CEOInterviews.AI
Start App
Get an email when Ben Horowitz gives a new interview
One email per new interview. Unsubscribe in one click.
Ben Horowitz
Co-founder of Andreessen Horowitz, Andreessen Horowitz

Ben Horowitz: Building Trust and Standards for the Agent Economy | AI Assurance & Governance Summit

📅 Oct 15, 2026 TrustModelAI 35 MIN 6 VIEWS 64 SEGMENTS · 3 SPEAKERS
A wide-ranging fireside chat on the evolving landscape of AI, venture capital, and why the agent economy needs standards now. Recorded live at the AI Assurance & Governance Summit, Stanford Faculty Club, October 1, 2026. Topics covered: Community and Culture Ben discusses the “Pay It Full” foundation, an initiative providing pensions to the original creators of hip-hop who never benefited financially from the genre’s early growth, and pushes back on fears of AI eliminating jobs, framing it instead as a tool for building new products and services. AI Investment and Moats Traditional software...

What Ben Horowitz said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Ben Horowitz discussed the agent economy, AI moats, open source AI, American manufacturing, and AI identity. He argued that worker fears about AI job loss are mostly unfounded, as AI creates more jobs than it eliminates. He said traditional software moats are weakening, but money, first-mover advantage, customer inertia, networks, and brand remain strong. On open source, he said China has taken a 'pretty large L' for the US, with Chinese models widely used in universities and enabling Huawei's chip ambitions, though he noted it's not over. He identified the biggest bottleneck for American manufacturing as a loss of expertise, not labor costs. He strongly endorsed cryptographically signed identities for AI agents, calling it 'fundamentally important' and necessary for trust and liability. He criticized AI safety practices, recounting an incident where an Opus 4.5 agent hacked a website and Anthropic blamed prompting. He supported AI model 'nutrition facts' labels as helpful for quantifying risks.

Key takeaways

  1. Money has become a moat in AI; first movers can buy more GPUs and advance faster.
  2. The biggest bottleneck for American manufacturing is lost expertise, not labor costs.
  3. Anthropic blamed prompting when an Opus 4.5 agent hacked a website, showing safety gaps.

Numbers and commitments

FigureWhat it refers toTypeAt
$100,000 a year Annual pension grants for hip-hop pioneers commitment 2:06
over 10x Price premium of frontier models over open-source metric 11:57

Chapters

  1. 0:00Paid in Full Foundation fundraiser
  2. 4:07AI job fears and product mindset
  3. 6:08AI moats and first-mover advantage
  4. 10:57Open source AI and China
  5. 15:25American manufacturing bottleneck
  6. 21:23Agent identity and crypto
  7. 25:51AI safety and liability
  8. 33:13AI model facts labels

Questions asked in this interview

10
  1. 3:14So, Ben, do we still have tickets left?
  2. 3:32If there is no people left, is there anybody to even buy the software?
  3. 5:40And you can hear me well, right, Ben?
  4. 5:51What is the moat that you are underwriting right now?
  5. 10:33Is your view still the same that was put out before or changed?
  6. 14:42Is it capital, talent, permitting, nerves?
  7. 21:10And so, do you think this is a venture scale category or you think that this will turn into some kind of a standards or compliance that nobody makes money out of it?
  8. 24:56What would be your advice to the corporate governance boards or to the government?
  9. 27:53... not into more regulations, but do you think that the companies have an obligation to materially disclose that to its shareholders about what AI risk they have and what agents they're deploying and whether those agents have the containment?
  10. 32:30Will that help solve the problem like very minimum self-disclosure that we're talking about?
Host 0:02 ↗
Our next session is a fireside chat with Ben Horowitz and Karl Mehta discussing the view from Sand Hill Road where the next great AI companies get built. Fortunately here, Ben is running late a bit, but I believe he's going to join on Zoom. So, Karl Mehta, thank you.
Karl Mehta 0:20 ↗
All right, thank you.
So, unfortunately Ben's running late, but we're going to get started. He's just sent us the Zoom, so we're going to get started.
So, Ben, I'm going to introduce you and then by that time the sound guy should be able to hear at Stanford, should be able to start. We got a whole room full here and we had a fantastic session with John Chambers and Aman Bhutani. So, Ben, as you all know, is the co-founder and general partner at Andreessen Horowitz, the best that I know of in the Valley. Today the firm has over a hundred billion dollars under management across multiple funds and invests across seed to growth stage technology companies in a broad range of industries including AI infrastructure, consumer and enterprise applications, bio plus healthcare, crypto, and companies building towards American dynamism.
Prior to founding Andreessen Horowitz, Ben was a co-founder and CEO of Opsware, formerly Loudcloud, which Hewlett-Packard acquired. He has also served as a vice president and general manager at Netscape Communications. Needless to say, he brings to this discussion experiences both building and investing some of the most consequential technology companies and has become a key voice in today's AI conversation. Welcome, Ben.
Ben Horowitz 1:40 ↗
All right.
Karl Mehta 1:45 ↗
That's great. Thank you, Ben. Well, before we get into AI, let's talk about the fun stuff. You and I are going to be at the fundraiser for Paid in Full in Vegas. Tell us a little bit about that. Been doing it for several years. It's amazing stuff that Ben does and I totally love it.
Ben Horowitz 2:06 ↗
Yeah, so it's a program to give pensions to the original kind of people who created hip-hop. So, you know, hip-hop became like a massive cultural phenomenon, the biggest musical category for a couple of decades. And you know, influenced fashion and everything else, but the people who created it, artists like Rakim, Scarface, this year Kurtis Blow and the D.O.C. You know, because it was so early, didn't make much money, I would just say. And so, what we do is we go back and we honor them by giving them pensions, essentially. So, grants of $100,000 a year. And then, we have a big event to honor them and it's unbelievable fun and just a great atmosphere if you like hip-hop. I highly recommend it. It's in Las Vegas, October 24th. paidinfullfoundation.org
Karl Mehta 3:14 ↗
Yeah. So, Ben, do we still have tickets left?
Ben Horowitz 3:18 ↗
Yes.
Karl Mehta 3:19 ↗
All right. Okay, good. So, you're a music lover, you love hip-hop like me, look it up, Google it, Paid in Full Foundation, October 24, Vegas, and come join us for the fun. It's an amazing event.
All right, so let's get started. So, Ben, a16z and we love the newsletter and a lot of opinion and a lot of great podcasts. So, one of the things that a16z put out back in April that the founders and the workers are afraid of the opposite things. Founders think that they are moving too slowly, whereas workers think that they are just not needed. I mean, the fear of losing the job. So, you have the two opposite things. You left the hardware open. If there is no people left, is there anybody to even buy the software?
Ben Horowitz 4:07 ↗
Yeah, so I actually think that the worry of the workers is mostly unfounded. Like, I think we get... So, the way that, you know, the AI community has kind of thought about AI is, you know, gets a little, I would just say, cultish from time to time. So, like, this is a technology built by humans for humans. Not, you know, and there's a great kind of motto from a new company called TypeSet, where they say, "Build prod, not God." And I think that's like, for the most part, what's actually happening is that these are products. They help you do many more things than you could before they came. But, kind of the least interesting thing, you know, at least from my point of view, is kind of replacing jobs. The most interesting thing is doing kind of new stuff and more stuff. And if you look at our business, way more kind of jobs have been created from new companies than jobs have been eliminated from old companies getting more efficient with AI. And from an investment standpoint, like, if your best idea is to, you know, use AI to reduce the number of accountants you have, or what have you, then like, you're probably out of ideas. Because, you know, that's a pretty mediocre idea.
Karl Mehta 5:40 ↗
Yeah. Well, makes sense. Absolutely. So, another piece that a16z... And you can hear me well, right, Ben?
Ben Horowitz 5:50 ↗
Yeah, we can hear you.
Karl Mehta 5:51 ↗
All right, cool. So, another piece you said that product runways have gone from 10 years to about 5 weeks. Yeah? If nothing stays defensible for 5 weeks, what is still defensible? What is the moat that you are underwriting right now?
Ben Horowitz 6:08 ↗
Yeah, so let me take a step back and kind of describe what's happening. So, like in my whole career and really in the whole history of the technology industry, the one thing that we knew was you can't kind of make up a technical lead by throwing money at the problem. You couldn't. In the old days, you can't catch the leader by hiring a thousand engineers. That's actually not going to speed you up, it's just going to slow you down because the mythical man-month, nine women can't have a baby in a month, like that never works. In fact, my favorite joke used to be what's a man year? 700 IPMs before lunch. You know, it doesn't work. However, that has changed. So, you can, you know, put money into, you can if you have enough money, you can buy TPUs and if you've got data or if you buy data, you can kind of train a model and or post-train a model to solve like very hard problems.
And so, software itself is kind of one of those problems. So, the idea that software and software's always been a moat and that you've got to write good software and then you have the data that the software contains and then you have a user interface that's kind of bespoke to your users and so forth. But, if the AI can write the software very fast, if the AI can migrate the data very fast, and if that's using the software not a human, agents don't really care about good user interfaces as much. They're like pretty adaptable to things, to AI agents. And so, that traditional moat is kind of not what it was. But I, you know, still I would say, you know, the SaaS apocalypse is overblown. And there's a lot of other things that go on. So, you know, the other thing that's changed is first mover advantage used to be a relatively small thing and now it's a much bigger thing because you know, if you're the first mover and you get the higher valuation, then you can buy more GPUs and then you can advance your product faster than the competitor coming behind you. You can build a bigger model. You can do things like that. And so, you know, one moat is just money. Money has become a moat in a way it didn't used to be before.
And then all of the traditional moats like, you know, in enterprise software and enterprise anything, possession is 9/10 of the law. So, you get the customer. Your ability to keep that customer is very, very strong just by inertia. And so, you know, that kind of thing matters a lot. I think networks, they'll matter. Like if you can build a network effect in your product, that's a very, very powerful thing and difficult to defeat. And so, you know, many of the kind of... brand is a, you know, it's a real moat. Like how different at this point is ChatGPT than a lot of other things, but you know, that's what people know. That's what people have. So, even with, you know, as differentiation gets thin and migration gets easy, having a brand, having a customer, having a network are all things that are very, very meaningful competitive...
Karl Mehta 9:37 ↗
Cool. So, the product differentiation is thinning down, but the brand, the first mover advantage, customer network, access to capital, all of that still somewhat contributes towards a moat.
Ben Horowitz 9:53 ↗
Yeah, I think it's very relevant. And we're seeing that, you know, we're seeing that play out. It's actually been extremely hard even though like it's very well known how to build a state-of-the-art foundation model. Like, there haven't been new architectural breakthroughs and those kinds of things recently. And so if you had the money, you could potentially run down OpenAI and Anthropic. But nobody's got the money.
Crazy amount of money.
Karl Mehta 10:33 ↗
So, switching gears, one of you had another viewpoint that a16z put out sometime back that America has lost the open source AI culture war to China with DeepSeek, you know, capturing the movements, momentum and the values that are embedded in the model weights. So, what's your view? Is your view still the same that was put out before or changed?
Ben Horowitz 10:57 ↗
Well, I think, you know, there's kind of good news, bad news. And there's an important economic kind of factor underlying this. So, the challenge with open source models is, you know, you can't charge as much for them because, you know, somebody else can run your model and serve it, you know, like if you charge what Anthropic charges, somebody would undercut you with your own model. And so, as a result, you can't spend as much money training model. And what the Chinese companies have allegedly done, allegedly, is kind of distill the large models and save a tremendous amount of money on training. That's been really good for them, but it does have the problem that they are kind of fundamentally behind the state of the art cuz they're kind of copying whatever model's out, and so they're always a bit behind.
It's still been a very good business and those companies are worth a lot of money, but it's enabled, you know, kind of the frontier models to still have a huge pricing advantage and that kind of thing. They charge, I think, you know, over 10x as much as the open-source models.
There's a few problems though with the Chinese models putting up in source. One is though all of our universities use Chinese models, not US models. And then, they're in very, very kind of broad use in applications because, you know, you have a lot of control over your products if you're using an open-source model. And as a result of that, they're like fundamentally important in the ecosystem in AI. I take two important ways. One way is just the, you know, kind of cultural influence. So, you know, if you ask the model invented anything, it's going to give you the answer that's okay by, you know, the Chinese government. And, you know, whatever you may think about that, like that's, you know, a potential issue. And then, the other one is I think more important, which is it kind of has generated like a very big opportunity for Huawei and the Chinese chip industry. So, we have a big huge lead on GPUs, led by Nvidia, like massive. And that lead, you know, because the open source models are, you know, and by somewhat dictated by the Chinese government are being kind of architected to run on the Huawei family. Huawei has a real opportunity to become a dramatic competitor in the chip market globally. And that, you know, if US open source had won, that would not be the case.
So, I don't think it's over. And, you know, Thinking Machines has a very interesting kind of open source approach. Mistral out of France has an open source model. Gemini's got an open source model and then Nvidia is very, very, very keen on kind of promoting those things. And, you know, potentially helping fund those things. So, I don't think it's over, but I would say it's a pretty big hole. Well, like if you had a US economic strategy, I would say that's a, you know, that's an L.
Karl Mehta 14:39 ↗
Yeah.
Ben Horowitz 14:39 ↗
We've taken a pretty large L on open source.
Karl Mehta 14:42 ↗
Yeah. Well, I mean, speaking of the Chinese dominance, I think a16z is probably the first, if not the only, to when you guys started the American Dynamism Fund. And I was super excited that you want to drive the American innovation, the American dynamism. And your 2026 thesis is about AI native industrial base, factories, energy, automation tasks, labs. I think Ryan McIntosh wrote that the nations that master the electro-industrial stack will define both industrial and military technology. So, what's the real bottleneck for the American dynamism? Is it capital, talent, permitting, nerves?
Ben Horowitz 15:25 ↗
Yeah, so I'd say, and you're really kind of talking very specifically about manufacturing, which is kind of the big challenge that we have right now. So, if you look at the new kind of offensive and defensive weapons that are getting built in the US and if you look at kind of, you know, how we're kind of supporting a modern military, you know, with energy and other kinds of things. The US, like the technology that we're building is definitely state of the art. Like we are, I would say, as good as any country on those fronts. The challenge we have is, you know, in war you need a lot of everything. And we're kind of seeing this with the Iran war, it's like, well, why don't we kind of go in hard if we're like, we're in already, we should finish the job. You know, saying a part of the conversation is like, do we have enough, do we have enough like, you know, missiles and kind of ships and everything to actually pursue the war to the fullest and still have kind of, you know, enough of a defense for all of our positions around the world? You know, and to help Ukraine and everything else we're trying to do. And it turns out like we lost our ability to manufacture at high scale the way we used to as a country. And yeah, this is, you know, in a way it's shocking because we invented high scale manufacturing. I mean, the US became the US that we know and love and live in today because, you know, going into World War II we were the best manufacturing country in the world. We won World War II with our manufacturing capability. We saved, I would say, Europe with our manufacturing capability.
Karl Mehta 17:14 ↗
Yeah.
Ben Horowitz 17:15 ↗
But over the years, you know, we kind of, I would say, got arrogant and literally forgot how to do it. And the interesting thing about it is if you go into the details, you know, I think the perception in America has been, well, we don't manufacture because labor costs are too high and they're so cheap in China. That's not actually true. So, labor is not the thing that makes it difficult for us to compete on manufacturing. We can train people. And then if you factor in the labor costs versus the cost of shipping something from China, like we're actually in a pretty good position.
The problem we have is that one, we have just made it really, really hard to manufacture things here. So, we have all kinds of corrupt bureaucracy around. And you know, there may be a valid environmental law, but the way it's enforced just means that, you know, building a plant is extremely delayed. A lot of, you know, kind of our position on clean energy has led to higher cost of energy. Energy is like a really important input on manufacturing. And so, you know, and depending on which state you're in, like that's maybe reasonable. But the biggest thing, the biggest bottleneck is we don't have the expertise anymore. And not, you know, like we can train people to work in manufacturing plants and we're very good at automating manufacturing plants, but we don't have a lot of people in the United States who actually know how to run a manufacturing plant. And I know it sounds a bit surprising, but you know, most of the people who you look at in the American dynamism portfolio running these things, if they don't come from like SpaceX or Anduril or Tesla, like they just don't know how to do it. And so, we really kind of knew how to do something as a country and then lost that ability. So, we're, you know, working hard to try and regain it, but that's a significant bottleneck right now.
Karl Mehta 19:17 ↗
Yeah. So, the physical AI and the robotics should bring the manufacturing resurgence back to America.
Ben Horowitz 19:25 ↗
I think it will help a lot. I mean, you know, like we have a lot of wood to chop. So, like we are furiously trying. I think that, you know, Doug Burgum as the Secretary of the Interior has been very, very helpful in kind of enabling us to do things that were difficult before. And we have very smart people working on it. But right now, like if you look at the manufacturing capacity from the new companies, like we're still, like we're working hard to ramp that and we don't need a lot, you know, we need like maybe two or three companies that are really good at it and we'll be fine. But you know, we're building that capacity now. And you know, the old school kind of prime have no capability in kind of the modern tech. They can't do AI. They don't know how to do AI and then they don't have the kind of capital structure where you can hire kind of brave AI researchers or the people necessary to work on these things. So, like it does, you know, basically need to be new companies.

25 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Horowitz, B. (2026, October 15). Ben Horowitz: Building Trust and Standards for the Agent Economy | AI Assurance & Governance Summit [Interview transcript]. TrustModelAI. CEOInterviews.AI. https://ceointerviews.ai/interview/3180461/

MLA

Ben Horowitz. "Ben Horowitz: Building Trust and Standards for the Agent Economy | AI Assurance & Governance Summit." TrustModelAI, 15 Oct. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/3180461/.

BibTeX
@misc{horowitz2026_3180461,
  author       = {Ben Horowitz},
  title        = {Ben Horowitz: Building Trust and Standards for the Agent Economy | AI Assurance \& Governance Summit},
  howpublished = {Interview transcript, TrustModelAI. CEOInterviews.AI},
  year         = {2026},
  month        = {oct},
  url          = {https://ceointerviews.ai/interview/3180461/},
  note         = {Speaker-attributed transcript with timestamps}
}