theCUBE + NYSE Wired: AI for CFOs & AI Leaders in Silicon Valley | Tony Maslowski, Scaleflux

Clips
More from theCUBE + NYSE Wired: AI for CFOs & AI Leaders in Silicon Valley

Anthony Maslowski

CFO

ScaleFlux

play_circle_outline Importance of components in AI data centers like retimer chips and memory controllers

play_circle_outline Maximizing AI Configurations: Optimizing System Performance and Utilization While Managing Wall Clock Time

play_circle_outline Maximizing Efficiency in AI Operations: From Training to Inference and Energy Cost Reduction

play_circle_outline Blurring lines in AI revenue and identifying key content providers in AI data centers.

play_circle_outline Impact of AI on financial reporting and business models

play_circle_outline Execution and efficiency in chip companies

play_circle_outline Navigating AI Transformation: CFO Advice for Understanding Technology and Business Model Importance

Info
Transcript

Tony Maslowski, Scaleflux

Anthony Maslowski

CFO ScaleFlux

Tony Maslowski, CFO at ScaleFlux, discusses challenges and opportunities in the AI industry, emphasizing the importance of components like retimer chips and SSDs in data centers. ScaleFlux focuses on memory controllers to address the need for faster data access. The conversation also covers the significance of wall clock time in AI processing and managing job completion, as well as innovations in AI training, inference, and reinforced learning. Inference is noted as a critical application, with companies like Amazon recognizing its importance. The financial s... Read more

explore Keep Exploring

What is the significance of ScaleFlux in the area of SoC for SSDs and memory controllers within the AI data center? add

What are some factors driving the need for better processors and the shift towards wall clock time in the current technology industry? add

What are your thoughts on the progression of training, inference, reinforced learning, and reasoning all coming together, particularly in relation to Amazon's announcement about their Web Services? add

What are some challenges with accurately categorizing and defining AI workloads in a data center environment? add

What are some considerations for a CFO when evaluating vendors and their AI tools in terms of increasing productivity and potentially reducing workforce? add

What qualities are crucial for success in the chip industry? add

What are some important factors to consider in implementing AI technology in a company? add

bolt Powered by CUBE AI

Tony Maslowski, Scaleflux

search

>> Welcome back everyone to theCUBE special presentation, here in the Palo Alto Studios of theCUBE. I'm John Furrier, your host of theCUBE. Tony Maslowski is here, chief financial officer at ScaleFlux, has a lot of CFO experience, Cerebras, Broadcom, Avago, others. Welcome to theCUBE.

>> Good to be here.

>> Great to have you. I'm super psyched for this conversation, because we were just talking before we came on about the ways of innovation, certainly at the chip level, but also just historical Silicon Valley perspective. We're here in the Valley, we've got our studio on the East Coast, NYSE, and we're creating a Wall Street to Silicon Valley tech TV and network connection with the open system of the Wired network from the NYSE, so it's fun to get these conversations out there.

>> Yeah.

>> CFO AI is the theme, and you've been at the game on the CFO side really building technology that's powering AI. All those chips are coming home, Broadcom obviously doing extremely well, the stock price is doing great, Cerebras has got a great traction, ScaleFlux has been on theCUBE. A distributed computing paradigm is coming, and the conversation has shifted to, okay, financially, it's how do I pay for it, what scale is it, is it available? I'm disrupting my operations with new business model opportunities. So there's a lot of disruptive enablement coming, and everyone's asking, "Okay, what's the strategy around this? How do I get my arms around this?" This has been the top conversation, because the demand for performance is coming in at an all time high, not just in the cloud, on premise, edge, this is what you guys are doing. What's your take on this, as someone who's been in the industry through many waves? We're at a new era.

>> Yeah, I think the most impressive thing about this wave is that everyone wants to write it off so quickly. They want to compare it to other waves, like a couple of the analysts say, "We're definitely going to go through a digestion period very soon." I don't see that at all. I think this new paradigm of the number of GPUs people are connecting, the 100,000, the people who are talking about connecting a million GPUs together, those are very difficult things and are on the drawing board right now. So there's no real slowdown that's coming. However, around the GPUs, you need some important components. You see what happened with Astera Labs with something like a retimer chip. I remember people saying, "Well, what do they make?" "They make a retimer chip." "Why is that important?" "Well, it goes into every NVIDIA DGX, that makes it important." And so-

>> But people who bought cars during COVID, you had that car on the lot. "It's my car right there. Where's the chip that runs the radio?"

>> Yes.

>> That's what you're getting at here, there's components in these things that matter.

>> Yeah. And then, with ScaleFlux, they have an interesting dynamic, because the area of the SoC for SSDs and memory controllers has really been underfunded for many, many years. Hao and his team have been at it for now 10 years, and they're just hitting their revenue ramp. And this is a very important component within the AI data center as well, because people need to get their data, they need massive amounts of data, and they need it quickly. So HBM really is just a temporary fix for getting the data closer. And so, over time, you'll see, with the implementation of PCIe 6.0 and some other things, that it's going to be important to do that. And Hao has pretty good IP, I would say world-class IP, in the areas of error correction, as well as compression. When people say compression, it's like, "Oh, everyone can do compression," but it's compression without latency, and that's very difficult because that ends up being a balancing act of how much do you compress and how do you affect latency. And he's spent a lot of time with it, 20 years in the data space working on it, his PhD thesis was on compression.

>> It was great to meet the team. When they were in New York, they did an evening happy hour event, Brian Baumann and Wired put together, they did a presentation. Some of the stuff might've been confidential, I can't remember which one, I don't want to maybe say the wrong thing. But what was very clear to me is that the new architecture of systems design, not the old school of designing boards, but clustered systems, so you need, one, the systems itself, the hardware in combination with other things around it reminds me of the old school PC server days where little things mattered, how much memory you had, where was it, sitting close to this processor, I/O, basic constructs. But for the most part, that's elementary. When you scale up these large systems, you have to leverage the power requirements, you have all kinds of new constraints that have to be factored in for designing this. It's not just getting the hardware right, what's the configuration? So there's a lot of stuff going on we're trying to get our arms around. This is affecting the scale, because price performance is coming back in vogue. It's never left, but it's never been more emphasized today, in my mind at least, I'd love to get your thoughts, than the '80s and '90s, where all you talked about was price performance, there was always new things coming. What's your reaction to that?

>> Well, yeah, I grew up with Hitachi Data Systems was one of my earlier jobs, and I remember at one point someone said that, "All the equipment we sell depreciates at 1.5% per month because it's replaced by something new." I think today, where NVIDIA's pushing their performance, they're pushing it hard, they're turning up the power, causing heat. They're doing that because of the fact that some of these 1,000-GPU or 10,000-GPU configurations are still running in the 30%, 40% utilization. So people really need better processors to get more done more quickly. The other most interesting thing is that, again, my past used to be all about response time, so how long does it take someone on a terminal to book a car, or something like that. And now, what's interesting is wall clock time is coming back into vogue, because when you start the model is when you finish it, and if a GPU goes down and you have to do checkpoints and all the rest, wall clock time is what matters these days.

>> Yeah. The term we hear a lot now is job completion.

>> Yeah.

>> Job completion is a legit benchmark, and if you have to redo the GPU cycles again to generate where you left off, that's more energy, that's more cost. And so, memory is being kicked around in other contexts, memory of the system, because you can accelerate with the computing, but if you got to... There's no offload. So offload is coming back, because I've got to store that state.

>> Correct.

>> We all see it in ChatGPT when you pick an old thread, oh, yeah, I was going deep on that one thread, you hit it, it says memory, and then you want to activate, it's got to recalculate.

>> Yeah.

>> That's actually spinning up more GPUs on your behalf, so behind the curtain, not really optimized.

>> No, no, no, no. And if you read, I believe it was a couple of weeks ago, they published a detail of the errors, I think, on a Llama run, not specific, but someone sent it to me, and it was just line after line, GPU failure, overheated, network issue. But over 50% of it was the GPU, and if you have one GPU in one of these clusters, things come to a halt, and that's when you have to offload on memory, memory gets burdened, you have to restart. And so, that's why I said it's interesting that nowadays, it's like you said, when does the job actually end? It's not about did you hit 40% utilization or anything, it's when did we complete.

>> Yeah. No, I did not know that term, wall time.

>> Wall clock time.

>> Wall clock time.

>> Old term.

>> Explain that, because I think it means it's something real.

>> Yeah. In the old days, when you were running batch jobs and all the rest, people would say, "Okay, well..." It would sit in the queue for a certain amount of time, they'd get into the computer, and you're always switching jobs and all the rest. So you would look at wall clock time and say when it's going to finish against the other competitive jobs that were out there.

>> Get a coffee.

>> One of my first jobs was the Cadillac division at General Motors, and I remember, one of the first things, believe it or not, that was an interesting job was she put me onto the job that creates the sticker on the car, and I said, "How important is that?"

>> What sticker?

>> The sticker price.

>> That's in the window?

>> Yeah.

>> Oh, yeah, great.

>> Yeah, that's right. I go, "How difficult is that?" She goes, "No, no, there's a lot of... It goes to a certain state. You have to calculate this, that, and the other thing." And she said, "And the other interesting thing is if your system has a failure, we lose $10,000 a minute in the code."

>> In the sticker? You said, "Can't I just write it, price, sale, and ask for attendant?"

>> Again, dating myself.

>> Yeah, but this was time-sharing systems, mainframes probably, right?

>> Yes.

>> You were probably working on a 3270.

>> Yeah, yeah. So this is the old IBM mainframes and Hitachi's old mainframes back then and so forth. And once again, it became system performance was really measured by this wall clock time. Then when you got into online systems, it was measured around how much throughput you could get through. And nowadays, like I said, since utilization is so low, and that's the thing that's, I think, somewhat under-reported a lot, that people think, oh wow, because they use so much energy, they're running as fast as they can. No, they're really idling for the most part.

>> Yeah, yeah. And there's a lot of unused, and this is where virtualization, going back to the virtualization innovation with servers, we don't have to go back to that, but that innovated the physical server, VMware's original product, Xen, open-source hypervisor. Is there going to be a moment, you think, like that? What's the equivalent... Hypervisor and virtualization was the server, server virtualization was a real win, that set the cloud up, basically.

>> Yeah.

>> What's the equivalent of that dynamic in the AI world, is there one or no?

>> I don't think there is on training, because I think there might be some innovation there on inference, definitely will be, because you need to load a bunch of weights, depending on the model and so forth, and whether or not systems will be fully dedicated 100% of the time to a specific application. So I think that's where that may come in. But in training, I don't see it at all. I think it's more important just how quickly you can checkpoint and get moving again.

>> Yeah, yeah.

>> And again, dating myself, I think Tandem was a company way back when, and you had the dual non-stop processing.

>> Fault tolerance.

>> Yeah, that's right. You can't introduce that in a training mode. Who's going to produce another 100,000 GPUs just to be the backup for the failure GPUs?

>> Yeah, resilience doesn't exist.

>> No, no.

>> This is actually a fascinating conversation. Because this brings up the whole point of, okay, training will probably have to be figured out, but from what we're seeing in our research side and talking to experts is training will have its day and people will have to inference more than train, and then this reinforced learning. So the progression of training, inference, reinforced learning, and reasoning, all are coming together. Inference is the killer app we're seeing. Amazon announced, as their Web Services, a building block, and they're so hardcore, they think it's as revolutionary as S3 or EC2 in terms of how they see inference, where you need to have this to be a critical party, it's just inference in particular. What's your thoughts on that?

>> No, I think what's happening now is that, obviously, the models are changing quite a bit. When I was at Cerebras, G42, one of their first models, was an Arabic model, we produced that model very quickly for them. And then, there was several, I would say, fine-tuning efforts on the training. And then, we finally presented it to G42 and the benevolent dictator there, and it asked us some questions that were economics, not based on the language specifically, it knew nothing. We had to go back and retrain it on statistics in the Middle East to make it more usable for the audience that it's made for, the Arabic language. So I think that those are the things that are just smoothed over. I already heard someone, I think it was last week, who said, "Oh, we're already past the agent phase, we're into the swarms of agents." And I'm like, "Okay, point to the 100 agents that are working today."

>> Where are they?

>> So everything seems to be speeding up a little bit too much.

>> Yeah. Well, on the hype cycle, for sure. But right now, in the conversation we're having here at this session, the Silicon Valley side of the AI/CFO conversations, practical AI, real ROI comes up. These are like, again, back in those agent areas, everyone sees the setup of the infrastructure. And the other conversation is cost. How do I do risk management and manage the cost? For example, if there's no supply chain, if there's a supply chain disruption, I better have alternatives so I can have a non-disruptive operation. So do I go to a CoreWeave? Do I go to a TensorWave? So when you think about the financial side of, it's like, okay, I've got to manage the model, I've got to finance it. Do I finance the CapEx? As a CFO, when you look at the challenges in the industry, what are some of the things that you see, being on the side that you're on, you're on the innovation side producing AI, now you look at the buyers, who are consuming the AI, who might do hybrid AI on-premise where the data is, there's an economic model. Will they stand up a bunch of GPUs? No, they probably won't, they'll use a managed service, or they'll buy some gear on-premise, connect to the cloud.

>> Yeah. I think that's the challenge they have, because everyone's CapEx is going up so high, and so who's the first one who's going to spend 100 billion on CapEx? And then, they look at CoreWeave and these other people and say, "Okay, why can't we just convert CapEx into OpEx?" And so, they go that route. And that's what's also interesting about this whole thing is that Wall Street will take a certain attitude toward CapEx, and then instead, they turn it into OpEx. Well, guess what? They still need the same number of GPUs in either case, one just happens to be financed by some other person as opposed to the first person. It's really interesting what's happening there, because people want to tap it down on the CapEx, but you don't want to show it's growing too fast, and then, yet all they're doing is creating these other entities out there to offload and turn into OpEx.

>> Yeah. Tony, one of the things that comes up, Dave Vellante and I were talking, we were at the NYSE last week for our Cyber Week and AI Innovators on the East Coast, I hosted a panel with AMD and TensorWave, and the AMD folks, we were talking about some of the things they're doing, and it was about the cloud option for the neoCloud, also they have the chips and all the performance. Dave Vellante was in the audience, there wasn't enough chairs to co-host with me on the panel, but he asked a question to AMD, and I want to get your thoughts, because you've seen this movie before, the ODMs in the white box days. Dells, the HPE servers, the classic model was build some servers, rack and stack rack switch, connect it, do all that infrastructure. Now, they're suppliers to the CoreWeaves, so the white box became an alternative to the branded servers, ODMs, OEMs, were all that ecosystem. What changes in the ecosystem, if we look forward and think about partnerships, economics, the business model of large-scale systems?

>> Well, I think where NVIDIA has taken it is to the systems level. And so, people are all trying to jump on the systems bandwagon. HPE comes out and says, "Oh, we have a system too,"" or AMD buys someone to say, "Hey, we're in the systems business and all the rest." And I think it's more of the fact that NVIDIA can do that. As soon as they wrap it with a whole bunch of other things, it gets pretty expensive pretty quick per GPU, and they're just trying to maximize that profit right now. The people who are attempting some of the white box things out there, I don't think they're as successful as people think they are. And as we all know, as soon as AI becomes the buzzword, there's the true AI revenue, and then there's the relabeling of AI revenue.

>> What does that mean?

>> Well, because people will be selling the same part to the same company, and then all of a sudden they say, "Oh, we didn't realize that 50% of it is going into your AI data center, so, oh, we're going to count that as AI, and oh, we don't really care if that's too accurate that way, so everything that goes to this facility is now AI." Even people talked about backlog that way, saying, "Oh, we have this AI backlog of X, Y, and Z." And I think it's getting very blurry, and people really want to know... I think it's interesting, I was talking to an analyst and he said, "Much like with the cell phone," they said, "What we want to know is what's your content," say, ScaleFlux or Astera Lab, "What's your content percentage for every 1,000 GPUs? So how much content do you have?" And it's interesting, because that's what the teardowns on the iPhones were, like, "Well, whose chip is that? Is that a Broadcom chip? Is that a Skyworks chip?" And all the rest. And people are trying to do that simplified version for the data center, which, again, is a little lazy.

>> You're saying because there's workloads going through, that's the key usage?

>> Yeah, they want to know who's getting the biggest piece of the pie so that they hear that Musk is going to do another 200,000 GPUs somewhere, they know automatically it's going to include this, that, and the other thing. So it's an interesting thing that's happening right now, everyone wants to know the content attach rate.

>> Yeah. It's interesting you brought up the counting the revenue, obviously CFO's perspective. I remember back in the day, in the dotcom bubble, SOP 97-2 revenue recognition was a big issue. "Hey, I'll give you cash, you be a customer, and we'll both call it revenue." They crack down on that in the dotcom bubble, not saying that's the same thing, but it sounds like a lot of that might be... You have to read between the lines in terms of where the revenue goes. Dave and I were talking about Snowflake and Databricks and how they report their revenue, a lot of us go to Amazon. So is that what you're getting at, or is it more of... I'm not saying it's fraudulent accounting, but there's loopholes. What is the CFO perspective on revenue, what's real, how does someone peel back the onion and recognize what's legitimate AI?

>> Well, I think that's the issue. We all know what revenue is, because again, the rules have been very, very clear, 606 rules, you go through and say, "This is revenue." What I talk about is when people represent it to the street as, well, who is this going to, or which customer is this going to, and that's where people, I think, are being, I would say, a little gray in some of their descriptions, because they want to say that, "Hey, we're experiencing a similar growth to NVIDIA." Because that's the key thing is if NVIDIA-

>> You want a magic ...

>> then we want to know that you're growing with them. And I think that's where it gets a little bit disingenuous, and there's, again, I think-

>> Is that because they want to jump on the comparable wave?

>> Yeah.

>> Say, "We're going to grow as fast as NVIDIA."

>> Yeah, that's the bottom line. And you've seen already in the marketplace some companies being hurt because they aren't keeping up. So NVIDIA has a great quarter and it's like, whoa, well, why are you slowing down? And that's a valid question. And so, I think that's what it's more about, it's not about that people are doing a new version of a shell game.

>> It's kind of a shell game. Microsoft certainly has a shell game going on the way they report revenue. I won't say shell game, let me rephrase. Microsoft bundles their apps in there, and Amazon Web Services doesn't have any apps, they have Connect, but it's a different thing. Anyone can slice a salami the way they want, you know the expression. Now, on the CFO side, let me ask you this question, because I think this what comes up a lot. Remember the old consumerization of IT, going back 20 years, consumerization of IT, that was just IT trying to be more friendly. Then you had digital transformation, that went on for a decade. And now we're in, I think, a legit business model transformation, business transformation. And that seems to be the buzzword around the AI. From your perspective as a CFO, if you were sitting in the chair of, say, I won't say JPMorgan Chase because they're pretty innovative, but enterprise, they've got to make technology decisions, the CFO has to lean in because now there's GenAI involved, another application, there's resilience frameworks, there's other mechanisms they have in place, but also they have competitive pressure.

>> Correct.

>> So there's cost reduction, which is drops to the bottom line if they can get it, and then revenue potential. So as a CFO is trying to squint through that general equation, in comes the vendors, "Buy my stuff. I'm a great tool for cybersecurity," or, "I'm the platform that does it all. We're a single pane of glass." You've heard them all.

>> Yeah, yeah.

>> So what is the best motion from a vendor's standpoint, and what's the best tactic for CFO to identify the players versus the pretenders?

>> Well, I think nowadays, especially with the software side or people presenting, let's say, quote, unquote, "agents" because it's the new term, but it might be the very simplified AI model that does something very, very specific, I think they're all interesting to look at. However, at the end of the day, are you going to make your person, A, X more productive, or the extreme, are you going to take a group of eight down to six? And that's the question that a lot of people are asking internally. And then, once they get there and they say, "I'm going to remove two people," it's the whole issue of governance, risk management, things like that. So the tools are coming, however, I think there's a lot of peeling back the onion to understand what are you really getting with some of these tools. And there's been some pretty impressive reports of programmer productivity, engineering productivity, and all the rest. And I think the CFO has to make sure that they understand what... They have to get a little bit educated on how AI works and say, "Whose data is it running on? What are some of the results I'm going to get? And what are some of the risks?" The funny part, it's just like driverless cars, is that people sit there and go, "We'd never want an agent to say that to a customer." Well, guess what? There's probably an agent saying that to a customer today. And so, people want to say, "You have to have that perfection in the error rate," because it's AI, but really, you're not at the perfection level today.

>> Yeah. If you're a customer service agent, no problem for Expedia because the bar is so low, it's nothing but upside, but if you're JPMorgan Chase doing $10 trillion a day and you've got a $17 billion IT budget, Lori Beer, the CIO, she's told me on theCUBE, "We have zero tolerance for failure," so they cannot have an error.

>> But I think it's funny, because if you talk to some of the big firms, some of them are implementing what I would say light agents. So one person I deal with very directly, I did some type of transaction across my accounts and my trust, and he says, "Tony, about a year ago, that would've taken me probably 15 emails and probably half a day to get that accomplished. Today, I get it done in 15 minutes."

>> Yeah, couple of clicks.

>> "Because we have an agent that's -"

>> Internal.

>> Internal.

>> Process-driven.

>> Yeah. And they say that's very, very, very helpful. Again, this zero tolerance for error I think is-

>> I think it's more of an external app regulated industry, obviously bundling in the private.

>> Yeah. And you can't say-

>> But if you're an internal enterprise, you've got work flows.

>> But the funny part is, remember what the buzz was a couple of weeks ago with JPMorgan Chase was people were doing this depositing checks on Saturday and taking their money out.

>> I didn't hear about that.

>> You didn't hear about that? It was some type of thing that went viral that we've all done it, you get to Saturday, you get a check, you put it into the bank, and it always gives you $400 or $500 back, saying, "That's what you can take out." Well, there's this error where you could take everything out, so people were just writing fraudulent checks, taking out $5,000.

>> Once that goes viral in Discord, it's over.

>> I forget what the Wall Street Journal or where I read it reported what the total amount was, but now they've got to obviously fix it and then go chase all those down, criminally or otherwise. And so, it's always funny, especially in the banking area. What was it, Citi, a couple of years ago, when they paid some entity the interest payment wrong, or double the interest payment.

>> Yeah, I remember that one. That was a human mistake and he got fired. But this is exactly it, the human in the loop is still critical, you have internal processes. I see those learning mistakes, it's a big scar tissue, but still, they get productivity gains. So your example about your trade, that's multiplied, that compounds.

>> Yeah. And it gives people the comfort that you're not trying to do something that is an act of God to get this thing moved through their system, and gives you confidence that their back room is very, very efficient.

>> Yeah. Tony, one of the things about your career is you've seen all the major waves of innovation. You mentioned the Cadillac example, use of the mainframe to open systems, inter-networking with the PCs, that brought LANs, internet, the rest is history. Is there a learning that you see, from a pattern perspective in those waves, that equate to it now? Me, I look at it and I say, "I've learned that the stack is the stack, and if you don't get the hardware right, the top of the stack doesn't really work well. You've got to get the bottom of the stack, performance, whether it's network, compute, the basic stuff right." Now, it's a little bit different now with AI, and we're seeing that, that's why agents are still coming, I agree with you on that. What patterns have you seen that you can point to and say, "We saw some things over here that match this from a scale standpoint that give us a..." Help us connect the dots on how companies and CFOs and other business leadership prepare. Because if this becomes a business model evolution, the business productivity is going to increase, not the IT, the IT certainly gets better, certainly it'll get better because we're always working on that, but the new aha moment is the business person can do work.

>> Yeah. I hate to be always the one that says, "This is different this time," that's the famous last words. But it seems very different this time, because, again, people are shocked by the amount of investment that's going on, the issues they're running into. You say, "Oh, wow, guys, is it really that hard to keep up this many server racks?" But it's different, because we're really pushing the edge of performance on some of these chips. So an overheated chip here really can cause the training run to stop, you've got to back it up and all the rest. It's more of the engineering IT side, where people are running the experiments, as opposed to the production IT side that says, "Oh, just do this, do that."

>> Yeah, operations.

>> You get the phone call, blah, blah, blah. I think there was something on LinkedIn, where I think it was a person at either OpenAI or some other company, he was in charge of the infrastructure and he says, "I never slept well on weekends. I knew the phone was going to ring at Sunday at 3:00 AM," so it's his life.

>> Yeah.

>> You've been around the block a lot. Think about in IT the last time it happened. Maybe it's a power outage, a big failure, 99.99% uptime, that 0.01% happened. But it wasn't like now, where people are kicking off these large training runs.

>> Last Thursday, OpenAI went down. Yeah, it was down for a long time, hours.

>> Yeah.

>> Tony, great stuff. Quick plug for ScaleFlux, love that company, they've been on theCUBE. What's the secret sauce, looking at the CFO, looking at the team, they're breaking out, I'm seeing some good performance from those guys, you guys, what's the key to ScaleFlux's success?

>> Well, I think what I've always looked at through my experience is that, first off, lots of chip companies, lots of chip companies getting funded, and it's really around execution. So people who have a repeatable effect on either downsizing the chip size or features or whatever, that's really, really critical, and there's very few that have great, great track records. And when I joined Cerebras Know was kind of the decision to go was they were going down to seven nanometer, the confidence, the execution that was there, and they went to five actually after that. And the same thing with Hao and his team, their execution around getting the chip out on time, they give it the normal number of months for testing, finishing that testing line with no major errors, something small comes up, they know how to fix it through firmware or something, that's really what makes them successful.

>> And market cycle times are huge.

>> Yeah, yeah.

>> Talk about the order of magnitude now versus, say, just going back 10 years, in terms of the lag and the time it takes to get a chip out the door.

>> Well, like I said, some of the chip design world was several years, and now it's in this cadence of two years. Typically, you lock down design, you get your tape out, and it's a year's worth of testing. People are saying it's going to speed up to 18 months.

>> Testing's testing.

>> Yeah, testing's-

>> AI might accelerate that.

>> Yeah.

>> But not much.

>> A small amount. So I think the two-year cycle, doing that consistently with good results is important, because you don't hear too much about the stories of someone doing three re-spins and never getting to the finish line. So that's what I look at, and I think Hao's team, similar to Cerebras' team, performs really well.

>> Yeah, great stuff. Final question to wrap us up for today. CFOs that are out in their enterprises in meetings, board level, top down, "Let's get into AI business, make some money, and lower costs. Bottoms up." "Boss, we need more budget. I don't know who to buy. We're refactoring our platform engineering team. Data's coming everywhere." What's your advice to the CFO to, one, maintain their sanity, but two, to re-architect and rewire the blueprints of the business model?

>> Well, I think it's very similar to previous IT implementations, you have to deliver on the return, but educate yourself on how the AI world works, so the importance of data, the importance of training, fine-tuning, the importance of speed and inference and all the rest, and show the results, at the end of the day. This idea that everyone has to just run out and just tomorrow, again, relabel a bunch of people internally and say, "Well, you were database yesterday, now you're AI, so go do it," that doesn't happen.

>> Everyone raises their hand, "I want to lead the project." Maybe it's this way.

>> That's right.

>> There's some companies where it's like, "I don't want it." Some people want it bad-

>> Yeah....

>> depending upon what's going on. Tony, great to have you on theCUBE.

>> Pleasure.

>> Thanks for taking the time for our program here, appreciate it.

>> Take care.

>> Okay. I'm John Furrier here at theCUBE. This is theCUBE plus the NYSE Wired community, where we're breaking down the East Coast and West Coast, connecting Silicon Valley to Wall Street with theCUBE and the NYSE Wired community. Thanks for watching.