# The Total Risk of Ownership (TRO) of AI Factories

https://www.youtube.com/watch?v=aAuzfl6-6xg

[00:01] Hello, welcome.
[00:03] Hello, welcome. Rich Lapenbush with Super Micro and I'm joined by Imran and Omama.
[00:07] Omama.
[00:07] Hi everybody.
[00:09] Hi everybody.
[00:09] Hi.
[00:09] We're going to be quite informal in this.
[00:11] Uh we're introducing a new topic uh to OCP uh the notion of risk and how risk is thought of and evaluated uh primarily towards AI factories, AI edge and AI deployments.
[00:28] So there's a lot of different ways of thinking about risk with data centers that have been done for decades.
[00:35] Uh and we're taking a more modern spin on that and I think you'll uh hopefully enjoy uh what we're offering up and uh really this is an introduction.
[00:45] This is a invite if you will to join us in a discussion that we started uh two two years ago.
[00:50] It's been two years.
[00:50] It's been two years in which we felt there was a der of uh consideration of the risks.
[00:57] So, uh if
[01:01] Consideration of the risks.
[01:04] So, uh if every single product you're deploying is only one month old, uh we're going to have different issues than if some of the products are a little bit more mature than others.
[01:12] And so, uh Iran and I were having that discussion and it it boiled over into well really we should think about risk in a in a broader sense.
[01:19] And so this is our opening volley to you the community to join us in that discussion.
[01:31] We started from a basis of the total cost of ownership which is well understood well accepted.
[01:38] Uh many finance professionals use this uh in evaluating uh data center investments whether it's capital in primarily capital investments um and we took a spin off that and said okay they're looking at just the cost they're looking at just the cost smoothed over time but absent risk so if the risk premium is super high on one investment and super
[02:02] super high on one investment and super low on another they they look exactly the same.
[02:04] low on another they they look exactly the same maybe we should consider uh handicapping each each investment,
[02:07] the same maybe we should consider uh handicapping each each investment, each risk and uh we started down that path and we and I I had recalled through a prior deployment years ago that ISO had had codified a lot of this in ISO 31,000.
[02:11] uh handicapping each each investment, each risk and uh we started down that path and we and I I had recalled through a prior deployment years ago that ISO had had codified a lot of this in ISO 31,000.
[02:14] each risk and uh we started down that path and we and I I had recalled through a prior deployment years ago that ISO had had codified a lot of this in ISO 31,000.
[02:17] path and we and I I had recalled through a prior deployment years ago that ISO had had codified a lot of this in ISO 31,000.
[02:20] a prior deployment years ago that ISO had had codified a lot of this in ISO 31,000.
[02:23] had had codified a lot of this in ISO 31,000 and a series and it's now 31,18 and it keeps going.
[02:25] 31,000 and a series and it's now 31,18 and it keeps going.
[02:28] and a series and it's now 31,18 and it keeps going.
[02:30] keeps going. There's a whole family of 31,000 series uh standards in there that all of you can go take a look at but it's very generic.
[02:33] 31,000 series uh standards in there that all of you can go take a look at but it's very generic.
[02:35] all of you can go take a look at but it's very generic.
[02:37] it's very generic. It's it's risk in the generic sense.
[02:39] generic sense. It's not applied to the data center industry per se or or AI per se.
[02:42] data center industry per se or or AI per se. So we started from that route if you will uh TCO and there's a great TCO uh working group already underway.
[02:45] se. So we started from that route if you will uh TCO and there's a great TCO uh working group already underway.
[02:50] will uh TCO and there's a great TCO uh working group already underway. It's been uh now 3 four years uh deep into the uh evaluation of cooling TCO air versus liquid.
[02:51] working group already underway. It's been uh now 3 four years uh deep into the uh evaluation of cooling TCO air versus liquid.
[02:54] been uh now 3 four years uh deep into the uh evaluation of cooling TCO air versus liquid.
[02:58] the uh evaluation of cooling TCO air versus liquid. So there is some basis uh for our work here and and so we took a a
[03:02] versus liquid. So there is some basis uh for our work here and and so we took a a
[03:04] for our work here and and so we took a a riff if you will off that TCO group here.
[03:07] riff if you will off that TCO group here at OCP.
[03:11] And when we think about risk, we think about the entire risk uh geopolitical risk, uh environmental risk, seismic risks, fire, smoke, you name it.
[03:21] A very broad the broadest array of risk you can consider.
[03:24] uh and and what might happen to you as you run your AI factory out in the wild wherever you plan on deploying that.
[03:32] And what we did was we took the ISO guidelines standards and literally created a mapping the first crosswalk that we we are aware of uh of how an AI factory maps to these ISO standards and we're plan on presenting uh some of our findings to that and sharing it with you.
[03:50] Um, it also introduces and and takes in a lot of assumptions from OCP.
[03:57] So, if it's not explicitly documented, I want you to know that basically we're starting from a route of TCO and OCP and then growing from there.
[04:04] Um, and really
[04:07] then growing from there.
[04:10] Um, and really the goal is to promote better decision-making.
[04:13] uh by owners and operators of data centers uh in everything from their concrete recipe uh all the way to how they manage their weather and how they consider their cooling systems.
[04:21] So that's just a quick overview.
[04:24] I'm now going to hand it over to who's going to share a little bit more about the model itself.
[04:32] Yeah, thank you so much.
[04:34] uh basically based on whatever the background which has given us why the risk is important or something.
[04:40] So we came across the conclusions that we built a model that is normally we call it cooling and energy risk assessment model serum.
[04:47] Basically we derive this word from the medical terminology you can say serum that used to treat or diagnose the treatments and this model is basically based on ISO 31,000 uh risk management which is basically to design implement the risk management process and obviously do the some recommendations as
[05:07] Obviously do the some recommendations as well, not only the ISO 31,000 as it is.
[05:10] Well, not only the ISO 31,000 as it is mentioned series, like for example, 31073.
[05:14] Mentioned series, like for example, 31073 and few other whole series we have.
[05:16] And few other whole series we have covered that and uh in this model just.
[05:20] Covered that and uh in this model just to give we have taken the latest version.
[05:22] To give we have taken the latest version which was released last uh I mean two.
[05:24] Which was released last uh I mean two years ago and two or three years ago, 2023.
[05:27] Years ago and two or three years ago, 2023, and obviously whatever the standard.
[05:29] And obviously whatever the standard comes in, it is valid for next five years, so.
[05:31] Comes in, it is valid for next five years, so whatever the model we have built in you.
[05:33] Whatever the model we have built in, you can say it'll last at least for next two.
[05:36] Can say it'll last at least for next two, two to three years, and obviously there.
[05:37] Two to three years, and obviously there will be minor updates as soon as the new.
[05:39] Will be minor updates as soon as the new versions comes in, so basically based on.
[05:43] The some methodologies which I will be.
[05:45] explaining in a little bit uh in the.
[05:47] Next slides, but we came across some of the risks and then we compare it with.
[05:50] The risks and then we compare it with the air and the liquid cooling.
[05:52] So this is what the main methodology we will uh.
[05:55] We have followed to build this model.
[05:58] First you can see we have set up the.
[05:59] Scoping of the our serum model and the.
[06:01] Scoping obviously we discussed about the.
[06:04] ISO model that we have taken as a.
[06:09] ISO model that we have taken as a background and we identify I mean background and we identify I mean identify or you can say in other words identify or you can say in other words we consolidated the main types of air cooling and the liquid cooling like six types of air cooling we consolidated.
[06:20] We consolidated that's you can say mechanical economizers and operators and I'm not going into the details that based on the cooling experts we shortlisted those uh cooling technologies and then obviously we came across some of the requirements as well.
[06:35] That is we came across 14 different types of requirements because risk is not necessary to that is required to every uh data center operators for example if you are trying to build design or your data centers in the Middle East region you need to think about more about the regulatory.
[06:50] You need to think about uh sometimes about the security in certain parts of the worlds as well so we came across 16 different types of requirements that for example the capacity it depends on the what type of loads you wanted.
[07:05] We came across sustainability, security and scalability.
[07:07] So you can tailor it based
[07:10] Scalability.
[07:12] So you can tailor it based on your requirements and then go and further explore this model based on the.
[07:14] Identify the risks that you are then we obviously traditionally followed the ISO.
[07:17] Obviously traditionally followed the ISO 31,000 risk assessment process which you can see that is highlighted in the gray.
[07:24] So that in that we follow the three elements of the risk assessment process.
[07:27] Elements of the risk assessment process of the ISO 31,000 that is risk uh identification, risk analysis and risk evaluation.
[07:32] Identification, risk analysis and risk evaluation.
[07:36] In the risk analysis obviously we identified the types of risks uh that is uh available uh identified between uh the different ways.
[07:38] Obviously we identified the types of risks uh that is uh available uh identified between uh the different ways.
[07:44] It's not only about the cooling, it's about the power and other elements as well.
[07:46] Ways.
[07:48] It's not only about the cooling, it's about the power and other elements as well.
[07:50] We did some analysis based on some assumptions that we made it.
[07:53] Some assumptions that we made it.
[07:55] We will be sharing these and discussing with you uh later on throughout this panel as well.
[07:57] With you uh later on throughout this panel as well.
[07:59] And by the way, I mean just a quick I mean these methodologies and other models as well.
[08:01] Just a quick I mean these methodologies and other models as well.
[08:04] We will be soon publishing a peerreview scientific research paper which is available open access to get more detail about this.
[08:06] We will be soon publishing a peerreview scientific research paper which is available open access to get more detail about this.
[08:08] Research paper which is available open access to get more detail about this.
[08:09] Research paper which is available open access to get more detail about this.
[08:11] access to get more detail about this methodology and to access to all these.
[08:14] methodology and to access to all these uh risk elements that we came up.
[08:16] uh risk elements that we came up different uh I mean different results as.
[08:18] different uh I mean different results as well.
[08:18] Then we did the evaluation.
[08:21] Evaluation we did the simple scoring.
[08:23] based on the risk management uh standard.
[08:26] that is about what is the level of the.
[08:28] risk based on the severity and the.
[08:30] likelihood.
[08:30] Uh moving on I mean this is.
[08:33] just to give you an idea based on.
[08:35] because as um rich already mentioned.
[08:37] that we have started two years ago since.
[08:40] this model is still keep evolving and we.
[08:42] are adding more and more risk based on.
[08:44] the industry experts and what is exist.
[08:47] at the moment.
[08:47] So we came across around.
[08:50] over 180 risks that we observed and peer.
[08:53] reviewed and obviously we also analyzed.
[08:56] through the failure modes and effects.
[08:58] analysis.
[08:58] Just to give you a quick quick.
[09:01] perspectives what does failure mode and.
[09:03] effects analysis means in terms of the.
[09:05] data centers.
[09:05] Basically what is failure.
[09:07] mode and effect analysis is that what.
[09:10] could go wrong and what would happen if.
[09:12] could go wrong and what would happen if it goes wrong.
[09:15] So if it goes wrong let's it goes wrong.
[09:18] So if it goes wrong let's say there's a leakage in coolant leakage or pump failure then there could be a overheating.
[09:23] there could be a equipment failure as well.
[09:25] So this is what a thorough analysis as well we have done in terms of identifying the risk as well and based on the 180 180 risks that we have identified we have categorized them into the 16 different risk types.
[09:39] you will come to know in the following slides what the 16 different risk types how we categorize them uh again to as I said in terms of the identification analysis and the evaluation phase and based on what is the ISO 31,000 standard sea users we make some 15 risk around 15 risk assumptions as well just to give an idea the one of the risk assumptions is that that it's the person who is filling that identifying the risk or scoring this should have some experience and they have the one mega a lot of loads it loads as well or the whole power load you can say and that data center is
[10:14] you can say and that data center is optional for the five years as well.
[10:17] optional for the five years as well.
[10:19] So these are some of the key assumptions that we also and it is not uh tailored to particular type of data centers but
[10:23] it's not only for the AI factories you can have it for the edge data centers as well you can have it for hyperscaler enterprises and cloud providers whatever
[10:32] it is it's so this risk is not uh restricted to one uh type of data center
[10:37] as well then we score them based on the scores that we that ISO 31,000 suggests
[10:43] as you can see from the screen that five levels of the risks that is 0 to 5 depending on the severity of the risk.
[10:50] So this is what uh ISO 5x5 uh risk model we followed and we based from the likelihood and severity then we assess the levels and you can see that how the low, medium, high and extreme length uh risks we came across based on the scoring that we have uh got it.
[11:09] And so these are I I will just share a couple of results that we came across while we identified the 180 risks.
[11:11] So in this on
[11:17] identified the 180 risks.
[11:17] So in this on your left side yeah left side you can your left side yeah left side you can see the graph.
[11:22] you see that how we analyze that over 180 rigs in that we noticed that top uh you can say top 48% of the risks are basically you related to cooling and climate system and building but the majority is in the majority of the risks that is uh around related to cooling and climate in this phase I would like Iran to chip in and see that uh to give your perspectives from the industry that why you think that the cooling and the climates have approached the highest uh point of the risk types.
[11:55] Sure.
[11:55] I think it's a great segue and I'm just going to give a little background before that that how we came up with this as rich had mentioned we were just discussing exactly what is it needed in the industry where the gap is um obviously we don't have a holistic um risk model that exists right now which can really encompass all these categories that um had mentioned right and be able to give the data center
[12:19] and be able to give the data center owners operators designers the entire uh.
[12:21] owners operators designers the entire uh players in the ecosystem an understanding of what are they building.
[12:23] players in the ecosystem an understanding of what are they building and exactly what foreseeable risks you.
[12:25] understanding of what are they building and exactly what foreseeable risks you know uh they might be seeing in the build of that AI factory for example.
[12:27] know uh they might be seeing in the build of that AI factory for example right it just happened to me and we were just building a very large data center a couple years back and uh we were literally thinking uh which AI hardware to deploy whether it's going to be CPU or GPU based for example and given that the limited knowledge we had at the time not knowing that the high bandwidth memory fab was already you know locked in or compromised for the next 2 years.
[12:31] build of that AI factory for example right it just happened to me and we were just building a very large data center a couple years back and uh we were literally thinking uh which AI hardware to deploy whether it's going to be CPU or GPU based for example and given that the limited knowledge we had at the time not knowing that the high bandwidth memory fab was already you know locked in or compromised for the next 2 years.
[12:55] memory fab was already you know locked in or compromised for the next 2 years.
[12:58] in or compromised for the next 2 years. It was very hard to get some of these GPU based uh AI nodes within let's say 6 to9 months.
[13:00] It was very hard to get some of these GPU based uh AI nodes within let's say 6 to9 months.
[13:03] GPU based uh AI nodes within let's say 6 to9 months. So that was a showstopper uh cuz we had to meet certain deadlines and obviously we know how much time to value time to token matters in the the current AI uh factory era.
[13:07] to9 months. So that was a showstopper uh cuz we had to meet certain deadlines and obviously we know how much time to value time to token matters in the the current AI uh factory era.
[13:09] cuz we had to meet certain deadlines and obviously we know how much time to value time to token matters in the the current AI uh factory era.
[13:14] time to token matters in the the current AI uh factory era. So this is why we came up with all these u categories.
[13:18] AI uh factory era. So this is why we came up with all these u categories.
[13:21] came up with all these u categories which really spans from obviously.
[13:23] which really spans from obviously cooling and environmental being one of.
[13:25] cooling and environmental being one of the top challenges and the risk areas we.
[13:28] the top challenges and the risk areas we see in the industry.
[13:30] I mean imagine if we are building a data center somewhere.
[13:32] in a dry climate versus a you know a.
[13:35] humid climate Denver Colorado versus.
[13:37] maybe Houston for example Texas right.
[13:40] there will be different design.
[13:42] requirements which can actually be you.
[13:44] know uh affected by the supply chain.
[13:46] issues for example right so it's really.
[13:49] important to understand and look at this.
[13:53] metric as a as a guide where as I said.
[13:57] every single stakeholder in this.
[13:59] ecosystem can use it to their advantage.
[14:02] It's so easy to use and by the way these.
[14:05] 180 risks that we have identified these.
[14:07] are not the only risk that will be used.
[14:10] um you know as a guide.
[14:12] This is a living.
[14:15] document and uh once this is going to be.
[14:18] uh you know made public or made uh used.
[14:20] to the public right obviously uh anybody.
[14:20] could use this to their advantage and be.
[14:22] could use this to their advantage and be able to build on the existing work that.
[14:24] able to build on the existing work that has been done uh by the co-founders and the authors of this u this very uh Sam methodology and really use it to the advantage.
[14:32] I use this pretty much every time I have a problem, right?
[14:36] And I talked to Rich Omayima and the other uh you know co-founders we have the co-authors we have on this and we talk to people in industry for example like I said this this memory issue is a big issue right and again um we know there is no such thing like one size fit all and more and more we are going towards liquid cooling we we we are seeing the road map from Nvidia you know it is obviously um you know going towards a megawatt rack and we know what it takes to cool a megawatt rack pretty much within the same form factor.
[15:07] This is going to be exciting.
[15:09] And as we talk about warm water cooling, we sometime may not be able to have the buffer that we might think we have.
[15:17] You know, a valve fails, you know, there's a leak.
[15:19] You know, we're looking at this very tight uh circuitry, the electronics on a
[15:25] tight uh circuitry, the electronics on a motherboard, which can really heat up,
[15:27] motherboard, which can really heat up, you know, due to failure of uh like I say, a water quality issue.
[15:31] for example, we're going through these very fine art orififices within the gold plate for example, right?
[15:35] So, so these are uh the risk which can really help you know reduce the TCO and we see this as a metric which is going to really make TCO look great because this is a level above TCO a TRO right this is going to encompass all the risk of course there's there's no crystal ball that we have that we can just talk about every single risk there's some unforeseen circumstances that might come up but based on the industry knowledge.
[16:07] I mean we have people who have been serving in this industry for years and this is a byproduct of that experience that you see here.
[16:12] So obviously cooling is a big challenge and we the the the graph that you see here we we can clearly see that uh liquid cooling is a champion obviously uh it poses less risk as compared to air when it comes to the
[16:27] compared to air when it comes to the entire uh way we we have calculated these risks.
[16:32] Right? So, so understanding the entire um sequence of this uh this system and this this product right uh the 5x5 metric which has the risk severity uh the the the uh the rating of the risk how we see this right is really important and you can come up with u the understanding of where the risk lies within that um within that uh you know uh framework.
[16:51] So this is how we see it.
[16:54] Um this has really been um an amazing tool and we are looking forward to the feedback from the industry as uh Dr. Mama had mentioned we do plan on uh you know having this submitted to a peer-reviewed journal which is going to be made you know uh available to a broader audience and uh feedback is really welcome.
[17:15] Thank you so much Iran.
[17:17] Uh yeah as Iran touched upon on the graphs if you see on the left hand side right hand side is just to give you a quick picture that how we normally score the risk that we have that we have identified.
[17:26] If you can see that we have scored them based on
[17:29] see that we have scored them based on the ISO 31,000 standard on the 5x5 ISO.
[17:32] the ISO 31,000 standard on the 5x5 ISO model that is scoring from 0 to 5 then.
[17:35] model that is scoring from 0 to 5 then you can see how we score them the.
[17:36] you can see how we score them the severity likelihood and then simply we.
[17:38] severity likelihood and then simply we multiply to get the level of the risk.
[17:40] multiply to get the level of the risk that we achieved.
[17:43] Uh this is again some of the quickest snapshots.
[17:45] uh basically at the moment in the model that we.
[17:47] at the moment in the model that we focused on uh on the director to ship.
[17:49] focused on uh on the director to ship liquid cooling and you can see on the.
[17:52] liquid cooling and you can see on the graph sides that uh how in terms of the.
[17:55] graph sides that uh how in terms of the risk severity if it is low to medium and.
[17:58] risk severity if it is low to medium and then high to extreme APC once we move.
[18:01] then high to extreme APC once we move into the high and the extreme the risk.
[18:03] into the high and the extreme the risk severity in the direct to liquid cooling.
[18:06] severity in the direct to liquid cooling chip is reducing as compared to the air.
[18:08] chip is reducing as compared to the air cooling.
[18:12] So most low medium risk are of in terms of the direct to liquid chip.
[18:14] in terms of the direct to liquid chip cooling is on the low and medium side.
[18:17] cooling is on the low and medium side but how the trend goes from the high to.
[18:19] but how the trend goes from the high to extreme.
[18:21] So as it says that when we scored them when we came across uh the.
[18:23] scored them when we came across uh the total of that like in terms of the air.
[18:25] total of that like in terms of the air cooling we came across let's say it's.
[18:27] cooling we came across let's say it's roughly the figure of 1300 sometime uh.
[18:30] roughly the figure of 1300 sometime uh scores and the liquid cooling was 1100.
[18:33] scores and the liquid cooling was 1100.
[18:35] So we came across that liquid cooling system shows an overall 15% lower risk
[18:38] level compared to air cooling and
[18:41] similarly when we talk about the
[18:42] likelihood it is 13% lower than the air cooling systems and obviously it's based
[18:48] on our experience primary secondary research post industry expert says and
[18:52] what is the current uh research at the moment is available and risk seability
[18:57] as well is 2% uh lower than air cooling systems and last but n is obviously we
[19:00] is that it's improved operational risk profile for direct to liquid J chip
[19:04] cooling uh as compared to the traditional cooling.
[19:11] traditional cooling.
[19:14] Uh this is last but not the least I mean some of other results as I said repeat again and that the more details you can find more analysis you can find in our review paper which is going to be soon published.
[19:23] But over here is when we divide the risk levels in and risk severity from low to extreme and in
[19:30] severity from low to extreme and in terms of the uh different types of risk.
[19:33] terms of the uh different types of risk types that we identified.
[19:36] If you can see that over here the building and the human and the particularly the software ones you can see we have more extreme risk uh in the air cooling as compared to the liquid cooling.
[19:45] Uh so this is an kind of thing that you can notice that where the more risk are available in terms of the air cooling when you're doing the particularly from the software perspectives from the human p human perspective means the set of risks that could be happens due to the human errors and due to the software means any updates or any software vulnerabilities you are adapting it and obviously the building in terms of the design construction and operation of the data centers as well.
[20:11] Uh yeah, Rich, over to you.
[20:16] And uh I'll just uh add that there's an additional author uh who's not with us today uh Dr. RT Wang uh also with Super Micro and uh he will hopefully uh be also uh presenting this same session in
[20:31] also uh presenting this same session in Taipei at the uh OCP
[20:34] Taipei at the uh OCP uh pack
[20:36] uh pack uh conference coming up in the next few
[20:37] uh conference coming up in the next few months. So, uh, there are four of us,
[20:40] months. So, uh, there are four of us, uh, and we're kind of that initial push
[20:43] uh, and we're kind of that initial push of people, uh, working together to come
[20:45] of people, uh, working together to come up with a model, something worthy of a
[20:47] up with a model, something worthy of a discussion, a a place to start. And, uh,
[20:51] discussion, a a place to start. And, uh, um, that's the four of us.
[20:53] um, that's the four of us. >> Yeah, I think uh, we are really open to
[20:55] >> Yeah, I think uh, we are really open to any feedback and suggestion because this
[20:57] any feedback and suggestion because this is not just our problem. This is the
[21:00] is not just our problem. This is the entire industry's problem. How we see
[21:02] entire industry's problem. How we see this entire ecosystem, you know, we are
[21:05] this entire ecosystem, you know, we are building data center as such a fast
[21:07] building data center as such a fast pace. You know, time to value like they
[21:08] pace. You know, time to value like they say time to token, time to revenue is so
[21:10] say time to token, time to revenue is so important. We are seeing more and more
[21:12] important. We are seeing more and more of these modular buildouts, right? So
[21:15] of these modular buildouts, right? So maybe there is not enough time to build
[21:16] maybe there is not enough time to build a brick and mortar data center, right?
[21:18] a brick and mortar data center, right? So obviously staying close to the edge
[21:21] So obviously staying close to the edge um poses different um kind of risk,
[21:24] um poses different um kind of risk, right? And this document that we have
[21:26] right? And this document that we have prepared uh uh this is again just as
[21:28] prepared uh uh this is again just as rich mentioned this is just a beginning
[21:30] rich mentioned this is just a beginning of uh you can call this a living
[21:32] of uh you can call this a living document and risk can be added on to
[21:35] document and risk can be added on to this as we go forward in time because we
[21:38] this as we go forward in time because we are coming up with a new technology we
[21:41] are coming up with a new technology we know that um the the AI models are
[21:44] know that um the the AI models are really taking everybody by surprise.
[21:46] really taking everybody by surprise. What used to be a model coming up in 12
[21:49] What used to be a model coming up in 12 to 18 months now literally coming up at
[21:51] to 18 months now literally coming up at 3 to 6 months and the size of these
[21:54] 3 to 6 months and the size of these datas is getting quadrupled in size you
[21:56] datas is getting quadrupled in size you know every 3 to 6 months which of course
[21:59] know every 3 to 6 months which of course is now posing uh you know rack density
[22:02] is now posing uh you know rack density issues is pushing to the limits. We are
[22:04] issues is pushing to the limits. We are now discussing really seriously about
[22:06] now discussing really seriously about the the DC voltage coming into the data
[22:08] the the DC voltage coming into the data center and most likely this is what the
[22:11] center and most likely this is what the future is going to hold with respect to
[22:14] future is going to hold with respect to uh the the power distribution and that
[22:16] uh the the power distribution and that itself poses a lot of challenge. Imagine
[22:19] itself poses a lot of challenge. Imagine 800 volt DC right next to liquid
[22:21] 800 volt DC right next to liquid cooling. I mean I can't even think about
[22:23] cooling. I mean I can't even think about it right. So really understanding the
[22:26] it right. So really understanding the risks beforehand uh is really important.
[22:28] risks beforehand uh is really important. It might save lives. it might save um uh
[22:32] It might save lives. it might save um uh that risk of getting electrocuted for
[22:34] that risk of getting electrocuted for example. So so this is is a model that
[22:37] example. So so this is is a model that we have um out um to the entire
[22:40] we have um out um to the entire community here. It's based off of a
[22:42] community here. It's based off of a real, you know, um, proven um, ISO
[22:47] real, you know, um, proven um, ISO framework which is really easy to use
[22:50] framework which is really easy to use and um, we we really would like you guys
[22:53] and um, we we really would like you guys to give it a shot. um and um really hope
[22:56] to give it a shot. um and um really hope that this can be used by many peers in
[22:59] that this can be used by many peers in the industry so we can take this forward
[23:01] the industry so we can take this forward and really apply this in the real
[23:03] and really apply this in the real situations cuz really um I have used
[23:05] situations cuz really um I have used this multiple times and I said it it
[23:07] this multiple times and I said it it really has helped us better understand
[23:10] really has helped us better understand how to design how to operate and at the
[23:12] how to design how to operate and at the end of the day you know uh we all are
[23:14] end of the day you know uh we all are looking at uh providing our customers
[23:17] looking at uh providing our customers you know a better ROI on their
[23:19] you know a better ROI on their investment and imagine you know a risk
[23:21] investment and imagine you know a risk which is not foreseen upfront can really
[23:25] which is not foreseen upfront can really cause downtime or other you know uh
[23:28] cause downtime or other you know uh mainstream issues right which can result
[23:30] mainstream issues right which can result in millions and billions of uh the you
[23:33] in millions and billions of uh the you know the the the damage with respect to
[23:35] know the the the damage with respect to uh uh a risk which may have been
[23:37] uh uh a risk which may have been proactively thought uh in the initial
[23:40] proactively thought uh in the initial stages even at the inception process. So
[23:43] stages even at the inception process. So um really encourage everybody uh to join
[23:46] um really encourage everybody uh to join um you know us in this uh moment forward
[23:48] um you know us in this uh moment forward right um and your suggestions are always
[23:51] right um and your suggestions are always welcome.
[23:53] welcome. Yeah, just to end what Imran says about
[23:56] Yeah, just to end what Imran says about the uniqueness of this model is that I
[23:58] the uniqueness of this model is that I mean I'm just thinking from while I'm
[24:00] mean I'm just thinking from while I'm wearing the head of the academic
[24:01] wearing the head of the academic researchers I mean know industry experts
[24:03] researchers I mean know industry experts have given their opinions but it's not
[24:06] have given their opinions but it's not only about this model it's just to sell
[24:08] only about this model it's just to sell as a product or it's just restricted as
[24:10] as a product or it's just restricted as we are using it at the open source
[24:12] we are using it at the open source platform and uh disseminating it to open
[24:15] platform and uh disseminating it to open compute projects. So there there is
[24:17] compute projects. So there there is currently if you talk about the
[24:19] currently if you talk about the literature as well currently there is no
[24:21] literature as well currently there is no such uh risk assessment framework that
[24:23] such uh risk assessment framework that is tailored to the data center cooling
[24:25] is tailored to the data center cooling there are few basically we have studied
[24:28] there are few basically we have studied over 50 research papers and over we
[24:31] over 50 research papers and over we analyze the industry's different
[24:32] analyze the industry's different services what type of major players are
[24:34] services what type of major players are in the data center industry if you if I
[24:37] in the data center industry if you if I talk about the current literature that
[24:38] talk about the current literature that exists we do have a risk assessment uh
[24:40] exists we do have a risk assessment uh framework for the construction we do
[24:42] framework for the construction we do have a risk data center risk assessment
[24:45] have a risk data center risk assessment framework for the security Particularly
[24:47] framework for the security Particularly we have it for locations
[24:50] we have it for locations focus on the locations as well. We have
[24:52] focus on the locations as well. We have for like some dedicated to particular
[24:54] for like some dedicated to particular sectors like pharmaceuticals you can say
[24:56] sectors like pharmaceuticals you can say or to the public sector as well. But as
[24:58] or to the public sector as well. But as a holistic approach that's the beauty of
[25:00] a holistic approach that's the beauty of this model is that it is the holistic
[25:02] this model is that it is the holistic and it is targeting the comparisons
[25:05] and it is targeting the comparisons between the liquid cooling as majority
[25:07] between the liquid cooling as majority of you have heard in the earlier in the
[25:09] of you have heard in the earlier in the keynote by the Nvidia that to build the
[25:11] keynote by the Nvidia that to build the AI foundation you need to think about
[25:13] AI foundation you need to think about the direct to chip liquid cooling and
[25:15] the direct to chip liquid cooling and they also giving some projections as
[25:17] they also giving some projections as well that by 2029 it's going to be 7
[25:19] well that by 2029 it's going to be 7 billion revenue in terms of the direct
[25:21] billion revenue in terms of the direct to chip liquid cooling. So this is what
[25:24] to chip liquid cooling. So this is what make the unique about this and this is
[25:26] make the unique about this and this is we give it is open to all of us. Please
[25:28] we give it is open to all of us. Please collaborate with us and expand this one
[25:31] collaborate with us and expand this one as much as we can and it can be more
[25:34] as much as we can and it can be more useful.
[25:34] useful. >> Yeah, absolutely. So uh next slide.
[25:40] >> Yeah, absolutely. So uh next slide. So, uh, what we'd like to ask is is that
[25:44] So, uh, what we'd like to ask is is that we coalesce around a common methodology
[25:47] we coalesce around a common methodology and we adopt the ISO standards that have
[25:50] and we adopt the ISO standards that have worked so well in other industries and
[25:52] worked so well in other industries and that we take this initial population as
[25:55] that we take this initial population as starter fuel. It's not intended to be
[25:57] starter fuel. It's not intended to be the one- all- beall for all of us. It's
[26:00] the one- all- beall for all of us. It's intended to be a starting place where
[26:02] intended to be a starting place where all of us can contribute and uh we'd
[26:05] all of us can contribute and uh we'd like to see the the risk roster grow
[26:08] like to see the the risk roster grow from 180 190 actually it's about 210 by
[26:11] from 180 190 actually it's about 210 by now
[26:12] now >> uh from 210 up to 300 we'd love to hear
[26:15] >> uh from 210 up to 300 we'd love to hear all the risks uh because there are edge
[26:17] all the risks uh because there are edge cases that we're soon uh as we deploy in
[26:21] cases that we're soon uh as we deploy in Riad as we deploy in Singapore as we
[26:23] Riad as we deploy in Singapore as we deploy in many other locations we learn
[26:26] deploy in many other locations we learn more and more and more about things
[26:28] more and more and more about things non-intuitive uh to somebody from
[26:31] non-intuitive uh to somebody from Seattle uh we have mold but not that
[26:33] Seattle uh we have mold but not that kind of mold and we don't have to do
[26:35] kind of mold and we don't have to do treatments like that.
[26:37] treatments like that. >> Um and these are interesting uh times
[26:40] >> Um and these are interesting uh times where we're all being called to deploy
[26:43] where we're all being called to deploy broader farther faster than and many of
[26:45] broader farther faster than and many of us uh feel in our comfort zone. So OCP
[26:49] us uh feel in our comfort zone. So OCP gives us a great place to standardize
[26:51] gives us a great place to standardize this and to work together in
[26:53] this and to work together in collaboration uh even across uh
[26:56] collaboration uh even across uh coopetition competitors directly working
[26:58] coopetition competitors directly working together. We all agree that a customer
[27:01] together. We all agree that a customer with less risk is more willing to invest
[27:03] with less risk is more willing to invest and so that's an easy place uh to place
[27:06] and so that's an easy place uh to place our bets together. Um, and ultimately
[27:09] our bets together. Um, and ultimately it's all about de-risking the AI edge
[27:12] it's all about de-risking the AI edge and the and the factory clusters uh so
[27:15] and the and the factory clusters uh so those deployments can work smoother,
[27:17] those deployments can work smoother, easier for everybody involved. Um, we
[27:20] easier for everybody involved. Um, we will we we don't have a specific working
[27:22] will we we don't have a specific working group within OCP, but we do believe this
[27:25] group within OCP, but we do believe this is part of the AI uh uh group that's
[27:28] is part of the AI uh uh group that's starting up through uh the AI uh uh uh
[27:33] starting up through uh the AI uh uh uh factory organization. It's uh it's
[27:36] factory organization. It's uh it's starting up right now actually I think
[27:38] starting up right now actually I think tomorrow. So we'll learn more about that
[27:40] tomorrow. So we'll learn more about that soon and we intend to join up in that
[27:42] soon and we intend to join up in that group uh as as an AI uh adjunct if you
[27:46] group uh as as an AI uh adjunct if you will. So at this point uh I'm just going
[27:48] will. So at this point uh I'm just going to open it up for a few questions to
[27:50] to open it up for a few questions to kind of have a more of a panel
[27:51] kind of have a more of a panel discussion then we'll open it up for
[27:53] discussion then we'll open it up for Q&A. So
[27:57] Q&A. So um I'll start with you.
[28:00] um I'll start with you. when you first learned about our crazy
[28:02] when you first learned about our crazy ideas about trying to assess this
[28:04] ideas about trying to assess this massive scope of risk uh tell me a
[28:08] massive scope of risk uh tell me a little bit about what went through your
[28:09] little bit about what went through your head like why did you see this of value
[28:13] head like why did you see this of value >> I think uh for me like we all talking
[28:16] >> I think uh for me like we all talking about the AI and high density AIdriven
[28:18] about the AI and high density AIdriven data centers as well so I think it's we
[28:21] data centers as well so I think it's we always talk about efficiency but we
[28:23] always talk about efficiency but we never thought about the risk as well so
[28:26] never thought about the risk as well so caring about the risk is more about uh
[28:28] caring about the risk is more about uh means designing thesis that are not only
[28:30] means designing thesis that are not only efficient but we need to think about
[28:33] efficient but we need to think about that they are resilient they are
[28:34] that they are resilient they are predictable and obviously they fit for
[28:37] predictable and obviously they fit for demands for the AI specific as well so
[28:39] demands for the AI specific as well so it's not only about the economic impact
[28:42] it's not only about the economic impact or the as Immran already touch upon the
[28:44] or the as Immran already touch upon the TCOS only we need to think about the
[28:47] TCOS only we need to think about the TTRO's kind of thing so it's not only
[28:49] TTRO's kind of thing so it's not only about the efficiency sustainability as
[28:51] about the efficiency sustainability as well we need to think about the risk as
[28:52] well we need to think about the risk as well
[28:55] well >> and Iran tell me a little you have a
[28:58] >> and Iran tell me a little you have a interesting perspective because you've
[29:00] interesting perspective because you've both
[29:00] both >> bought, sold, deployed,
[29:04] >> bought, sold, deployed, >> decommissioned, commissioned. I mean,
[29:06] >> decommissioned, commissioned. I mean, you you've done literally the whole life
[29:07] you you've done literally the whole life cycle throughout your career. Tell me a
[29:10] cycle throughout your career. Tell me a little bit about how you view the what
[29:13] little bit about how you view the what what is the premium of having this kind
[29:16] what is the premium of having this kind of risk model in advance of investment.
[29:19] of risk model in advance of investment. >> I think this is essential and I think I
[29:21] >> I think this is essential and I think I call this being inevitable at this
[29:23] call this being inevitable at this point. I mean imagine
[29:26] point. I mean imagine um so so a cooling vendor for example
[29:28] um so so a cooling vendor for example just like where I am right now I mean so
[29:31] just like where I am right now I mean so a cooling vendor shouldn't just be
[29:32] a cooling vendor shouldn't just be selling a hardware it should be helping
[29:35] selling a hardware it should be helping the customer derrisk their AI factories
[29:38] the customer derrisk their AI factories right so which is a function of a good
[29:41] right so which is a function of a good product right and really working with
[29:43] product right and really working with the uh designers and the the owners
[29:45] the uh designers and the the owners right or developers to help them
[29:48] right or developers to help them understand how that product can derisk
[29:50] understand how that product can derisk that AI factory cuz we know we are
[29:53] that AI factory cuz we know we are talking about time to scale, time to
[29:55] talking about time to scale, time to revenue as as I said before, right? So,
[29:57] revenue as as I said before, right? So, this is where there is a there's an
[30:00] this is where there is a there's an opportunity where something might go
[30:01] opportunity where something might go through the cracks cuz we know everybody
[30:04] through the cracks cuz we know everybody is talking speed, speed, speed, right?
[30:06] is talking speed, speed, speed, right? Speed is the name of the game and how
[30:08] Speed is the name of the game and how fast these AI factories needs to be
[30:10] fast these AI factories needs to be built. So, imagine having a holistic
[30:14] built. So, imagine having a holistic model like this which is already having
[30:17] model like this which is already having pretty much a good coverage of the risk,
[30:19] pretty much a good coverage of the risk, right? Um obviously there's always room
[30:22] right? Um obviously there's always room to add more risk as we learn. Uh and in
[30:24] to add more risk as we learn. Uh and in in fact if this becomes uh you know a
[30:27] in fact if this becomes uh you know a work stream where people from all over
[30:29] work stream where people from all over the globe are going to put their
[30:30] the globe are going to put their perspective on oh by the way you didn't
[30:32] perspective on oh by the way you didn't think about you know certain risk which
[30:34] think about you know certain risk which which the original um you know uh form
[30:37] which the original um you know uh form did not have for example because it only
[30:39] did not have for example because it only existed in certain geographic for
[30:41] existed in certain geographic for example. So really understanding uh and
[30:44] example. So really understanding uh and using this tool can really be very
[30:46] using this tool can really be very helpful for all aspect of the data
[30:49] helpful for all aspect of the data center at every intersection. I see this
[30:51] center at every intersection. I see this being used by the designers by the
[30:53] being used by the designers by the operators right. So it is something
[30:56] operators right. So it is something which as I said is is is a layer above
[31:00] which as I said is is is a layer above the TCO which is capex and opex. So
[31:03] the TCO which is capex and opex. So really using this which ultimately will
[31:06] really using this which ultimately will result in better efficiency, lower TCO
[31:09] result in better efficiency, lower TCO and hopefully more tokens per job. So
[31:12] and hopefully more tokens per job. So this is how I see this um value added to
[31:15] this is how I see this um value added to um any stakeholder in this industry.
[31:18] um any stakeholder in this industry. >> Yeah. And I'll just uh with that launch
[31:20] >> Yeah. And I'll just uh with that launch into a shameless plug for OCP's TCO
[31:22] into a shameless plug for OCP's TCO group, which I encourage you to go check
[31:25] group, which I encourage you to go check out as well. That's a foundational
[31:26] out as well. That's a foundational element. uh in many ways we kind of
[31:29] element. uh in many ways we kind of assumed but didn't weren't that explicit
[31:31] assumed but didn't weren't that explicit here that your total cost of ownership
[31:34] here that your total cost of ownership is a model you probably should run
[31:36] is a model you probably should run already in addition to a TTRO model as
[31:39] already in addition to a TTRO model as well and the two can complement each
[31:41] well and the two can complement each other and suggest to each other uh uh
[31:45] other and suggest to each other uh uh better quality uh and and and we but we
[31:48] better quality uh and and and we but we we know that with having these models
[31:50] we know that with having these models people can make better decisions in
[31:52] people can make better decisions in advance and even when they make a
[31:54] advance and even when they make a decision that might be a little risky at
[31:56] decision that might be a little risky at least They know everyone was was at the
[31:59] least They know everyone was was at the same level and making that same and
[32:01] same level and making that same and accepting that same level of risk at the
[32:02] accepting that same level of risk at the same time.
[32:03] same time. >> Yeah. And I also say adding to that rich
[32:05] >> Yeah. And I also say adding to that rich I mean anybody who's been burned down by
[32:08] I mean anybody who's been burned down by you know some data center event you know
[32:10] you know some data center event you know they will really appreciate having this
[32:12] they will really appreciate having this proactiveness uh using this model right
[32:15] proactiveness uh using this model right as as we know not everything can be
[32:16] as as we know not everything can be really u you know u thought when he has
[32:19] really u you know u thought when he has a crystal ball like I said but it's
[32:21] a crystal ball like I said but it's really important to have this unified
[32:24] really important to have this unified model which will be a product of the
[32:27] model which will be a product of the intelligence or the input from all over
[32:29] intelligence or the input from all over the globe. This is really what our hope
[32:31] the globe. This is really what our hope is that this becomes a model which will
[32:34] is that this becomes a model which will have risk that um would be you know
[32:37] have risk that um would be you know added by many data center experts across
[32:39] added by many data center experts across the entire intersection of this data
[32:41] the entire intersection of this data center ecos and imagine how helpful this
[32:44] center ecos and imagine how helpful this could be to a data center developer or
[32:47] could be to a data center developer or somebody who is looking to fund a data
[32:50] somebody who is looking to fund a data center right they they can see uh you
[32:52] center right they they can see uh you know the risk from a different
[32:53] know the risk from a different perspective and they're they're you know
[32:56] perspective and they're they're you know they're they're might change or they may
[32:59] they're they're might change or they may just doing something even bigger. So
[33:01] just doing something even bigger. So really um I see this as a great tool for
[33:05] really um I see this as a great tool for everybody in this system who could use
[33:07] everybody in this system who could use it to their advantage.
[33:11] >> Okay, we have uh time for just one or
[33:13] >> Okay, we have uh time for just one or two questions.
[33:18] I
[33:19] I >> think we have two.
[33:20] >> think we have two. >> Hi, my name Kawashima. I'm from Entity a
[33:24] >> Hi, my name Kawashima. I'm from Entity a service provider in Japan and this is my
[33:26] service provider in Japan and this is my first time to know the concept tier
[33:28] first time to know the concept tier role. I'm very excited about this
[33:30] role. I'm very excited about this concept because for two reasons one is
[33:33] concept because for two reasons one is this will be very important concept when
[33:35] this will be very important concept when we propose a solution for uh sovereign
[33:38] we propose a solution for uh sovereign AI ar in
[33:41] AI ar in the second reason is uh this concept is
[33:44] the second reason is uh this concept is very important to express the inherent
[33:47] very important to express the inherent value of open architecture and in tico
[33:51] value of open architecture and in tico world we used to say that the multi-
[33:53] world we used to say that the multi- vendor interoperability is very
[33:55] vendor interoperability is very important because Telico operator should
[33:58] important because Telico operator should not rely rely on a single oper single
[34:01] not rely rely on a single oper single vendor. Bender locking is a risk but in
[34:04] vendor. Bender locking is a risk but in reality many service providers rely on
[34:06] reality many service providers rely on Cisco products because
[34:09] Cisco products because vendor locking is one risk but also if
[34:13] vendor locking is one risk but also if we try to integrate products from
[34:15] we try to integrate products from multiple suppliers then maintaining
[34:19] multiple suppliers then maintaining infrastructure engineers in house is
[34:22] infrastructure engineers in house is very difficult. That's another risk. So
[34:24] very difficult. That's another risk. So the the difficulty of this type of
[34:26] the the difficulty of this type of concept is how to comp compare several
[34:30] concept is how to comp compare several types of leaks. So how would you score
[34:34] types of leaks. So how would you score each risk item
[34:37] each risk item quantitatively
[34:39] quantitatively and in the fair way?
[34:41] and in the fair way? >> It's a big challenge. Obviously you'd
[34:43] >> It's a big challenge. Obviously you'd love most most customers would
[34:46] love most most customers would appreciate one throat to choke is the
[34:48] appreciate one throat to choke is the horrible American term that we use. uh
[34:51] horrible American term that we use. uh but they want single point of
[34:53] but they want single point of accountability. Uh on the other side you
[34:56] accountability. Uh on the other side you want competition to drive innovation to
[35:00] want competition to drive innovation to drive cost competitiveness to drive
[35:01] drive cost competitiveness to drive better service. So there's this
[35:05] better service. So there's this continuum and you as a manager a
[35:07] continuum and you as a manager a decision maker in whatever corner of
[35:10] decision maker in whatever corner of this industry you're working in have to
[35:12] this industry you're working in have to draw that line and you have to defend
[35:13] draw that line and you have to defend that line. So I would suggest to you
[35:15] that line. So I would suggest to you that you will use this model to help
[35:18] that you will use this model to help illustrate the the map on how you drew
[35:22] illustrate the the map on how you drew the line. It's not so much for us to
[35:24] the line. It's not so much for us to tell you. We're just setting the tape.
[35:26] tell you. We're just setting the tape. >> I understand it.
[35:27] >> I understand it. >> Yeah.
[35:27] >> Yeah. >> Okay.
[35:27] >> Okay. >> And we look forward to working with you.
[35:29] >> And we look forward to working with you. Thank you.
[35:29] Thank you. >> Yeah. Thank you very much.
[35:30] >> Yeah. Thank you very much. >> Yeah.
[35:35] >> Okay. Hi, I'm Andy Young. Um I developed
[35:38] >> Okay. Hi, I'm Andy Young. Um I developed uh with a co-lead the TCR model that
[35:41] uh with a co-lead the TCR model that Rich was referring to. So thanks Rich
[35:43] Rich was referring to. So thanks Rich for the shout out. Uh my question is uh
[35:46] for the shout out. Uh my question is uh mainly based on the results that I'm
[35:49] mainly based on the results that I'm seeing and the importance of cooling and
[35:51] seeing and the importance of cooling and climate
[35:53] climate um how do we level that against the
[35:55] um how do we level that against the known risks around power continuity
[35:58] known risks around power continuity so why do we which I guess was in the
[36:01] so why do we which I guess was in the electrical score quite a way down the
[36:04] electrical score quite a way down the table why so different and I can't help
[36:08] table why so different and I can't help but think uptime institute have this
[36:10] but think uptime institute have this term continuous cooling which is a
[36:13] term continuous cooling which is a little and the definitions and the
[36:15] little and the definitions and the actions around that is are quite
[36:17] actions around that is are quite ambiguous and therefore how can we use
[36:20] ambiguous and therefore how can we use the insights from this tool to make a
[36:24] the insights from this tool to make a more clear definition of continuity of
[36:27] more clear definition of continuity of cooling which I only presume we've
[36:29] cooling which I only presume we've achieved in the industry around
[36:31] achieved in the industry around continuity of power
[36:33] continuity of power um or is it so case dependent based on
[36:38] um or is it so case dependent based on local climate based on local um
[36:40] local climate based on local um structures of data center that you can't
[36:44] structures of data center that you can't make continuity of cooling guidance any
[36:47] make continuity of cooling guidance any clearer. So why is power so low and
[36:51] clearer. So why is power so low and cooling so high? I accept that cooling
[36:53] cooling so high? I accept that cooling is a high risk but do you think that
[36:55] is a high risk but do you think that this tool can give us greater insights
[36:57] this tool can give us greater insights to make clear what action you take to
[37:00] to make clear what action you take to derisk continuity of cooling?
[37:02] derisk continuity of cooling? >> That's a great question and I think both
[37:04] >> That's a great question and I think both power and cooling go hands in hand.
[37:06] power and cooling go hands in hand. Obviously no power you don't need
[37:08] Obviously no power you don't need cooling, right? So I think the the the
[37:12] cooling, right? So I think the the the whole idea behind this tool is for let's
[37:16] whole idea behind this tool is for let's say if you're building a data center for
[37:17] say if you're building a data center for example right every site as you said is
[37:20] example right every site as you said is going to be different right based on the
[37:22] going to be different right based on the weather demographics availability of
[37:24] weather demographics availability of power maybe the the budget available for
[37:27] power maybe the the budget available for providing certain kind of redundancy in
[37:29] providing certain kind of redundancy in the system for power and cooling for
[37:31] the system for power and cooling for example right so so I think it's it's a
[37:33] example right so so I think it's it's a function of how deep you want to go on
[37:35] function of how deep you want to go on the design and that can really determine
[37:38] the design and that can really determine the risk that will be involved. For
[37:40] the risk that will be involved. For example, and this is why good thing is
[37:42] example, and this is why good thing is that this this tool has a rating based
[37:44] that this this tool has a rating based on the severity and imagine if you're
[37:47] on the severity and imagine if you're constrained by by the investment or the
[37:49] constrained by by the investment or the availability of power. So that is going
[37:51] availability of power. So that is going to have um your score based on that and
[37:55] to have um your score based on that and at the end of the analysis you can come
[37:57] at the end of the analysis you can come up with the score and and really know
[38:00] up with the score and and really know the severity of that risk. So for
[38:02] the severity of that risk. So for example if if there is funding available
[38:05] example if if there is funding available right for example or funding is made
[38:07] right for example or funding is made available and we know that we can add
[38:09] available and we know that we can add more redundancy in power for example
[38:12] more redundancy in power for example which in turn is going to enable
[38:14] which in turn is going to enable continuous cooling as you said right
[38:16] continuous cooling as you said right which is a function of maybe having n
[38:18] which is a function of maybe having n plus1 cdus or chillers for example so I
[38:21] plus1 cdus or chillers for example so I think it's all dependent on the entire
[38:24] think it's all dependent on the entire requirement uh for a particular AI
[38:26] requirement uh for a particular AI factory or data center which is then
[38:29] factory or data center which is then going to be um you know um this tool can
[38:32] going to be um you know um this tool can really use understand the entire um
[38:35] really use understand the entire um requirement set and be able to gauge the
[38:38] requirement set and be able to gauge the risk rate and see what comes up with um
[38:41] risk rate and see what comes up with um the ultimate um outcome of that.
[38:45] the ultimate um outcome of that. >> Yeah, I think uh just to follow up on
[38:47] >> Yeah, I think uh just to follow up on that uh very very briefly I think
[38:50] that uh very very briefly I think there's probably something in the
[38:51] there's probably something in the distribution of power and the
[38:53] distribution of power and the distribution of cooling and the risk
[38:55] distribution of cooling and the risk inherent in that network.
[38:56] inherent in that network. >> Yeah.
[38:56] >> Yeah. >> That is probably uh pivotal to it. But
[39:00] >> That is probably uh pivotal to it. But it'd be good to talk uh more offline
[39:02] it'd be good to talk uh more offline about this.
[39:02] about this. >> Happy to talk after.
[39:03] >> Happy to talk after. >> Thank you very much.
[39:04] >> Thank you very much. >> Thank you.
[39:05] >> Thank you. >> I want to thank OCP for uh giving us
[39:07] >> I want to thank OCP for uh giving us this opportunity on behalf of my uh four
[39:09] this opportunity on behalf of my uh four co-authors and three presenters. Uh
[39:11] co-authors and three presenters. Uh we've appreciated this and and we look
[39:13] we've appreciated this and and we look forward to working with all of you. Uh
[39:15] forward to working with all of you. Uh we'll be writing back uh or online uh
[39:18] we'll be writing back uh or online uh until next time. Thank you.
[39:20] until next time. Thank you. >> Thank you all.