Panel Discussion | AI Computing - Corporate VC's AI investment strategies

12:00AM May 30, 2024

Speakers:

Keywords:

startup

ai

companies

investing

models

cvc

investment

data

compute

qualcomm

investor

venture

side

market

question

term sheet

generative

capital

samsung

tushar

Okay, all right, absolutely I need to just go on. Yep. I

Hi, hello, thank you for sticking around at this late hour. We have a great panel for you today. And my name is Shi Hui Xiao, a managing partner of Foothill ventures, early stage JPEG fund based in downtown Los Altos, just a few hours down from here. And I'd like to give a quick intro of our esteemed panelists today. You can read the details bios on the slide. Jeff is the head of corp dev from AMD, in charge of MNA and also CVC and daedae is the head of Samsung Catalyst Fund, deep tech and AI and Elena. First time new job. Just joined a Dell CVC today or this week, after a 14 year span of illustrious investment experience from Intel Capital and then, last but not least, the Tushar, Senior Director from Qualcomm ventures, also deep tech and AI investment. So have you seen any patterns here while Nvidia is grabbing all this headlines, but this panel, if you have invested in this panel, the stocks you'll be doing really well. AMD, after a huge run last year, still up more than 12% this year. I'll talk about cents on just a moment. But Qualcomm up 45% this year, and Dell after 5x and then this year alone, up 120% and then Samsung, I wish we could buy Samsung, but it's not traded here, but that's probably the best value in this group, if we can buy the stock. So we have everyone here that is in AI. So Joke aside, last 12 months have been crazy and continue the momentum generative AI. And as a deep tech and AI investor, I've been investing for last almost 10 years and but this wave of ai ai innovation has a much broader and deeper impact to everybody's life, everyday life, as well as our investment landscapes. I'd like to ask our panelists, what have Gen AI changed your investment focus and strategy in the last four months? I'll start with Jeff.

Thanks. I think in general, we've we've had to pull in a lot more capital, because there's just, you know, we, we AMD historically has not been incredibly active on the venture capital side, but that has changed over the past couple of years. So with the rise of AI and some other things, with respect to just a big deal that they did in 2022 there's a lot more focus on that now. So, you know, I'd say that just in terms of capital allocation around venture efforts, just given all the sheer amount of opportunities, that's one area that's changed. And then, you know, honestly, there's we're just investing more. So we shifted a lot more of our focus to AI where I'd say, you know, three or four years ago, it was maybe 25% now it's north of 80. So just more capital, more focus, is really what's changed. And the other thing is, we've just had to move quicker. There's, there's a rush on a lot of these deals that are, that are being raised. So you you have to just kind of be able to react sooner and and get answers out because, and get decisions made sooner because, you know, there's more competition. There's, there's my peers on this panel that we've, we've, I think we're in the cap table on a number of different, uh, investments. So, you know, we all play nice. But there, you know, there are others that maybe, that I think you alluded to at the beginning, bigger companies that are on this panel, that might not play that way. So you have to be able to make decisions a lot more quickly in this environment.

Yeah. So I, first of all, I'm told that for a panel to be interesting, you need to argue, right? So let me start with a small argument. I when I joined venture capital space, it was, you know, I took the decision in year 2000 and I started in 2001 and I was told at the time that the most important capability of a venture capitalist is to run and put the term sheet first. I mean, this was like 99 and the beginning of 2000 and we had a similar thing in 2021 so I really want to disagree in the sense that, you know, rushing is not a good thing in venture. You need to take your time. And if things are happening too quickly, just let them go, and maybe you'll have the opportunity to do them again in the future, maybe at lower evaluation, and maybe not. But you know, we need to be prudent, especially these days. And you know, if we look at what has happened over, you know, the past several years, you know, with the decline, you know, the, you know, declining, the amount of capital that's being invested in general, right, and down rounds and things like that, actually, AI, is the only space that's climbing up so, tons of opportunities, but, you know, there's a bifurcation between the fantastic companies and so, so you need to be careful. I'd like to say just in terms of what, how Gen and AI has changed the kind of things that we're doing. So we've been investing into AI as a venture team for almost a decade, right? We invested many years ago into graphcore and Habana labs and also some bonova and other companies. And I would say AI was always a challenge, because, like, if you play soccer or football, the goalpost is moving all the time, right? The models are changing and so forth. Gen AI dramatically accelerated that. So it's even way, way worse. And because of that, if traditionally, we were investing into companies that have the, you know, the unique architecture for accelerating AI or things like that, we shifted to addressing, of course, we continue to do that, but also to addressing the issues around that. Maybe we'll, we'll discuss these issues with later. So I don't want to grab all the, all the topics. Thank

you, Dede and Elena. Are you a rush in with the term sheet person or more on the patient side,

I kind of echoing both of your points. So in a way that a few we have been dow capital have been deeply focused in enterprise software, SaaS, AI, ml, so we've been looking into the Al field, machine learning, AI field for a decade now. And when the chatgpt, when this wave, regeneratively come, personally, I felt like there's a sandstorm. I could barely open my eye with everything that's happening. And so, in a way that, how do we take the time to look for the true value creation. Where's the value going to go to that takes some market development, really. There will be in the long run, we believe in generative AI will change the next generation. There's going to be a lot of winner emerging. But in the short run, how we can, how can we pick the winner, the category creator? That's something we're debating deeply inside our firm.

Thank you and Tushar, you're the tiebreaker. Well,

data is telling us that don't rush to write a term sheet, but I'm sure he's going back and writing a term sheet right away.

He's just telling us to be patient.

I That's how it works. That's the truth.

So for us, you know, for for Qualcomm ventures, in terms of capital, amount of capital that we deploy, it's for at least for the last decade, I know that it's been around one $50 million every year with Gen AI. However, over the last four years, of course, a large chunk of that capital is going into a lot of AI companies. So we've been investing in a lot of AI infrastructure companies and also some AI application companies. The other thing for us, like besides AI, we Qualcomm has businesses in automotive and IoT and networking and communication, what we're seeing is other than just core Gen AI companies, we're seeing Gen AI disrupting a lot of these other adjacent areas. So in fact, most of the deals that we invest in these days are in some way, shape or form, Gen AI enabled. So that's something that we're seeing across the board and excited about it.

So the next question is more designed for some of the entrepreneurs in the audience, is that there are many different aspect generative AI is touching from infrastructure side, software infra, hardware infra, to all the way to application layer. And in the middle, it's the pipelines and data, and also, whether it's cloud AI versus edge and device AI, what's one area you and your firm prioritize or focus on you're really excited about, personally, data you're looking at,

I'll say something. It's not because I'm Samsung and we're a memory player, but going back to what I mentioned earlier, so in generative AI, you know, one of the issues is that the size of the models is growing very rapidly, right? Like, if you know, we used to have computer vision models like ResNet 50 and this, you know, like 20 million parameters, 50 million parameters, you know, things like that. And it was growing like maybe 2x a year. Now is generative. Ai, you know, it's growing. Could grow up to 4040, 40x a year. You know, GPT one was maybe 100 million parameters. GPT four is like 1.7 or 8 trillion parameters. It's just incredible. And the implication of that is that a lot of data has to move between the compute and the memory. So, so that that's, that's one aspect. The other aspect is that in some of the traditional AI models, like computer vision, for example, there was, there was a lot of data movement for training, but inference was like more minimized, right? And in generative AI, you have the same issue at inference, right? You need to move a lot of data. And so our thinking, of course, there are several ways, if traditional we were investing primarily in Compute companies, we still think that there are fantastic opportunities in Compute. And we actually invested, we haven't announced all the investment in some companies that have incredible compute architecture for that. But one of the derivatives of that is that we're looking at ways to solve the data movement. So data movement becomes a critical thing, right? And it's another if you think about Compute, compute capabilities, you know, with lores more and all that, with Moore's Law and all that over several decades. You know, over the past 20 years, maybe compute capabilities grew like by four orders of magnitude, right? But the memory bandwidth grew by only one or two orders of magnitude, which, I mean, there's some sort of disconnect, right? So you have the compute it used to. The problem of AI used to be compute bound, but today it's actually memory bound. It's the bandwidth of the memory. It's a latency. It's the power right? Companies are measuring how many picojoules per bit it cost them just to move the data, so we're highly interested in that. And as a result of that, we've started investing way more into interconnect technologies, like we recently led the financing round of Elian, which is an electrical, incredible interconnect uh, phi technology. We've invested also in silicon photonics and optical, you know, companies like celestial that have, together with you guys, a company that has an optical interposer, or Avicenna that is leveraging other optical technologies like micro LEDs, just to lower, to increase the bandwidth and to lower the power consumption, lower the latency, because by the end of the day, you know, we have fantastic models, we have these engines of Nvidia and other great companies, but you need to move the data. So that's a big deal for us

who want to go next.

And we have been investing a lot of enterprise SaaS, for instance, MongoDB from the database side, JFrog from the DevOps side. We had about nine IPO in the last 12 years and routing on that. What we know, which is enterprise software now, looking to generative AI, there's, in a way, there's a similar kind of horizontal building block that's enabling the generative AI movement, including from the data infrastructure side, right before for micro Kubernetes or micro service structure, moving to this parallel compute side of things, there's going to be software, horizontal platform enabling that infrastructure so generative AI could be deployed in a scalable, scalable way. In addition to that, we see a lot of peripheral DevOps and even sort of data centers as solving the power issue with different kind of cooling, different kind of hydrogen power, all sort of different peripheral element that's enabling the power efficiency of the data center for the generative AI side, from the hardware side, as Didi mentioned right that there's a bottleneck between the memory and processing. So definitely, for the there's a lot of startup focus on inferencing, solving that generative AI deployment at edge, type of thing. We've seen a slew of opportunity there, just to touch on a few area, but also the customized end to end. We call it vertical AI application, definitely with more with more power, compute, sort of capital expenditure in the field, we definitely see we're here in a perfect storm to enable more and more end applications.

Thank you. Yeah, anything to add?

Tusha, yeah, sure. I mean one of the things so I agree with the scaling law and data movement challenges with larger language models, one of the things that is of a lot of interest to Qualcomm is these hybrid edge and cloud models, so we're spending a lot of time on that. So one of the examples over there is, we announced this big partnership with Microsoft last week. We get build we think that the future of PCS is going to be arm nobody from Intel here, right? You're not. So we think that the future of PCS is going to be arm and then. So we announced this partnership with Microsoft and announced these features that are essentially running hybrid local models on the edge, where your data doesn't need to leave these PCs. So they have this feature called rewind, such that you can remember whatever you've seen on the PC. So essentially, it's processing language tokens and image tokens on the edge running on the NPU. So these types of hybrid approaches, even for startups, are very, very interesting for us. So we're spending a lot of time there. Yeah. I

mean, I'll just add I think we we we do investments in the hardware side of things we've as data mentioned, but we see a lot of innovation happening, really, in the model layer. So and that type of innovation can actually drive more architectural decisions on the hardware side that can, that can result in, you know, some power, you know, tokens per watt, improvement, and maybe less reliance on some of the the bottleneck challenges that you deal with, with processing and memory and sort of the back and forth so, and a lot of innovations happening there right now. So, you know, we we're focused more, I mean, we do more, way more software investments, and we do hardware investments. Hardware is great. There obviously have been a lot fewer hardware companies, but that's changing. But for us, you know, really understanding the key components of the stack and how we can best adopt our hardware and modify it to to run those that's really, really solid feedback that we can get through some of these investments and partnerships.

Well, with two of the three largest cab companies being hardware companies, I think interest on building hardware companies just increased. But to get to the more meat of the question, after the warm up, there's a exponentially escalating arms race on compute resources, and we see companies spending billions to buy GPUs and build data centers. We see startups raising billions of dollars because they need to reserve that compute resource. The latest being, you know, Xai raising $6 billion so lot of people worry that this game is so far out of reach for most of the normal startups and academia, academia or r&d teams. On top of that, largest large language models have become inaccessible because, you know, they cost, you know, millions and 10s of millions to train. And people don't want to share. They share the weights. Maybe so to source, open source or not to open source. How do we break through for most of the small companies? Is there still an opportunity for small startups? I know that Tushar is passionate about some of the open source personnel open source questions, I will start with you.

I think that's a good question, which is, Will, Will closed source models win the race, and will there be no room for open source models at all? We can debate that, but I think what we're seeing in deployment, so clearly, yes, there's enough surveys out there that the most the model that is in deployment with most of the enterprises today is open AI, open AI's models. But we, as we're looking at a lot of the startups. So one of the trends that we're seeing across the board is enterprises want to bring models to where the data resides, versus bringing the data to where the models reside. So and of course, OpenAI is also enabling that these days, but for a lot of the startup deployments, what we're seeing is they're still using open source models, whether it's Mistrals models, whether it's llama three, whether it's some other open source model, and they're fine tuning, and we're seeing lots of deployments across the board where a lot of these models that are open source models that are getting deployed in production. So that's another trend that we're seeing. So I think both will coexist. There's no doubt in my mind.

So from to your point, right? It does take a lot of capital to develop the next generation of foundation model. But in addition to that, we see that lot of those models are based on the transformer technology, essentially. And that transformer technology, when you scale the contact lens, the amount of compute resource you need go exponentially up. There is a scalability issue fundamentally here. So we are one of the China we're looking into. It's a startup that's doing disruptive technology to transform our base. So that's an alternative base, arguably, that will help us with that scaling, scalable curve, that could be some lower capital requirements startup in a more potentially disruptive way to transform.

Just a quick comment, because we focus on lower layers of the stack in our investment, I would just say we had a period like maybe six years ago, when you know this data is the new world. Was like the new wave, and we were looking and investing into companies that are leveraging data. We always thought like, I mean, do you have access to proprietary data? I mean, because if you don't, it's going to be extremely difficult to fight with the big guys, if, for any reason, you see, for healthcare, for like, you find a way to get your hands on proprietary data, then maybe you can, you know, one way to do it is to use, you know, models developed by the big guys, open source, but then go and create solutions for enterprises, for, you know, for doing the derivative, which is the private model for the enterprise, or things like that. Because otherwise, you know, again, it's not our expertise, but it looks like a very difficult endeavor. Anything to add? Jeff, no,

I just think we all understand the accessibility issue with hardware, and there's been a lot of projections out there with respect to, you know, Elon's talking about 100k you know, GPU cluster that you know he wants for Xai. And, you know, Sam is shopped around even larger amounts of money for, you know, for training. There's a, you know, we're all chip people for the most part here, so we can understand some of the supply chain. But, you know, the powering of those is really significant. And I think you're going to run into a challenge, which is with the sheer amount of power needed to support the compute that is currently being projected from an AI perspective,

yeah, it's not just a chip, right? You know, as we deploy those high density racks and the data centers, even the power consumption become the issue, right? The latest rack is like 120 kilowatts average power consumption, and then you have hundreds of them. And even the computers now needs its own batteries. The battery guys are really happy, because they can sell more batteries. And Microsoft is applying for nuclear power plants to power their latest data centers. And then, if you compare to nature, I just did the quick, you know, back envelope calculation the other day is like seven, six orders of magnitude higher than what nature is doing, like our brains, maybe 20 watts. And the insects, milliwatts, right? And some, in some cases, well, it's microwatts versus try to accomplish the same kind of perception, navigation, you know, motion control, needed for modern computers that you need 100 kilowatts. So 200 kilowatts, right? That's not sustainable. And then cooling has become a problem, right? So I think what we're looking at is like peripherals, right? Invest in Transformers, solid state transformers. Invest including technology. Invest in, you know, batteries and because, you know, Nvidia has done only the central part of the GPUs, but there's so many other issue networking, where we're talking about moving data, right? So there's a lot of other opportunity for us. But back to, you know, our this panel is about cvcs as a deep tech VC. We love to partner with corporate ventures because they are very they are very deep in their own domain. And more importantly, we believe they bring far more than the capital to the table. They bring a lot of resources, a lot of know how, lot of strategic help. But I want to, you know, hear from the heads of the respective CVC themselves, like, what do you believe the unique differentiation you bring to the table to startups? I want to start with Elena this time, because you have visibility up too.

That's great. First of all, this is my second day at Dell So, but from Dell technology capital perspective, we are more leaning toward the financial VC side, in a way that we have a single LP, but we make our decision on investment based on financial return, less so with the business unit side of things, but from our focus area of enterprise, SaaS, from the AI side of things, we've been investing for the last 12 years, and we have actually introduced A lot of collaboration between, for instance, J frag or mango dB, back to the DAO corporate side from the value, value add perspective, definitely, we invest in area that we know we understand well, and we introduce that relationship from the DAO and also outside, from the our entrepreneur network to the startups.

Yeah. I mean, I can say that first of all, as as you've mentioned, when not all cvcs are made the same, right? So you have cvcs that are like, not purely strategic. You have purely financials, or almost purely financials, and then you have people in the middle, I would say Samsung Catalyst Fund were probably somewhere in the mill, right? We operate like we have the VC bar, but then there's always a Samsung angle, like, you know, how it could make sense today or in the future, I think, with the perspective of, like, two decades in venture like, many years ago, you know, fine, and I used to be a financial VC because before transitioning to the other side, we were extremely suspicious towards cvcs, like, you know, because at the time, many, many cvcs were approaching ventures like almost an expansion of their r, d, right? So it was, like, very risky. You don't take CVC money before the company is really mature. And all that you know, this has really changed over the past years and and it's a certain art right for the CVC to be like the ambassador of the startup in the organization, you know, help. We literally work really hard for the CEOs of our portfolio to help them, you know, navigate into this huge Samsung thing. But on the other hand, in certain domains, you know, companies like the companies represented here are, like, almost a good representation of the market. So we can really help the startup also. So it's a fine balance, right? You have a startup with 100 employees, and then you have Samsung with 300,000 employees, and it's like, you know, that's our job, right? To bring the value and and to help.

Yeah, yeah. So for us, really, I mean, we operate with the discipline of a financial investor, but our value add. A significant part of the value add comes from Qualcomm. So there's three things for value add for us. One is in cases where a startup that we're investing in. In some cases we're not even investing just because it's not meeting the power for investment. Yet we enable, and if a startup can be a customer of Qualcomm's product, then we enable a stronger partnership with Qualcomm and startup. In that case, second is go to market. So Qualcomm has businesses across mobile, automotive, IoT, networking, XR, compute, and a bunch of other businesses. And each one of these has strong go to market channel partnership opportunities. So in a lot of cases, we also enable that. And then the third one is we have we've been around for about 25 years, so we've got an active portfolio of close to about 200 companies and then so, so there's a lot of value add in terms of partnership across portfolio. So, so,

yeah, I think we're very similar to Qualcomm in the sense that you can leverage your channel to help a partner, an investment partner, go to market. I think the other thing that we can we help out on is look early access to hardware and tools, especially in this environment. I, you know, if I had a dime for every startup that was asking us to, you know, can you send us the latest MI, 300 exit, I wouldn't be here right now. I, you know, I so it's, it's a, you know, we have that ability where we can provide some of that to our key partners, and that includes a startup community. We are to day days point in terms of the spectrum of corporate VCs, we are in the middle. We have a rigorous sort of financial metric that we need to meet, and we have demands from our financial leaders that we have to exceed certain return metrics and cash metrics, we have to invest with the strategic rationale. And if you look at some of the work that we've done with partners over the past nine months, since we launched our latest GPU product in December of last year, we've had three CEOs of companies that we've invested in on stage at larger events to showcase the partnership. So I think that's another thing that we can bring is like, look, we we want to, when we have these larger product announcements, we want to showcase partners that we've invested in and we're working with, and I think that helps in terms of getting them, you know, getting them some, some face time with, with our developer base and our community, and, you know, promotion from, you know, from you know, what they're doing and the solutions that they're solving. So that's, that's another thing that we bring to the table, but very, I think, very similar to all three, in terms of just, we all work for large companies, and we have a lot of different ways that we can help out with the partners.

Thank you. And then one of the other element of this is nobody have mentioned is exit strategy, because there's some theory on the street that with the current administration, FTC chair that looked at scrutinizing every public company acquisition companies are resorting to acquire private companies. And look at Microsoft just recently did with one of their investors. Is that true that now you have a mandate when you're investing in startups you keep an eye towards maybe future acquisition is that, is that something that you think it's a it's a change that that startup need to be mindful of, who want to start with? No comment. No comment. Okay,

thanks for I would just say,

you know, you have these cycles or periods of time where IPO is the path. Then we had the spec, yeah, and sometimes markets are closed, so you have to go for m&a. But, you know, no comments.

But in general, right from startup perspective, to your point, there's cycles. We always encourage our startup, regardless what cycle you're in, to be cash efficient, to really, of course, is rule of 40 generally, that growth, while having that margin in in the down side of when the market is tightened and the more cash efficient startup can run on the board, where as we're encouraging that momentum, the startup measures through the tough market. When the time pick up, then this really the time the startup that survived can fly. So from that access perspective, from from venture investor looking at it, we definitely in looking to big enough market. Look into the product market fit, look into startup that could grow into have a pass to 50 million. AR, 100 million. AR, have the pass for potential acquisition exit or a pass to IPO, that's definitely one of the sort of one of the role will play, as board member, as investor with a startup, is always reminding everyone was original purpose. Why are we in in through the tough market doing this, but in the end, helping a startup to be more cash efficient will have that growth engine

for us. I mean, we definitely look at whenever we're investing in a startup, we look at what are the potential exit opportunities, whether it's an M A or whether it's an IPO from from the standpoint of Qualcomm being a potential acquirer of with that's not something that we put a lot of weight towards, because there's only a very, very few investments in the history of Qualcomm where we've invested and acquired those companies. So it's not a lot of weight on that. We want the startups to build sustainable, long term businesses, really, yeah,

yeah. I might be the m&a person on the panel. I spend more time, more time in m&a than than ventures with my role. But I, you know one thing that entrepreneurs can look at, just to kind of get a line of sight for what you know from a corporate VC is corporate VCs love to market themselves. They love to talk about the deals, the companies they've invested in, and also the exits. So if you do your homework, you can actually see which companies were bought, that they had investments in, and you can kind of, you know, figure out from there, if that's a, you know, something that's a try before buy strategy that the corporate VCs that you're working with use, or maybe it's not something that they use, but, you know, to echo to Shaw's point, like for making an investment in a company there, there has to be a view that this is a sustainable business and that it's Not something that is off balance sheet, R and D to an extent, because if you have that, it's you just write off the investment as soon as you make it, and it's worthless. So, you know. But again, I do your homework and figure out who you're partnering with, with respect to the exit.

The last question is really about more education, because, you know, there's a lot of stereotypes or misconceptions about working with CVC, that they're more picky, they're slower. There's too many cooks in the kitchen make a decision, and you need to have really mature product before you talk to them. You can't talk to them too early, because there are a lot of things going on in their lives and and they scrutinize your technology. They run you through the ringers, through their BU's. I'm sure most of these are not true, but as a takeaway for the audience, what will be the biggest misconceptions, or what are the biggest takeaway for aspiring entrepreneurs if they want with CBC, what would be your advice?

I think what we're trying, I can talk about us. We're trying to behave like a VC, right? So I can say, for example, before I joined Samsung Catalyst Fund. Of course, there's interviews, and we also have this sort of exam. Let's say, without saying what exactly I mean, how it looks like, what is pretty intense for me. Also, it was very important to talk to the people, but one of the things I asked them to do is to see the term sheet. I mean, how does a Samsun Catalyst Fund term sheet looks like, I wanted to make sure that it's, you know, plain vanilla, no strings attached. You know, no exclusivity is no I mean nothing. I mean it's really, really very simple, straightforward. I mean, other than the fact that we always need to have information, right, just it's a regulatory thing. So I think this is the kind of things you need to be careful and make sure. Because sometimes you're like, I mean, as was mentioned, sometimes your cash constraint, you need capital and so forth. And, you know, getting, you know, taking money from, from an investor, it's like, it's like, marriage. It's actually worse than marriage, because marriage, God forbid, you can divorce, but when someone is your shareholder, it's very difficult to separate, so you need to be to think very carefully and do your diligence like the investor is doing a lot of due diligence on you you know, the investee, I mean, the startup company, you know, should do a lot of diligence on, you know, the fund that's or the CDC that's going to invest, you know, the partner that's leading the deal, and so on. I mean, and making sure also that talk to other companies that raise money and see that there are no swings attached, no limitations.

I think there's a couple things to maybe add to that. One is, there may be a misconception that if I have a strategic investing in my round, that I'm somehow obligated to to, you know, only buy their product or, or, you know, comply with their software, etc. There's a lot less, if you think about it from our perspective, if, if we ring fence your startup to only work with our stuff, then we're, we're lessening your market value and potentially damaging your business. So that, I think there were probably stronger terms around that, maybe in the year, maybe a decade ago, but it's less so now, partly because there's more competition, but in the CVC side, but I think also there's, there's just more of an approach around you can't, you can't get all the strategic rights that you may want without, you know, the company going kaput. So that's one second. One is going back to maybe the first question about moving quickly. To maybe go back to DD on this when, you know, cvcs in this environment are moving faster. And I would say historically, we would not move very fast. You get tied up with, you know, trying to get the BU aligned on this being, you know, a worthwhile investment. And then the BU may think, well, this better than what they can do. And it's this NIH mentality. So given, given, exactly given, given Gen AI, I think it's putting a lot more pressure on just, you know, cvcs need to up their pace a little bit to remain competitive. So, you know, the that, I think, is a misconception that is not the case. Now, if you have a if you're have a good team, you have a good product. Cvcs will not be the long pole in the tent when it comes down to closing around

good team. Yeah,

like to echoing what Didi was saying earlier. I do believe the venture industry, the long term investment industry, we know each other. All the forms are pretty well connected. The startup have words of mouth as well how each firm treating the startup, the portfolio companies exit pass the whole thing the industry always have words around it. So we try to behave like good VC trade our startup CEOs with respect. And really, I do encourage the CEO out here, in here the startups, when you choose investor, right? The investor always asks you for your customer reference, for your management team reference, do the same for your investors. Check your investors reference, what did how did they behave when company are going well? What value did they add? How did it behave when things didn't go well, right? How did they behave when they potentially change a CEO on a board or different tough situation? Do your homework, to your reference, check as well before you choose your investors.

For us, actually for us, the decision making for Qualcomm Ventures is fairly independent within Qualcomm ventures, so in cases where we actually take a startup through a BU ringer, it's mostly to get conviction within the group whether the technology works, or whether the product works, or whether we can help them through a go to market channel and actually help them scale better. Scale better. So so the decision making, yeah, it can range from anywhere from the shortest during covid days was two days to two of six. Typically it's around five to six weeks, is what I would say our time horizon from making a decision. But, yeah, it's fairly independent within our group, though.

So we have a few more minutes. I want to do a quick poll. The current gen AI wave started November, the year before. It's been almost 18 months. What's your each of the panelists prediction? When will this wave crash, and we can check the answers next year in the Gen AI Summit, see who is correct or closer to truth. Who want to start I know the exact

answer. I keep you to my stomach.

Predictions, quick predictions, how many months

collapse or hit like an air pocket

goes? South. Yeah, once you go south, it's pretty quick.

Well, I think that there could be an air pocket in 25 with respect to just, you know, inventory levels and but I don't think it's a crash, per se. So that would be my comment, like

a true semiconductor veteran.

I honestly don't know. I think that investors always need to be prudent, and they need to be prudent right now as well. You know, we think that there are lots of jewels. We hope we found some of them, and when you find a jewel, you're willing to pay more. But you know, it's, you know, thinking about, you know, each, each cycle is very different, right.com? Was one thing financial, you know, 8000 I was something else. You know, something happened in the past several years. And AI is an exception. And we'll see you.

Think we'll be as enthusiastic, excited about Gen AI this time next year.

I think next time, maybe, yes, but if you ask me, like, three, four years, it will take a different turn. I mean, again, you know, it's like playing soccer or football. The goalpost is moving all the time. You know, it's very strong runway. Now it will continue, but it's maybe a completely different direction.

I'm going to answer your question going around about a way, and then answer your question. So first of all, doing startup are extremely difficult. You need to do continuous innovation. You need to adapt to the regulatory environment change. You need to face the challenge of the financial pressure. You need to deal with supply shortage from time to time. And those are all true for generative AI companies. So it's never easy, even when things are going well and and when things start to I don't know if it's going to crash or it's going to slow down. But in general, we would advise our startup, even when they're raising this round in this good environment, when there's a milestone you think you can raise the next round and that runway, we would encourage everyone to have two to three years of runway to be prepared for any kind of crash,

you know, crash, I think so, I'll tell you, and I could be wrong tomorrow. I don't think, I really don't think we're going to see a crash in the next 12 to 18 months. Here, the reason why is because of the pace of innovation. So the pace of innovation and this technology has just been so fast, like just this morning, you know, the first application of state space models launched, and that's exciting. And, you know, striped hyena would be ideal, and something else will come up. And next GPT five happens, and then you get another 24 months of hype. So, yeah,

so it seems like the panel is very optimistic, at least by the same time next year, we'll still be very excited about all the new things come out of Gen AI and innovation and all these investments. So to that point, I want to end on high note and thank our esteemed panelists and give the whole team a round of applause. I don't know if we're at Okay, one, one questions, yeah.

I think this whole thing will will be crushed once we opened this black box, the black box, the technology, yes, we can use Billings. This parameters to change model mathematically. Yes, you can do it. Suppose you have a sphere and only few data on the CRP. You can use a three dimensional box to enclose it. But from my mathematical research, I'm a mathematician. I already see many, many of the neurons that we are using for nothing, no use. I can show you how this happened. And I used to develop the geometrically unified learning. You can see it explainable, and it's supported by National Science Foundation. I can open the black box you will see all this investment that you are doing by such a big computers resources will be wasted.

Yeah. Well, thank you for the comment. I think, I think Elena mentioned this. I also agree that to some extent that we're using, brute force transformer model has shown amazing results, but at the cost of huge computer resource. But there are researchers looking for alternative modeling architecture, and that will be proximate in the next 12 months.

If I can make a comment or a question, how much is 40 725 times, 80, 233

no idea. No, yeah, right.

We don't do mattress multiplication in our, in our in our 15 watts here, right? So we're doing brute force, as you say. And you know, as a venture team, of course, we hope to find a company that will do it completely differently, right? Of course, they're fantastic companies that have amazing technology to crunch the numbers. But you know, if you ask me, not in 12 months, but in 20 years from now, it will have to be different. It will have to Yeah, it's amazing

that I just attended my PhD advisor's retirement party. I finished my PhD 25 years ago, but my topic, surprisingly, is still relevant today, because I was researching model complexity, the error bound, it has two terms, the bias and variance. Bias with so many parameters, like trillion parameters, you can get any training data set to zero and but the variance is a tricky term that is the model complexity based term. So anything that you change the complexity a little bit, the error bound on some of the old training examples or new training examples start to go offline very quickly. So I can, you know, really, you know, emphasize that comment from the true mathematician, but it's too technical for the time we left today, which is,

you shouldn't be too technical. All right,

thank you, and give it a round of applause to all the Oh, one more. One more. Okay, one more.

Hi, really great panel. My name is Anu. I'm a partner at KPMG, and I actually lead our studio. We also have a CVC that's part of our business, and one of the questions that I had for you was a big part of cvcs, is to be able to learn from the startups that are doing things differently than the incumbents are right? All four companies up here are incumbents into the market. What mechanisms do you use to get the learnings from these startups back into the business.

Well, I can acquire them and then put them into production.

Maybe I'll be the no comment on

this. Okay, first of all,

I want to say KPMG is our auditor, so I don't know if I can

different different team,

I would say this is a very sensitive topic. We're extremely careful about the data, information, the IP, the know how, the whatever of the startups. We're very, very careful operating with Chinese walls, making extremely careful that information will not flow from the venture team. We're a separate legal entity. We only share company material, raw material, with business units, with the permission of the startups. Having said, right, we meet hundreds of companies. There's some level of integration that you see, right? So it's not like this specific company is doing this and that, let's call the BU and tell them do it. Absolutely not. But after you meet hundreds of companies, and also we meet the other cvcs, we meet VCs, you know? So it helps you form some opinion and understanding on trends, on where things are going. We're not trying to help the BU how to design a transistor or something, right? It's more on the higher level of where things are going. What could be the opportunity? It's really an integration stuff. It's not like specific material. Sometimes it is. We ask the permission of the company. If the company is willing, then, you know, sometimes they want to do it right, for the sake of collaboration, but you need to be extremely

careful. Yeah. It's increasingly about talent, right? And it's no coincidence that almost half of the interactivity happened in the state of California, northern and southern combined, right? Because non compete is not enforceable in California, but it is in the other states, right? So you wonder why that happened. All right. Do we have one more questions or Okay, last question? All right.

Election, yes. Do

you guys have any no problems? Yeah, election, concern, politics, religion,

these are the things we don't.

All right. In the back, I think, yeah,

so the questions is, for Tushar and the Jeff, what do you guys think about the consumer electronics play in a Gen AI space?

I can go, I mean, we're seeing so something that for us, it's very, very important. We're seeing a lot of startups that are building dedicated AI devices. We're also investors in humane but so it's, it's definitely very exciting. The question to ask is, I think a lot of people are asking, which is, do you need a separate device for AI, or will your smartphone be the AI device that you use for a lot of the use cases. I've got a humane device myself. Of course, it's got some issues as a gen one device, but there's a lot of use cases that are enabled by being more context aware with the camera that you can't do with your smartphone. And at the same time, have you tried using the assistance on your smartphone? All of the assistants in your smartphone are still very, very dumb, having said that, you can see Google and Apple probably releasing smarter versions of those assistants, but it's something. It's a space. We're also very excited about AR glasses, so similar to the AR glasses as well. So there's several companies innovating in that space. So I do think, like in a year or two from now, we might see more proliferation of some of these devices.

Yeah, from where we play today in consumer electronics, which is primarily your PC, your laptop, or if you own an Xbox or a PlayStation, we're powering those systems. I think what we're excited about is AI is just going to play a bigger role in those devices. We're not necessarily going to be playing in your glasses. We're not as far on the edges are as my friend here on my left, but where we where we currently play today, I think you just look at it as AI is going to become more and more of a layered offering on those and we're excited to see those markets develop based on having that technology embedded.

I recommend you purchase one Samsung Galaxy, you know, s 24 ultra super before the end of the quarter, please.

Yeah, all right, we need to get out of here, I guess. Thank you, everyone.

Thank you so much. Thank you so much for your time.