AI Investment: What Makes This Time Different? Hani Enaya, Brad Gerstner, Jeff Horing, Umesh Sachdev

If you have more than a five-minute conversation with me, you'll quickly learn that my favorite word is 'interesting.' But I think it's fair to say that this is an interesting panel, mainly for our distinguished speaker we have here today, but also for the topic at hand. AI has been top of mind for investors, executives, and founders. I believe this is in no small part due to the 100 million users, users like you and I, who started using ChatGPT. We've seen firsthand the utility, but also the impact it can have.

By AI winters, where development in AI seems to greatly stall. So why now? Why is this time different? If you allow me to share a few data points from McKinsey's Global Survey on the state of AI: one-third of respondents said their organizations are using generative AI regularly in at least one business division. Nearly one quarter of the surveyed C-suite executives say they're personally using AI tools for work, and 40% of respondents say their organizations would increase their investments in AI because of the advancements in generative AI.

And as another... Investment Chief Investment Officer at Saabel Investments, an investment company focused on private investments like venture capital and growth investing in founders and fund managers across the globe. Brad Gerstner, founder and CEO of Altimeter Capital, a lifecycle technology investment firm. Jeff Horing, co-founder and managing director at Insight Partners, a global... I'll ask you to lead us and set the context. How did we get to the current wave of AI that we're seeing right now?

Well, first, thanks for having me. An incredible conference. I hope you can hear me; it's a little loud in here, but that's only because there's so much excitement and momentum in Saudi right now. As I said to H, the last time I was this tired was Burning Man. So we're covering a lot of territory, as I'm sure everybody on the panel can attest. Those of us who've been investing in the internet and search and software for the last 20 years have been thinking about augmented intelligence, when we're going to get answers instead of links. With large amounts of unfettered data that is now set free in the cloud, coupled with a Transformer model that allows us to semantically understand data in a way that we couldn't before, it has created a kind of Cambrian moment. I think your first point is the most important one: every super cycle in the world, whether it's internet, mobile, cloud computing, is only a super cycle because it makes consumer lives better and enterprises businesses better. The reason this moment is so profound: Satya, the CEO of Microsoft, who... faster than the Explorer browser. And that's because businesses are seeing transformations, productivity increases, step-function productivity increases among engineers, call centers, and other administrative areas that they haven't seen before. So I've been waiting for answers for 20 years. It's exciting to get to this moment, but to also remind ourselves that we are at the first inning, the beginning. I like to think of it as augmented intelligence, not artificial intelligence. You know, we haven't removed the pilots from the cockpit of the plane, but autopilot has changed how we fly our planes.

Amazing. Just to set the context, when people speak about AI, I think it became a household name mainly because of ChatGPT. It's like the spark that lit the fuse. But let's always remember, AI started... what's the core of AI? It's literally compute or cloud compute. And when it comes to computers, PCs, CPUs, and GPUs, that started six years ago. So the wave started six years ago. Then the internet came 40 years ago, and then 20 years ago, when Brad started waiting, mobile and internet came. And this all converges in what we know now as generative AI, which might lead to an AI singularity in the near future. So Jeff, Brad mentioned we're in the first inning, and we've sort of seen... Look at the internet and mobile as two phenomenal, revolutionary data points. I put this closer to mobile than to the internet. I think the internet actually took about 10 years. There was a lot of excitement and enthusiasm in the late '90s, but in fact, the real business models didn't start to emerge for almost a decade after that. Whereas Apple, certainly the second, had introduced its phone. Overnight, that was a sensation. The apparent use cases were obvious, and the adoption curve was tremendous. I think this will be somewhere between, but closer to mobile. I think there's still a lot of evolution to go on the products, but I still think that the adoption rates are going to be closer to mobile than they will be the internet in terms of how this picks up.

You know, it's interesting. We can talk about the trends of AI investing in AI and the evolution, but here's how it's actually done. We have today over 1,500 enterprise customers around the world, with over 750,000 users within those enterprises using our products. And like you said, we've been doing this for the last 16 years. What started with a focus on speech and conversational AI and natural language gave way to deep learning, transitions, and more recently, Transformers, diffusion models, and generative models. But the true opportunity for multimodality, having AI models using computer vision to see... close to what? One of the largest coverages globally of call center applications over time. We then expanded to other use cases in the enterprise, sales automation, etc. And today, every part of the enterprise that can impact the customer, the entire front office, is ripe for disruption. So are other parts of the enterprise, but the front office, particularly the way customers are used to interacting with enterprises when they're thinking of buying something, when they have a problem after they've bought something, and reaching out for customer care. The fact that we do this in 20 different countries, over 100 languages, we've seen this vantage point, and we've seen customer maturity evolve. From our lens, which is a service provider's lens in AI, the opportunity unlocked earlier this... etc. But now the CIOs care a lot about data platforms, knowledge, and learning, feeding those generative AI models. Therefore, they are very thoughtful about how to not have one AI in the call center and one AI in marketing. Now, the CIO is beginning to say, 'I want a cluster of applications on a single platform, and I want to pick the co-pilot example of Microsoft.' So if it's a coding copilot, there's going to be one type of platform within the enterprise. But if you think about call center, sales, summarization of content in general, that's going to be a singular platform in the enterprise. And probably there's going to be a third, but people are going to start to consolidate AI investments.

You think value will accrue, and how will value capture look like in this market?

Yeah, so again, if you think about the sequence of events, the primitives to AI are data and compute. So we talked about the choice of data platform. Unless your enterprise has all the data in the cloud, right? And this is a challenge. Most large enterprises use hundreds of databases, some of which are still on-prem, some of which are in the cloud. They don't talk to one another, and you can't leverage the full power of your data unless that data is unified and cleaned in the cloud. So we've seen a re-acceleration in the hyperscalers. We're a large shareholder of Nvidia. And if you look at Nvidia's price, everybody says, 'Well, the stock has tripled this year.' I would take the other side of that and say, 'Why was it ever valued at $125 a share?' The consensus estimate for Nvidia this year was that data center revenue growth was going to be -6%. Right? That's the pros. They got that wrong. And what they got wrong was that AI, to Jeff's point, is being adopted faster than they ever forecast because the benefits they're delivering are much more significant. If you're in a competitive industry and you're not using AI in code generation, to run your sales centers, to transform every... more than all the compute ever deployed by humanity combined. That's how much compute is going to drive these training models. In terms of durable value capture, there's a lot of speculation. There are a lot of question marks that people are asking about. Have they pulled forward all the demand? When we move from training to inference, will GPUs be as important in that business as they are in training? I think, as we saw yesterday with the ARM announcement, Nvidia will continue to evolve. When it comes to models, I think of it a lot like search in 1998. There are dozens already of foundation models. In 1998, everybody knew... Vista, Ash, Geves, Infoseek, Excite. Go through the list of all the search businesses. They got the internet right, they got search right, and all of those businesses demonstrated they did not have durable value capture. And Google ran away with it. So I think it's very hard right now to look at a closed model, whether it's Anthropic or OpenAI or Cohere, go through the list, and say that the revenue you're seeing today is durable. When we look at the future, clearly Meta has a different opinion. They hope to commoditize the foundation model layer with Llama 2 and open-source models. There are a lot of people who think that there'll be many, if not... You didn't have to invest in search until Google's IPO in 2004 to capture 95% of all the profits ever generated by internet search.

Amazing. And Jeff, so I want to ask you: a lot of the value seems to be accruing with the incumbents, right? So Nvidia, the Microsofts. How do you view that in contrast with how can startups start capturing value?

Well, first of all, I agree with Brad. I think there are two obvious winners: the infrastructure players and the customers. I think we could both say they're going to win. I think it gets a little trickier when you start to look at the moats, as Brad describes, of which of the earlier, less mature companies are going to win. I do think an incumbent could be defined as a big... businesses like Microsoft or others adopting AI at a pace that we've never seen before in a new technology launch. Cloud computing largely revolutionized the tech stack, and most of the players today in cloud computing were not existent in the early 2000s. I think it might be different here. I would not be surprised if the majority of value capture is with the incumbents, the established players who add this feature set. There will be white space, inevitably, that comes out that no one is really positioned to solve today, that new entrants will solve. So I think the startup community will do very well, but I don't think it'll be a wholesale switching of leaders as it had been in prior technology generations.

I want to take it to a mesh...

One thing I'll slightly disagree with Jeff's original comment of this being similar to mobile. I think the transition that we are going through right now is a lot like internet and mobile in that it is transformative. But the one key difference is that the pace of evolution with this AI transformation is one that we've never seen, even in the software industry in the last 20 years. Every 48 hours, the pace of innovation is intense. So we speak about GPUs. We are a big partner to Nvidia, obviously a big consumer of the technology, etc. But once again, this compute is going to move to the edge very fast, and that's going to open a whole new paradigm that none of us on this stage today are thinking about. If you think about the opportunities with... that customer data flowing through our systems for model improvements, so now there's an opportunity to create a foundational model exclusively built on enterprise first-party data. And that is a whole new paradigm. None of the existing foundational models are thinking about that. So the point I'm making is that the pace of innovation with AI is so fast that if you take the next six months and try to predict six months out what's going to happen, the compounding effect of every two days learning something new about compute, infra, models, or applications, that's a fool's errand to try and imagine what the world will be six months out. Now, of course, for a company like ours, we don't have cloud, we don't have infra, so obviously we partner with... related to the front office, we've been doing that for 16 years. Our models are purposefully trained for those use cases. That gives us a different advantage versus somebody trying to do a coding copilot, where my model won't be as strong as somebody else's. So it's one where truly the comparison is this is the internet of 1998, except it will move much faster than internet or mobile ever moved.

I'll take it back to you, Jeff. So he is saying it's going to move much faster. Do you agree with that? Are we headed very quickly toward that singularity, or should we brace ourselves because another AI winter is coming?

I'll do a little combination here. So I think that's why when I contrast mobile to the internet, I thought mobile was a shift. You know, I think there's a chance that we're going to have an aha moment here in the other way. I feel like we might be at peak expectations. And we've had this with a lot of other technologies. I think this is real, it's going to work. But if you look at my favorite to pick on, self-driving cars, we had that seven years ago where we were all out in the back seat reading our books. And here we are, I think maybe not a year closer to that solution seven or eight years later. So I think there's definitely some concern, risk, that maybe we think that some things that are going to get replaced are more likely to be co-pilots. Augmentation is a good word that I would use, where we're not quite going to get to the level of automation that we all think. So I think there's a little bit of... comprehension, but getting to 100% might actually be a 20-year journey.

So Brad, how do we navigate that 20-year journey? How do you distinguish what's real and what's hype as an investor? Do you have a framework to generally deal with hype cycles, and is there any sort of specific AI framework you're using to evaluate the opportunities?

Well, part of the reason we're scaling so much faster is when I started investing in the internet, and Jeff did, there were probably 30 million people connected to broadband internet. Today, we have 3 billion people with supercomputers in their pocket. So changes that occur are rather instantaneous, and we push them out... day in Microsoft. 117 billion people have inhabited the planet. For 110 billion, they never saw a single invention in their entire lifetime, because their life cycle was shorter than the invention cycle. And so I think that this will be the fastest thing we've ever seen. It will have more impact on humanity than the internet itself. And the reason it will occur quickly is because of that. I'm not that interested in how long it will take us to get to AGI. I mean, I'm interested in it because we sit around on the All-In Pod and we debate the singularity and is it good or is it bad, and you know, does Elon really think that we're all going to become cats? Ideas not well understood can lead to excess regulation, and we don't want to over-regulate something that I think can have this much positive impact on humanity: breakthroughs in life sciences and medicine at a rate we've never seen before, energy storage and electrification at rates we've never seen before. To me, I think this is a very optimistic moment in time. In terms of investing, our framework doesn't change. It does help to have the pattern recognition and the scar tissue that Jeff and I have, because you can be right but too early, and it's equivalent to being wrong. So... don't think those are good founders or maybe even good business models. But the idea that we could forecast out for 10 years the durability of the revenue, like we can with a software company, would be foolish. And so the valuations on many of these companies have gotten ahead, at least from a returns perspective, where I think the durable value capture is. So you have to hold these two uncomfortable simultaneous truths: AI will be bigger and more impactful for humanity than most people currently think, but we also may be at a moment where we have peak excitement from a venture capital valuation perspective, the equivalent of everybody scrambling.

About riding that or guiding the founders in your portfolio through that wave.

So I think much like my colleague next to me, I think the opportunity for existing companies to adopt is a slam dunk. I mean, that's the most obvious starting point. You've got some ability to actually transform okay businesses into very interesting businesses. And if you look at what, take Adobe as an example. Photoshop probably appeals to 3 to 5% of the population on a good day. Maybe 3% really know how to use it. If they could use an AI front engine to open that market up to 15 or 20% of the population, wow, what a home run. Same product, just a better way of interfacing with it. So I think our existing... really profound and successful, and they've got to reposition themselves to a world that says, 'Wait, that's going to be commoditized now.' You know, understanding conversation in a Zoom call is not where I'm going to differentiate myself. I need to have something very different. So I think the portfolio is going to stand to benefit tremendously from what's happening. At the same time, a lot of our initiatives are focused on getting them operationally more efficient. So I've got a thesis and an effort inside of Insight to take over the next 24 months: how do we create 10% productivity across every operational function in our software companies? I would challenge anybody who has a traditional business; they could have the same upside. I think there's real opportunity to make companies more efficient with this technology, and a whole bunch of...

How do you think about your product roadmap, your market roadmap, your fundraising roadmap as a CEO and a founder of a company?

Well, I'm going to try and make it real for everyone in this audience. I'm going to talk about two of my customers. One of them is one of the world's largest cybersecurity companies. They have 2,000 people in their customer care group, and the CEO has gone to his CIO and said, 'I want our AI in my customer care group to go down from 2,000, not 1,500, not 1,000, not 500, to 20 people.' Okay, that's the goal. I want you all to come up with a strategy, use tools and technologies over the next couple of years for over 90% headcount reduction in customer care. I have another customer... to make sure that as the sales team scales, all my sellers are well trained and they're saying the same thing, and my sales efficiency is top-notch. Saying now, 'I'm too large. The traditional sales approach won't work. I want to use AI, and of my 150 people in RevOps, that RevOps function would be completely automated.' Both of these companies trade on Wall Street in the tens of billions of dollars. And all it will take is one of them to come very close to their goals: either have sales efficiency just break out from their peer group, or have margin because customer care cost has dropped by 95% break out from the peer group. And one thing Wall Street knows how to do is when it smells money, it spreads the message to every constituent. So the reason I bring this up is that the financial benefits would drive this motion forward at a pace that we're not used to. And then the other thing that most people are ignoring, which is surprising to me, is the difference between an AI-native software business versus a software business. The COGS of a software business is well understood: SaaS gross margins are 70-80% plus. An AI-native business inherently requires compute costs at a pace that software businesses did not. So AI-native businesses, how do you make sure those businesses still have gross margins where SaaS investors were comfortable? So the business dynamics are going to be different, new unit economics...

And maybe there's some hype cycle in some innovations, but this is a train that's left the station. Hany, I'd love to get the regional perspective from you. How do you think, or what opportunities do you see in AI or because of AI that are specific to Saudi or the region?

Oh, great. I believe there's a tremendous opportunity in Saudi for AI and because of AI. I believe this is driven by the presence of a very strong political will, the gila... and also the growing digital infrastructure with a very savvy, digitally savvy population. And I believe to spread the mass adoption of AI in the region of Saudi, we also... compute might seem like the primitive core for AI. I believe there is an even more primitive core that we discussed yesterday. So Brad mentioned compute next year will be used in a way that's louder than we ever used as humanity. But this needs energy. Where will this energy come from? Right? So I believe in the next phase of innovation, the presence of abundant, cheap, renewable energy will be a must, not a nice-to-have. And I believe Saudi and the region can play a critical role in that.

Brad, you helped us open. I'd love for you to help us close. So peering into the crystal ball, what excites you the most? You seem to be net positive on AI in general. What excites you the most?

Chance that it's an existential threat. He's happy to be alive to watch because he said either way, it's going to be pretty damn exciting. I agree. I think in many ways, over the last 40 years in technology, we've been laying down the tracks to increase the pace of invention in ways that will improve the lot for people. And so here we are. We think about medicine as an example. Life expectancies have increased on almost every dimension. People are living healthier lives on a global basis. And I... limiting factor will be energy consumption for these models. They're going to need to build their own sources close to the data centers that are powering this compute. But at the end of the day, I think the question as to whether or not it's gotten ahead of itself comes back to this very basic element that we've talked about up here: businesses are already benefiting. This will not be an option. The question will seem silly in a few years. Would we ever ask an enterprise whether or not they use the internet? If they don't use the internet, they're not going to be in business. Right? AI will be as essential to the... that are driving every ad recommendation that you see. So it's already threaded through so much of our day-to-day lives. And I think now we're going to see purpose-built applications, whether it's in healthcare, whether it's in national defense and security, that will be leveraging these technologies.

Amazing. And to your point, we're already seeing it in the hands of the consumers, right? So people have the ability to be better writers, better researchers, all because of AI, and to recapture time into their lives. At the end of the day, there are a lot of jobs that are going to be displaced by AI. I think social contracts are going to have to be... appears in places like Doordash and TikTok. Their objective is to get that down to below a thousand by the end of next year, leveraging AI. One product, one company, 35,000 fewer people. So we have to think about what are these folks going to do, right? Because the technology will not keep up with the pace of the dislocation. And I think that there are folks who work in call centers, content moderators, etc., that are especially vulnerable. And that's why governments are going to be critical in terms of being involved.

Amazing. With that, I'd love to open it up for questions.

Very soon, I want to get your opinions on how you think that's best going to happen. Do you wonder? I guess there's two sides to that. One is, a lot of these AI models are based on somebody else's copyrighted facts, if you want to call it that, or content. And there's a big open issue with a lot of litigation to come around: is it fair to scrape a newspaper and incorporate it in my AI model? Is that the same as a human reading the newspaper and incorporating it in their knowledge base as it relates to facts? I don't know that there is an answer to that, and I don't know that that's a fair assumption to say that it will happen. I think it's a nice assumption to think that it will happen.

Perfection of the AI models is because it's only as good as what it's been trained on. And ultimately, that's going to be not as good as the best human beings. I tend to think that the models and applications that produce the best answers, the marketplace will choose them. We're not going to have a single model that gives you an answer. Any enterprise that has front-facing, customer-facing answers that go out, they're going to have other AIs fact-checking it. So you're going to chain together a series of models to ask the question whether or not the answer that a certain model propagated is in fact agreed upon by a whole group of models.

Of that corpus, then they won't be incorporated because the model has no ability to produce answers outside of those that it's been trained on. I add that just as the past few trends, internet, mobile, have given rise to newer categories of software toolkits, etc., in the world of AI, given everything that we've discussed, explainability tools: how does your model not become a black box? But when challenged in the court of law, you're able to point to how did my AI model come up with that answer? How much was fact, how much was fiction? Observability of data flowing through those models...

More. I'm curious about, as you think about the deployment of models, if we look at, for example, TikTok, they use a single recommendation engine but for multiple geographies with slight differences. When we look at UAE developing Falcon, maybe Saudi developing something similar, with companies that are using these models in their products, how does this end up getting integrated? Are we going to see an OpenAI model deployed globally across systems? Are we going to see each company adopting different models in different geographies? How do you see that playing out at the large scale?

I mean, I can give... there's a view emerging between countries and large enterprises that ultimately points to... a company that is 60% owned by a French-domiciled engine. So Google and Microsoft or their clouds have joint ventures with French entities like Dassault and others. So to think that countries or large companies will not be very thoughtful about sovereignty of data and AI models, that's not going to happen. And we're beginning to see some of that thinking take shape. It will ultimately result in forms of regulation and so on and so forth. And so what's important is architecturally to develop these models to support those future requirements of country or company or specific use case sovereignty, while giving it the benefit of the global nature of these models.

Saying that this is evolving, but to my mind, sovereignty is a thing that will become louder. I think a lot of what happened this year is ChatGPT gets launched last fall. CEOs go home for the holidays, and all their kids are asking them what they think about ChatGPT. 'Have you used ChatGPT?' The CEOs show up for their first quarter board meetings and they say, 'What are we doing in AI? Are we using ChatGPT?' And what I've heard from one of the largest banks who's here, in the first quarter, they quickly entered into a deal with OpenAI to start testing the model in... enterprises. They understand that they have to begin testing. Nobody wants lock-in, just like they don't want lock-in to a single cloud or a single data provider. They don't want lock-in to a single model. So I think we're still in this test-and-learn phase. And I think it's unknown and unknowable whether or not you're going to have a single omniscient frontier model emerge as the winner. I think it's much more likely that you're going to have different models tuned to different use cases. If you have a model that's trained on sales centers, it just may be better, right? If you have regional models, national models, they may be better. Both... that we're looking at. And what we found is that we use OpenAI for GPT-4 for certain things, we use Claude for certain things, we use Llama 2 for certain things, right? So I suspect at this stage of evolution, you're going to see most large enterprises using many, many model providers.

I think you'd agree that that was interesting. Please join me in thanking our amazing panelists with a round of applause. Thank you.