Capital and Distribution in AI, and the Rise of Neoclouds with Kevin Zhang

Kevin is an investor at a family office, where he leads AI investments across asset classes. His career has spanned roles as a venture capitalist, startup founder, and software engineer, with experience in both Silicon Valley and New York, before moving to Asia. He brings deep technical and product expertise across domains from machine learning to enterprise software. In his spare time, Kevin writes East Wind, which is focused on technology investing.

In this conversation, Kevin Zhang shares his insights on the evolving landscape of AI investments, the implications of hyper-scaler capital expenditures, and the future of AI model training. He discusses the cultural differences between investment ecosystems in the US and China, the valuation of private market companies, and the role of neoclouds in the AI sector. Kevin emphasizes the importance of capital and distribution in determining the success of AI companies and reflects on the future of work in the context of AI adoption.

In today’s world, there’s no shortage of information. Knowledge is abundant, perspectives are everywhere. But true insight doesn’t come from access alone—it comes from differentiated understanding. It’s the ability to piece together scattered signals, cut through the noise and clutter, and form a clear, original perspective on a situation, a trend, a business, or a person. That’s what makes understanding powerful.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

For more information on the podcast series, see here.

Chapters

00:00 Kevin Zhang's Journey From Software Engineering to VC to Equity Investment

02:04 The Hyper-Scaler Capex Debate

04:31 The Capital-Intensive Nature of AI Models

07:49 Future of AI Capex and Market Dynamics

11:43 Understanding Private Market Valuations

14:49 Consensus Capital and Investment Strategies

17:10 Cultural Differences in Investment Ecosystems

21:50 The Future of Chinese AI Companies

23:50 Capital and Distribution in AI

27:47 Open Source vs. Closed Source Models

32:22 The Role of Neoclouds in AI

40:40 Investment Opportunities in AI and Beyond

Transcript generated by AI

Grace Shao (00:01)

Hi everyone, this is Grace Shao. Joining me today is Kevin Zhang. Kevin is an investor at a family office where he leads AI investments across asset classes. His career has spanned roles as a venture capitalist, startup founder, and software engineer, with experience in both Silicon Valley and New York, before moving to Asia. So he’s now based in Asia. He brings deep technical and product expertise across domains from machine learning to enterprise software. And in his spare time, Kevin writes a blog on Substack called East Wind. Go check it out.

It’s focused on technology investing. So Kevin, thanks so much for joining us today.

Kevin Zhang (00:33)

Hi, great pleasure to be here.

Grace Shao (00:36)

Yeah, tell us about yourself. You’ve had quite a journey, you know, from Silicon Valley and now into your base in greater China. You’ve worked as an engineer and now an investor. You know, that’s quite unique. Tell us about your professional journey.

Kevin Zhang (00:50)

Great. So I guess going all the way back to my college days, ⁓ studied computer engineering, both in Canada and then states for grad school. And then spent most of my career in the States ⁓ in Silicon Valley and New York. So started as an engineer at a company called Salesforce. So they make CRM software before ⁓ transitioning to a couple of venture funds, ⁓ one in the Bay Area, one in New York. So primarily the focus has been on

early stage software, AI investing. And in these days, I look primarily at public equities, specifically focusing on the companies that have exposure to AI or ⁓ businesses who ⁓ will see a re-acceleration of growth because of AI.

Grace Shao (01:39)

Awesome, so we’re gonna go straight into this. I think we have a lot to cover. ⁓ First, let’s get into the hyper scalar capex kind of story. We’ve been seeing jaw dropping capex commitments, alphabet raising, AI capex too, 85 billion, meta committing hundreds of billions, ba-ba 10 cent I think over the next couple years, also committing like 300 billion upwards. It’s just these crazy numbers.

Do you see this as a necessary investment or do you think they’re absolutely overspending right now and they’re creating a bubble?

Kevin Zhang (02:11)

I think if we look at past ⁓ technology cycles, there will always be, or there has always been ⁓ some overspend across the industry. ⁓ So I think ⁓ bring it down to on a per company level, I think things get more nuanced, right? So looking at, for example, alphabet, right? How much of that is internal workloads for search, right? For their Gemini answers, ⁓ looking at open AI where how, ⁓

whatever their, can we maybe start this again? Let’s do a rewind. Okay, great. ⁓ So I think across the industry, if we look at past ⁓ technology waves, ⁓ most famously in the ⁓ initial tech kind of a dot-com bubble, the industry has overspent, right? But as we zoom out, ⁓

Grace Shao (02:49)

Just go ahead, just restart your answer.

Kevin Zhang (03:11)

the spend becomes more normalized and then the demand ⁓ ends up catching up. And so I think the real question is on a per company basis, right? Whether it’s Alphabet, Meta, Azure, with Microsoft, AWS, how much of that is overspending? How much demand can they generate, right? With their ⁓ whale customers? And then if they end up overspending, how many of these players can survive, right? So for some of the smaller players,

who are comparatively more leveraged, who don’t have the cash flows to support some of the CapEx, right? They’re maybe in a little bit more of a dangerous territory than one of the Meg-7, where they’re generating a significant cash flow to ⁓ fund their operations.

Grace Shao (04:01)

And I think this kind of goes into also one of your writings recently. You were saying that, like, look, it’s all these big tech that are able to afford ⁓ spending on, like, LLM training and inference, as well as whatever infrastructure that’s needed to really build out, like, sophisticated LLMs. And you said it’s basically foundation models are a rich company’s game. Why is it the model layer is so capital intensive? And do you think that means we’re going to see the startups just kind of

in this field kind of just die out one by one or acquired or what’s trajectory going forward?

Kevin Zhang (04:38)

I think there are two paths that the industry can take. So the default path is if we look at ⁓ the progression of ⁓ costs for model training, whether it’s pre-training or post-training, ⁓ each generation has been significantly more expensive. So ⁓ many years ago, it might have been ⁓ several hundred thousand dollars. Then it went to the millions, tens of millions, hundreds of millions to train a model.

So assuming that trend holds, ⁓ we’ll see kind of billion dollar training runs, right? So taking a billion dollars to train, let’s say GPT-6 or GPT-7. And so if that is kind of where the world goes, then these companies will need to raise more and more capital to fund their training. They will raise more and more capital to fund their inference, right? So once you train a model, how do you serve it to...

⁓ and users, right? That’s also very, very expensive. ⁓ However, if you’re of the view that there will be, or transformers, which is kind of the models that are used in ⁓ LLM or ⁓ used in things like ChatGPT these days, ⁓ if you believe that there will be other model architecture paradigms ⁓ that are going to be significantly cheaper, then maybe there will be another kind of startup that comes and disrupts.

the entire business model of an OpenAI or Anthropic or any of these labs.

Grace Shao (06:07)

and was DeepSeat one of those that kind of disrupted the whole model.

Kevin Zhang (06:11)

⁓ I think not necessarily. think ⁓ going a little bit into the weeds, the, I think, five, six million dollar ⁓ final training run touted in their paper, ⁓ that was only for the final training room, which is not inclusive of the GPUs that they’ve acquired, ⁓ their human capital.

all the prior training runs and experiments that they’ve run. And then also within kind of AI training, if you basically train last year’s model today, right, it’s significantly cheaper than if you want to train a frontier model. ⁓ And I think Anthropic had a paper where, or had a blog post where basically ⁓ for if they want to train a similarly similar model to the DeepSeq R1 a year ago, it would have been 10x more.

Right, so essentially the costs track. And so it’s less so disruptive ⁓ than I think some of the folks in media might have thought. And certainly they have ⁓ made certain architectural improvements as well as inference improvements at DeepSeek.

Grace Shao (07:32)

So like if that’s the thinking, like looking ahead next three to five years or even a longer run, is this like the hyperscaler arms, I guess, if you put it that way, is it gonna just keep on climbing up that capex or will it eventually plateau or is this question we will never know.

Kevin Zhang (07:49)

⁓ it depends on like, like how long the labs, right? The, the open AI anthropics, even Google can sustain, ⁓ this pace of model improvements, right? So we we’ve seen a little bit of plateauing in, in, in the past year or so, ⁓ as, as we’ve kind of reached the limits of pre-training, right? So, ⁓ right now, a lot of the emphasis is on like these like thinking models, right? So, ⁓ when you actually type a prompt,

into ChatGP, you know, think for a while and then that tends to ⁓ generate better answers for you. ⁓ And so it feels like the like assuming compute requirements continue to wait, what can we start the question again? What was the specific question? Like what is the arm? ⁓ Okay.

Grace Shao (08:43)

Yeah, I’ll just redo it.

So if we use this logic, right? Like, does that mean that in the next three to five years, we will in the longer run that the capex numbers will just continue to climb? Or eventually we will see this hyperscaler arms race kind of plateau out a little bit because I don’t know the day, like these are already like crazy numbers, right? Like we’re looking at a couple hundred billion dollars put into training in the next three to five years. That’s the plan. But how does one keep up with this kind of money?

Kevin Zhang (09:13)

Okay, so there’s two things, right? There’s training and inference. ⁓ And so on the training set of things, the assumption is ⁓ labs will continue to require more and more compute to train more and more expensive models, right? So let’s say the next model takes 10 billion to train, right? And the model after that takes 50 billion to train. Then theoretically on the training side, that tracks. And then I think where a lot of this capex is going, especially if you look at

⁓ the Meg 7 where they’re putting 70 to $100 billion ⁓ per year per company. I think a lot of that is the expectation of inference demand. So as you put these models to production, whether it’s large language models, whether it’s recommender systems, image models, video models, that demand will catch up. So as it stands, there’s a mismatch between ⁓ the capital outlay

into these data centers versus the revenues that Gen.ai companies are generating. So if we look at OpenAI, Anthropic, they are the primary beneficiaries in terms of how fast their revenues have grown and the absolute scale or the relative scale of their revenues relative to even companies like Cursor, who grew very, very quickly to 500 million in ARR.

And so ⁓ in terms of like software revenues, we’re kind of in the tens of billions range, whereas ⁓ for ⁓ hardware CapEx or data center CapEx, we’re in the hundreds of billions, right? So assuming that ⁓ software revenues, two to three X year over year, then eventually it will catch up to CapEx if end users.

⁓ enterprise customers find that they’re not getting ROI from these Gen.ai ⁓ apps, then I think that’s where the house of cards ⁓ collapses.

Grace Shao (11:14)

Yeah, actually, let’s just like talk about the private market valuation quickly. Like right now, OpenAI is valued at like over 500 billion, something like 180 billion, right? Like startups like Hercer, Lovable, Chasing Billions, or what they’re calling trillion dollar ambitions right now. ⁓ I think the Lovable CEO said they want to become the first trillion dollar business in Europe, right? How should we make sense of these numbers? Like, I’m not a quant person. They just sound like humongous numbers.

Can you explain this to us, like how to make sense of this? these are just, ⁓ does it make sense for these companies to be valued at this high in the private market right now?

Kevin Zhang (11:52)

Yeah, so I think that’s a really good question. ⁓ So the ambition for a frontier lab like OpenAI Anthropic is to be one of the big boys at some point in the future. And so taking ⁓ OpenAI as an example, right? So if an investor is of the belief that they will eventually build their own cloud, they will get into robotics, ⁓ their ⁓ core lines of business, right? Chatchi PD as well as their APIs become

really large businesses, right? So let’s say TriGPT is embedded in various enterprise ⁓ kind of customers. And if they’re actually able to charge, let’s say two, 300 bucks a month, right? As ⁓ more and more white collar workers are reliant on open-air technology, that tracks to a market, even in the enterprise side of things, that’s several hundred billion dollars, right? And then if you’re also of the belief that

Google search will be disrupted and Gemini somehow fails to catch up to OpenAI, then they could also run ads on the consumer side of things. So once you add all these kinds of lines of businesses together, an ⁓ optimistic person might see kind of a line of sight for OpenAI to be this like three to five trillion dollar company that some of the Megs have been at already.

are an investor at the $500 billion ⁓ mark, then I think that’s the return profile that you’re looking at ⁓ before kind of taking into account all the dilution from subsequent funding rounds, stock options, et cetera. And then moving down to the application layer, I think these companies are making is if we’re able to replace broad swaths of labor,

⁓ and you are able to command pricing that’s at some proportion of the ROI that you deliver relative to just replacing like a human headcount, then the exit values for these become enormous, right? So then cognition at 10 billion might sound really reasonable. I think as the issue right now is the exits will be very spiky, meaning we’ll see a lot of zeros.

⁓ and you’re going to see a lot of companies really become those $10,000,000,000 companies. And then for a VC fund, you have limited shots on goal. And so as the entry valuations ⁓ increase, you have less shots on goal. And so on a per investment basis, ⁓ your risk ⁓ increases quite a bit.

Grace Shao (14:40)

So in that sense, you don’t think we’re nearing the ceiling of model layer evaluations or anything. We haven’t hit the peak of the bubble or anything yet.

Kevin Zhang (14:49)

⁓ The markets are definitely frothy, but the winner will be much bigger than we, ⁓ I think, originally estimated. And so if you are one of those investors that are in these assets, I think you’re going to be fine. If you’re not, then I think, which is probably the majority of these funds, ⁓ I think they’re going to be hurt.

Grace Shao (15:10)

So in your writing on Substack, you’ve argued that, you know, consensus capital is crowding into foundation models and fra robotics, I think. But, you know, are there areas that you think are under invested and still in the private market? Like, where do you see, like, overlooked opportunities right now?

Kevin Zhang (15:26)

I think it’s less so ⁓ maybe overlooked opportunities in AI, right? So like a generalist VC is able to allocate capital across different things, right? So that could be psychedelics, that could be robotics, that could be biotech. so, or consumer as we’ve talked about before. ⁓ so figuring out kind of what the market dynamics are for those industries where

you’re just non consensus enough to be that first check in, but you’re consensus enough that at the next round, ⁓ whatever you’ve invested becomes consensus. And this was like the subject of some Twitter debate ⁓ with Martin at Andreessen where ⁓ he was arguing it’s not bad to invest in consensus deals because like in the end, like some of these deals ⁓

end up generating huge returns. And we know that even with an AI, like if you’re in a consensus bet that ⁓ pans out, assuming OpenAI is that company, then you’re still seeing like a 10x gross return, right? Assuming one of these companies becomes like five trillion.

Grace Shao (16:43)

Yeah, because consensus, guess, it’s for a reason, right? I was speaking to a few VC investors in the Bay Area a couple of weeks ago, and they were saying, like, some of them are kind of complaining that their bosses are just chasing logos rather than the differentiated bets. But I think in some ways, like you mentioned, if it’s an open AI and it’s still going to be the market leader, market winner, you’re still going to come out on top, I guess. ⁓ Yeah. I want to hear, OK, taking a step back from...

Kevin Zhang (17:05)

Yep.

Grace Shao (17:10)

these questions, think from a cultural perspective, you you worked in Silicon Valley, New York, and now like, you you moved around in greater Asia, greater China. ⁓ What do you think, like differentiates the two ecosystems the most in terms of like the investment space and then maybe even just like some kind of high level work, cultural tech tech space observations?

Kevin Zhang (17:33)

Yeah, I mean, there’s a couple of things, right? So one is the abundance versus the scarcity of capital. ⁓ And so ⁓ in the US, there’s still a relative abundance of capital where ⁓ as an entrepreneur, it’s comparatively easy, easier to be funded versus ⁓ a similar entrepreneur in Europe or Asia. And so and the other thing is,

you have capital at every stage, right, from seed through growth. ⁓ And so the market as a whole has more shots on goal, more opportunity to experiment versus China, right, where ⁓ there is a comparatively ⁓ or significantly less capital, especially US dollar funds in the past couple of years. And so on the investor side of things, ⁓ they are

also more risk off, right? Because for some of these ⁓ Chinese VCs, they might not be able to raise another fund, right? So each shot on goal ⁓ is a very heavy bet, right? Versus entries, and if you deploy like a fund very quickly, could probably just raise another one very quickly as well. ⁓ And so ⁓ if, you know, entrepreneurs can be more risk on in the US, investors can be more risk on than

one of these bets will pan out and then that becomes the next big company versus in China where you could raise less capital at lower valuations, less capital at growth. ⁓ And I think where the domestic VCs might be ⁓ extremely careful and maybe not having the same kind of venture parallel mindset. I think that’s...

that reflects on the products that you can build and the scope of ambition for entrepreneurs. And I think broadly this is why more more companies are trying to do the true high model, right? Where they might raise their first round of funding ⁓ in China, but then quickly pivot to a Singapore or Canada, the United States, right? And then raising capital in the US.

Grace Shao (19:54)

Do think it’s like really affected the dynamics between investors and founders as well? Or do you think that relationship actually is still quite similar?

Kevin Zhang (20:02)

⁓ For the US, I think it’s still an entrepreneur’s market for the best entrepreneurs. ⁓ Like given the abundance of capital, that has not really translated to kind of a linear increase in top founders. think capital is still chasing founders every year or

taking a step back, there’s a limited number of great founders per year that can build these generational companies. And so when AI investing becomes consensus and these founders ⁓ tend to be in the US, then you have quite a bit of capital chasing these select founders. ⁓ And then within China, like there’s probably higher pricing ⁓ or

there’s more power on the buy side, right? Where if you’re one of the 10, 20 funds that still have dry powder ⁓ at the early stages or one of the five to 10 funds at the growth stages, then comparatively it’s less competitive ⁓ for the investor versus the US.

Grace Shao (21:18)

How do you think that’s affected, I guess, this generation of AI entrepreneurs coming out of China? you know, like actually, if you look at the six dragons or four tigers, whatever you want to call them these days, ⁓ you know, like they’re definitely nowhere as like valued as high as the American counterparts. ⁓ There are a lot of them actually finding funding from the BATs instead of, you VCs. So like has this affected their

future trajectory or the entrepreneur’s mindset or their business model.

Kevin Zhang (21:50)

Yeah, I mean, I think it’s ⁓ all TBD based on or TBD, like given the pace of innovation, right? So to our earlier conversation, like can one of these labs build ⁓ something that’s post transformer, right? Or ⁓ can one of these win in other modalities, right? Whether it’s video or image or something else. ⁓ But from a capital perspective,

it’s really hard to raise like another 2-3 billion for any of these companies. And so, like my hypothesis is ⁓ some of these companies going public is a way for them to get another turn of the card, right? So it’s like, hey, if we’re able to raise 500 million a billion ⁓ on the kind of Hong Kong stock exchange, that gives us another two to three years, right, to figure things out versus OpenEI or Anthropic where every round is

heavily oversubscribed and they’re able to ⁓ basically chase everything, right? Chase capex, chase kind of their core products, ⁓ as well as some of the moon shots, With opening at going into robotics, ⁓ going into ⁓ video image, et cetera, et cetera. ⁓ Obviously, maybe not to the same level of success as some of the other kind of image players or video players.

Grace Shao (23:12)

When we last spoke, you kind of made this like big statement thinking that eventually these Chinese ⁓ companies will potentially be irrelevant or they’ll stay within their ecosystem, right? It’s a big statement. But I was kind of curious if you could like elaborate on that a bit more. Like, I know you’re very bullish on the American leaders right now, market leaders right now, but why can’t a deep seek or a moonshot, you know, maybe

do well globally as well, or would Alibaba’s open source eventually kind of take on a leadership role?

Kevin Zhang (23:50)

Yeah, so it’s two things, right? I think it’s capital and it’s distribution. On the capital side, like we’ve talked about, that ⁓ model training inference follow the current trajectory, then it’s a capital game. And so the companies that are able to generate the most revenues the fastest and is able to also raise the most capital will win, right? Under this kind of paradigm. And then secondly,

Even for companies like OpenAI, think they realize that there’s a limit to, for example, how much you could charge for APIs. And so they’re definitely moving into applications, whether it’s ChatGP or something else. And that’s where you ⁓ can grab higher kind of margins than purely API revenues. And then as you go into workflows, as you go into the replacement of human labor, as you...

Take advantage of your initial tech advantage to embed yourself into kind of the fortune 500 clients They’re not gonna kind of rip you out versus and then swap swap potential your Chinese competitor in right so I think the speed of distribution and how fast you’re able to embed yourself into ⁓ Customer workflows where those customers have a very high likelihood to pay or a high willingness to pay I think that that’s kind of the name of the game

So capital and then distribution.

Grace Shao (25:18)

Because you’re basically saying monetization still has to be on the enterprise end. What about China’s AI application on the consumer end? Do you see that being one of their advantages or something that they’ve done really well in terms of adoption rate? I deep sea hit like, not deep sea, dobao has hit more than 450 million in the MAU. Deep sea even higher than that. It’s kind of crazy in terms of scale. think about just like compared, it’s almost comparable to the leading American.

⁓ applications right now.

Kevin Zhang (25:49)

Yeah, I think from a usage perspective, for sure. ⁓ The more interesting thing is how do these companies and these products monetize? If we look at Alibaba, Tencent, guess like Baidu, ByteDance, and then even like some of the newer players like Xiaohongshu, how do they monetize? It’s through ads, it’s through e-commerce. And so I think for these,

new Chinese companies, it’s figuring out how to monetize this chat interface ⁓ catering towards a Chinese audience. And I think the company that’s able to figure it out, and it could be one of the giants, existing giants, they’re going to capture that slice of revenue. But as far as the willingness to pay for a Chinese consumer, I think that’s significantly less than ⁓

an equivalent US consumer where the number of people in China who are willing to pay 20 bucks a month or 200 bucks a month for kind of the higher tier of opening, that’s gonna be limited. So you have to win on ⁓ scale, right, scale of users and you have to win on these other potentially less obvious sources of monetization.

My guess is it’s gonna be advertising, it’s gonna be commerce.

Grace Shao (27:17)

Yeah, it’ll be like value added services or like invisible ways of making money. It’s not going to be like a subscription model or anything. I agree with that. Yeah. Let’s just actually double click on China. ⁓ know, Bill Gurley, I think has been one of the more I would say pro China, but more outspoken investors in Silicon Valley. That’s not anti China, at least. I think he recently just talked about going to China as well, doing a big trip with his daughter. ⁓ You wrote about Bill Gurley being wrong.

Kevin Zhang (27:20)

Yeah.

Mm-hmm.

Grace Shao (27:47)

about China’s open source models. Can you explain to the audience what was your whole thesis and why do you think he’s not analyzing the space correctly?

Kevin Zhang (27:58)

Yeah, I first off, like I think where he is right are a couple points, right? So the entrepreneurs, investors, ⁓ folks in technology and finance definitely pay way more attention to the US ecosystem and learn from the US ecosystem way more than vice versa, right? And so that creates a huge blind spot for US entrepreneurs and investors. And then secondly, I think there’s

Quite a vibrant and talent dense kind of ecosystem ⁓ in China here as well ⁓ I think where he is maybe not ⁓ Where he might have like missed the mark one is this kind of like grasses greener on the other side syndrome, right where a US investor might feel like like China’s like the land of opportunity because of potentially lower valuations because of

⁓ very competitive entrepreneurs, ⁓ etc. Whereas, and they might have this feeling that ⁓ China is like kind of the buyer’s market, right, or the investor’s market. ⁓

Grace Shao (29:12)

But wait, that was

what happened with Internet error, right? So it’s not like he has no basis with this mentality or thinking.

Kevin Zhang (29:20)

I think that’s partially correct. I guess the question is how does an American VC ⁓ monetize ⁓ this ⁓ insight if they are potentially unable to invest in these ⁓ underpriced but potentially competitive assets given some of the limitations for American investors? And then secondly,

⁓ I think the open source closed source debate is more interesting. So I think ⁓ Bill Gurley is a proponent of open source. And so the framework is like, if you are a front to your lab and you are training models that cost ⁓ tens or hundreds of millions of dollars per training run, you have to monetize it in some way.

And so if you are a lab that gives it away for free, that truly limits your ability to monetize. ⁓ And that is like very different from Facebook open sourcing their models because they have a very, different business model than some of the labs here in China. So it goes back to even opening in Anthropic where initially they have the tech edge, they have the best models.

eventually folks are going to catch up. So they’re really in this race to maintain being one of the top players, right, top two to three players, and then ⁓ winning distribution, right? So if you are at this Google scale as an opening eye where you have multiple software assets that people are using daily and your model is, let’s say, top one or two or three, that becomes very, very hard to displace compared to a lab that only has a model but no product.

And so I think that confluence of how do you build product on top of your model layer while you have the lead is really the name of the game versus, hey, I’m going to train a pretty good model and then I’m going to monetize via API and then just give it away for everyone else. I think that’s probably not the winning game in this era.

Grace Shao (31:38)

Then what’s the game that Deep Seeker moonshot or any of these, ⁓ the kit moonshot that’s behind Kimmy, what, what, what should be the game for them, I guess, for them to be able to monetize eventually.

Kevin Zhang (31:51)

That’s a great question. I don’t know.

Grace Shao (31:53)

Yeah, I think that’s why like for Baba, it made sense, right? It’s kind of like the meta llama business model because end of the day, they own distribution and they own the infrastructure. but but like it’s interesting to see because even the deep seat narrative like, he’s a billionaire, he funds himself. But then that’s actually not that much money to fund. like we just talked about capex is like this is hundreds of billions. This guy’s got a billion. Like, there’s still a pretty big gap in between that he’s not going to bankrupt himself to fund this. Right. So

Kevin Zhang (31:54)

Yeah.

Yeah.

Grace Shao (32:22)

It’ll be interesting to see how he monetizes or make this into something bigger.

Kevin Zhang (32:22)

Yeah.

Yeah,

I mean, the other thing is like the market is very dynamic and for every one of these labs, you’re a one hit product away from monetizing, right? And I think it’s never a wise thing to count a player out, especially if like a founder is very good and very thoughtful. ⁓ And so if some of these like companies

end up going public and raising, let’s say, their few hundred million or like a billion, they’ll have another two, two, three years to figure out the monetization piece, right? While they try to catch up in some model modality, right? So they might, they might figure out, or mini-macs might figure out, our video models are really good and truly world-class, and we’re going to build a bunch of these workflows and we’re going to be adopted by the world, right? That could be kind of interesting as well.

Grace Shao (33:22)

It could be a business model that we haven’t even seen before. could be something completely innovative, right? Something completely new.

Kevin Zhang (33:28)

Yeah, potentially.

Grace Shao (33:30)

Yeah. Okay, I think I want to zoom out a little bit. When we talked, you said you are looking at Neo clouds, and this is a space where I frankly know very little about. So I want to hear from you and please do explain things in layman terms and dumb it down for me. But alongside the hyperscalers, we’re seeing Neo cloud players like Coreweave, Nebius emerge, right? Like they’re making headlines. ⁓

How does the dynamic really work between the Neo clouds and our old traditional cloud players?

Kevin Zhang (34:03)

Yeah, mean, for some of these, so defining Neo clouds like the way I think about it is it’s GPU as a service, right? So whereas a traditional cloud provider like AWS or Azure or GCP, they might provide a bunch of services, right? So they might offer databases, they might offer compute, they might offer a bunch of these other pieces of software storage. ⁓ Neo clouds kind of simplify what they offer and

The core of what they do is they help companies with training models, doing inference on the models that they train. And so the dynamics so far has a couple of dimensions. So if we look at ⁓ the upstream, NVIDIA, so they basically supply GPUs to both the hyperscalers and the Neo clouds. Their incentive is to have or to not have the hyperscalers be that big because they know that

hyperskillers like Google, like ⁓ AWS, ⁓ or Azure, they’re building their own hardware. so, NVIDIA is very incentivized to build credible competitors to hyperskillers where they reduce their customer concentration. And then for the Neo Clouds themselves, like a few of them were kind of in the crypto mining space before pivoting to kind of this like AI workflow, like AI compute.

And I think the game they’re trying to win is, we have this wedge where Nvidia will give us some allocation of their ⁓ latest GPUs because of those dynamics that we talked about a couple minutes ago. And we’re going to use this as a wedge to eventually become a really big player ⁓ in the AI space and maybe ⁓ for the players with even grander ambitions.

to become kind of the next hyperscaler. And so you could look at like Oracle a few years ago where their cloud business was a very, very distant forth, right? In the U.S. But because of these recent contracts ⁓ with ⁓ these larger customers like OpenAI, that they’re able to ⁓ see significant appreciation in their ⁓ market cap ⁓ and kind of this like...

like exponential increase in their RPR remaining performance obligations because open eyes saying, hey, we’re, we’re, we’re willing to spend a couple hundred billion dollars on, on compute with you, right. Versus, ⁓ within Azure. So, so I think the dynamic is hyperscalers, ⁓ trying to maintain their, their lead, ⁓ while trying to build more, more and more of their hard, hardware in house. And then on the Neo cloud side, it’s, Hey, we have a wedge right with.

⁓ a ⁓ demand for AI, both on the training and inference side. And we’re going to use that to grow our revenues very, very quickly, potentially take on a lot of debt, right? And then become kind of one of these large dominant players tomorrow.

Grace Shao (37:12)

So you said that there are essentially service providers. Could you actually elaborate a bit more on that in the sense that, ⁓ again, this is really new to me, this whole sector, but people have said, NeoClouds are a real estate business. Others saying it’s a software layer value ad service. How do we actually see this? Can you explain that to me?

Kevin Zhang (37:35)

Yeah, so at its core, like, Neo Clouds are just clouds with like GPUs, right? And so how do they monetize? And so there are basically three ways to monetize that we’ve been kind of diving a little bit deep into, right? So one is, hey, I am an OPENAI and I’m just gonna rent your GPUs, right? I’m gonna rent your infrastructure and I’m gonna do everything ⁓ myself, right? And so for the Neo Clouds,

That’s kind of the lowest margin type business because you’re basically taking some profit ⁓ or taking some margin on top of your cost ⁓ to provide that compute, right? So part of that is your initial cost to acquire that hardware. Part of it is your kind of ⁓ ongoing electricity costs or costs to run the data center. And so that’s kind of bucket number one of

monetization for Neo clouds and then bucket number two is kind of the managed services, right? So if you want to do, if you want to provide software for training, if you want to provide software for experiment tracking, for AB testing, things of that nature. So that gets you much closer to software margins, right? Traditional SaaS margins. And so I see more and more Neo clouds going there, right? And that BS included, uh, uh,

CoreWeave as well with their acquisition of weights and biases. ⁓ And then the third part is what if we provide, excuse me, APIs as a service, right? What if we give you inference, we’ll host the models ourselves and then you just pay based on the tokens generated. And so that’s like kind of the third category. And so ⁓ how this space ends up playing out is what’s the purpose

portion of revenues that these Neo clouds will be able to generate from each bucket of services and products, where if they’re able to generate more and more revenues from kind of the bucket two and three, that makes Neo clouds a much higher margin business than comparatively under differentiated, you know,

GPs as a service infrastructure, although running these data centers isn’t that easy. And there’s a lot of nuance between ⁓ even some of these more leading edge cloud players.

Grace Shao (40:09)

I see. ⁓ I think I’m going to actually ⁓ ask you one last question on investment, is I think energy is something that people have been talking about being affected by AI, obviously. And then like you said, software businesses, you know, we’re obviously already spoke about LLMs, cetera. So what do you think is a sector that is going to be affected by AI and you’re looking at it as an investor? ⁓

but are not as obvious to the public or not as obvious right now.

Kevin Zhang (40:40)

Yeah, I think that’s hard. ⁓ So just given ⁓ most of our efforts are on the ⁓ kind of public side of things, we are kind of looking at every layer of the sack, right? So from ⁓ semiconductors, kind of ⁓ the companies that build the underlying kind of infrastructure, right? So your transformer companies, your... ⁓

power companies ⁓ to your application companies. And I think it’s more ⁓ what’s undervalued relative to kind of market consensus and, ⁓ you know, rewinding the clock back one or two months, Nebius was one of those players, right? And Oracle was one of these players. So it’s more which player within which kind of layer of the sandwich is undervalued and then how do we think about

our risk adjusted returns, which is a little bit of a cop out answer. But our objective is to build a basket of these securities across the stack where we believe that these businesses will offer outsize returns. then going maybe a little bit deeper is for a lot of these businesses, they don’t have kind of a pure exposure to AI. Meaning even if you invest in a company like

a Microsoft, right, or an Amazon. They have ⁓ their traditional minds of businesses that you have to price. And then those tend to have, especially if they’re more mature, those tend to have a slower growth rates, right? So even if your AI revenues are exploding, they might be dragged down by some of these at scale, ⁓ mature ⁓ business units ⁓ or products.

And so how do we think about the blended returns for, let’s say, a ⁓ much, much more or much kind of like traditional player in the kind of transformer space?

Grace Shao (42:47)

All right, Kevin, ⁓ anything else you think you want to share with the audience? ⁓ Mindful time or wrapping up our episode? Is there anything you want to talk about that I did not touch on?

Kevin Zhang (42:59)

Sure. I think a very interesting thought experiment is how fast AI will displace work. So I think my perspective is going to be much slower than folks might think in Silicon Valley and much faster than the rest of the world thinks. Meaning next year, 90 % of the code is not going to be written by

or at least like code and production is not going to be written by AI, right? Or I don’t see like broad swaths of ⁓ workers gets automated. ⁓ But I think is AI adoption going to be a 20 year kind of journey? think, no. I think ⁓ for a lot of these professions, it’s going to be this like five to 10 year disruption.

And then we’re already seeing some of this ⁓ for new grad hires, right? Where a combination of AI giving more experienced workers a higher leverage, ⁓ as well as some of the broader macro headwinds ⁓ in the US ⁓ affecting kind of new grad job placements. And so I think my intuition is it’s gonna spill over to kind of the mid-level folks.

as well, right, comparatively soon.

Grace Shao (44:26)

Yeah, I think actually I have an episode coming out literally today, which is with ⁓ Diana David. She’s the director of features at ServiceNow. And she was saying kind of like, instead of thinking about how it’s going to displace jobs, it’s going to change where the workforce will go towards. It’s just that we need that time to kind of figure out where it’s going. But like you said, right now, unfortunately, it’s hitting the young junior staff the most because their work is usually, you know, more

you know, the hands-on kind of like the grunt work that is really easily done right now by AI. Anyway, thank you so much for your time today. I have one last question for you, which is a question I ask everyone. What is a view you hold that is unconventional or you think it’s against consensus? It could be something related to investing or anything, you know, in life.

Kevin Zhang (45:49)

Let me try to think something that’s unconventional.

Okay, ⁓ so I think something that’s unconventional, especially folks ⁓ within finance, think folks in finance tend to be very sensitive to kind of market fluctuations as well some of the macro headwinds or tailwinds. And so, especially right now, where I think we’re in potentially a more challenging part of the cycle.

with employment, with kind of the degradation of ⁓ kind of global connection. think in the long run, ⁓ collaboration will probably win out, at least that’s my hope. ⁓ And that bodes well for entrepreneurs, right, who want to play in multiple ecosystems. That bodes well for investors who want to play at ⁓

these multiple ecosystems right across across Asia Europe and in North America and I think ⁓ the game is Right now is especially for for some of these venture firms in China is like who can survive right because I think those that can survive and Those who can continue to raise they’re gonna be fine and they’re gonna they’re gonna see some very very good vintages ⁓ and I think the same goes for ⁓

when the US also sees their down cycle, the folks that have a lot of dry powder who are able to deploy through those tougher periods, think that’s where you’ll see good ventures as well. so the question of like, given some of these macro headwinds, is that the death of the buy side in Asia, I think that’s probably a little bit overblown.

And then like I five years from now, 10 years from now, like the ecosystem will be mature and healthy.

Grace Shao (47:53)

You think that USD denominated funds will see a revitalization or do you think that we’ll just see a complete different kind of ecosystem from five, 10 years ago?

Kevin Zhang (47:59)

Maybe?

I think it either might be a completely different ecosystem or ⁓ there’s other global capital that’s interested in the broader Asia ecosystem. And I think they’ll come. It may not be at the same scale as ⁓ the era of like Hulianwang, right? But like as long as the ecosystem continues to be vibrant, as long as ⁓ entrepreneurs are able to build

good products and generate or build large businesses, then there will be some capital somewhere in the world that’s willing to back these entrepreneurs.

Grace Shao (48:46)

There’s a lot of Middle Eastern money coming into China, actually, and I think Southeast Asia, family office money, and even, I think, European money. But it’s definitely not the same scale as what we saw during the Indian era with institutional US investors, right? But that’s encouraging. That’s positive note to end on.

Kevin Zhang (48:52)

Thank you.

Yeah.

Yeah, yeah. I think like for some of these LPs, right, there’s definitely a learning curve, right, in terms of LP sophistication and being very long-term oriented. Whereas ⁓ some of these ⁓ OG US LPs, they’ve had decades of experience investing in the VC asset class, right? And so when that learning curve catches up, then... ⁓

At least on the LP to GP side, ⁓ we’ll see some revival.

Grace Shao (49:39)

Great, thank you so much, Kevin. Kevin Jang, thank you.

Kevin Zhang (49:42)

Great, thank you.

AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Get full access to AI Proem at aiproem.substack.com/subscribe