00:00We're rebounding in the technology sector when it comes to public markets.
00:04Remember, chip stocks fell into a bear market Friday in part because of questions posed by
00:09the release of Kemi K3 by China's Moonshot. We're off session highs right now, but the
00:14NASDAQ 100 adding almost a percentage point, the SOC's up 2%. And it's a lot of the compute
00:20names, NVIDIA, but also memory names that are driving some upside in those public markets.
00:26Since Moonshot released its Kemi K3 model last week, everyone's debating what it
00:29really means. Does more efficient AI mean less hardware? Is China closing the gap faster
00:35than we expected? Where does AI's real value ultimately accrue? Joining us now is Crosslink
00:41Capital partner, Mary D'Onofrio, focused on venture growth investments, but across software, AI,
00:48and infrastructure. And for the last three days, I've said Kemi K3 many times. But it was interesting
00:55to take a sort of straw poll of VCs, founders, people in industry of how they react to it so
01:02differently. Some people are very fearful, defensive. Others say this is very good for
01:08AI as a whole. Very difficult job first thing on a Monday, but your reaction to what happened?
01:12Well, Kemi K3, as you said, was phenomenal. And it's similar to the DeepSeek moment we had last year.
01:19You know, 2.8 trillion parameter model. And what it demonstrates, I think, is that, you know,
01:25China and frontier, pardon me, China and open weight models are closing the gap with frontier models.
01:30You know, obviously, it's done, it was done based on their own benchmarking. But according to the
01:35benchmarks, you know, they're fourth. And while, you know, you can debate the merits of those benchmarks
01:41themselves, the productivity gains are tremendous. And I think that in general, it is part of the
01:48pathway towards more widespread open weight model usage. And in the context of enterprise budgets
01:54ballooning in AI, I actually think that it's probably a boon for AI ubiquity over time.
01:59When you think about your existing portfolio of companies, were you sort of like, OK,
02:05how does this impact their trajectory? Is anything actually fundamentally changed for the different
02:11pieces of the AI five-layer cake that you're invested in?
02:14I don't know if it changes the cake itself. But I do think it's another demonstration that models
02:21keep leapfrogging one another. And if companies don't out-innovate, they're going to be leapfrogged
02:26too. Whether it's, you know, every single model seems to outperform the last. And it's a matter of
02:32out-innovating by creating modes of data, of performance, of customer resonance. And otherwise,
02:40unfortunately, you know, the innovations in both frontier and open weight models is tremendous.
02:47You're on the crossover investment team, but basically focused on private and maybe later
02:52stage or venture growth investments. Seven years at Bessemer as well, prior to being at Crosslink.
02:58Sort of introduce us to your overall investment focus, your investment thesis.
03:02Yeah, of course. So as you mentioned, I'm a partner at Crosslink Capital. I help to lead the
03:07mid-stage venture practice. I spend an asymmetric amount of time in vertical AI and AI infrastructure.
03:13What are some examples of what you call vertical AI and AI infrastructure?
03:18Yeah, of course. So AI infrastructure is a place I spend a lot of time. That's,
03:22you know, the physical and software systems required to build, train, and run models.
03:26And in the first wave, it was about how to build a scalable model. And now we have those,
03:31as we're talking about right now. And so in the next wave, it's how to deploy them quickly,
03:36quickly, fastly, cheaply. Agents, of course, being one of the ways to do so in an efficient way.
03:42You know, taking models as the reasoning layer, adding tools and infrastructure around it to
03:46connect outside data and apps, and using memory in order to do it in a way that produces outcomes.
03:55And agent infrastructure itself has become a sector that I've spent a lot of time in,
04:00whether it's agent harnesses or the identity layer around it.
04:03One of my portfolio companies, Teleport, is a leader in that space. And it's a place that I'm
04:07spending an asymmetric amount of time.
04:09I wanted to get to that kind of background on your focus to try and answer some of the
04:14questions that the Kimmy K3 moment posed. One idea is that basically by lowering the margins at
04:22the model layer, it's an enabler. It opens up the market a little bit. It's a driver of demand,
04:28essentially. Do you agree with that kind of thinking in response to this?
04:32I do think that it demonstrates what, you know, a lot of even the inference infrastructure layer
04:38is demonstrating, which is, you know, as cost has gone down by 95%, does that displace the total
04:44dollar value expended by customers? And I would argue it probably won't, because the migration will
04:51just be from workloads that are at the frontier, that are most expensive, and of course, that require
04:56the most reasoning capabilities and are the most advanced. But they'll go to cheaper models that,
05:02that, you know, still provide the same amount, the requisite amount of compute for the question
05:08being posed, but do so at much cheaper cost while not sacrificing on things like latency and performance.
05:15It's interesting because, like, Kimmy K3 kind of pricing $3 per input, $15 per output on the token
05:23side. Like, it's quite in line with, like, what a lot of leading models are also priced at. I found
05:30that interesting. Maybe we should talk a little bit about valuations. It's not a sophisticated argument,
05:35but when this all happened, a lot of people came out and said, well, if a $20 billion Chinese startup
05:42can do a $2.8 trillion parameter model with these economics, why are we assigning OpenAI and Anthropic
05:49a valuation of a trillion? You know, they looked at the sort of fundamentals of that. Is that the right
05:54way to think about it? Frankly, I don't know yet, and I don't think anybody knows yet.
06:01Interesting.
06:01Even in the context of those questions, Anthropic and OpenAI are still producing billions and billions
06:07of revenue, and it's not just from their, you know, they have first-party businesses, they have
06:12third-party businesses, and I think, you know, obviously the API opened for Moonshot, but the
06:19open waits have not yet. Yeah, 27th of July. And it's worth noting that because it's open wait doesn't
06:24mean it's free to run, and so I think that's a dynamic to these models that people potentially
06:29underappreciate as well. The argument you just outlined is the same that Greylock saw Modern
06:34Medi made on Friday on the show that, you know, these are very robust businesses that have multiple
06:39revenue streams. You are focused on the private markets, but we've had some really interesting
06:46IPOs of late, you know, not just SpaceX, but also SK Hynix doing its ADRs. I was in New York
06:51City for that.
06:52What does that signal to you? I think it's two things. Obviously, those are fantastic
06:58businesses, and I think when it comes to what it says for the markets is that investors are really
07:03looking for anything having to do with the AI build-out. AI infrastructure, obviously, you made the
07:08reference to the performance of public market stocks today, but I also think it's one other thing for
07:13the capital markets, which is that, you know, we have this narrative that potentially the IPO window is
07:18reopening, but is doing so with these long-standing businesses with billions of dollars of revenue
07:22that are, frankly, older. You know, they're at the cutting edge of innovation, but they've been
07:28around for decades. And so I think it's saying, you know, there is a desire for things having to do
07:34with AI, but nonetheless, investors are still, they're hardly speculative. Investors are looking
07:39for things that are relatively secure.
07:40Investors are looking for things that are relatively secure.
Comments