Seize the AI Moment with Juniper Networks

0:10 I am manour Karam and the global uh VP for data center here at at Juniper and

0:17 uh I know it's a bit early but I hope you're as excited as I am I mean that's

0:22 just been a feeling of excitement these days that's Happ is happening pretty much every day there is always new stuff

0:29 happening every day when was the last time we had 30,000 people cheer for a

0:34 new Asic I mean the last time I remember was with x86 po possibly like the x86

0:42 and then we had two 286 we were excited and 386 and then when penum came out I know I'm aging myself uh but you know

0:50 that's the feeling I'm getting here it feels like it feels like that same level

0:55 of uh of excitement and indeed every morning you wake up there is something new and exciting that

1:00 you can use like I've installed GPT 40 on my phone and I've become kind of

1:05 dependent on it I ask it questions every day and and I didn't know that thing existed just a a couple of months ago uh

1:12 I went to LinkedIn to look at my blog uh that I written back in January I'm sure all of you guys read my blog in

1:18 preparation for this and that was um that was the blog we I

1:24 wrote when we launched our AI native platform and uh what I saw was like this little Square on the side and that was

1:32 LinkedIn kind of uh nicely summarizing the blog for me and I read it and it

1:38 felt like I couldn't have written it any different or any better even like the

1:43 highlighting the fact that we had were the first to launch an 800 gig platform it was right there uh written by AI I

1:51 mean it's just one of those things where every day there is something exciting and it feels like we should be all uh

1:59 surprised that we are surprised but at the same time we shouldn't be I mean that's just been that kind of incredible

2:04 cycle of innovation that we've had since uh we've had Mo's law uh 50 years ago or

2:11 TW 6 50 60 years I can't count uh ago to the to the to the day uh where you know

2:17 uh uh more predicted that we would be doubling compute power every two years

2:23 in frct what happened is that we kind of doubled every 18 months and what it means is incredible Innovations when

2:28 once in a while amazing crazy breakthroughs and uh I was very uh

2:35 grateful that since I came to the us back in 1995 and that was in September

2:41 of 1995 which was actually the same month Netscape went IPO um I've been

2:47 witnessing some of these inflection points some of these great cycles of innovation starting with of course the

2:53 internet you know every company trying to go online that was the web Revolution and then the cloud Revolution when these

2:59 services became a lot more sophisticated and then the mobile Revolution we always remember Steve Jobs holding the iPhone

3:06 introducing the iPhone it became clear at that point that everything we do we can do it just from our phone and then

3:13 after that we had the digital uh transformation it was accelerated by the pandemic where we realized not only can

3:20 we do everything from our phone but now we can do it from the comfort of our own homes and all the way to today where

3:26 we're seeing this massive great AI Revolution and so and bringing it back to networking uh look at the evolution

3:33 of ethernet during this period when I uh was at Stanford in 95 100 mag was kind

3:39 of like the screaming latest exciting uh state-ofthe-art and then one gig kind of

3:45 fueled the web Revolution and then 10 gig fueled the cloud Revolution and then after that we had 40 gig 100 Gig just I

3:53 remember it was a year or two ago we sitting with analyst saying H 400 gig uh is U is arriving but are there use cases

4:01 for 400 gig you know maybe it's going to take longer for adoption of of 400 gig

4:06 and then we get AI not we skip 400 gig we go all the way to 800 gig that's shipping today and we're talking about

4:13 1.6 and 3.2 terabit per second and so if you're like me at kind of at the

4:18 intersection of networking and AI or like Juniper I mean of course Juniper in was founded in 97 uh as part of the web

4:26 Revolution we've been experts at networking for many many years and then since the acquisition of mist and and

4:33 abstra Juniper realized the importance of AI I would say you know very early and we're all be became AI expert so if

4:41 you're at the intersection of networking and AI it's an exciting place to be because really AI at the core of it the

4:47 engine driving AI is the data center and the core of that data center is the network the network is kind of the

4:54 neural network in connecting all of the components that make up these AI

4:59 clusters but not only that AI is also transforming the way we operate these

5:06 networks and so again if you're at the intersection of networking Andi it's a really exciting place to be and it's in

5:14 the right time to Kind of Revolution revolutionize networking and lead the AI

5:19 networking wave and hopefully what you'll see today is a lot of examples of how we're doing that and in fact you

5:26 know back to that blog that's when we launched our AI native platform which was kind of

5:32 like one big step in this direction and so this AI native platform is quite

5:37 ambitious but at the end it starts with being AI experts really understanding

5:44 first of all what data one has to collect and at Juniper we've been collecting data for many many years and

5:50 we've become experts at knowing what specific data one has to collect and then becoming experts at what models to

5:57 feed the data into in order to to provide the right insights and uh to the to our Network operators or to automate

6:05 the network in the most powerful way but it's also about building the right

6:10 infrastructures to support AI workloads so having the right secure infrastructure for AI and I'd say it's

6:18 pretty it's all encompassing you know we are uh we play in many domains all the

6:23 way from campus to Data Center which is the focus today to the to the W and AI

6:29 native uh spans across all all the way from client to cloud and I'm going to

6:35 talk about these two pillars today AI for networking and networking for ai ai not for networking is how we're

6:41 leveraging AI to transform how we operate networks that's kind of like the

6:47 first pillar and the second pillar is networking for AI which is what type of

6:52 infrastructure networking infrastructure is needed to deliver on these AI use

6:59 cases and so let's start with uh AI for networking for data center in the

7:06 context of data center for jiper the journey started with the acquisition of abstra and I was uh the co-founder of

7:13 abstra uh we started Abra back in 2014 we celebrated recently our 10y year

7:19 anniversary and we've been part of juniper for more than three years now and what abstra uh set itself to do is

7:27 to deliver this power powerful automation of data center networks end

7:33 to endend the entire life cycle from designing to building to deploying and to operating data center networks and we

7:40 did it with this Innovative approach uh that we invented called intent based

7:46 networking we'll get a bit more into that but before I describe the technology what I want to say is that

7:54 let me just tell you some of the outcomes we're seeing with our customers starting with I would I would like to focus on reliability first the reason we

8:01 started abstra is we wanted a an automation tool an automation solution

8:07 that Network operators could trust software that they could trust will run

8:13 their networks with them with utmost reliability like an autopilot on a plane

8:18 you tell it to go to 10,000 ft and then it will deterministically go to 10,000 ft you know it's going to work in fact

8:25 if you build your software right you should get much better reliability in order of magnitude more reliability than

8:32 if you were doing things manually and indeed that is the outcome we're seeing with our customers in order of magnitude

8:38 Improvement in reliability that's number one number two automation should let you

8:44 run a lot faster not just more reliably but also a lot faster and so we've seen

8:49 deployment times up upgrades times which have been accelerated 10-fold we've had

8:55 customers upgrade their Global networks worldwide across mult multi vendor

9:00 multiple vendors just by a click of the button within one morning uh and then

9:06 last but not least we want to empower these Network operators so they can do a

9:11 lot more with less we have uh network teams uh that maybe are just like a

9:17 handful of of folks running networks this is a bank that I'm thinking about in Europe running a network that is

9:23 supporting thousands and thousands of users and these are stories that we see in and out and so that is the type of uh

9:31 that is the type of outcome we see with our AI based Solutions with abstra uh as

9:38 part as part of our automating as part of us automating data centers 10x

9:43 outcomes manour um you you talked about experience first or you experience first

9:49 I think was on the on one of the earlier Graphics you showed and then here you have operations first which made me

9:55 suddenly wonder if you mean operator experience first or if you mean application user experience first a

10:03 great question in fact we so when we say experience first uh and this is kind of

10:08 like a juniper uh slogan I would say we're focusing on both we're saying uh

10:13 it is about the network operator so it's the experience of the network operator you know so we don't want a network

10:19 operator to be spending way too much time and then doing things with the

10:24 worry of and and and and in exposure to risk of making errors right so like

10:30 we're improving the network operators lives by give them giving them solutions that allow them to run very fast but

10:37 doing it very reliably but also at the end the network is there to your point to support the users and the apps and we

10:44 want the user experience and the users of using those applications we want

10:49 their experience to also improve dramatically yeah I don't I don't claim they have a huge audience but anybody

10:56 who listens to me for 10 seconds knows that um I think that the functioning of the

11:01 network and its management is fundamental if not foundational to the user

11:07 experience absolutely I mean I mean the network touches every component and we'll talk about that a bit more and

11:13 it's so frequently ignored it's I mean by by the lay person let's say the late Tech person it's it's taken for granted

11:21 or not thought of I agree it's it's one of those things networking is one of those things you only remember your

11:28 network or you when things go wrong right otherwise you're taking it for granted this is why we say you know the

11:33 best thing for the network is that you know people don't know about it because then you know it's working well right

11:40 and so but you I completely agree with you the network is critical to making the user experience and the app

11:47 experience be what it needs to be yeah absolutely and so uh back to abstra and

11:54 just want to talk a bit more about intent-based networking and where we're taking this technology uh the idea

11:59 behind intent-based networking was that it's a layer of software the way you interact with it is that you tell it

12:05 what you want uh like for example I'd like to add a new Rack or I'd like to add a new virtual uh virtual Network or

12:12 I'd like to add a security policy and then what the software does is that it makes it happen 100% reliably it does it

12:21 by uh configuring the devices of course but it's not just configuring the devices it's collecting all of the

12:27 Telemetry the state from your network building a whole representation in the distributed data store that captures all

12:33 of the relationships in a graph and then by having all of this knowledge it knows it can prev validate whether a

12:40 configuration change is going to deliver on the right outcome and then it can test post deployment whether or not in

12:47 uh whether or not a change that you've made the indeed at the end delivered on your outcome so it's a closed loop

12:53 system it's a closed loop system that is collecting Telemetry and interacting with your network in real time does it

13:01 have a roll back to that then thank you because everything is described in software every time you make a change

13:07 we're able to keep a snapshot of the entire network if you're familiar with juniper roll back function that's on a

13:14 per device basis this is a roll back function at the entire network uh level

13:20 the entire network and not only can you roll back to where the the network was prior to this change you just made you

13:27 can roll back to whichever version of your network you want you know if you want to roll back to the way your

13:32 network was it Tuesday a week ago it's it click of a button so how Backward

13:38 Compatible could you get uh so like if you had an

13:44 update H how do I put this side if you have a uh an update that all of a sudden

13:49 these backups just can't work because you have to roll back the software to make the backups work yes is H is plan

13:58 for that or no well so essentially the this is a great question so what I'll say is with abstra um it it used to be

14:05 in the past like's say 20 years ago that when you build automation it was you're still manually interacting with your

14:11 automation it was like through the web or something of course we have that with abstra you can do this you can you have

14:16 a web interface it's uh it's easy to use Etc but really it's an API layer

14:21 everything you do is through the apis and the type of thing you just described is orchestrated through the API this is

14:27 why I have here terraform and service now and what you're going to see through the demos today is how we've leveraged

14:33 those Technologies to essentially Implement some of the workflows you're describing and so you would have let's

14:39 say you're making a change because you want to upgrade your applications so you could use something like service now and

14:45 you could orchestrate all of that and then there is a component of that which is the network that's going along with it and let's say your network is

14:51 upgraded and you're done with that upgrade but the application team is not ready and so even though everything was

14:58 successful on the networking side you have to roll back because your application guys hav't tracked with you

15:04 and that's all orchestrated through something like service now and that's the importance of having this API layer

15:10 and that's also why what's really important here is that you want this API layer to be vendor agnostic you want a

15:18 Loosely coupled architecture between your choices southbound and your choices

15:23 Northbound and with Abra even though we're part of juniper we are still and we will continue to be a multivendor for

15:30 exactly that reason you don't want your choice of uh service now or terraform in

15:36 the north uh bound and all of the implementation you've done to be locked into the choice of the vendor that

15:43 you've chosen southbound okay and so part of that technology is giving you

15:48 that loose decoupling that is so critical when you have an API layer like you've described so my S two questions

15:56 um there you mentioned that you know you're talking to this solution almost in a natural language is that what

16:02 you're using to say I want to deploy some portion of the network is that or is that something that's uh you

16:09 predicted my next slide okay I mean a second question is is are

16:15 you simulating the network change before it's made or you you checking the network change after it's made to see if

16:20 it's still viable or are you doing BS I'm loving these question so we do both

16:26 you know how I said we have distributed data store in fact I like we have two of them right so there is one which is for

16:32 the staging uh of your changes and we have one that is for your Active network

16:38 they unless if you're not making any changes they're exactly equal to each other they're mirrors but then when you

16:44 start making changes you're making changes in the staging uh uh area and

16:50 when you're making changes in the staging area this is where all of the prevalidation happens it's making sure

16:56 that in the context of everything it knows and the context of the intent you have the changes you're proposing make

17:03 sense and it will not let you commit if it doesn't compute if these things don't

17:09 compute if the prevalidation doesn't pass and if when the prevalidation passes it will allow you to commit and

17:15 then you press commit and hopefully today in the demos you're going to see some of this when you press commit at which point the state from the staging

17:23 distributed uh uh State goes into the active State and which point the

17:28 configuration are pushed at which point the Telemetry is collected the tests are run on your real Network to post valid

17:35 validate that indeed the networks changed as you the network change happened as you expected them yeah great

17:42 questions and so and can can I jump in there to finish that thought there how

17:48 much control does the operator see and have over each step in that that process

17:57 you know we don't want our AI overlords just like rolling through here and taking over so when

18:04 you're so at that same point when you're when you're about to commit you have

18:10 like a whole description to every last detail of every change that's occurring

18:16 in your network and you can go and review these yourself in fact we have customers that you know they have like a

18:22 step there where a human goes and you know as part of the process they review everything and it's only when they

18:28 review everything that they press commit right and then again after you commit you have an ability to roll back when

18:34 you roll back you have a change you have an opportunity to say okay well actually I want to bring I want to do this again

18:40 but I want to do this extra change so you have tons of flexibility in how you do this and I could go in there and say

18:46 these changes I'm I'm confident in the system's ability to detect that it's correct and I'm going to just whatever

18:52 it says is okay these require an operator to say yes you can continue

18:59 yeah I mean at the end it becomes very much based on the policies that you as a an organization set up right so we can

19:05 let it go as fast as we want and stop at the critical Paces where we're worried

19:11 about going wrong ex exactly for example when I showed service now as one example like we have some customers that use

19:17 things like slack where there is also like you know like communication that's going on with the various teams to slack

19:23 and there is protocols in terms of like who has to approve Etc that is going that is going going on as part of the

19:30 workflow right and then uh in terms of change Windows and things that we're required to maintain those can be

19:36 programmed into so we make sure operators don't kick things off outside of legal change Windows yeah so again

19:43 this is done through the AP I would say this are done through the API we have tons of role based access controls so like we have different users we have

19:49 different that have very granular we have if we have a very granular control of in terms of what they can access

19:55 right these specific users so you have super admins you have admins you have like the folks that are only doing design and

20:01 then you can be very granular in in in in those authorizations uh but then using the

20:06 apis you can also restrict like access to the platform yeah but you know actually with change

20:13 Windows one thing I'll say is there was one customer just an anecdote that had this change window on Saturday uh

20:18 through the whole weekend and they started deploying on Friday and then they called me on Friday afternoon I

20:24 mean clearly they were a bit I would say you know they were very excited in Emotion emot and they were like Mansour

20:30 I just want to call you because I had reserved the whole weekend to do this but then it's Friday noon I'm done you

20:36 know going home and spending the weekend with my family so maybe you know sometimes change Windows are for the

20:41 week no I'm kidding it's I love that you're talking about um

20:49 network devops uh a lot of the methodologies around continuous delivery and continu well continuous deployment

20:55 in particular would say there was no such thing as a change window we deploy changes including Network changes yeah

21:01 no I mean I always I mean I was a joking I I would always recommend chames windows but I mean in fact with

21:07 technology like this you can do things a lot lot lot faster the reason you have changed Windows is because you need to

21:12 coordinate amongst multiple teams right so so that that's part of you know so I'm not

21:18 recommending no change Windows um so so I was going to just I I I skipped the

21:24 slide so one of the things we're doing well as I mentioned we are uh you know using intent based networking to operate

21:31 your network day in and day out right um and so like this is the tool that you

21:37 use to deterministically get outcomes from your network make changes like an autopilot

21:42 on a train as I on on on a plane as I mentioned but then in the process you're collecting all of this data and would it

21:47 be nice if that you if you took all this data and send it somewhere in the cloud in some AI Ops layer that is not just

21:54 collecting data from the data center but also from your campus from your when from your sdw and this way you have like

22:02 a one repository for all of the state like as I mentioned collecting the right data and having the ability to then

22:08 apply models on this data and then this is where the chat B bot can come in and then you have an ability to through a

22:14 conversational interface interact with your network and also this is you know

22:20 AI is probabilistic and so these things are very compliment deterministically you're making the right changes and

22:26 operating the network but then you can provide probabilistic insights in terms of hey you know I'm giving you

22:32 visibility I believe you know like if you're having an issue it's probably in in that realm right so having that

22:40 deterministic approach combined to that probabilistic AI approach is extremely powerful and this is essentially part of

22:47 what we launched in January and we've delivered and we started delivering in the first half of this year okay and

22:54 we've not we haven't stopped there the third aspect and we delivered that in Q4

22:59 of last year is uh is uh application

23:05 awareness and uh the first step towards that is I'd say the most comprehensive

23:10 flow uh flow Telemetry available on any platform but that's now part of abstra

23:17 part of that same distributed data store you can now uh collect a Telemetry

23:23 application uh level Telemetry and combining it with combine it with network

23:29 Data Network State data and this way as a network operator not only can you

23:34 answer questions about whether or not your network is working well but you can do that in the context of every specific

23:41 app in fact that's really like the what the job of network operator when the application team comes to them and say

23:46 hey my application is not working well you want to have the ability to tell them whether or not it's your fault it's

23:54 because of the network or not meantime to innocence Ain so it's never the network

24:00 it's never the network it's DNS or bgp yeah so here the summary

24:07 right for AI for networking we started with intent based networking deterministic control we're adding aiid

24:13 driven probabilistic insights and then we're combining this with application awareness and I dare any competitor to

24:21 have a solution that as comprehensive as ours with the outcomes that we are able to deliver to our customers and one of

24:27 the proof points I'd say is the number of customers that are upgrading to our premium licenses we have standard Advan

24:34 and premium licenses and 68% of our customers in q1 purchase premium premium

24:41 licenses to me this tells me the amount of value the value our customers are

24:46 getting with our Solutions okay so that was AI uh for networking so let's now

24:53 talk about networking for AI essentially the role networking infrastructure we

25:00 need for AI workloads well before you move on we did most of the talking in

25:05 that first section about the deploy side do you have a a similar story you could quickly share on the troubleshooting

25:12 side of of that uh implementations where this data actually raises issues and how

25:18 that how The Operators interact with that yeah no absolutely so well hopefully some of the demos will show

25:24 you that but we're collecting Telemetry on a continuous basis and you you can

25:29 with uh abstra collect any Telemetry that you want that you can otherwise

25:35 access and so once the Telemetry is in this tool you can run any uh analytics

25:42 pipeline against this Telemetry it's completely custom so of course we have a bunch of things uh that are TurnKey in

25:48 the solution uh that are there out of the box but then you can customize this to your heart's content and so if there

25:56 are any uh areas that like for example if for a given application you're seeing

26:01 packet loss on a port or you're interested in seeing there's going to be any packet loss on a port then you can

26:07 get alarms and events for those specific uh occurrences and so that is that goes

26:12 a long way in helping you uh troubleshoot so that's number one and number two we're now shipping all of the

26:18 data into the cloud and there if we see any patterns that could indicate some sort of trouble kind of lurking in the

26:25 in the uh in the future like for example uh for transceivers uh we collect the

26:30 state for these transceivers and we know a signature of a transceiver that is about to fail right so we don't wait for

26:38 the transceiver to fail for you to go and to give you an alarm so that you go and replace the transceiver we notify

26:45 you prior to the transceiver failing so that now you have an ability to go and replace your transceivers before they

26:51 fail okay so that's all um something I can select for automatically at the

26:56 beginning which ones of these alarms we're going to raise and push up to the to the to the knock operations for

27:03 escalation and they're all um pre prepackaged based on your your look at

27:09 the data from other sources correct yes just be clear AB support Arista Cisco

27:17 Del any of these other Solutions as well you yes absolutely Dey manage absolutely and we have many

27:23 customers that use them across uh these different vendors yes and is it just

27:29 limited to that list or kind of anything you can key on with an API or so we have a whole uh if you if you just actually

27:36 probably if you ask uh GPT for it'll tell

27:42 you well we it's all documented uh publicly and there is a a validated uh

27:49 list of uh and qualified list of devices which operating systems Etc we're uh

27:54 we're very detailed in in in specified in uh uh which platforms we support yeah

28:02 so so how far you can go in like orchestrating these different vendors and connect them together so I'm talking

28:08 for instance like VXL GPO stuff like that where I will be able to like connect like different vendors together

28:15 are you able to build like multi sight multi data center sort of a fabric

28:21 multiple different vendors you know so you know I'm a networking guy you know I've been in networking my entire career

28:27 and wasn't that the promise of networking that you should have that ability to connect these hey back here in the real

28:34 world is it isn't there isn't that why we've spent so much time on standards

28:40 like bgp and evpn vxlan and this why I feel so strongly about this vendor

28:45 agnostic thing right like why would you throw all of that away by bringing in a management solution that only works with

28:52 one vendor and locks you in completely we might have as well like implemented completely proprietary version of bgp

28:59 why do we need bgp if you're going to do that right I mean so so so absolutely

29:05 you can do that to the extent that the the standards allow it right so uh you

29:11 know we've seen for example VPN vxn is a good example everyone straight away a stay and kind of it wasn't really

29:16 consistent I think we're seeing a lot a bit more consistency um but like to me we're

29:22 limited by the ability for the standards to allow for that right so that's essentially the limiting factor but if

29:29 the standards are allow allow it then there is nothing in the solution that will prevent you so so let's say like

29:34 you know I'm a client and I'm acquiring like you know companies and my data center is now Cisco I'm going to acquire

29:41 another company they are doing Arista and I'm going to on board Arista into into abstra can I then like literally

29:49 use the abstra to sort of like connect these two data well you're describing exactly you know like the use case that

29:54 one of our customers has has used apps for yeah and Abra runs

30:00 in the cloud or runs on Prem no so abstra runs on Prem and there is a really good reason for that by the way the session is grunning longer that we

30:06 originally in sped but I'm really excited about it because because of all of the questions that I'm getting and I

30:11 would like to encourage them so this is good but uh what I'll say is with

30:17 uh uh the sorry repeat the question I running the

30:22 cloud there is a reason why we're r abstra on Prem the reason is we need to

30:28 hit the highest levels of reliability remember reliability is the number one goal we require an outof band Network

30:34 for that reason you want to be as close to the network as you are the last thing you want is you're trying to make a

30:40 change and you can't access your network right having said that the AI Ops layer I described is in the cloud because

30:46 that's more about providing you insights that's more optional right if you have you lose access to that that's not going

30:53 to affect your network okay so that's that's the answer okay so let's talk

30:58 about networking for AI let me motivate this by you know so we talked about the

31:04 role of networking and how it's so critical for users and applications well in the context of AI clusters that is

31:11 even more the case the network is a critical in AI clusters if you think of

31:17 an AI cluster it's a really complex machine it's made of like all these different components different layers of

31:23 memory SRAM Dam flash uh attached storage different kinds of compute you

31:30 have CPUs you have gpus you have accelerated compute different types of networks you have the front end the back

31:36 end the storage and in fact in one of the sessions Prof is going to dig um into the details of that you have the

31:42 scale up the scale out Network you have all of these components and they all have to work in this harmonious balanced

31:50 way if you want to deliver on your an optimal job completion time job

31:55 completion time is the measure the iCal measure where these clusters are

32:00 measured it's the the how long it takes you to train a model right and if you're

32:07 familiar with some of these models they can train they sometimes take months to train but you know what if your gpus are

32:14 underutilized or like your model is your your AI cluster is out of balance then

32:19 you can add many many more months to that and these gpus that you spent like $35,000 to

32:25 $50,000 for each of these they are going to run completely underutilized which would be sad right and so the network is

32:34 critical to making sure that you get Optimal Performance uh to reduce to to

32:40 kind of reduce your job completion time to a minimal to to a minimum it's kind of like the system of Highways and

32:46 streets like if there is congestion there if you're not load balancing your traffic the right way if you're having

32:52 accidents I mean that's you know my analogy for packet loss right then then

32:58 essentially uh you're you're dead in the water but then you may ask well you know that's true like the network is really

33:04 critical but like what does it have to do with juniper isn't aren't these networks specialized isn't it like something like infin band uh that is

33:11 running in these uh in these clusters and to me uh it's it's actually you know

33:16 uh dja vu in so many ways because you know being in networking for a long for

33:21 a long time we've had these battles you know this is not the first time we see this right I remember back in 2008

33:28 whether it was oil and gas or uh or high frequency trading you know there was a

33:35 infin ban came in but at the end you know you know the saying never bet against the US I think in networking

33:41 this the equivalent would be never bet against ethernet right even um M MFE had

33:47 to go and eat his word literally because he expected the collapse of ethernet

33:53 which Never Never Happened he's by the way the in one of the co-inventors of of ethernet uh and so and the reason

34:01 is pretty clear I mean ethernet is the most uh wide deployed technology where

34:07 you have most of the in most of the Investments it has the widest ecosystem as I mentioned it's really critical to

34:13 these to get these AI clusters working well and critical to that is visibility debuggability you have hundreds of

34:19 management tools uh with with ethernet you can go and tap on the the link and

34:24 you will know exactly uh what's going on on that link you know this Frame format has understood for decades and and

34:31 decades uh you also have a massive ecosystem it works with any GPU I know

34:36 Nvidia today has the line share of the market but for specific use cases uh you

34:42 may get much better performance with other gpus and we've seen customers deploy you know kind of many different

34:48 options including AMD Intel and newcomers like cerebras where for some use cases you get you know orders of

34:54 magnitude difference in performance and infin ban only works with in Nvidia whereas ethernet is going to work across

35:01 the board and last but not least because again it's ubiquitous technology and there is so many investment at the end

35:07 the pro the price plummets much faster in fact today the total cost of ownership for ethernet is 50 time 55

35:15 times lower than with infiniband Monsour um so Ultra ethernet is being

35:21 developed um particularly for this kind of high speed

35:26 the highest speed con it and the latency issues in the systems that you were talking about um but are you talking

35:33 about that because Ultra ethernet initially is pretty targeted in terms of

35:38 its use cases right so are you speaking broadly about ethernet as as it exists

35:44 right now or are you signaling towards Ultra ethernet with no so I'm talking

35:51 about ethernet broadly and Ultra ethernet is yet another standardization effort on top of ethernet we saw r ma

35:58 back in the U over ethernet back in in the days yeah Rocky Etc so like there is

36:04 always going to be improvements to ethernet and that has been part of the journey with ethernet and yes uh you

36:11 know we are part of the ultra ethernet itra low latency ethernet Consortium uh

36:16 and we expect there will be some good Innovations coming out of there and I suspect propul is going to talk a bit more about it in in the session today so

36:24 just give you an example with with juniper of like the breadth and the performance of these ethernet Solutions

36:30 with juniper we both have broadcom based Solutions we're partners with broadcom and we deliver some of these uh

36:36 solutions to Market and then we have our our custom uh Solutions a wide range

36:41 just looking at broadcom for a second here just over the last 12 years look at the kind of incredible improvements in

36:47 performance back in 2010 we had 64 ports of 10 gig 640 gigs in one chip at the

36:53 time it felt like an incredible amount if you look at today we have 64 ports of

36:58 800 gig 51.2 terbit per second of performance in just one chip I mean this

37:06 is kind of the level of performance you get with ethernet but you're not stuck just with that you can if you want

37:11 deeper buffers if you want more programmability which you're going to need in some use cases then you have solutions for that and that's just in

37:19 the context of juniper you have other vendors plenty other vendors out there and so hopefully I've convinced you that

37:24 ethernet is kind of like the right solution and in fact even Nvidia in their earning calls a a a couple of

37:31 weeks ago admitted that ethernet is going to be even for them a multi-billion dollar market and so this

37:37 is a really important market for us to go after and we're indeed very focused on it and this is what analysts are

37:44 saying uh deloro on the on the on on the left they're predicting that this

37:50 backend network will be growing at 65% year toe and you can see the impact on

37:56 the total addressor Market Market on the right here this is 650 group you can see how it's turbo boosting the total market

38:03 for Ethernet from $22 billion today to $32 billion in just a few years and so

38:10 of course with juniper we haven't stood idle uh since actually you know I got in the role here uh back in September we've

38:17 really accelerated all of our efforts towards kind of delivering solutions for AI cluster and remember that 51.2

38:24 terabit per second chip we were the first to announce and that's was I was describing back in January in the blog

38:30 but now it's shipping and it's being deployed in some of the largest clusters on the planet today and of course that's

38:37 just one switch amongst many switches we have a whole portfolio whether you have use cases that require that type of

38:43 performance or different types of use cases we have all of those switches for our customers that's just one aspect of

38:50 the of the of the strategy the other aspect is operations right as I

38:57 mentioned you need to fine-tune these clusters you need deep visibility into these clusters

39:03 and of course we have abstra right and Abra is extremely valuable in the

39:09 context of AI clusters for two reasons one is that as I mentioned you can collect any Telemetry you want and run

39:16 any tests you want if you care about visibility to understand exactly what's going on you need that capability and

39:23 and abstra provides it uniquely the second is that as I mentioned these comp these systems

39:29 are complex and you talked about ultral low ethernet uh Consortium and there is lots of new features and knobs you have

39:35 in these in these networks and you want an ability to tune these networks to get

39:41 Optimal Performance and you may want to do this dynamically in real time and in fact one of the demos today is going to

39:46 show you how you can use abstra in in order to tune these networks based on

39:52 what you see so essentially a closed loop system leveraging abstra to find you in the Network for Optimal

39:58 Performance so that is also something that we have uniquely and that we're delivering as part of our solution and

40:04 last but not least and you're going to actually have the chance today to visit our lab here we are becoming as I

40:10 mentioned AI experts and in that context we've built a whole Lab here for AI

40:15 we're learning every day in this lab this all all of the folks you're going to see here they've spent so much time

40:21 in the lab including myself I love to visit and spend time in that lab and

40:27 uh we are learning with our customers our customers are spending a week at a time with us tuning their networks

40:33 tuning their models in in the in the lab and we are sharing these learnings with our customers and packaging them in what

40:40 we call validated designs so we're making it our business to deliver endtoend performance not just think of

40:47 the network as one component but figuring out how to deliver these endtoend solutions for our customers um

40:56 and we're packaging them in these validated designs if you have a large llm we know like this is the net this is

41:02 the the the cluster you need this is the cluster design you need you have a smaller uh Network maybe you wanted for

41:08 for for to to to find to fine tune a model that's uh a EV validated design a

41:15 self-driving Network model that's EV validated design inference EV validated design think of them as blue PLS for

41:22 these different use cases turn key validated designs so that you don't have to choose between either I have to

41:29 assemble it all myself because I'm choosing the components or I get a turnkey solution you can get a turnkey

41:34 solution while getting all of the flexibility of being GPU agnostic or

41:40 vendor agnostic and so I'm going to leave you with this uh this is a a survey that

41:46 Barkley did uh it has been you know it's it's been conducting this survey over many many years and in fact it was

41:53 Michael Dell that shared this on his uh on his Twitter uh on his Twitter field um many years ago uh uh when you asked

42:02 cios if they were thinking of repatriating their workloads only 43% said that they were if you ask this

42:08 question today 83% are saying they want to repatriate their workloads into the private Cloud into un Prem and there are

42:16 two reasons for this one is with AI data is becoming of primordial importance and

42:24 customers are a lot more hesitant to ship away this data in areas where they

42:30 lose control of it and it become super expensive if one day they would have think about bringing that data back and

42:36 so they are increasingly building private clusters for their AI uh for

42:42 thei the AI workloads That's number one and number two with Solutions like we've

42:47 described today like abstra like AI for networking it has been it is much easier

42:53 than ever before to manage and operate uh private clouds um and so actually one

43:00 of the demos today will be about how you can manage a a private Cloud as simply

43:06 as you would a public cloud and so for those two reasons we're seeing this massive interest in with CIO in building

43:14 out their own private data centers and so it's a really uh interesting time it's an exciting time we're seeing

43:20 progress every day we're now deployed in more uh than 70 countries and we're

43:26 seeing more and more more opportunities every day in this in these very exciting times and so that's