AI Unbound, Your Data Center Your Way with Juniper Networks

Intro

0:00 my name is Proful Landani and I lead product management at Juniper for our data center business unit so the

0:05 delegates in the room and for everybody else who's tuning in uh welcome to Juniper and thank you for being here we

0:11 do appreciate the valuable amount of time that you spend with us today now if you've been in the

Who is Juniper

0:17 networking domain for a while I would suspect you know a little bit about Juniper but those of you tuning in and

0:23 wondering who the hell is Juniper about this is a onelide primer on our history right so we've been around for 25 years

0:31 uh we have deep roots in networking and security a global customer base and we

0:37 serve enterprise cloud and service provider verticles customer segments we

0:43 have products and solutions that power networking and security in multiple places in the network starting with

0:49 wired and wireless in campus and branch van routing number two and number three which is the focus of today's discussion

0:56 is we have networking and security solutions for data center as well in fact data center is where Juniper is

1:03 seeing the maximum momentum these days these in the last few years with our revenues in data center crossing $1

1:09 billion in networking revenues alone that's not even including security in 2024 and we have more than 10,000 happy

1:17 customers who have deployed Juniper in the data center some of those logos in fact just a fraction of those logos are

1:23 on the screen uh in front of you who are some public references that we have but today uh the focus of today's

AI is increasingly distributed

1:30 session we're going to dive into the realm of AI data centers and from our

1:35 vantage point in the industry as an infrastructure provider for AI clusters we see one very clear trend that AI is

1:44 increasingly distributed let's explain how why so right so if you're an

1:49 enterprise you're going to have some of your applications that you're going to host with your favorite hyperscaler

1:55 but in the new world of AI there's a breed of GPU as a service providers and

2:00 AI as a service providers that are coming out there with niche offerings and we call them neocloud providers so

2:06 your application your AI workload may be hosted with your favorite hyperscaler but very likely is also hosted in some

2:13 other cloud provider like some of these neo cloud providers now if you're an enterprise for various business reasons

2:19 whether it is cost whether it is data privacy reasons or whether it is regulatory reasons you may also decide

2:25 that some of these applications you want to host them in your own private cloud or onrem and then there is also edge AI

2:32 because for some inferencing applications latency matters and you want to be as close to as possible to

2:38 the end user for those inferencing use cases now at Juniper we believe that in

2:44 order to support this distributed architecture we have we are uniquely positioned with a comprehensive

2:50 portfolio that spans data center fabrics it spans data center interconnect to connect all of these various clouds

2:56 together and security natively built in so wherever your AI is happening whether

3:02 it is in the cloud whether it's in private cloud or at the edge Juniper is right there to provide solutions for you

What makes data centers exciting

3:10 now this is what really makes us excited about data centers in general because this new wave of AI if I take an example

3:18 of a generalpurpose server right for general purpose computing reasons even today two ports of 25 gig for a total of

3:26 50 gig bandwidth per server is considered more than sufficient somebody may be going to 100 gig but

3:31 predominantly 50 gig per server you look at an Nvidia node even the last generation of hopper series with eight

3:37 GPUs you're looking at 400 gig connectivity per GPU you look at

3:42 multiply that by 8 GPUs front-end networking storage networking that's a total of 4 terabs of terabits per second

3:49 of networking per server that's 40x sorry 80x more than what you need for a

3:54 typical general purpose server right so this is what really makes us exciting because AI is lifting all boards and

4:01 networking is no exception to the rule over here and that doesn't stop in fact it

AI is lifting all boards

4:07 goes on fire in the next few years this is a slide straight from Nvidia's GTC uh Jensen's presentation and what Jensen

4:14 showed is that in 2025 when the Blackwell 300 series of GPUs comes out

4:19 is going to be powered with 800 gig of networking with the CX8 nicks and doesn't slow down there if you fast

4:25 forward another you know 12 to 18 months in 2026 2027 when the Reuben GPUs come

4:30 out they're going to need 1.6 terabs per second of networking per GPU from 400

4:37 gig today to 1.6 6 terabytes per second in just the next few

Why networking is important

4:43 years and so now I want to kind of impress upon why networking is important

4:48 in your entire AI life cycle and no rocket science in this slide so far over

4:53 here i think people understand the life cycle of an AI model but just to baseline it I'll list a little bit right so you start with what you call as

4:59 foundational model training this is the purview of large cloud providers or large enterprises who have resources of

5:04 thousands and thousands of GPUs to train a model from scratch but most enterprises are essentially going to be

5:10 doing fine-tuning where you take a generalized model which has been trained on a generalized data set and fine-tune

5:16 it for your with your proprietary data so that it can be customized for your own specific use case and then the

5:22 rubber hits the road in the case of inferencing where you take that trained model and deploy it in a distributed way

5:29 across many many AI clusters so that it can serve end users now we know that training was always a multiGPU problem

5:36 but what I want to impress upon today is that with the rise of things like reasoning models and agentic AI uh even

5:42 inferencing is becoming a multi-GPU communication problem uh while networking is important

5:48 for all of them what the the metrics tend to change between training and uh

5:54 inferencing in training the most important metric is job completion time how much time did it take to train my

5:59 model and with inferencing the metrics are slightly different you're looking at throughput uh which is my I have an AI

6:05 factory that is doing inferencing how many tokens per second or what's my throughput that I can generate for all

6:10 of my users and the second metric is because there's a realtime user sitting behind that LLM waiting for a response

6:17 the latency or the time to first lo token is a important metric for inferencing as well and like I said

6:23 networking plays a critical role in maximizing all of these metrics and we'll you know today we'll show you

6:29 how now diving Moving a little bit into training uh you know the training

Training cluster

6:36 cluster the green box in the middle think of it as a cluster of GPUs typically has three different networks

6:42 involved there's a front-end network there's the backend storage network but the important one that we want to focus

6:47 on today is that backend GPU training network that provides GPU toGPU connectivity across that cluster and

6:54 this is where you want to provide high-speed low latency congestion-free networking because any congestion in

7:01 these in this infrastructure can lead to traffic drops and anything that slows

7:06 down GPUs due to retransmission of packets or any other reason is really criminal in a in the case of AI

7:11 infrastructure because these GPUs are expensive you're spending millions and millions of dollars for these GPUs uh

7:18 for your AI infrastructure and having them sit idle is not something that you want so the networking plays a critical

7:23 role in improving the job completion time for training jobs

Inferencing

7:28 you look at inferencing like I said it's is no different you start with the left hand side the most simplest case is

7:34 single node inferencing right single client request comes in a GPU response sometimes it may even be a CPU right but

7:41 as I said with the rise of new types of models for example reasoning models where there's an internal dialogue

7:47 happening in the model before even the first token is sent out to the user there's a lot more east west traffic

7:53 happening as the model tries to reason its output before the first token or the first response goes out to the user uh

8:01 again there the you need that becomes a multiGPU problem the same thing with agentic AI frameworks so what we are

8:06 seeing is that increasingly uh inferencing is becoming a multi-GPU problem as well and requires a need for

8:12 a back-end inferencing network for east west communication and la lastly we have rag or retrieval augmented generation

8:20 where idea behind retrieval augmented generation is that you're supplementing an LLM's output to give it to enable it

8:27 to give more precise responses by giving by supplementing proprietary data that

8:32 it can use to to generate a response uh again in the case of rag you're basically looking at an environment

8:38 where you need to deliver very low latency because again there's the end user sitting at the end of the response

8:44 and you need to have you know those shared storage nodes that can be accessed at very very high speed so

8:49 networking again is plays a very important role over there now you know stepping one step

Challenges

8:57 above uh you know what are the challenges that anyone who's deploying an AI cluster faces i hopefully by now I

9:04 have stressed upon the fact that performance is important because GPUs are a critical asset but these AI

9:10 clusters are increasingly more operationally complex to manage as well think about it i told you that it was 4

9:16 terabs per second of networking per server that's so many more optics that can fail so many more links that can

9:22 bounce so many more BGP sessions that can flap and on top of that you have the need in AI clusters to have realtime

9:29 visibility of congestion events and the ability to manage and remediate those congestion events as well so

9:35 operationally way more complex than typical data center networks security was always important in an AI cluster

9:42 now in the world where your proprietary data is being used to train these models you face new rest new threat vectors

9:48 around model excfiltration data excfiltration etc and we're going to spend some time on that and then finally

9:53 in the in the world of AI clusters what we are seeing is that customers are facing this new dilemma should I pick

9:59 open best of breed technologies based on standard Ethernet or am I locked into

10:05 proprietary technologies just because they offer uh you know a closed ecosystem or a tightly coupled ecosystem

10:10 and we're going to spend some little bit time on that aspect as well how much are you seeing the shift these days from

10:16 scale up mega clusters to more scale out uh more efficient ones that only use

10:21 resources as needed is that something that that you all are seeing and you're planning for or do you already have that

10:27 so we are in the scale out domain right now and uh the scale up domain today

10:33 with Nvidia is a pretty locked in and we switched architecture but as AMD GPUs become more relevant that scale

10:40 up scale up opportunity opens up to other vendors as well so today we are mostly scale out but we are also looking

10:47 at scale up with non- Nvidia GPUs so it seems to be opposite you have the dense

10:52 models that people started with and the that are vertically but you've started out smaller with the more efficient LLMs

10:58 so I mean look until NVL72 came out uh most of the hopper class of GPUs was

11:06 sold as discrete at DGX and NGX systems right it's only with the blackwell where the NVL72 is coming out right so most of

11:12 our customer deployments even there's a customer who's deployed us for 200,000 GPUs this was all scale out it was not

11:19 deployed with scaleup scaleup is in fact is emerging recently since NVL72 became more uh popular starting in 2024 GTC

11:28 And do you have at some point a slide and maybe I missed it earlier that shows the whole topology of a of a AI network

11:34 yeah we're going to spend some time on that later okay great thank you

Performance

11:40 so uh starting with performance uh no shame in saying the first thing that we did was to throw bandwidth at the

11:46 problem right and we were the first vendor in the industry to ship a 64 port 800 gig switch we were the first OEM

11:52 vendor in the industry to ship a 64 port 800 gig switch that can uh in a two

11:57 compact form factor that can actually have two ports of 400 gig connectivity to the hopper class of GPUs and is

12:03 futurep proofed for the Blackwell class of GPUs as well so you can actually build a pretty large size cluster 8K

12:09 GPUs cluster in a three-stage design with just the QFX 5240 in a leaf and spine roll but if you want larger we did

12:16 not really stop there we also introduce a PTX 10,0008 chassis that has 288 ports

12:22 of 800 gig so you can build clusters again in a three-stage design going up to 32,000 uh

12:28 GPUs and now because the PTX is also a deep buffer system like I said there are

12:34 multiple AI clouds they need to be interconnected together using data center interconnect this deep buffer

12:39 platform is a is a superior choice even for your data data center interconnect use cases in fact we were the first

12:45 vendor in the industry to have a deployment of 800 gig using coherent optics for ZR lens so we were first with

12:53 800 gig on the fixed switches we were first with 800 gig on on the PTX with the debuffer switches both for data

12:58 center fabric as well as data center interconnect uh use cases now um now we were obviously busted to

Market Share

13:07 the market like I said with 800 gig and that's hopefully a testament to our engineering innovation in inity but

13:14 doesn't hurt that we got some market recognition out there as well this is something that c you know nobody expects

13:20 to see but as reported by 650 group just a few months ago in the year 2024

13:26 Juniper had a leading 800 gig market share the reason for that is we had the products first to market uh but

13:32 obviously this AI way was right picking up right at that time so it was a good product market fit but we had a market

13:38 share that was larger than Nvidia larger than Arista larger than Cisco in this data right here are we looking at any

13:44 specific verticals or regions or would that be a slide talking about global market share this is global market share

13:50 but data center alone it's not outside of data center it's just focused on data center

Dynamic Load Balancing

13:59 now like I said first thing we did was throw bandwidth at the problem but we all know that standard Ethernet was not

14:05 designed to be lossless uh it is not standard Ethernet without any

14:11 advancements advancement that we're going to talk about today is not suitable for AI workloads and the reason

14:16 for that is that Ethernet switches make a random hashing decision on which path

14:21 a flow should take so if you look at a picture on the top right two two flows two 400 gig flows are coming in from the

14:27 leaf from GPUs hitting the leaf and even if I have built up enough capacity between my leaf and spine it's a

14:34 possible that that Ethernet switch makes a decision that both those flows go to the same path to the spine and that link

14:40 is going to get congested so you have a situation where some links are underutilized others are overutilized

14:46 and need and you start hitting congestion and failures out there similar situation happens on the picture

14:51 on the on the top left as well so we started looking at this problem

14:56 you know way back in 2022 2023 and the first thing that we did was introduced a capability called dynamic load balancing

15:03 what dynamic load balancing does is that it makes a path forwarding decision not

15:09 just based on what available paths that routing tells me but based on the quality of those links in real time is

15:15 making a real-time decision on the quality of those links and making a forwarding decision you'll see in the

15:20 next session that that gave us tremendous improvement in performance over standard Ethernet but again we did

15:26 not want to stop there uh we introduced this feature in 2024 called global load balancing the idea behind global load

15:32 balancing is that it's like a Google maps for the network when I'm driving I see the congestion in front of me but I

15:39 only Google maps can tell me the end to end path quality so I make a decision now to turn left versus right the same

15:46 way global load balancing what it does it gives this leaf device that it needs to make a path decision end toend

15:53 quality information of the entire path all the way to the destination leaf and using that information it can make more

15:59 educated decisions to avoid uh to avoid uh the both the local link

16:05 as well as the remote congested link but did we stop there no and today what we are announcing is something called RDMA

16:13 aware load balancing this is a capability that enables a switch to see RDMA

16:21 subflows be aware of RDMA subflows and map them to specific paths through the

16:27 fabric it is gives us the highest performance as you'll see shortly it

16:32 gives you the lowest variability in performance so if you run that AI job 10 times you're going to get the same

16:38 performance each and every time dlb and GLB still give you some some amount of variability around a mean and it

16:45 requires a least amount of tuning to optimize the performance of your infrastructure

16:51 so today we are announcing AI load balancing uh which is these umbrella term that we are using for all of these

16:57 load balancing capabilities that deliver a 50% or greater than 50% improvement

17:03 over standard Ethernet now at this point if you are thinking that's you know you

17:08 made me curious but did not satisfy my curiosity completely i need to learn more I need to know more uh this is what

17:14 I was designed to do uh Ali Vikram Singh who's joining me or joining the next session is actually going to cover these

17:20 in a lot more detail so stay tuned for uh his session but I want to put a little bit plug over here um and this is

17:26 an important one you will hear our competitors who have proprietary

17:32 networking technologies for AI clusters constantly say things like this proprietary technology is 60% better

17:38 than standard Ethernet 70% better than standard Ethernet what they're comparing against is unoptimized Ethernet that was

17:45 not designed for AI clusters you know with the capabilities and the improvements that we are bringing to the

17:51 table we feel very very confident that customers deploying these clusters will get the performance and like comparable

17:57 to any proprietary technology out there including Infinipad so in this case what does open mean as far as proprietary

What Does Open Mean

18:04 does this mean that people have so anyone has access to the protocols is it

18:10 is it it's based on BGP it's based on Ethernet i mean all and know it can connect to an AMD GPU it can connect to

18:16 Cerebra Samonova it's not it's agnostic to the GPU you know it's not locked in

18:21 yeah industry standard industry standard yeah i have a question about Bitcon um

18:28 uh on the previous slide you said no tuning i would expect that uh that you

18:34 need to do some tuning definitely with RDMA if you're going to go all the way into the stack of the operating system

18:41 the servers as such tuning on day zero like uh you know you don't need to tune

18:46 uh like you it's not like you have to tune constantly like if I'm running a different model model A versus model B I

18:52 don't need to tune the network for that DLB and GB do require some amount of tuning there's some timeouts that you

18:58 have to configure appropriately to get the maximum performance uh But you know this technique will not and I think

19:04 Vikram session will make make that a whole lot more clearer okay thanks so and some of the things that you're

Packet Spray

19:10 talking about are are these part of the ultraethernet consortium work these so

19:15 uh so part some of them are uh some of them will become standardized as part of the UEC uh so for example our switches

19:22 already support packet spraying right what what is UEC standardizing the standardizing the ability of switches to

19:27 packet spray and the ability of the nicks to be able to to handle those out of order packets that's something we already support today okay right because

19:34 I knew that was part of it when you were talking about the the load balancing i just wasn't sure if this was tied into

19:41 to what's actually going to be standardized or not yeah some of these things so honestly from a from a UC

19:46 perspective the standardizing on packet spring when it comes to congestion avoidance right and that's something we

19:52 already support in our platforms

Power

19:58 now uh we can't really have an AI data center discussion without actually talking about power and uh if any of you

20:05 attended Jensen's uh keynote this year in GTC he eloquently stated the problem

20:10 statement especially around optics right he said a 100,000 GPU cluster has 400,000 optics associated with it you

20:18 know if we think that those optics consume 12 watts of power depending on whether what the reach is uh that's

20:25 total of close to 5 megawatt of power that is consumed just by optics alone

20:30 right and what Jensen said that you know he's introducing technologies uh like CPO that will be available in 2026 we

20:37 are working on CPO as well but what I want to say the operative word is that we have technologies today in something

20:44 called linear pluggable optics that take some of the DSP functionality away from the regular transceivers into the switch

20:51 the tomahawk 5 switch itself on the chip itself so that we can actually deliver that reduction in power consumption like

20:58 67% in the case of VR optics again this is available for our customers to take advantage of even today while we are

21:04 working on other technologies like CPO aim for

21:09 2026 now I did mention operational complexity as the one of the bigger

21:15 challenges in AI clusters And uh we have a whole session talking about Abstra and how how it is the

21:22 simple antidote to that complexity but just a little bit of an introduction on what Abstra is abstra is our day zero to

21:28 day end full life cycle management uh solution for fabric management and assures so right from design to deploy

21:35 to day to assurance it can do all of those things for regular data centers now obviously in the last year year and

21:41 a half uh we have been spending time enhancing the capabilities of abstract

21:47 to be ready for AI clusters and we'll see a lot of that today but I just want to point out two capabilities that I've

21:52 you know close to my heart because in talking to our customers one of the things that we realized was congestion

21:58 management was the hardest part of operating an AI cluster so we have delivered some capabilities that provide

22:04 not just switch visibility but switch to nick visibility end toend visibility from all the way from the switches and

22:10 the nicks to give you that view of where is the congestion in my network in my in a in a singular view but again we did

22:17 not stop there seeing congestion hotspot is one part of the problem operators

22:22 could sometimes spend days in remediating those congestion hotspots

22:27 and what we do with Abstra is this intelligent mechanism that is constantly monitoring for congestion and making

22:33 recommendations to operators that here's the hot spot this is a recommendation to tune the network or change the

22:39 congestion handling profiles so that you can al remediate that congestion as you

22:44 switch from one model to the next

Security

22:51 so the third pillar of our solution is around security and you know again Kedar

22:56 will join us later to talk about the security portfolio but like I said security was always important in any

23:02 data center in this new world of AI where your proprietary data is being

23:07 used to train a model you want to make sure that you are protected against those new threat vectors around model

23:13 excfiltration around data excfiltration etc now let me be very clear over here that when your proprietary IP is at risk

23:22 this is not just a technical issue this is you know a hit to your business value if your entire model gets excfiltrated

23:28 out and that's what your and that model was trained on your proprietary IP uh more more of that to come but what I

Cases from 2024

23:35 want to point out over here is that this is not just theoretical risks these things are happening in the industry as

23:41 we speak and I just want to touch upon a couple of cases just from 2024 in an example on the top in early 2024 there

23:48 was a case where hugging face models were infected with malware you know data scientists or users who downloaded that

23:54 malware of that model essentially had that model model running some you know

24:00 some suspicious code that opened a back door to a bad actor in the cloud and from there you know that bad actor can

24:06 now do anything nefarious with your models your data etc the second one actually blows my mind completely is

24:13 because you know this is a case where an AI as a service provider was uh had a vulnerability where a tenant in that in

24:21 that in in that infrastructure could be a bad actor and they run a malicious model once that malicious model is run

24:29 that model now laterally moves across to other tenants in that infrastructure so it's not just you who are you know maybe

24:35 did something wrong uh maybe you know maybe by intention or or not but now

24:41 you're affecting every other customer that was residing in that AI as a service or GPU as a service uh

24:47 infrastructure so I think that was pretty crazy uh so that's why at Juniper we recommend a defense in depth approach

24:55 you protect at the edge with our secure with our SRX portfolio you do not want that threat to enter your infrastructure

25:01 but you know sometimes even with the best efficacy in the industry day zero threats do penetrate and what you want

25:07 to do is that then to provide that east west security within your fabric itself so you assume that you know you have

25:12 some day zero vulnerability some day zero threats you know did enter your infrastructure you want to prevent that

25:19 lateral movement of those threats across your infrastructure with our both our QFX multi-tenency capabilities that

25:24 provide isolation for tenants in the with the QFX as well as our SRX portfolio then you have data center

25:31 interconnect you want to encrypt every communication between your between your

25:36 clouds to prevent from snooping style of attacks so we have max IPSec and uh

25:41 encryption capabilities available on both our SRX as well as our MX series of data center internet portfolio and then

25:48 finally if your applications are distributed and are running on public cloud then the SRX portfolio and more

25:53 specifically the virtualized SRX or the containerized SRX provides the same level of security for your public cloud

26:00 applications as you would see uh for your uh private cloud uh infrastructure and Kedar will spend a

26:06 whole session you know talking a little bit more in depth about this

Open vs Proprietary

26:13 so I said I'll touch upon the topic of open versus proprietary a little bit more and u and what we see is that you

26:21 know in every engagement I we have with our customers they're facing this dilemma right you know Ethernet had

26:26 become the open de facto stat standard for pretty much every place in the network whether it was campus whether it

26:31 was van whether it was data center and now comes this new world of AI and now they have to they're dealing with this

26:38 new dilemma should I go with best of breed technologies But then I have the have the challenge of kind of start

26:43 having to stitch it together and making sure it performs and works well or I go to this single vendor out there that

26:48 provides a GPU also provides a network and I have this allure that going with

26:53 this tightly coupled s you know ecosystem is the only thing that you know gives me that performance so I have

26:58 to make this trade-off between lower cost with a best of breed versus proprietary solutions that are more

27:04 expensive but have that allure of of performance so we really decided to take this challenge head on and the way we

27:12 did this was built an AI lab where we bought in all of these best of breed

27:18 components right we spent a lot of money honestly uh but we brought in all of these best of breed components a GPUs

27:24 from Nvidia and AMD storage from our partners like Vea and Vast nicks from

27:30 Broadcom AMD's Pensando sorry Polar Nick as well as Nvidia's connectex 7 six and

27:36 seven nicks and We brought them together stitched it together into an AI cluster

27:41 rigorously rigorously tested it for the for few months uh against you know real world training and inferencing workloads

27:47 and we're constantly doing that that kind of exercise never really stops as technology keeps evolving and what we

27:53 produce were these validated designs like the Juniper validated designs so if you do a Google search even now for

27:59 Juniper AI validated designs what you'll hit are two assets one is a validated

28:04 design that is anchored on AMD's GPUs another validated design that is anchored on Nvidia GPUs so what we are

28:11 what we what we are hoping to offer our customers is something very simple that they do not have to make a trade-off

28:17 anymore between best of breed flexibility and performance on the other hand they can get the best of both

28:23 worlds so I'm a little confused about Okay so you're doing a validated design based upon this stack here of Kimberly

28:29 Bates with HR group and I'm not a network person so just really not

28:35 um are you're doing a validated design using those that criteria as opposed to

28:42 what as opposed to uh as opposed to Nvidia

28:48 saying that hey if you buy my GPU you have to buy Infiniband or my buy my proprietary networking technology so in

28:54 Apple stack it's like an Apple because some of those things are also in the stacks that is in with so so what you're

29:00 basically saying it's Nvidia's um networking and Nvidia's GPUs which you

29:05 have Nvidia so it's basically you've taken Nvidia's networking out put yours in built that stack out have a validated

29:12 design so I've got confidence in a ship is are you seeing Well that's fine no

29:18 that's enough i'll get exactly that yeah uh and in the case of Nvidia we've

29:23 actually followed their super pod design right the only thing that we've replaced we even use slurm for orchestration

29:30 everything is the same we use we only thing we replaced is infin with Ethernet that that's pretty much the only thing

29:35 that we changed over there so with as validated designs you're saying hey we've tested all these you can mix and

29:41 match do you have reference architectures that say this is exactly how you can connect these and yeah and

29:46 sizing guides and all sort absolutely that that validated design actually is our reference guide so has all the best

29:52 practices down to which knob to configure how to tune your DLB dynamic load balancing everything right is is

29:58 all out there uh and not only that we've actually codified that into blueprints

30:03 in Abstra so if you don't want to you're not a CLI junkie and you want Abstra to deploy that all of those best practices

30:09 that we have learned and deploy them in one click reliably repeatably Abstra can do that as well abstra can deploy that

30:16 entire reference design so what part of Broadcom did you use what what's Broadcom element in that the Broadcom

30:21 obviously the the the switch is powered based on Broadcom's uh switching silicon that's what I figured okay but Broadcom

30:27 also has nicks they have the Thor 2 400 gig nick was that in there as well yes yeah okay mhm

Gartner Magic Quadrant

30:37 okay so um you know here at Juniper we are a engineering company and we do

30:43 enjoy solving some of the most pressing needs for our customers but once in a while it feels nice to get recognized

30:50 and this is something that was uh recently announced by Gartner just three or four weeks ago gartner reintroduced

30:57 the Gartner magic quadrant for data center in 2025 after a 5-year gap and we

31:03 were squarely in that leaders quadrant along with the other giants in that space right so the Gartner magic

31:09 quadrant being in the leaders quadrant I think speaks for itself so I won't linger on this slide uh anymore uh but

31:15 it is powerful what I do enjoy uh and I like even more is that there is a

31:21 companion report to the Gartner magic quadrant called the critical capabilities report and here is where

31:26 Gartner goes into details of a vendor's capabilities to assess their capabilities against a specific use case

31:33 and we were number one in enterprise data center buildouts like I was talking about those 50 gig servers for

31:39 enterprise data center buildouts we were number one uh out there ahead of Arista

31:45 ahead of you know the other you know typical vendors you would expect to be out there and u uh this is a testament

31:52 mostly to our fabric management capabilities with Abstra that you're going to see in a few

31:57 minutes and the topic of today's discussion was AI data centers uh I did say we have a leading market share but

32:04 Gartner recognized that as well uh we are only number two to Nvidia right

32:09 Nvidia is a behemoth out there that forces you know their customers to bundle in the whole solution and look at

32:16 the scoring like we were only 02 behind Nvidia for AI networking right so again

32:23 great testament to a great validation of our technology and the momentum that we

32:28 are seeing with AI data centers out there with our networking and security solutions

32:36 And it's not just analysts who are uh you know raving about us uh we have

32:42 customers as well uh this is just a small fraction of the customers who are actually public references and all of

32:48 these whether you're looking at some GPU as a service providers like digital ocean vulture ion stream xai who's

32:54 building this massive cluster in uh you know called colossus in me in in Memphis or some other uh enterprises out there

33:01 like sambonova cerebras wyoming PayPal uh they all have trusted uh juniper to

33:06 build their AI cluster whether it is for storage networking backend GPU networking or front-end networking so um

33:12 again that was my final slide on uh the traction that we're getting uh with customers and I'll take any few few

33:18 questions before turning it over to

33:26 Vicram going back to your slide where you showed the the uh hybrid AI cloud

33:32 how are you defining that compared to multicloud are you seeing more of multi cloud environments so I think of hybrid

33:38 I think of onrem and on the cloud and then multicloud of course being you've got multiple clouds within that not

33:45 onrem stuff yeah so so we are seeing so I think hybrid cloud maybe have been loosely

33:51 used term it is multicloud as well uh we are definitely seeing a lot of repatriation of workloads happening due

33:57 to the AI wave all right because you know you you need the models are being

34:03 trained on your proprietary data right so whether so we are seeing that applications moving closer to the data

34:08 rather than data kind of moving to the applications in public cloud so we are seeing a lot of private cloud buildouts happening uh with uh the know AI wave uh

34:17 and yes it is actually you could say it's a it's multiloud in a in a sense that some applications are running in private cloud some applications are

34:23 running in in public cloud in that sense yeah so data gravity and latency and data gravity kind of pulling

34:29 applications towards private cloud yeah I would assume security too for particular industries absolutely like privacy data privacy security reasons as

34:36 well thanks another question uh Um until now

34:43 what I heard most about Juniper was Mist how does it fit in this big picture here

34:49 mhm so Mist started out as our um you know portfolio for managing campus and

34:57 branch right data in data center we always so had abstra for managing and

35:02 operating uh data centers right but over the last few years we have uh integrated

35:08 some of these technologies together so Mist is basically a leading AI ops engine like for you know today we're

35:14 talking about networking for AI but that's more of AI for networking right and abstra started out as an on-prem

35:20 instance management instance but now abstra also has a cloud presence and that cloud presence is powered by mist's

35:27 AI ops engine to provide more predictive capabilities uh for data center as well

35:32 so simple answer mist was for campus abstra was for data center but going

35:37 forwards we are seeing more and more of their integration where we're bringing this KI ops capabilities from Mist into

35:43 Abstra in a cloud instance as well okay thanks

35:48 so Jim Sprincy CDC um your customers you know you've laid out your three or four

35:54 pillars is it the security aspect that most people are interested in i mean I'm

35:59 just interested if you kind of did a pie chart yeah so I would say that when it comes to AI clusters I talked about four

36:06 pillars uh you know let's say performance operational complexity security and open versus proprietary if

36:11 you don't have performance the rest of the pillars don't are not nobody's going to consider you right so you need to

36:17 first show that you know I'm neutralizing every other proprietary technology out there with an Ethernet

36:23 lowcost solution lower cost solution that delivers the performance so I would say everybody's in there for the

36:28 performance otherwise you know the remaining pillars do not even matter but yes security becomes a a differentiator

36:33 for us our operational capabilities with abstra and our security do become the differentiator once we have neutralized

36:39 that performance is is good okay