How the AI Talent War is Rewriting Deals ft. Josh Wolfe, Dan Primack & Dan Grossman

Hi everybody, um, thank you for, for sticking around. Uh, my name is Dan Primack, uh, business editor at Axios. I am with Josh Wolfe, who I think you probably all know from Lux. I'm with another Dan, uh, Dan Grossman, uh, who is, uh, VP of Corp Dev at Amazon. I apologize ahead of time, I've got notes here. I'm not like checking, uh, TikTok or something like that. Um, so guys, so, uh, we're going to talk kind of the business of AI a little bit and all the acquisitions and all the stuff you guys are doing over at Amazon. But Dan, I might as well, since you are up here, Amazon is the company of the hour this week, it seems. I guess I'd say, how's your week been?

It's a long week. No, but it's been a, you know, the outage to me is a, it's a story in part about complexity. These systems are complex, but it's interesting how there were different impacts for different customers. Some customers had virtually no impact if they had resiliency and redundancy with the design of the system. Others had different impacts, but it's something that we would like to learn from and get better about. Obviously, we want to have the most stable systems we can.

Is there— I'm curious, some of the kind of commentary post, leaving aside kind of the technological piece and kind of why it happened and the impact on particular companies, has kind of been this, is Amazon too big to fail? Not from a financial standpoint, I guess arguably from financial standpoint at the end. Does this speak to that? Is Amazon now too big to fail?

Well, again, I'd say that the interesting thing is how there are different impacts and they're really a function of design, not necessarily a function of if the provider was big or not.

Josh Wolfe, I mean, do you, is it?

I look, I, I'm a huge Amazon fanboy. Uh, I'm a Dan fanboy too. Uh, but you know, but by the way, I think it's one of the greatest capital allocators in tech corporate history. Dan himself, many people may not know this, but he was on my side. He was on the dark side. He was a VC a few times, and then he was a portfolio manager at SAC, which is one of the great hedge funds running TMT, and now is responsible for investments, M&A, a whole slew of things in corp dev for Amazon. Many of you may sell your company. We have done a lot of deals with Amazon. They have acquired Covariant. They've acquired Zoox in self-driving cars. They've been a co-investor with us in Osmo. They've been a co-investor investor in Control Labs, uh, and then Jeff has been a big investor in Physical Intelligence, who you heard from, uh, earlier. Uh, so I'm, I'm a big Amazon bull.

Thank you.

Does that mean it's too big to fail?

I think so. I, I, I just don't think it's going to fail. I mean, there's—

I don't mean that it's going to fail. I, I mean, but I, I mean, but maybe I don't mean the company's gonna fail, but when it goes down, I mean, is it, is it a problem?

Of course.

Yeah.

Not as a problem.

No, look, they, they, they made a decision which was an extraordinarily strategic decision to effectively become a technology utility in part of the business. So it's a, it's a fail when New York City, you know, electricity goes out, you know, during Hurricane Sandy or something, there's a— to me, it underscores the dependency that so many people have on Amazon, which I think is a net positive thing for them and for shareholders. But for sure, you know, and I think to Dan's point, you have resiliency, you have redundancy, you know, it's a good thing.

There's kind of been— the reason I was interested in asking that is there's been the kind of this bigger narrative around AI over the last— and it really seems to have accelerated in the last month, this circularity argument, right, that everybody is connected. The hyperscalers are connected, the chips companies to the data center providers, to the energy companies, to the stakes in various companies and obviously orienting tons of policy around it. The concern in all this is what if we— and this is probably collective for this particular group— what if we're wrong? Not completely wrong. Not that this is vaporware and, you know, but 5% wrong, 10% wrong. Is that a valid concern? And as somebody who's investing, is this a legitimate fear or is this Chicken Little?

Well, this is what makes markets, you know, people are wrong. There's going to be massive misallocation of capital. and it doesn't mean that we as consumers don't benefit. I mean, I, you know, if you look at a few different dimensions of this, we decided very early on we were going into open source. So we were early investors in Hugging Face. Uh, we did not back the Frontier models. We're not in Anthropic. Uh, they own 12, 15% or so of Anthropic, give or take. Um, we're not in open—

and you have your phone, you can get the exact numbers while he's talking.

And, uh, you know, but, but we are all beneficial users of that. Now I will say I, at any given point in time, I'm using at least 5 different sources when I'm, you know, prompting, and I would just copy the same prompt and put it into OpenAI ChatGPT. I'll put it into Anthropic, I'll put it into Gemini, I'll put it into Groq, I'll even put it into Perplexity. Each one of those companies is basically looking for 100% market share for me. They're getting roughly 20%, but they're all collectively, not for accounting Perplexity, 500% in CapEx. And the history of CapEx in most tech is that it benefits the users in the end, but, uh, there's enormous amount of capital destruction and misallocation of capital as everybody individually rationally does something that is collectively irrational. So I believe that there's no difference this time, uh, and that is of course the most dangerous words this time is different. So I do think that there will be massive misallocation of capital. Uh, I have crazy predictions about how this might end up. If I were a betting man, which I am, uh, I would bet that Anthropic ends up owned by Amazon at some point. I don't know if this ends up as a standalone, you know, business. OpenAI today, 70% of their $13 billion of revenue are from $20 a month subscribers like us. Um, and there's an enterprise, you know, version and so forth. But the dark horse here is Google. Remember a year and a half ago, Google with Bard was like the laughingstock. And today Gemini is one of the most performative models, a large context window. We're all waiting Gemini 3. 3.1 is amazing and we're in Runway ML and I think they're amazing, but it's, it's very serious. And so Google could Microsoft a lot of the large frontier labs, by which I mean what they did in the early 2000s or mid-2000s when Microsoft basically had PowerPoint and Excel and Word and the entire Office suite they gave away for free. What would happen if Gemini was free for a year or two, subsidized by their own cash flow? Uh, would it hurt the business model of some of these models that have spent huge amounts of CapEx? $100 million for the first Anthropic model, billion for the next, $10 billion for the next. And at some point, does revenue hit an asymptote or face competitive threats and suddenly it's no more economic and it has to be absorbed by someone?

For all that to happen though, Amazon's got to figure out a way to make Gemini via search actually make money when it's on search, because that's where all that money comes from. If they kill their own golden goose, then they don't have the money to give it away.

Yeah, that, that is true. But look, the venture investors that are pouring billions and tens of billions of dollars into this are really accruing, in my view, to two main people, which are NVIDIA and AWS.

One, one of the things you said, so I had you on stage in New York, I don't know, 5, 6 months ago. And, and when we were talking about kind of this issue, uh, we talked a little bit about the data centers and the energy, uh, buildout. And you suggested that you thought the buildout was too big or, or was kind of getting ahead of itself. Uh, you talked about the hyperscalers as a whole, Amazon being one of them. You said kind of individually everybody's plans make sense. As a whole, there's too much going into this. Dan, there has been some pullback lately. Do you agree? Has there been too much going into data center buildout and, and then kind of the, the energy buildout, which is being used to, to prop that up?

The challenge with the energy buildout in part is that the lead times are long. So to me, it's tough to take a short-term perspective and assess, is there too much or too little? I think it's more appropriate to think about what is the longer-term outcome and what conviction do we have about what is actually going to be. Inevitably, I think in a portfolio of these types of bets at this scale, when they're as horizontally applicable as they are here, there's going to be some that work spectacularly well and some that, that don't. But, but I think to me, the thing to focus on is really the long-term scenario. But you can't look past the, the, the issues and the challenges from the failures.

They're going to be there.

The big view that I had, we were very lucky 10, 11 years ago to be in Zoox, which was early autonomous vehicles. And people have heard me tell the story before, but they had the physical vehicles that were basically ingesting all this information from LiDAR and vibrational sensors, thermal sensors, optics at the rate of 1 second per second, what they call reality. And I got pissed because we had $25 million invested early on. And I looked at people that looked like they were playing Grand Theft Auto. And I'm like, what the hell are these guys doing? They're screwing around and wasting our money. He said, no, no, you don't understand. They're training in silico, running high physics simulations on things that look like Grand Theft Auto. Oh, well, what are those things running on, uh, and they're doing thousands of simulations a second. Well, they're running on these Nvidia chips that aren't even on the market yet. That was at a time when Nvidia was a $15 billion market cap company and Intel was $150 billion. And I remember pitching this publicly at a charity event and said, this is the pair trade of a century. I feel the same way today about some of the memory players and the edge inference chips that Lux is backing, where the, the view 10 years ago was that the cyclicality of the gaming console market, whether it was, um, Xbox or PS5, was where Nvidia, you know, would basically, uh, ebb and flow. And then suddenly it became the soul of the new machine, you know, in proverbial terms. Today, people would feel the same way about the commodity memory players to basically say, um, SK Hynix, uh, Samsung, Micron, and then some of the other memory folks are, you know, sort of commodity memory players making flash and NAND, uh, and SRAM and DRAM. And I believe because of a paper that came out a little over a year ago that said we could do large language models on device using flash memory, that you will see probably 50% of inference, which would really shock the assumptions of endless demand for hundreds of thousands of H100 clusters, endless demand for data centers, endless demand for novel forms of power, whether it's modular reactor or other things, uh, nuclear reactors. And if you end up doing 50% of the inference on device, uh, and 50% is still going to the cloud for search, it will change the economics of what everybody's, uh, underpinned assumptions are.

Do you agree with that thesis?

I don't know. So I think the next 2 years we have pretty good visibility into both the available supply in, in thinking about the energy, availability, but also how it will be consumed. I think beyond 2 years, I don't have a working crystal ball. I think for what Josh Wolfe said to come true, it would take some extraordinary advances. It's in the realm of possibility.

Dan, let me ask— and Josh brought up Anthropic. I'm curious, obviously Microsoft makes its deals with OpenAI. You guys make your deal with Anthropic.

Did you—

when I say you, did Amazon have an opportunity to make the deal with OpenAI?

I don't think I can talk about the past opportunities, but if you've got a question on Anthropic, happy to take that one.

All right. Why Anthropic?

With, so let me start with how we think about AI. We think really about 3 strategies. First, the underlying infrastructure. For us, that includes NVIDIA GPUs. It also includes Trainium and the specialized hardware and software that we have to run all those workloads. The second layer is the model presentation layer. We have an inference service called Bedrock that exposes the models to customers. And it provides all of the tooling that you need to make those models work really well. The third is the application layer. So those are the new services that will hopefully delight and astonish all of us that are built on that underlying stack. With Anthropic, we had specific activities and goals that tightly dovetailed with each of those three. Plus, we thought the company had unbelievable talent and an ability to innovate.

Is Amazon content— is maybe the wrong word, but I'll use content— content to have Anthropic be the it's not an affiliate, but the partner, I guess, that's really working on frontier models, whereas you guys are working on kind of more niche stuff. Is that, is that the thinking?

Well, it's funny that you say content because what we talk about internally is how we really try to pay attention to what customers want and they're divinely discontent. They are always looking for something else. We love what we're doing with Anthropic. Our mission though is to offer customers choice. So we will keep looking for that. We'll keep trying to find that divinely discontented customer. We'll try to understand what it is that they don't like, and then we'll try to serve it to them.

This is a philosophy, by the way, that I think is really important, particularly as people think about partnering with, getting investment from, or ultimately selling to Amazon. Um, you had a quote, I think roughly when the big, one of the big Anthropic investments were made, uh, you know, in billions of dollars that was like, this represents the scale of our ambitions. You know, it's not just the investment itself. The word that he used before of choice is really important because I've always admired when Jeff Bezos used to say, I don't know. And he would do this with a little bit of false humility. I don't know what technologies are going to matter in 2 years or 5 years or 10 years, but I know for sure that people are always going to want at least 3 things. More choice, lower prices, more convenience. And you can see that they will make investments in service of that. I suspect that it's the same sort of thing in AI.

It, it, it totally. And one of the things that's so liberating about thinking from the perspective of what won't change as opposed to really focusing on what's transitional or what will change is it enables you to make big investments and be less focused on when they're realized. If you know you care about lower prices, you can deeply invest in that.

Speaking of big investments, Amazon has this reputation, particularly on the acquisition side, but I guess you could argue Anthropic is not an acquisition, but multibillion-dollar deal, it might as well be. You have kind of this barbell that you do some small deals, but you also, when you guys go to buy something, you generally try to buy something fairly large. There's some exceptions to that. Is that still how, when it comes, particularly when it comes to AI, is that still how you're thinking about it? You've done at least one or two of these kind of reverse acquihires. Kind of where are you thinking when it comes to AI in terms of actual Amazon acquisitions or acquihire acquisitions?

Our strategy really hasn't changed fundamentally over the years. There are details that, of course, change regularly.

The—

on the barbell model that Dan mentioned, though, where we think about either smaller or larger, more mature partnering opportunities, really what we're trying to do is avoid integration risk. When we've studied what has and hasn't worked, That's where acquisitions often have problems. So with small companies, there isn't as much to integrate. With big companies, you have something that is relatively mature, stable, that has a path. So you don't necessarily need to integrate in the same way.

So when you say big, like, I mean, okay, Whole Foods was whatever, 30 years old, right? When you guys bought it. Big, as you say, big, but also mature. Right now though, you have these multi-billion dollar valuations, maybe 10, 20. They're not big. They might be small in terms of people. They're definitely small in terms of age. I mean, does it still follow the same model then?

It's still, but you're referring more to valuation and what I was referring to is more related to the operational stack. But the valuations right now are there. I haven't seen anything quite like it before. And, and just on that front, I think our view has generally been we, we like making big investments, but we like to also have commensurate conviction in the outcome. The challenge here is that there's a pretty broad spectrum of possible worlds that I think we could see when you look into the future. But it's not irrational to make big bets when you have big belief.

There is this gap where they do a lot of smaller sort of tuck-in acquisitions, $100 million, team technology, you know, and then sort of billion plus. Zooks was a billion plus. Ring was a billion plus. Whole Foods, obviously. Yeah.

But we also, you know, we do many type of partnering activities. We, even if we wanted to, there is no possible way we could build or own all of the things that we would ultimately like to deliver to customers, the things that customers want from us. So one thing that I suggest, uh, for folks to keep in mind is to be creative in thinking about how we could work together, if that's something that's of interest to you. And one framework for it is what Josh Wolfe mentioned before. It's to focus on what won't change, at least for us. And then that often can inform the dialogue that we have, the levers that we pull, and how we set up whatever the relationship is.

Josh Wolfe, for you, you know, venture capitalists traditionally, I mean, whether if there's big strategic acquirer, that's kind of an old path, right? It's happened for years. The acquihire thing is relatively new and not, not just acquires. They've been around for a while, but big acquires, right? You know, the scales of the world, that is new. Pulling somebody out of a company that still actually theoretically still exists as a VC. Is this— are you— do you think this is great because we still get it? We— any way we can get money in this day and age where there's basically no liquidity, good for us. Or does it worry you? It feels like the— and sorry, this is, I guess, all of you— it feels like the AI founder cohort to a certain extent feels a little different than, say, the internet one or the SaaS one. Folks who would want to stay with their company no matter what. Maybe if we sell, we sell, but I'm not just going to walk away from it. People seem much more willing to walk away right now.

Historically, our best exits for our limited partners have been oligopolistic industry structure, 3, 4, 5 people that are irrationally going to compete with each other, both because they want it and they want to prevent their competitor from getting it. And so that is the best sort of setup for us. And we've sold businesses that did not have great fundamentals at great prices, uh, that were great outcomes and it ended up being a win-win-win for everybody. But, um, that, that's a great structure. What we've seen in the past 9 months especially, uh, which I've dubbed poachapalooza, which is Zuck's real infiltration and saying, okay, wait, we're behind in some of these domains. We were going—

400 of them are getting fired today. You can get some of them back.

Uh, but like, you know, seeing the— which is amazing in an era where it is all about artificial intelligence intelligence, how much the dollar values really are about human intelligence. People that were at Meta and then left and went to Scale or went to SSI and being poached back to Meta and like, it's just, it's crazy, you know, TBPN or whatever does a great funny thing like the trading cards of like all the, the human capital that's going back and forth. It is an issue like in the Scale and some of the other acquihire deals, there's a shell of a company that's left there. Meta's acquiring, you know, 49, 50, 51%. It's an implied valuation that gives the investors a multiple. The founders are taken care of, but it, it's upsetting the, the traditional relationship between a preferred investor and a founder with common stock. And I suspect that like every legal term that exists in every deal that's ever been done as a sort of proverbial Chesterton fence, you look at this and you're like, well, why was it there? It's because somebody got screwed in the past. And I think that investors will probably start putting terms in.

I'm starting to hear about it. Are you starting to see it?

We are, and we're thinking about it as well, which is we want to be aligned with our investors. But if you have a dynamic which itself was provoked by a very non-permissive DOJ and FTC in an acquisition environment, and Zuck in particular was very savvy about figuring out how do I acquire these people or these companies or these talent and getting around that kind of review. I do think that there will be terms put in to anticipate that this is not only a possibility but a probability or an inevitability and, and trying to thwart it.

Dan, I'm curious, do you feel particularly as a corp dev guy, was Zuck smart doing this to get around FTC? DOJ and FTC can investigate whatever the hell they want. They're choosing not to, it seems, or at least they're publicly choosing not to. Is it, is it that he was smart in doing these sorts of deals, or was it that he happens to be in a Trump DOJ and FTC that's just really not— doesn't care?

I think time will tell. I think he chose an option that he saw as a viable strategy. It obviously involves some risk. They're making changes that you just talked about today. I think it's early. We'll see.

Can I ask, we are 9 months though into this kind of antitrust regime. And again, that is relatively early given where you sit. And Amazon's obviously had issues in the past with DOJ and FTC. What is your general takeaway from what you've seen so far? The, the, the assumption was it was gonna be much more permissive. Is that fair?

You know, our focus is really on pursuing our strategy, which includes fully abiding by the law and preparing for extreme scrutiny. That hasn't changed. We haven't changed our approach. And so far I'd say, uh, we're happy with that. So we haven't seen, um, a regulatory change that, that I think is noteworthy, uh, insofar as, you know, we're changing our strategy. We're still full bore ahead looking for opportunities, and we expect, uh, intense scrutiny.

Josh Wolfe, last question for you. First, will OpenAI go public in the next 24 months? And secondly, bigger picture as a VC, will some of these AI companies actually start going public? Are we going to see IPOs again?

I think the demand is high. I mean, you can see it in some of the value— demand's been high for a couple years now, but I think the demand, uh, the valuations of some of the companies that sort of are AI adjacent, the deals that are getting announced, even when they do deals with OpenAI, and you see the stock pops, uh, for people like AMD and beyond. Uh, so I think that the demand is there. I think that OpenAI could be a viable candidate. Uh, we're not investors. Uh, they have had the runaway success in the consumer. Uh, I think it'll— if it does go public, you know, probably some multiple of where it's at today. Uh, we are already in the trillion-dollar range of multiple companies.

And so they need to wait until Dan buys Anthropic so they have a comp and they have a multiple to the street?

I think they'll move first.

Guys, thank you very much.

Thank you.

Thank Thank you.