ストレージのスループットがAIワークフローの成否を左右する――Everpure Cameron Bahar氏に聞く

How Storage Throughput Determines AI Workflow Outcomes - with Cameron Bahar of Everpure

The AI in Business Podcast25分
ストレージのスループットがAIワークフローの成否を左右する――Everpure Cameron Bahar氏に聞く

ストレージのスループットがAIワークフローの成否を左右する――Everpure Cameron Bahar氏に聞く

The AI in Business Podcast

0:0025:08

要約

Everpure(文字起こし上の表記は Everpeer)で Scale AI の GM を務める Cameron Bahar 氏が、AI インフラを数ラックから数百ラックへ拡大する際の運用課題を語った。ケーブルやスイッチの故障、SRE の負担など、従来のエンタープライズ向けストレージが想定していない問題が表面化すると指摘する。テレメトリと自動化、電力・スペース・接続性の事前計画、ベンダーへの依存が重要だと述べている。

この回のポイント

  • ●数 TB/秒のスループットを出すには多数のシステムを並列にネットワークでつなぐ必要があり、大規模クラスタ化が避けられない。大規模環境を模したテストは難しく、スケール時にこそ想定外の問題が出ると Bahar 氏は述べる。
  • ●ハイパースケーラーではチームの5〜7割が運用・SRE(KTLO)に回る一方、従来のエンタープライズは管理者1人で5年動く世界だった。このギャップが課題だという。
  • ●学習ジョブは1ノード故障で全体が止まるため、ケーブルの半挿しなどを検知するテレメトリが重要。AI を使った SRE 的な監視にも言及した。
  • ●スケールでの運用には、計算からネットワーク、ストレージまでの可視化、自動修復、マルチテナントのガバナンスとトレーサビリティ、需要予測に基づく先行増強が必要だとされる。
  • ●GPU を1つの建物に入れ切り、後からストレージを足せず GPU を撤去した例など、データを後回しにした計画の失敗が紹介された。
  • ●企業は AI プラットフォーム部門を作り、小規模実験で利用パターンを把握しつつ、スケールと簡便さを解決済みのベンダーに頼るべきだとの見方が示された。

章立て

  1. なぜストレージのスケールが重要か

    AI ワークロードの容量・性能要件により、1ラックから数百ラックへの拡大とグローバル名前空間が必要になる。大規模環境のテストが難しく、運用体制が問われると説明する。

  2. 運用負荷と自動化の必要性

    ハイパースケーラーの SRE 負担と従来型企業の運用の違い、ファームウェア差異など見えない問題への対処として、ベンダーが自動化と可視化を提供する責任を語る。

  3. テレメトリとケーブル故障

    1万 GPU 規模ではケーブルが故障点となり、1ノード停止でジョブ全体が止まる。AI で多数のシステムを監視し診断時間を短縮する考えを示す。

  4. スケール運用の条件

    可視化、自動修復、マルチテナント、トレーサビリティ、需要予測に基づく先行増強など、スケール運用に必要な要素を整理する。

  5. 顧客事例:建物・電力・接続の拡張

    1万 GPU から5万 GPU への拡大で建物が足りなくなる例や、自動運転・金融・創薬企業の成長パターン、xAI の事例に触れる。

  6. データを後回しにしない計画

    ストレージを後付けして電力や空間が足りず GPU を外した事例を挙げ、IT 内の AI プラットフォーム組織や多職種での計画の重要性を議論する。

  7. ベンダー活用と結び

    サービスとしてのストレージと運用負荷の低減を強調し、AI ネイティブでない企業はベンダーに頼る必要があると述べる。最後に3つの要点で締める。

解説記事

Everpure で Scale AI 部門の GM を務める Cameron Bahar 氏が、The AI in Business Podcast に再登場した。テーマは、AI インフラを数ラックから数百ラックへ拡大するときに、ストレージと運用で何が起きるかである。なお文字起こしでは社名が「Everpeer」と表記されているが、番組情報に合わせて Everpure とした。

規模が壊すもの

Bahar 氏はまず、AI ワークロードが求める容量と性能が、データとメタデータの両面で既存のエコシステムを壊していると述べる。数 TB/秒のスループットを出すには、多数のシステムを汎用ネットワーク上で並列に動かさなければならず、クラスタは必然的に大きくなる。しかもマルチテナントで、利用者は科学者、推論の担当者、エージェント型ワークロードの運用者など多様だ。そのためクラウド規模のアーキテクチャと、テキスト・動画・音声を含む全データを見渡せる単一のグローバル名前空間が必要になるという。

課題は、そうした規模を模擬するテスト環境を用意しにくい点にある。ハイパースケーラーでも、大規模リージョンに入って初めてソフトウェアの問題が表面化することが多いと同氏は説明する。

運用負荷という見えにくいコスト

ハイパースケーラーではチームの50〜70%が運用と SRE(Site Reliability Engineering、サイト信頼性エンジニアリング)に取られ、新機能開発は少数派になる。これを同氏は KTLO(keep the lights on)と呼ぶ。一方、従来のエンタープライズでは、機器をつなげば管理者1人で5年動くのが普通だった。

50ラック規模で RDMA ネットワークを使い、他社製のスイッチやクライアントが混在する環境では、同じ運用は成り立たない。そこで同氏は、ベンダーが自動化とテレメトリ(稼働状況の計測データ)を提供し、顧客の運用負担を減らす責任があると主張する。自社で運用チームを抱えられるのは一部の企業に限られ、AI が企業全体に広がるにはシンプルで拡張しやすい仕組みが要るという論理である。

ケーブル1本が止めるジョブ

具体例として挙がったのがケーブルだ。1万 GPU をつなぐと、ケーブルやトランシーバーが故障点になる。学習ジョブでは1ノードが落ちると全体が止まるため、半挿しのケーブルのような不具合をいち早く検知し、診断までの時間を縮める必要がある。同氏は、人間が千台のシステムを監視するのは難しく、AI エージェントを運用ガイドとして使う手があると述べている。聞き手の Dan 氏は、製造ラインの途中の機械が止まれば全体が止まるという例えで、予知保全と計画外停止の痛みに結びつけた。

スケールでの運用の条件として、同氏は次を挙げた。

  • 計算からネットワーク、ストレージまでの可視化
  • 問題を直す自動化
  • マルチテナントでのガバナンスとログ、エージェントが暴走した際の追跡能力
  • 6か月先、1年先を見据えたサプライチェーン管理と需要予測

建物・電力・データの計画

顧客事例では、トップ25クラスの AI 企業が1つの建物に1万 GPU を置いて稼働させ、3〜6か月後に5万 GPU を求めてくる例が紹介された。建物が満杯なら隣に別棟を借り、棟間を光ファイバースイッチでつなぐことになる。つまり成長は事前に計画されたものではなく、データセンターの空間、電力、接続性が基本的な構成要素になる。自動運転、金融トレーディング、創薬のスタートアップも同様に、控えめに始めて指数的に拡大するという。GPU は互いに遠く離せないので、場当たり的な増設は避けたいと述べている。

もう一つの教訓は、データが後回しにされがちなことだ。GPU を先に買って建物に入れ、あとからストレージが必要だと気づくと、電力も空間も足りない。実際に、ストレージを入れるため GPU を建物から外した例を見てきたと同氏は語る。

企業はどう備えるか

エンタープライズの GPU ファームはまだ少なく、多くはハイパースケーラーや NeoCloud を使っていると同氏は見ている。企業側には、IT 内に AI プラットフォームの組織を置き、リファレンスデザインを持ち、少人数の実験で利用パターンを把握することを勧める。大規模 LLM 事業者のように優秀な人材で手作りするのは現実的でなく、スケールと簡便さを解決済みのベンダーに頼るべきだという。スイスアーミーナイフを渡されて自分で加工するようなやり方は避けたい、という比喩も出た。

まとめ

番組の結論は3点に整理されている。第一に、規模が数百ラックに達すると、障害は稀な出来事ではなく日常の運用課題になる。第二に、ケーブル1本の状態まで把握するテレメトリと自動化が不可欠になる。第三に、データとストレージを最初から計画に組み込む必要がある。

日本の読者にとっての示唆は、GPU の調達量だけでなく、ストレージ、電力、スペース、運用体制を一体で設計する視点の重要性だ。自社に SRE 部隊を持てない企業ほど、可視化と自動化を備えたベンダーの選定が成否を分けることになる。ただし本回は Everpure のスポンサー回であり、ベンダー依存を勧める主張には立場がある点も念頭に置きたい。

文字起こし(英語・自動生成)

Welcome, everyone, to the Emerge AI in Business podcast. Today's guest is Cameron Bahar, GM of Scale AI at Everpeer. Everpeer provides enterprise data storage and management software that consolidates block, file, and object workloads across on-prem and cloud environments. The platform focuses on efficient data reduction, consistent performance at scale, and unified operations across major cloud providers. In today's episode, Cameron discusses why operating AI infrastructure at scale changed storage systems in ways legacy enterprise deployments never face. From networking and cabling to the growing operations burden of keeping thousands of GPUs running. He outlines how visibility, automation, and forward planning for power, space, and connectivity help leaders avoid the unplanned downtime and costly rework that come with rapid growth.

Today's episode is sponsored by Everpeer. In this episode, we cover what it takes to operate AI storage and data infrastructure at scale. To go deeper on this topic and learn how consultants are winning business with evidence-based AI ROI and build long-term capabilities instead of chasing short-term gains, Download our pre-PDF report, Three Keys to Thriving in the Coming Era of Automation at Emerge.com.com. That's Emerge.com.com to download your copy. Now the conversation with Cameron. So Cameron, welcome back to the show. It's great to be back, Dan. Thank you. Totally. We're speaking today about something we touched on very briefly the last time we chatted, which is the importance of scale. So much growth in sort of compute use. Data and storage architecture sort of had to level up with that.

I want to get a sense from your vantage point about why scale is so important. Some companies are going through it right now. They already get it. Some are maybe feeling the rustles in the breeze, but they're not there quite yet. Why is scale so important for storage? I think, you know, as we talked last time, the size and performance requirements of AI workloads, you know, the capacity and performance both on data and metadata are breaking ecosystems. And to support serving, you know, multiple terabytes per second of throughput requires many, many systems to work in parallel over a general network. And so you end up having to scale to very large cluster sizes, if you will, to run a particular workload. And as we mentioned, they're multi-tenant. They have lots of different workloads on them. And the provider of these AI platforms isn't usually the customer. It's usually a scientist or somebody doing inference or somebody running agentic workloads.

So as we mentioned last time, it has to look like a cloud-scale architecture. And what happens is when you go from one rack, which is a typical enterprise deployment, to 10 racks, 20 racks, 50 racks, hundreds of racks, and needing to see a single system image, like a global namespace with all the data, whether it's text, video, audio, we talked about multimodal last time. All of that has to be available, right, to training, inference, and agentic workloads. And as you scale, there's networking, there's top-of-rack switches, there is operational issues where things fail. And what you find is that it's very difficult to have a test environment that mimics such a large size. So when you're buying a storage solution, it's often not tested at that scale in an environment that you have. So imagine if you're at a hyperscaler of choice. When you check in some new code into production, those environments are effectively your testing environment.

You'll do some testing as best as you can, but most of the issues in the cloud are when you go into large regions where the scale ends up breaking the software. So while the systems are small, it's not too complicated to test and fully have high quality. but when you go into large environments, the scale ends up causing unforeseen issues. So then it leads you to have an ops, an SRE, we call it SRE, Site Reliability Engineering. You know, most of the hyperscalers have that. So for people who are deploying these AI platforms that are increasingly looking like a neocloud or a mini cloud, do you have that muscle to operate? because in the hyperscalers, given a team, often 50%, 60%, 70% of the team is working on ops and SRE just to keep the systems running. A minority is working on features and new things. We call it KTLO, keep the lights on. That's a big team.

And when you look at the enterprise, usually you plug your system in and you have one admin managing it and it just works for five years. That's just not the case here when you scale a 50 racks and you're running RDMA networks and it's somebody else's switch, somebody else's client, somebody else's data node? How do you make that complex system work well? So that's sort of what I mean when I say the challenge is to run at scale, is how do you provide a system that is automated enough that takes the ops SRE burden off of the customer and gives them enough telemetry and visibility into the system so they can manage it with as few people as possible? You're bringing up also this notion that much of what you run into in the real world is going to turn into a hurdle when you start to achieve scale, things that you wouldn't have seen beforehand. and be like, well, I don't know, is scale that big of an issue? And then all of a sudden, you double or you double three times, and all of a sudden things are massive issues. You're seeing some companies that are extremely sophisticated,

who probably still run into bottlenecks, but they sort of a little bit ready for what the next five doublings look like because they done 50 before And then you seeing other companies kind of leap in Maybe they in financial services Maybe they doing computer vision whatever the case may be who aren't quite as prepared for what hurdles are going to naturally emerge when scale happens. And they may be a little bit naive to that. Is there any kind of way to get people prepared to think through logically? Sort of, hey, where do you want this to be? What are the bottlenecks you're going to run up to? or do people end up having to learn the hard way? What's been your experience there with that element? Well, typically people who launch hyperscalers or neoclouds end up rolling their own ops and SRE teams and building automation and tools to manage this complexity. And, you know, I think, you know, if you look for AI to permeate through the enterprise and grow as we all expect it to grow, right, that system doesn't scale because that's for a few players who can hire engineering teams and invest massive amounts into ops and SRE.

So the vendors have a responsibility to design systems that are simple and scalable and provide visibility and all the tools and automation. Also using AI themselves to make these systems easier to wrestle with, if you will, at scale. It's often what you don't know that's hurting you at scale. One cord has the wrong firmware version. It's one release behind or it hasn't been updated or that didn't have a security patch. So, you know, having enough automation and visibility and telemetry helps will help customers overcome some of these scaling challenges. Yeah, you were talking about even using AI to make this stuff easier to use. And you were just hinting at a few of those right now. What have you seen as the components of telemetry that have made scaling less painful, let's say? What if seeing what customers actually use and what they value to help navigate and be nimble about their decision making? What's been helpful for them? I think this is useful to know. Imagine when you connect lots of GPUs to lots of storage.

One example is, are all the cables plugged in to the right ports? Are transceivers all working? Do you have telemetry to tell you if a cable is bad, right? You know, when you have 10,000 GPUs and you're connecting all these systems together, cabling becomes a failure point. How do you get telemetry to tell you if a cable is halfway plugged in and it's firing? Because often when these GPU workloads, when one node fails, the whole job stops. It's not like you have 1,000 nodes and one of them fails, the other 999 just keep going, right? If you're doing training, for example, and one GPU node goes down, the whole job stops. And so it becomes more extensive. And so the sooner you can resolve, figure out what the problem is, and diagnose a time to diagnosis, often we're using AI to do SRE type of things, which is use AI like an operations guide just to look at a thousand systems.

It's hard for humans to monitor a thousand systems, but maybe AI agents can peel into that a bit easier, right? It's that sort of thing that once you've got your file and object or whatever storage fabric and the protocol in place, as you scale to a thousand node cluster and running an exabyte, you'll run into these edge cases. and figuring out what that problem is quickly and having enough automation to do it will be a sort of a game changer because of the OPEX costs of operating side systems at scale. When you bring up telemetry, you bring up cabling, you're bringing up these challenges of, to be frank, it's funny because you'd mentioned one element goes down, all of a sudden your whole workflow is going to stop. The simplest business analogy here is in manufacturing, right? Your third machine of seven goes down. Well, you don't skip it, right? The whole workflow is going to stop. You're backing things up, customers, inventory, everything.

And of course, there's an equivalent in the digital world that you're bringing up right now. And what's curious for them is they've got to use telemetry to figure out predictive maintenance. When do I need to shut these things down? When do I need to check in and replace parts and whatever? Because unplanned downtime, which is what you're talking about in that space is increasingly painful. So basically, we're seeing the same kinds of things in this digital world that you're articulating. A hundred percent. Yeah, it's like if an assembly line goes down and you queue up, queuing through, he gets you, all the jobs back up. That's a simple analogy. I almost wouldn't think it would hold, but the way you were explaining it was like, wow, okay, that's the audience that's going to be able to understand exactly what you're saying there. So, okay, so we're getting into some of what scale looks like. Now we've got some mental models for how consequential this can be. The idea of telemetry tying in here is pretty interesting as well. I want to talk about what operating at scale looks like. There's companies that are scaling pretty well and I think have a feel for it, some that are starting to scale, some that aren't quite there yet. When people are truly operating at scale, what has to be true?

How do you think about that? Yeah, that's a great question. Operating at scale needs new muscles. And really it's about being able to see what's going on in your data center from compute all the way through the network to storage. So having visibility and then having enough automation to be able to remediate issues. So it's a similar angle about ops and SRE. As you launch these services, right, understanding the workload so that you right-size the problem and you make sure you have enough infrastructure to support that user base, as well as basic capabilities in the system that allow you to have a multi-tenant system so you could offer these services to different tenants in a secure way that's governed and logged so that you can have accountability on what happened there.

If an agent goes rogue, you'd be able to trace. So tracing ability. So it being able to run and have a nice dashboard if you will that AI enabled so that you could actually see what the data flows look like what multi look like and where in the system you're seeing pressure points. And then having an architecture underneath that allows you to scale those points ahead of demand. There's a bit of supply chain management and forecasting that has to go on when you're running these large systems to figure out what do I need six months from now, What do I need one year from now? Then you have to stay ahead of that curve and make sure you've got enough capacity lined up in the right places. You could offer this effectively token service to your users. A lot of the large LLM providers figure out how to do this stuff on their own because they're building large models. And they hire lots of smart people from leading companies who handcraft these things.

When you go to the enterprise, you need a repeatable, simple model. right and and so these systems have to become simple observable and easily expandable yeah let me try to nutshell this Cameron on the one hand we've got to have an architecture that we understand well enough and kind of know how to modually you know expand when and how we've got to so we're not really operating at scale unless we have a nimbleness to expand capacity and to know how that works so there's a operational sure it's like do you got room to grow do you know what growth would look like. There's an element of that. But then there's also the cockpit of what are we looking at to make those forecast decisions? What are we looking at to understand performance, energy use, cost, whatever the case may be? What are we looking at to know how well this is servicing our customers, et cetera? So once you've got your dashboard fully understood, and once you're standing on solid ground where you know how to open it up and turn it to something bigger and bigger as necessary, maybe now somebody could say you may be operating at scale once you

sort of playing off at that level. Is this somewhat safe to say? I think so. Yeah, I think that's right. You need to be able to have a predictable environment and be able to plan ahead so that you're not caught flat-footed when a workload runs away from you. You certainly will run away with us as AI is given the progress and pace here of AI. With that said, we're going to make this kind of tangible. I want to click into operating at scale. You've given us principles. Presumably, you get to see this at some of the biggest companies in the world. You guys work with the Hyperscalers, you work with the Legacy Enterprise, you work with different industries. Are there some customer examples of what scale looks like in practice that you could just touch on for people to see this in their mind? Sure. I've seen in my career some top 25 AI companies doing a lot of training and now doing a lot of inference. And what happens is you start with a building and maybe you put 10,000 GPUs in it and they start running, and they run into all these issues we talk about.

And then literally three to six months later, they come back and ask for, hey, I need to go to 50,000 GPUs. Well, that building is full. So you're going to talk to your real estate people and find another building next to it. And then now you've got 50,000 GPUs across two buildings, and you have an interesting fiber optic switch between them to send data across. So the world ends up growing like that. It's not fully planned up front for end GPUs. It grows. So that data center space, the power and the connectivity becomes sort of the infrastructure building blocks. And we've seen this with startups doing, you know, self-driving. We've seen it with financial trading companies. We've seen it with drug protein folding companies doing drug discovery. They often start with modest ambitions and then they grow exponentially. and if you're operating in that environment, you need to have a good handle on your power and space and growth

so that you don't end up with a hodgepodge because these GPUs can't be too far from each other. You can't take one and put it 10 miles away in another building and hope to have one big cluster, right? And so often I think you've seen XAI publicly go to Tennessee and 100,000 GPUs and lots of cabling issues, right? So it's that sort of pattern that people should be prepared for. And, you know, I think the enterprise is still in an infancy in the sense that there's probably not a lot of GPU farms running in the enterprise. There's not a lot of AI factories at scale. They're mostly going to hyperscalers and neoclouds. And, you know, it'll be interesting to see that transition over to the large enterprise and where that's going to go with inference, how big that's going to get, and what kind of scale they're going to really scale to and pay for to get the value out of it. Yeah. Some of that's still up in the air, but to your point, some of the hurdles have already been observed with some of the people doing this scaling and doing this growing.

Perhaps a little bit of the warning or wisdom you're sharing here is, hey, if you have some idea of what scale is going to look like and you want to be operating at scale in the way that you, Cameron, have shared with our listeners to do, you might want to have a lot of forethought about those real estate concerns and about not having these switches between wacky buildings and this kludge because that's not exactly going to be performant for what this stuff looks like at scale. So does this mean more ardent thinking about data center planning and power access? Totally. I think one thing I've seen is often the data element is an afterthought. So people are buying GPUs, plugging them in a building, and all of a sudden find out that, oh, I need to load this with data. So I need to add storage, at which point the building is full, there's not enough power. And so how do you right-size that? So proper planning, I think, is important because we've seen people take out GPUs out of the building to put storage in because it was not designed together as sort of one block.

You see what I'm saying? I totally see what you're saying. And I going to return this to another enterprise AI example that many of our legacy enterprise folks are going to understand but also I think the Neoclub folks will at least have some empathy with which is getting any kind of major AI thrust going inside of a large corporation is going to involve input from multiple parties You got to have your end users there You got to have your folks who understand the data and sort of have some AI chops You want to have your business leadership there that understand the mandate and what we're trying to achieve and what we have for budgets and how this aligns with strategic outcomes, et cetera, et cetera. You've got to have multiple perspectives because if you don't, then you end up with a project or a big investment that isn't good at all for users or totally doesn't tie to an ROI that the company cares about. And the point that you just said, when the right people aren't there thinking about how are we building out this space to handle this compute, including storage, if the right, let's call it committee of multidisciplinary people aren't sort of thinking seriously about that, you end up schlepping stuff out of your darn building

just to make sure storage can be part of the mix. Who's got to be part of that brainstorm to make these be smarter decisions, Cameron? I think you need sort of an AI platform group or practice within IT in these companies to have reference designs and forecasts and do a bunch of experiments with small amounts of users to get a sense about their usage patterns and what they want to do with AI in their company. And then based on that and their size plan ahead, depending on how many employees they have, right? I mean, everybody is going to get access to these AI models to augment and make them more productive. So what's that look like? And are you going to be, let's say, right-sizing And for that, are you going to be efficient? What's the ROI? All these questions are being asked now and enterprises are learning. And the costs are dropping on various different models. So what's that reference architecture look like? And how do you go about relying on mostly vendors that have solved the scale and simplicity problem?

because you don't want to be in the, how do you say, build everything myself type of crowd, which is what you invariably do when a nascent industry just starts up. But for it to scale and transition to the enterprise, you need to rely on quality vendors that simplify your life and take on some of that burden for you. So rather than hand you a Swiss Army knife and have you hack it. Yeah, I got it. But famously not a great solution for enterprise AI in many different other areas in regards. Any closing notes for us here, Cameron, for you in terms of leaders who are leaving? Some people might be pretty well plugged into the hyperscaler, neocloud world. Others might be more in the earlier stages of scaling up. But they're thinking about scale because they've listened all the way through to the end of this episode. Any other sort of tidbits of advice about being scale ready you want to leave the audience with to think about? Yeah, I think it's, as I try to portray, it's not just about speeds and feeds and one dimension, but it's really about storage as a service and providing intelligent software on commodity, you know, latest and greatest servers that provides you massive scale and throughput.

and it's easy to use and easy to expand. So the operational burden of all this is low and pushed down to the software and the vendor. I think if you think holistically through that without just focusing on one of the vectors, you'll have better outcomes because that's what most people who serve these environments end up wrestling with day in, day out. And I spend a lot of my time trying to build systems to simplify and provide these solutions to a broad set of customers who aren't necessarily experts at AI and shouldn't have to be. They need outcomes for their business to get productivity gains and real value and not be in the trenches worrying about infrastructure. Even as a market research firm, where I don't have a dog in the fight, I think different people will take their different approach. But I would say it seems outlandishly challenging that companies that aren't wildly AI native with the smartest people poached from the best companies are going to be solving this problem without relying on vendors to make this element of it. simple. So something to be said of that and some good tangible rules of thumb for you listeners

about what operating at scale looks like. I like those two components we talked about, Cameron. I feel like almost everybody can say, where are we on those two and kind of know where they sit. Maybe it's better than you thought or worse than you thought, but at least you know something else there, dear listener. So Cameron, really glad you were able to join us here for a second time in the show. Thanks so much for being with us. Yeah. Thank you so much, Daniel. It was a pleasure. Wrapping up today's episode, I think there are three key takeaways from our conversation with Cameron. First, as AI infrastructure grows from a handful of racks to hundreds, failures start being rare events and become a daily operational reality that teams have to plan for. Second, telemetry and automation, down to knowing whether a single cable is seated correctly, become essential once thousands of GPUs are networked together, since one failed node can stall an entire job. And finally, treating data and storage as a first-class part of the infrastructure planning,

not an afterthought bolted on once computers are already in place, prevents the costly rework of bringing GPUs back out to make room for storage that should have been planned in from the start. In this episode, we covered what it takes to operate AI storage and data infrastructure at scale. To go deeper on this topic and learn how consultants are winning business with evidence-based AI ROI and building long-term capabilities instead of chasing short-term gains, download our free PDF report, Three Keys to Thriving in the Coming Era of Automation at Emerge.com. That's Emerge.com to download your copy. For further executive level analysis and to join our network of leaders delivering workflow impact with AI, visit Emerge.com. on behalf of the team and emerge we'll see you on the next episode

番組の概要欄(原文)

Scaling AI infrastructure from a single rack to hundreds introduces failure points that most enterprise storage systems were never designed to survive. In this episode, Cameron Bahar, GM of Scale AI at Everpure, breaks down what it actually takes to operate AI storage at scale, from cabling and telemetry to the operations overhead of keeping thousands of GPUs running. The conversation covers the operational burden of scale, why visibility and automation matter as deployments grow, and how companies across NeoClouds, financial trading, and autonomous vehicle development are navigating similar growing pains.   This episode is sponsored by Everpure.   Learn how consultants are winning business with evidence-based AI ROI, and building long-term capabilities instead of chasing short-term gains, download our free PDF report, "3 Keys to Thriving in the Coming Era of Automation," at emerj.com/cok1

出典・この記事について

番組
The AI in Business Podcast(Daniel Faggella)
元の回
How Storage Throughput Determines AI Workflow Outcomes - with Cameron Bahar of Everpure
配信日
要約作成日
公式サイト
emerj.com

この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について

X でシェア

関連エピソード