Martin Casado on the Demand Forces Behind AI

What's going to happen to central buyers and platform teams and IT teams if agents are making the decision? It's very clear that coding is pretty much dead. But engineering is very much not. Every time you have a technical epoch, you have to redo everything and we forget that every time. I don't think people even have a common definition of a bubble.

0:00

If AI demand is real and accelerating, why does everything still feel constrained? Why does a technology that's clearly delivering value also feel harder to scale than expected? We've seen this pattern before in early technology shifts. It was easy to assume the hard problems were solved. Infrastructure was treated as finished. Then usage surged. Systems built for a smaller world began to fail. Networks drained power. Physical footprint and coordination became first order constraints again. Each new technical epoch forced a rebuilding of the stack. AI is creating that moment now. The demand is not speculative. Companies are deploying models, budgets are moving, real productivity gains are already showing up. And yet nearly every part of the system feels tight. Compute is scarce, Data centers take years to permit and build. Power is difficult to secure. Regulation moves far more slowly than the technology itself. This has led to two dominant stories. One says we're in an AI bubble. The other assumes scale will smooth everything out. Neither fully explains what's happening. Demand continues to outpace supply, and the biggest bottlenecks increasingly sit outside the models themselves. This is especially visible in enterprise software. AI is often framed as a threat to SaaS. But SaaS was never hard because of the interface. It was hard because it encodes business processes, compliance and operational reality. Those needs do not disappear. What changes is how humans and increasingly agents interact with those systems and in how software is priced, bought and controlled. That shift raises a deeper question. If agents are writing code, provisioning infrastructure and selecting tools, who's actually making the decision? And what happens when that decision making layer becomes less visible? This conversation helps clarify where the real constraints are and why infrastructure is not fading into the background, but moving back to the center of the story. This is a feed drop from the Six Five podcast featuring a 16Z general partner Martin Casado in conversation with Patrick Moorhead and Daniel Newman.

0:28

Let's go off the record. I know we don't do these as often as we probably like, but when we have the opportunity to bring someone in, that can really change the trajectory of the conversation here.

2:30

Pat.

2:41

Or just someone that's got really interesting ideas and things to talk about. I know we love to do that and we got one today that you met when you were doing your professional modeling and hosting. That was Fun. It was really fun for me to sit there and watch you because I saw you working and sweating at GGTC for like two days. Let's, let's have you introduce our guest, which I'm super excited to have here today on the pod.

2:41

Yeah, it's great. So I'm proud to introduce here Martin Casado, general partner at A16Z joining us today. Martin, it's great to see you.

3:04

Great to be here. Thanks so much for the invite.

3:14

Yeah, the lights aren't as hot, but the audience is actually bigger than the in person audience that we had in Washington D.C. for GTC. I really appreciated it. But hey, why don't we start off just to make sure we know who we're talking to. We know you're famous, but for those who don't know who you are, what do you do and what does your portfolio, what do you, you focus on?

3:15

Yeah, so I, I'm a general partner in recent Horowitz. I run the infrastructure fund, so the infrastructure fund. So our definition of infrastructure, this is computer science infrastructure. This is anything with a technical buyer. So think compute, network, storage, databases, of course, any of the low level AI stuff, dev tools, that sort of thing, security. And so I run the team. I've been here for 10 years and then prior to that I was actually portfolio founder for a 16Z.

3:42

Now that's wonderful. And isn't it funny how infrastructure was pretty much left for dead about five years ago and the meme of hardware was a. Hardware is an undifferentiated commodity and look at us now. And I know infrastructure spans not just hardware, but also software.

4:09

Yeah, yeah, for sure. So listen, it's been an evolving tale. I mean hardware has historically actually been pretty boring like networking silicon. Now software infrastructure, especially in data has been pretty exciting. Think like Snowflake or databricks. And so like there has been a lot of exciting, you know, GitHub, but AI has blown everything up. Like I, I don't remember the last time you had a lot of excitement around, you know, a silicon chip. And you know, Nvidia just bought Groq for, or, or doesn't buy Groq. They, they hired the Groq team.

4:32

Exactly.

5:02

Bought them. Everybody. You know, we're seeing that, we're seeing networking companies get funded again because AI requires new networking fabrics. And so times are very exciting again, you know, kind of early Internet esque.

5:03

And Martin, by the way, just a little props to Pat and maybe myself, but you know, both Pat and I actually did advised Grok and we decided when we did this back in 21, Pat, 21, 22, we decided to take stock from them instead of cash. So we're not like leases and cool quite like, you know, we're not the cool, cool kids. But we actually saw what was going on because Pat and I laughed. Like in 2019, we would make this joke, silicon Elite, the world, in the next decade. We kept saying, silicon, delete the world, Semiconductors, the world. And we had, we had journalists that would said, don't write any more op EDS about chips or we aren't going to cover chips.

5:16

Exactly.

5:57

No, just five years ago they were like, we're not even, we don't even want to hear about it.

5:58

Yeah, yeah. I mean, it just turns out every technical epoch requires you to redo the entire stack. We actually saw this even with like 5G, like a lot of the intel processors, like the Xeon was like actually used for 5G. We saw this with the data center. This is when you saw a network revolution. So I, you know, I worked in software, defined networking. That was kind of my focus. We redid the network. You saw this, of course, with the Internet, which gave rise to Cisco and Juniper. So every time you have a technical epoch, you have to redo everything and we forget that every time. So when you get to the end of the last epoch, we're like, oh, hardware's dead, or whatever. And then the next wave comes and we've got to kind of get back to building the stack.

6:03

So I have to ask you, just because everything you're saying here, I think I've seen you speak about this a bit, but are we in this AI bubble here, Martin?

6:37

Well, bubble's a tough. You know, I don't think people even have a common definition of a bubble here. Here's what I will say from a productivity standpoint, demand is real. You have real users paying real money, getting real value, and that's incredibly clear. The data couldn't be more clear. So is there a, is there like a, a demand bubble, meaning like we're a demand will come. Do we have a supply overhang where we're building out, hoping it'll come? No, the answer is absolutely not. We do not have a supply overhang. We have a supply underhang. The demand is very real. Now, speculatively, if you look on a deal by deal basis, sure, some deals are overvalued, but some deals are undervalued. So what I've Learned in this, 10 years of investing, 30 years in tech, is that markets are actually very rational in the long term and broadly, but it's uneven. So if depending on how you look at it and how you squint, you're going to see things that seem overvalued and things seem undervalued. But I would say if you take it all in, my true belief is it's all undervalued in the long term. This stuff is so, so disruptive. Demand is so real, it's monetizing so well, it's driving so much build out that we should all be incredibly excited for the future.

6:49

I'm just going to give this guy a high five.

8:06

Yeah, I mean Daniel and I are absolutely in there and Daniel and I do a lot broadcast. That's literally the only question we got for months. Daniel, what are your thoughts on that? I know they're very much aligned.

8:07

I, I did, I think I did my 50th TV segment this week where I got asked the, the question, are we in an AI bubble? And so I think, you know, I'm so tired of answering it. You know, my, my comms guy said to me, you can't laugh at them when they ask you this because you make it look like they're asking a stupid question.

8:19

I feel the same way.

8:36

I know, but like I, I just keep saying like, look, we're constrained in every part of the supply chain right now. Like we have than supply. Now there's a monetization question that hasn't been answered yet about the size of the TAM beyond subscriptions to LLMs. Like, you know, we'll get into enterprise right now because made is a perfect segue here for you, Martin. Like the AI disruption of software. So right now every SaaS company is adding agents, but at the same time, like I've got five terminals of Claude code open on my desk and I can't program for crap. I'm not a programmer, but I am building stuff and playing inside of, you know, IDEs now and actually seeing what can be done. And we're building our own applications that can do things that our CRM used to do and our, our ERP did and that enables project management. I mean, is software maybe the actual biggest risk here is that where the real disruption is happening is now that enterprises can sort of build anything really quickly with a few smart people. Is that maybe where we're going to see the, the bubble pop?

8:37

Yeah, so it's a very timely conversation. So we actually as a firm, my team in particular did a deep dive on this in the last week. And so we actually have a lot of data. Let me just predicate this entire thing by saying it's very early. And so a lot of what I'm going to say is purely speculative, but I do think that we are a couple years in and so you can say a few things. So let me just say a few things on this. So the first one is it's very clear that coding is pretty much dead, but engineering is very much not. And so you can clearly say the floor has been lowered so everybody becomes a developer. There's almost no indication that the ceiling has also been lowered. In fact, the companies that are the most aggressively using AI are also hiring the most. And so the question is, how do you reconcile these things? Right, right. There are many things that even like the best, you know, version of, of AI coding can't solve. Right. It's not very good at solving large, complex staple software code bases. It doesn't do anything with operations. Like, let's say, you know, these days actually writing software is running software because it's all SaaS. Like the operations is not, it's, it's not a solved problem because we haven't closed that control loop, like watching what it does and then refining based on what it does. And if you actually look dollar weighted, the majority of dollars that are coming in on AI coding are professional coders, not casual coders. And so my belief is this is widening the aperture of the people that can code, which is going to require more code, which means more operations. And so the tent gets a lot bigger. I think the ceiling actually goes up. It doesn't come down because now the problem become much harder. You're going to have professional developers and engineers and they're going to be, you know, put to work and then you're going to have a bunch of new coders that are coming in. There's a separate question which is, what does this do for SaaS? I just want to submit something before the next question, which is, I've been investing in infrastructure and enterprise for 10 years. I will tell you, SaaS has never been a technology problem ever. It's just not hard. Like if you have, if you build a SaaS app, it's not, it's not hard. It's never been hard. Right. And so the question is, is why do people buy SaaS? And the answer is you're buying a business process. Right? It's a business process that's been understood by another company that tells you how to run your business. It's never been about the technology or the software. So I don't think it changes that dynamic much either.

9:39

Yeah, it's really interesting you brought that up. I was sitting in Google Cloud next probably two and a half years ago and they put up a slide, agents for marketing, agents for finance. And I was aligning enterprise SaaS companies with these down the line marketing. Right. Human resources. And you could see the leaders in these businesses. And what's interesting is we have seen, at least from a markets perspective. I know you deal with a different market, but the public markets, the valuations are not doing well, probably with the exception of folks like ServiceNow. You know, you've seen, you know, folks like Salesforce and Workday and folks like that decline here. But talk me through your thesis on. Again, you'll never have the perfect data. Right? You have to have your data management in place. But I do think AI will help with that. So you have a data fabric that's accessible by the enterprise and these agents can hit that data fabric with the right security, the right access. Talk to me about your thesis of how this rolls out. We've seen some public battles between, you know, Marc Benioff and Satya Nadella. Right. That seem to go on for a year. So there is tension in the system for sure. And I do a lot of CIO roundtables and they are literally telling me we are finding a way to get off of this software vendor. You know, we are.

12:12

Yeah, yeah, yeah. So listen, let me, let me just provide kind of maybe like the historical sober view of this. Right. So why are valuations and growth lower from traditional companies? It's because we're seeing the largest movement in budget we've seen since the Internet. And when budget moves, it goes to new places. You know, are we seeing mass replacements of traditional software? No, we're not. Right. And so like historically software tends to get, it's not zero sum, it tends to get layered or you know, budget will move or it tends to slow down but you know, it doesn't tend to get replaced. Now why would you start replacing something like a system of record? Well, the answer is, is it doesn't evolve with the new technology. Meaning there were a number of companies that didn't make the Internet transition and they could have, they just decided not to. So like we as consumers are going to evolve our opinion on what it means to interact with software. Like when I interact with Salesforce, for example, I'm going to want to call it up and talk to it. I'm going to want to have an LLM. But this is a consumption layer change. It's not a fundamental change. The business process is still there, the guarantees are still there. So Salesforce 100% has the opportunity to evolve the consumption layer to be what the expectations of the user are. So I don't think, think there's something fundamental end to end that's going to disrupt all of SaaS. That said, it's up to the SaaS providers to evolve to meet these, these moving expectations and this moving budget. So again, I use the Internet as, as my anchor example when thinking through what's going to happen.

13:49

Yeah, I think that's really interesting though too. I think what you said is the most prescient, which is the experience has to change. And the reason that I think a lot of people, and I can say this as business owners, Pat, you share this probably sentiment with me is these systems of records are supposed to help you run your business, but they're actually not that easy to use. How many people you have to hire to be like, oh, I want to report on last year's business versus this year's business with this cut and this slice and this angle. And you need like a special person to like come sit at your desk with you. Or they can, thanks to SaaS, they can do it in another desk. But the point is, and drum up a report when in reality what we can do with ChatGPT or with anthropic now is we just ask it a question. The same way I'd ask you a question. And if it can actually go through the APIs and through the traps and find the data it will spit out and you can say, I want it in a, in a pie chart.

15:22

Yeah.

16:12

And you tell it, oh, I want it in a bar chart or I want it in a narrative that sounds like Eric Clapton singing, singing, you know, Tears in Heaven. And it can do that for you. And.

16:12

Right, but let me, let me just provide the other, the other side of. So I agree with everything you're saying, but let's look at the Internet. Right. So what did the Internet provide? It provided me the ability to like connect to software from my house, for example. So, you know, so it provided this fluidity with consumption, with access. What does AI do? It does the same thing. Like I can talk to it, I can use natural languages. But you still have all of the compliance, you still have all of the integrations. There's formal reporting things. There's also the business process on top. I still need to do pipeline reviews, I still need to do rollups. Like structured data is not going away and we Structure it to limit complexity. And the complexity is driven by the operations, it's not driven by the software. So I think the right way to view this is there's a reason that there is this complexity in this software. It's because we have complex business processes, we've got complex environments that they sit in, we've got a complex regulatory framework. And so you're going to always have that structured data. But it frees the individual from having a new consumption layer to work with. And so the right thing in my opinion for these SaaS vendors to do is to evolve with the evolving expectations of the users. But that doesn't mean it obviates all of this work of having to integrate into a complex operational.

16:21

Yeah, I just keep saying that when you have less total human users and you have more agents, they all probably need to refactor their business models to some type of consumption based on tokens and actions and not so much based on seeds.

17:36

Huge. That's a huge topic. And so do you. I mean you both probably remember the move from perpetual license on trim to the recurring. I mean that that gave rise to companies killed companies is one of the most disruptive things. Not all companies are even there yet. There's still companies today that are talking about this move. Now we're seeing another pricing change which is from recurring to consumption basis. And that's going to be a whole massive disruption at the same level of that. And we're seeing that right now. Absolutely.

17:51

Yeah. So Martin, I want to want to flip back to infrastructure. I think you talked about an infrastructure inversion driven by AI and cloud. Is there a contrarian view that you have on enterprise infrastructure that the market hasn't fully internalized yet?

18:23

I mean, I guess my contrarian view is like it really hasn't had an impact in the enterprise yet.

18:43

Listen, 100%.

18:50

That's a whole lot to come. I really think it's all on the. I mean there's very open questions. Like for example, I'll tell you like one open question is what's going to happen to central buyers and platform teams and IT teams if agents are making the decisions. Right. So let's say I'm a developer right now. So let's say pre AI. If I'm a developer and I want to use a database there, you know, like the IT team provides a set of infrastructure, a set of docs. I get onboarded and I know about them and I make these technical decisions that are in line with the policies of the organization right today. So I code every Night and I code every night. It's the most fun thing ever. And by the way, the fact that you're coding just shows that this TAM is expanding. We're all, we're all coding now and we didn't code before. So who is making a technical decision like if you're using cursor or if you're using cloud code, what's making the technical decision of infrastructure to use? The AI is making that decision. And so infrastructure is a multi trillion dollar business and you've removed the human by and large from actually making the decision of what to use. We have no idea what that means internally. We have no idea what that means to the industry. And so I think a lot of the real disruptions from AI are still on the come and we're just seeing very, very early glimpses of that through secular adoption by individual users.

18:52

So we've got only a minute or two. And I really appreciate. And Pat, we really appreciate. What about the, you're the infra guy and Pat kind of shifted us back here a little bit. But I mean I just keep hearing everything is constraint. You know, this is 26 is going to be the year of memory. You know, we're, we're inking nuclear deals for tech that doesn't even technically work yet in most cases because we're going to need to energize these racks. We got two ton racks, you know, now being delivered.

20:22

At least a one up did with a 2.6 rack.

20:53

Right.

20:57

I mean Jensen's breaking physics laws. But like, what is, what is your sort of read on compute capital efficiency, all these different constraints and like how is this going to reshape the, you know, kind of the way we're going to scale this whole thing going forward?

20:58

There's only one constraint and that's regulatory. You know, it's very interesting. So do you guys hear about data centers in space?

21:16

Oh yeah, absolutely.

21:22

It's, it's a ridiculous concept. You know, like you have to go all the way to space. It's so stupid. So it turns out it's not stupid for exactly one reason. Can you guess why that is?

21:23

Regulation. Less no rules.

21:33

So it's 100%. By the way, the numbers pencil out just because of regulation. It is so onerous to break ground in the United States. It makes more sense to send the data center to space. And so listen, we are a very, very innovative species. We're a very mature industry. If we need capacity on bandwidth on chips, we know how to do it. The issue is getting the Bureaucracy out of the way to do it.

21:36

That makes sense.

22:01

That was too fast, though. He didn't even give me a few good sound bites about like, you know, the fact that we can't build enough. Because your point though, in the end is like, we need to build more fabs. That takes time. You need to build more data center. That takes time.

22:03

You need to. Right, but, but the, the, the issue really is, is that they can't listen, if you went to Google today and you're like, you can break ground tomorrow, we would have the capacity we need. We actually have the latent capacity as an industry to do this. That the take time is purely a bureaucratic and regulatory morass.

22:14

Power is very much. Power is very much a challenge, whether it's SNR fusion that doesn't work yet at scale. Daniel, you're involved in a couple projects as well.

22:33

I advised Trump on the TAE deal. But the, the thing I was gonna say is I heard you can open daycares very quickly.

22:48

So.

22:56

If we can't do a data center thing, guys, there's other ways to make money.

22:58

Certainly you get them funded very quick. Exactly.

23:04

Well, you actually don't need to open them to make money. And that's the beauty. So, Martin, I just wanna say thanks. It was a lot of fun.

23:07

Absolutely.

23:14

Our off the record segment unfortunately isn't as long as this conversation really needed to be because I could have drilled in quite a bit longer. So I'll send that note back to the prod or that we got to figure out ways for these longer forms to, to let us keep going. But a lot of fun to chat to you. And by the way, that was the best answer on the regulatory. Best answer I've heard on something in a long time. Because I just, I've just been thinking about it too much because I'm thinking about like I'm not actually kind of going back with the right message is like, guys, if we just fix regulatory, you can fix everything else.

23:15

Because I mean, just, I mean, again, maybe we're rolling. Maybe I'll tell this. I talk to these people all the time. We have a full portfolio. I am telling you, the long pole by far, by order of magnitude is breaking ground. That's it. We know how to solve power, we know how to build foundries, we know how to do these things. It's not a technical issue. And the people that are sitting on the top of these big data centers know that. And by the way, is China smarter than us? No. Do they have more production capacity than us? No. Are they ahead of us? Yes. Why? Because it's like full throated endorsement of, of building out and this is what we need to do too.

23:46

That like a coal fired plant a week there is like what I, what I've heard or something like that that's going on.

24:20

Unbelievable. The other thing is we know we don't do this stuff. We know how to aggregate capacity, we just need to do it.

24:26

I started working with the Chinese in the mid-90s and they would show me schematics of where a metal bending factor would be and it'd be a forest and there's no roads. And then a week later forest has been cleared, people have been moved and roads were put in and then a week later power came in. I mean it was absolutely on. Unbelievable. So yeah, Martina, I really appreciate you coming on the show, would love to keep in touch as we move forward. I don't know if you're sending a contingent to Davos, but a lot of your compatriots are actually going the other way. No, no, I hear you. It's funny, a lot of your, a lot of your peers are headed there for the first time in years and Daniel and I, for the first, you know, we just started going last year when it looked like we were actually going to talk about technology and yeah.

24:33

So yeah, to them, brother, I was.

25:28

Thinking about your regulation, your regulation commentary, but quite frankly the, the biggest discussion there last year was we over regulated ourselves and these were people from the, from the eu and DJT is going to be there and Vance and, and.

25:31

Folk, I mean, did you just see like, like Italy just fined cloudflare, what, $17 million for like something that they basically couldn't fix and Matt Prince is going to, I mean I just feel like we're in this crazy thing where we're extracting like we're slowing down companies taxation.

25:50

This is what this is.

26:08

That's the, the, the EU's fundraising mechanism is, is, is, is, is basically regulatory regulating US companies. And by the way, if you want to go slowly, you want to feel like we're going fast, just look at what Europe does and you'll realize that we're going really fast here in the United States because that's, every time I.

26:09

Feel bad about the United States, I just think about the EU and I.

26:27

Feel better and that's there. So China makes us feel slow and the EU makes us feel fast and you know, we're going to land somewhere in between. That was, by the way, Pat, my worst ever exit of an interview. Like we exited and then we did five more minutes, but it was totally worth it. Martin, thank you so much for chatting with us. Let's have you back on the show again soon, gentlemen. And there you have it, everyone. That is off the record.

26:29

Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X16Z and subscribe to our substack@A16Z substack.com thanks again for listening and I'll see you in the next episode. This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product. This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with A16Z. Such advertisements, companies and individuals are not endorsed by AHC Capital Management, LLC, A16Z or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.

26:55