express what we call their intent. So I'm building an application and it needs to access some dataset somewhere. This is what I wanna, I wanna request from the network, and I want everything to get connected behind the scenes, uh, without me having to go and do all these, uh, point-to-point connections or setting up virtual circuits or whatnot.

And if you, you, you mentioned an environment where you have multi edge and multi-cloud, and we have one of our projects, uh, newest project Paraglider is dealing exactly with that, with abstracting the consumption of the network connectivity between multiple clouds and cloud and edge, uh, and, and making it consumable by, by the developers and providing, uh, unified API. So no matter where your data or, or workers are, whether they're in Asia or, um, in AWS or on-prem or on some, uh, edge Cloud, you want the same API to set up the connectivity and, and you wanna let to free up the developers to develop the business logic, the application, and not worry about how it's interconnected. Fair.

Um, so Rita, you know that the Linux Foundation obviously champion of open source champions of open source. I wonder how does that, if you could compare and contrast that to the, uh, CloudFlare solutions for, you know, the connectivity cloud, if we could call it that. Uh, and I'm sure there are some things that are the same that rainey's talking about in the open source, uh, you know, uh, model, but you guys have also had 20 years to play, you know, to play in this arena.

And I, I assume you've learned some lessons and would love to hear from that. Yeah, I mean, first of all, the, the two are not at odds with each other. And actually, um, yeah, a a lot of open source projects are, uh, built on top of CloudFlare and, uh, v vice versa, right?

Um, CloudFlare itself is built on a lot of open source technology. Um, and yeah, the, the way that we kind of approach it is, you know, as you mentioned, uh, you need this connectivity between all of these different endpoints, um, across device, across the edge, then you have the cloud, um, and we, we've, what we've really done over the past 15 years is build out this network, right? Um, and so with the network, we have different angles or different ways of looking at it in order to connect all these pieces.

So a lot of people, when they think of Cloud layer, they think of, um, C-D-N-D-N-S DDoS. And in that capacity, actually we've been, uh, you know, using machine learning and AI or, you know, it was called machine learning before. Everyone got super excited about it, um, for, for basically the entire time to observe things like DDoS attacks, right?

And understand, okay, this is what the pattern looks like. We build a model on top of that, and then we have it running on our network in a way that can, uh, allow through traffic that is legitimate and prevent traffic that's not legitimate. Um, so we, we started thinking about it from that angle.

Then there's, uh, the developer platform angle, which we touched on briefly before, but I, I think from that standpoint, it's like, okay, what are all of developers' needs when it comes to developing an application? And how do we bring all of those to a single place so that you kind of have, you know, I like cooking, so I like to think of it as a MIS and plus, right? Um, but you kind of have everything that you need in one place, whether it's running your compute, running your, uh, ai, running a database, running your storage.

Um, and, and that's kind of our take on connecting things there. And then, um, obviously, uh, we have companies that are using us to secure their networks, secure their own devices and employees. Um, and there's actually a fair amount of AI that's involved in there.

And actually with, um, more and more companies deploying ai, uh, solutions like DLP have started to become more prevalent where people are worried about, you know, I have, uh, I have all these employees and maybe I put these policies in place, um, that say, you know, you're not allowed to use these tools because we're worried about our data leaking out, and then models being trained on top of that. Um, but obviously if, you know, getting in the answer is as easy as typing a question to Chad, GPT, you're gonna have people that go and do that. And so having that visibility through our network becomes really, really important.

Um, so all of those things are kind of what we've observed, uh, in terms of problems that our customers have. And again, the way that we've tried to, or the way that we approach solving it is, okay, we have this network and what are the different ways that we can deploy it in order to tackle that, if that makes sense. It makes perfect sense to me.

I mean, and, and, and I like that you really hit the second piece of which, which is the security aspect of it, right? Because if we're not comfortable with that data on the edge, and then we're not comfortable with that data, you know, zipping around Cloud Edge endpoint and back around again, it just doesn't work and it just doesn't work for us. Um, Randy, interesting.

I know Linux found, obviously the Lennox Foundation is a foundation of foundations. There are like 40 some odd border foundations of lf, one of which is, for instance, the open, uh, open source security foundation, OSSF, there's several different security themed organizations, as well as security sort of built in security projects built into all of the various order foundations. You heard what Rita mentioned, especially, especially around security.

How important is, well, I know it's important, but what, what's the LF doing on that end as well? Yeah, so, um, maybe it's the, uh, opportunity to mention that indeed, uh, telling Foundation now hosts over 1000 different open source project, and some of them are organizing to these, uh, uh, dozens of, of sub foundations. And, uh, one of them is open, uh, SSF Open Secure Software Foundation, which provides tools, best practices, and methodologies for developing, deploying, and managing software across the entire life cycle.

So, uh, the, uh, the product of the open SSF can be consumed by anyone, uh, developing or deploying software. But another thing is that, uh, many of our, uh, other foundations under the Lung Foundation, like networking, like Edge, like, uh, energy and others are actually working closely with the open SSF and integrating these tools as part of our software development pipeline. So, uh, things like software build materials, uh, are already embedded into the pipeline of building many of our open source projects.

So if you are using some open source technology as, as a service provider like CloudFlare or as a technology vendor, uh, you get software that is secured using the most, uh, recent, uh, and broadest set of tools, making sure the software has no back doors, your, your supply chain is secure, uh, and so on and so forth. So I think that we, again, developers are always tend to kind of leave security, uh, for the end and, uh, try to, um, edit as an afterthought. But what we are trying to educate the open source communities is that it's not the best way to do it.

And you better start integrating these tools from open SSF and other places early on in, in the lifecycle of the software that can save you a lot of headache further down the road when you have to deal with some fire drills or breaches and stuff like that, where it's way too late to, to deal with that. So I think, again, I want to debunk the meat that open source software is an alternative to the service providers or vendors. We are just, our communities are actually, they consist of, of contributors from all these service providers and vendors that are building, commonly building the technology and then using it in their commercial offering.

So everybody who's, who's collaborating with the Learning Foundation on on the those collaborative project actually benefits from, uh, using the software that has these built-in, uh, security mechanisms. Are we perfect on that front? Um, obviously not nobody is, uh, we try to learn from our experience and to constantly evolve and as new tools and, and best practices evolve, we try to incorporate them into our open source projects and, um, make the projects, the products and services that are built on top of those projects more secure.

They're great. I mean, we're all one community here. Um, I got two other issues I wanted to bring up.

One, I, I'd like to have your thoughts on it, Mitch. You know, look, the title of today's episode on the last Great Cloud Regression Transformation is running AI inference at the Edge, right? And, and we're giving that as the reason why we need AI on the edge, right?

I don't know if everyone out here understands what we mean by AI inference and why we think it could only run on the edge. Mitch, can you give us the, the 4 1 1 on that? Sure.

Um, well, first of all, we're, we're not, probably not talking about training large language models that happens in great big giant, you know, GPU data centers where you need those things very close to each other. We're talking more of applications or use cases where, um, AI might, AI might not be the application, but it's in built inside of your financial banking application. Maybe you, you've got local data on your computer or your, or a laptop or your, uh, cell phone or some other device that you want to be able to analyze without passing it up, right?

And some of this also goes into the regulatory community, right? The edge is kind of an amorphous thing. What the edge is today isn't what the edge is tomorrow, and it doesn't just stop at the provider level.

It's at what are, what's consuming content at the edge. And AI is one where there's a lot of concern. Um, read dimension, uh, data loss prevention, uh, is an issue.

Um, even in some cases, you don't want it left outside of the hardware. You don't want it leaving the GPU and the memory that's, uh, that's processing that data, um, because of the sensitivity of it. So depending on the application, it can be proximity to the user and the data.

It can also be security that drives it, that we don't want that data pushed because maybe that data center isn't in a regulatory compliant area where we can be sharing that information or centrally housing it. I think, I think there's a lot of innovation happening too, about how we do use ai, AI at the edge. You mentioned, uh, you know, the new iPhone and Apple's inference or Apple in intelligence, and we have, uh, copilot plus and a lot of things that are being added to devices.

And we're really kind of at the beginning of what some of those apps are gonna do. I mean, I think in a year or two we're gonna go, wow, we never thought about that. That's very cool.

We may not even know it's inside of our app that that's actually AI performing those functions. So, you know, you used to have this debate. Is AI a product or a feature?

Well, I think it may be an embedded capability. Sometimes we don't even know it's there. Uh, so there's some unique requirements we have, uh, I think in this world where we have so much happening at the edge, uh, and devices communicating across it.

I think, um, if I can add something to that, the really interesting thing to me there is, you know, I, I think the thing that really wowed everyone about AI and that really gets people excited about AI is the potential for a boost in productivity, right? So once you start using these AI tools, ideally they kind of augment your work in a way that allows you to, you know, maybe code faster, right? So I've been using, uh, copilot and cursor and all these tools and, uh, you really do become so much more efficient as a result.

And, um, similarly, you know, things like chatbots, which is again, to your point, I, I don't think that we've figured out quite the end use case there, or you know, what that ends up looking like. But I, I think unlocking productivity is one of the things that is so exciting and so promising about it. And that's where I, I see productivity and performance as two things that are so tightly connected, where if you are, depending on ai, you know, is it a feature?

If, if it is a capability, if you're, uh, relying on AI capability for everything that you're doing several times a day, um, every single bit of performance, every single bit of latency really starts to matter a lot. Um, and so that's where, you know, as to why AI inference on the edge, we started to see, okay, well if it's not running on device, it's running on a cloud provider that's really far away. And right now in this experimental phase, everyone has a lot of patience for it, right?

It's a new technology. So you sit around, you wait for 30 seconds for AI to generate a response, and you go, okay, that's about what I expected. But as you really start to incorporate that into your workflow, you're like, actually, this is slowing me down more than it's speeding me up.

And so that's where, you know, I I think as it becomes more and more ubiquitous and embedded into everything, uh, we're gonna see the edge play a huge role in unlocking that next phase of applications and inference. Agreed. Yeah.

We, we also keep talking about that, whether the data is on device, on the edge, device on the cloud, what I'm starting to sense is that maybe there's a desire to blur the lines a little bit and maybe keep a mix of that and make the data accessible from anywhere. So we have one interesting, uh, open source project that we launched recently called Edge Lake, where it lets data stay on a distributed edge, but be queried from anywhere either the edge or, or central location. So that kind of addresses many use cases where the data might not all reside on the edge or on device or in the cloud, but again, whoever is developing the application needs to access it wherever it may be.

So we kind of, at least in the open source world, we kind of see a desire to, uh, make this access, um, kind of independent from where the data is and, and open up the door for new type of applications that don't care where, where the data is. Agreed. Guys, I'd like to bring up one other topic 'cause we're, we're running low on time.

I mentioned this term connectivity cloud earlier, and we really didn't dive more into it, but, you know, it really kind of describes what we're talking about, a new kind of cloud that in encompasses the edge, the core, the endpoint, the connectivity between those places and the security. There aren't a lot of players on the global stage who can, who bring all of that to the party themselves, right? CloudFlare happens to be one of them, right?

CloudFlare has the data center, the edge, connectivity, security. I'm wondering if we might not see, and, and there's probably, you could count on one hand the amount of providers who could do that besides CloudFlare worldwide, I think, right? Quite frankly.

Um, are we going to see, and maybe, and Rita, this is outside of your, you know, sweet spot at CloudFlare, but is CloudFlare partnering with other providers, maybe with some of the hyperscalers and stuff to enhance that connectivity cloud to make it available on a wider scale worldwide to, to others? And you know what I'm saying? I do, and we very much, uh, see our role here.

You know, the, if you think about the word connectivity, uh, it's about connecting two things, right? And so the, the, the more, uh, the more organizations out there we can partner with, especially, uh, if we, so as a product manager, I always go back to the customer, right? And from a, the customer's perspective, it's like, okay, this is my world.

These are all of the vendors that are used today. This is the cloud that I used today. This is what my, uh, employee network looks like today.

These are the devices that we have today. And so we very much see it as a big part of our role to partner with all these different entities. Um, whether it's, um, again, more on the, uh, zero trust side where maybe we partner, you know, with, uh, companies like CrowdStrike that do endpoint, um, application protection, right?

And we can take care of the other part of the security or with the hyperscaler. So yes, um, if you are working on an end-to-end AI deployment chain, uh, the training part, you're not gonna run that on us, but we wanna make it easy for you to connect, um, your data that might be in R two on CloudFlare, right? Um, this is where, uh, even things like, uh, free egress really come in for us, where we find it really important for you to be able to, uh, then connect out to maybe different cloud providers for training if you want to broker across them and not have to spend an arm and a leg there.

Um, and then be able to again, connect to CloudFlare to do the inference part. And so we, we really see ourselves here as, you know, if you, the, obviously the more of CloudFlare you use, the more advantage you get to take of where, you know, the, the pieces really play nicely with each other, but at the same time, we always wanna meet the customer where they are today. And, uh, realistically that's, you know, using a few different providers in culmination.

Yeah, absolutely. Yeah. And, and I, and I think that obviously an open source neutral, uh, organization is the, the place to work on these unifying the way we use and consume services.

There are always, I mean, it's not simple. There's always that striking a balance between, uh, the need for member companies to stay competitive and, and differentiate and provide value and the need to make things more standard and make it maybe. So the shortsighted approach that is now, I'll make all my interfaces unique so nobody can, can ever live.

There's an, uh, uh, a well known cloud provider that was notorious for, for being kind of this Hotel California approach where you can ever live. But I, I think in the long run, it, it doesn't work and it doesn't pay because you're losing a lot of business that way. And if you kind of align with the more standard interfaces that maybe are created by the opensource community, there's still a lot where you can differentiate yourself on per performance, on cost, on, on outreach, and so on, so forth.

But, uh, the, the interface and API are probably not the way to differentiate. I mean, it's better to align with some standard or defacto standard of, of interface and API and, and differentiate yourself on, on the true value that you can provide. Fair, fair.

You could check out, but you can never leave. I, I get it. The World Garden.

Anyway, guys, I'm looking at my watch. We, we are outta time. I think this was a great discussion about, you know, the, as again, I don't wanna call it AI two oh, but you know, as this AI generative AI works its way through our technology, recognizing that we need to be able to run these on a distributed edge network and what that kind of network needs to look at, like, or we want to call it a connectivity cloud, what that needs to look like, what kind of functionality it has.

Rainey, Rita Mitchell, I want to thank you for joining me today on this. Again, many thanks to CloudFlare for, uh, helping us produce and, and distribute this. I, I think it's a, as we go on, we're gonna dive deeper and deeper into these topics.

I'll also remind you that we have a live round table version of this one coming up where we will have people, hopefully you watching this log on and ask the our questions and contribute to the commentary and to the discussion, because it's great having experts like this on, but we really want to hear what you are doing out there. So stay tuned for that. But for now, this is Alan Shimmel.

On behalf of the last great Cloud transformation, CloudFlare and Techstrong, thanks for joining us.