Pod Insights Decoder AI is a money pit — here’s why investors don’t mind About this episode AI investment is massive, but AI profits are not — and yet investors seem confident massive AI fundraising will one day translate into sizable AI profits. To break it down, Verge Deputy Editor Alex Heath guest hosts this episode of Decoder featuring Menlo Ventures partner Tim Tully and AirStreet Capital founder Nathan Benaich.
Links:
2024: The State of Generative AI in the Enterprise | Menlo Ventures
State of AI Report | Nathan Benaich
AI Index Report 2024 | Stanford HAL
How companies are spending on AI right now | Tech Brew
OpenAI Is growing fast and burning through piles of money | NYT
Amazon to invest another $4 billion in OpenAI rival Anthropic | The Verge
Agents are the future AI companies promise — and desperately need | The Verge
Anthropic’s latest AI update can use a computer on its own | The Verge
OpenAI reportedly plans to launch an AI agent early next year | The Verge
Is AI hitting a wall? | Command Line
Credits:
Decoder is a production of The Verge and part of the Vox Media Podcast Network.
Our producers are Kate Cox and Nick Statt. Our editor is Callie Wright. Our supervising producer is Liam James.
The Decoder music is by Breakmaster Cylinder.
Learn more about your ad choices. Visit podcastchoices.com/adchoices
Transcript 0:00 Speaker 00 Give you and your team more time to focus on the work that matters. This is AMEX Corporate, one intelligent platform for company spending. With smart expense management to help capture details, categorize spend, and streamline expense submission and approval, there's less manual work for your team and greater visibility and control for you. Built to simplify, designed to scale. Get started at go.amex/ammexcorporate. Terms apply.
0:32 Speaker 01 Hello and welcome to Decoder. This is Alex Heath, not Eli Patel. I'm deputy editor at The Verge, and I'm guest hosting a couple of episodes of Decoder this month as we head into the holidays. And I'm very excited for what we have coming. I'll be back later this month with an exclusive interview with the CEO of chipmaker ARM, so stay tuned for that. And if you want to sneak peek, be sure to check out Command Line, my weekly newsletter about the tech industry's inside conversation. But today, we're talking about a topic I've been focusing on a lot lately. It's at the heart of the most ambitious and most costly gamble the tech industry has arguably ever made. That is, of course, artificial intelligence, and specifically the idea that one day soon, big companies across all kinds of industries
1:15 Speaker 01 will be spending hundreds of billions of dollars on AI products. Today, though, that's just not happening. Meanwhile, AI companies have taken in tens of billions of investment this year alone. OpenAI raised a staggering 6.7 billion, surpassing XAI's 6 billion fundraise five months earlier, and Anthropic just raised another 4 billion from Amazon. The list goes on. We've heard endless hype from CEOs on Decoder about what this technology is supposed to do and why this investment is justified. But so far, the actual spending on AI products nowhere near matches the level of investment being made into the models themselves. According to a recent report from VC firm Menlo Ventures, AI spending is growing fast, but hit only 13.6 billion in 2024.
2:01 Speaker 01 That barely covers the year's two largest AI fundraising rounds from OpenAI and XAI. So how is this money actually being spent on AI? What are companies actually buying and what are they doing with it? And why do investors think the return on investment here will be worth it? To find out, I caught up with two AI investors. Tim Tully, a VC at Menlo Ventures, who co-authored that report on AI spending, and Nathan Benaich, author of the State of AI Report and founder of Air Street Capital. We dove into the data, the big trends they're seeing with AI in the enterprise, and where we think all this is going next, including when these AI companies are going to start generating the kinds of profits they'll need to justify the money they're spending already. But first, to put everything in perspective, I wanted to start with Tim,
2:44 Speaker 01 the co-author of Menlo's major AI enterprise report, and what he thinks the data tells us about the state of the industry today. Tim Tully, welcome to Decoder.
2:54 Speaker 02 Hey, Alex, how's it going? Thanks for having me on.
2:56 Speaker 01 Your firm, Menlo Ventures, puts out this report every year about the state of AI in the enterprise. And it just came out. Up top, I want to disclose that your firm is a pretty large investor in Anthropic, which is a very key player in the model race and where spending is going in the enterprise. But it's a really interesting report. You talked to a lot of people to gather a lot of unique data. I want to get into all that, but first, can you just give us a little bit of background about yourself and Menlo as a firm? And what you guys do, especially as it relates to AI?
3:32 Speaker 02 Yeah, Menlo is a 47-year-old firm in Silicon Valley. We've been investing throughout enterprise and consumer for a number of decades now. We've done investments in companies like Uber, Carta, going all the way back in the old-school Hotmail, things like that. More modern companies are like Benchling, Chime, obviously Anthropic. And then we're in some pretty interesting AI companies like Pinecone, Neon, and Unstructured.
3:57 Speaker 01 The big headline out of your all's report about the state of spending in AI this year is that spending surged 500%. So it's hit about $13.8 billion, up from just $2.3 billion in 2023. When you drill down past that headline number, what are you guys seeing? What is changing, especially as it relates to last year?
4:21 Speaker 02 Last year in 2023, a lot of the dollars were going into simply infrastructure, sort of picks and shovels, you know, places where VCs like myself like to invest early on when they feel like there's a big, big inflection point in what's happening in technology, like the mobile phone, and we see it being similar to that. You see a lot of these infrastructure platform level investments. Now it's turned to the application side. Like you saw in the report, there's a lot more application spend this year. It increased pretty significantly and it went to places that, you know, would really surprise you, like legal, fintech, healthcare. I mean, that was really, really surprising to us. I mean, we heard it anecdotally, but to see it in the data was, was, you know, pretty grounding for us.
5:01 Speaker 01 And when you say applications, you mean like hyper verticalized AI applications for certain sectors?
5:08 Speaker 02 We talk about it as departmental spend. There's departmental spend where IT departments, R and D departments, you know, go to market departments are buying. And then there's more verticalized spend in healthcare, legal, financial services. So people are buying legal software that's heavy on AI. People are buying clinical documentation software for healthcare. That's, that's heavy in AI and then accounting software that's heavy in AI.
5:30 Speaker 01 Who are the biggest spenders and what buckets of spend are growing the most? I think you guys called out coding. That's something I just hear and see anecdotally. Coding assistants are growing pretty rapidly. But I guess aside from coding, what did you guys see in your research?
5:48 Speaker 02 Yeah, from a use case perspective, it was, is obviously coding. I mean, this is probably making me 70 to 80% more productive as an engineer. Cause you know, I'm not really just an investor. I, I write code every day and the productivity for me is off the charts from the code generation standpoint. So it's, it's very helpful. Second after that would be what we call support chat bots. And then enterprise search and retrieval after that. So there'd be companies like Glean, for example, where they're indexing documents in the various departments throughout the enterprise.
6:18 Speaker 01 So obviously, that 500% surge is a big surge. At the same time, when I see 13.8 billion spend in 2024, that's, that's big, but it's really when you just combine OpenAI and XAI's funding rounds this year's together, it's that number. Like in terms of the actual money being plowed into these companies, especially at the model layer, we're still not really seeing anywhere near the level of ROI as what's going in. And I'm curious, as an investor and as a engineer, someone who's been in the AI space for a while, how do you look at that?
6:56 Speaker 02 Well, I think this is just the tip of the iceberg in terms of what that revenue is going to look like. You, you hit the number correctly. It's, it was six and a half billion for foundation models, but I think myself and others, including who invest in the space, strongly believe that these are the APIs of the future upon which the next generation of enterprise software is going to be built. So that 6.5 number, what if it multiplies by five or six X again next year and becomes 36, $40 billion, right? Like, then you're starting to approach valuations that you mentioned in your point. So I think it's still early days. And when you see growth like this, you get excited and you want to invest. And, and this is the result.
7:33 Speaker 01 And you guys don't see a ceiling in the near term in terms of appetite for budgets to shift more towards these AI applications as they get better and better. You think that we will see another, you know, maybe not 500%, but triple digit percentage jump potentially next year?
7:50 Speaker 02 I think so, because again, the architectures are, they're still settling, right? How you build the software is, is still evolving and it's going to get better and better. And I think as that software gets better and we learn how to interact with foundation models in even better ways, you're going to see better results in the software and you'll see more software sales.
8:09 Speaker 01 And on the foundational model layer, you have an interesting point here, which is that OpenAI, you say seeded market share and enterprise AI usage declining from 50% to 34%, while Anthropic, a company that you guys backed, doubled from 12 to 24. Why is that happening?
8:26 Speaker 02 We didn't ask questions like, why did you switch necessarily? But what I hear anecdotally is that people are using Anthropic heavily for code generation. It's that use case that we talked about earlier. You talk to folks and I talk to the same folks that you probably do and you ask what model you're using and why. And you hear that there are certain use cases that people prefer Claude for, right? Code generation, some content generation as well. Structured data tends to do really well with. So I think people just are starting to prefer the output that they get from Claude and they see it as a better model for certain sets of use cases. And that's why you saw it double this year.
9:02 Speaker 01 Are the sectors of startups that are already doing really well, making a lot of money, kind of the same as where big enterprises are spending? Does it map directly? So like coding or are the coding startups that are supplying that tech, the ones really benefiting right now at the earliest stages?
9:20 Speaker 02 I think everyone's benefiting, frankly, I mean, across the board. I mean, you're seeing every department and every vertical sort of thriving, to be honest, on what they're doing with AI. It's pretty breathtaking, to be honest. So it's not just, you know, obviously the coding guys are doing quite well. It's also, it transcends into other departments also, right? You see legal tech companies doing quite well.
9:42 Speaker 01 What specific vertical is growing faster right now? Is there a category that you're seeing really starting to uptick?
9:49 Speaker 02 Definitely legal tech is growing very, very fast. Customer support as well. And really what that is, is the early movers who deal primarily in text, right? Because the foundation models do text to text translations effectively is what they're doing, right? And so what you saw is legal, fintech, digital marketing, these are the ones that were the early adopters of these of these things. And so because of that, obviously their revenue profiles are going to be a little bit higher because they were at the gates earlier. So it's really those, but you know, I think growing very, very quickly is obviously anyone who's doing code generation as well.
10:27 Speaker 01 I think photo and video generation has been largely held back from being at scale by compute bottlenecks. And we've seen that with OpenAI and Sora, and the fact that they still haven't released it. And I'm curious how much you think about the multimodal aspect of AI and how much that will affect the enterprise when that gets scaled out.
10:47 Speaker 02 I think about it a lot. I mean, we have some investments in companies that are doing, you know, diffusion type models as well, not just in the, in the, in the video and image side, but also on the biotech side as well. So, you know, we're definitely investing, but I think you hit the nail on the head, right? The compute cost on those is a bit higher. And so it's a little bit more challenging to have some of the success that maybe the text to text folks have had.
11:14 Speaker 01 We need to take a quick break. We'll be right back.
11:25 Speaker 03 This episode is brought to you by Subaru. You know that place your mind always wanders to? The lake, the mountains, the homes of the people you love most? The completely redesigned 2026 Subaru Outback is built to help you get there. With legendary capabilities, standard symmetrical all wheel drive, and the confidence to go further. The 2026 Subaru Outback. Love never looked so good. Visit Subaru.com slash Outback to learn more.
11:54 Speaker 01 We're back discussing AI spending in the enterprise and how it's supposed to turn into big profits in the future. Before the break, you heard Tim discussing Menlo Ventures' latest report and how one of the big findings was that companies are still mostly spending on access to foundational models like OpenAI's GPT series and Anthropic's Claude.
12:13 Speaker 01 Spending on these kinds of models jumped from 1 billion last year to 6.5 billion this year, accounting for close to half of all AI enterprise spending. And that spending is growing more diverse, with Claude eating into OpenAI's market share because of its capacity for code generation. But most of these AI model companies are not yet profitable and far from it. OpenAI is on track to lose roughly 5 billion this year, in fact. So when I caught up with Nathan Benet from Air Street Capital, I wanted to know what it might take for these AI firms to flip from burning billions a year to being profitable. Nathan authors a massive deep dive into the industry every year called the State of AI Report. And one of his key findings was that while AI revenues are rising across the board,
12:56 Speaker 01 there are some serious questions about the long-term sustainability of what's happening. So in your report, you write, "Many of these companies currently have no identifiable path to profitability. However, this isn't true for everyone as the biggest model providers see revenue begin to ramp up." So who are the haves and have-nots right now?
13:17 Speaker 04 Well, there's certainly the big model companies, which are not profitable and by far, those are kind of like the have-nots with regards to profit margins. And then the haves would be companies that probably provide much cheaper outputs, like whether that's images, Midjourney being famously quite profitable business. Audio generation is much cheaper than it is to run like massive models because the models are smaller and the data is easier to work with. These companies are pretty profitable. And in some ways, you're just in the early days of this technology. So you're just trading off profit margins and profitability for growth. And I think that makes sense for so long as the market's willing to provide you those tools. I will say though, that a year or two years ago,
13:59 Speaker 04 the path to profitability was a lot fuzzier than what it is today, in part because there just weren't that many optimizations that came into how the models were built, how they're run and all the like data infrastructure and cloud infrastructure around it. And to some degree, it's because the people contributing to progress in AI were pretty much AI research people, not people who worked for decades on cloud infrastructure and optimizing like this software runs really fast and is cheap on all sorts of devices. And all those people have started to move en masse into AI. And so I think you can bet just by the sheer capability of individuals who have done optimizations in other software settings and other technical settings, moving into AI will actually bring a lot of cost improvement, which will eventually like help margins.
14:44 Speaker 04 So it's one of the reasons why I think you see like an order or two orders in magnitude reduction in cost per million tokens, which is the unit economics of these models from one or two years ago to models that exist today for systems that are of similar intelligence, if you will.
14:59 Speaker 01 Do you think there's going to be a moment where OpenAI, which famously loses billions of dollars a year, Anthropic, et cetera, these big leading players, the model providers, become very profitable? And have you thought about what that scenario would be? How they get there?
15:17 Speaker 04 It's tough. I think you really need to see like hundreds of billions of dollars of revenue, I think, for that to be true.
15:25 Speaker 01 In aggregate or each of those companies making hundreds of billions?
15:29 Speaker 04 Probably say each of those companies. If you assume that the race for bigger and bigger models continues, like if we stop now, I think we could, we could have fairly profitable companies. And if there were more price discrimination with the services, these companies sell one license to somebody in a hedge fund for the same price that they sell a license to a student doing their master's degree. So that, you know, that doesn't really make a ton of sense. So it's a better price discrimination, maybe some more vertical products. And then we stop on this race of making ginormous systems that require a huge compute, then I think there's a path. If we don't, then you're going to need like hundreds of billions based on the compute spend.
16:12 Speaker 01 Do you think it's likely that they get to hundreds of billions? That seems like a tall order.
16:17 Speaker 04 It really just would require huge parts of actual labor getting translated into model jobs.
16:26 Speaker 01 There's early signs that, you know, you see like Klarna say they've let go of hundreds of customer support roles because of AI. There's signs that's starting to happen. But yeah, when Klarna cuts hundreds of jobs, that doesn't make OpenAI more profitable. There's a direct correlation. Maybe it's indirect, but you're suggesting that is that Klarna example needs to happen a lot for the margin profiles of these model providers to really change.
16:54 Speaker 04 Yeah. And in that use case too, it's heavily competed. You know, if it's not OpenAI, it might be some other company that is providing customer support.
17:03 Speaker 01 Dare I ask, Nathan, are we in a bubble?
17:06 Speaker 04 I think there I think I think there isn't a wall.
17:10 Speaker 01 You're you're of the so we're talking about this is the scaling wall conversation that's happening.
17:14 Speaker 04 Yeah.
17:15 Speaker 01 Uh, where everyone, you know, having questions about can we keep just plowing more money into these new models and continue to get better results and you are of the there is no wall camp.
17:26 Speaker 04 Yeah, I think we'll keep going for a while. I don't know if it's five years, but I think for sure, like next year and probably the year after that.
17:35 Speaker 01 So even if the revenue is not growing at the same pace, you see companies continuing to up their investments in new models and ship orders. And the the race is going kind of unabated.
17:51 Speaker 04 Yeah, I I think the revenue key is is still growing like OpenAI is still really growing anthropic according to reports is like eating into enterprise quite well. I mean, Nvidia had another blowout quarter. I mean, the fact that it does or doesn't meet Wall Street estimates is kind of like neither here nor there, I think, because Wall Street has a tough time as being a company anyway. And then Big Cloud is still printing money on these services, so I think it's still it's still real for sure. And we're going to see way more of this like sovereign investment in AI like every nation state is increasingly getting convinced that they need to be investing in this and
18:28 Speaker 04 you know, I think big finance hasn't really come to play very much yet with regards to financing new data centers and really providing the capital for this kind of scaling. So I think it's still pretty early.
18:40 Speaker 01 And the CEOs driving this, do you think they see a pot of gold at the end of the rainbow that we all can't like they see? Oh, you know, I get to this point and the profit is going to be just so immense that we'll have all been worth it. Or is this more FOMO driven? And hype driven and kind of wisdom of the crowd driven?
19:02 Speaker 04 I think they're more motivated by pushing the boundaries of technology capability and product experiences than they are seeking GAAP profitability.
19:13 Speaker 01 Yeah.
19:13 Speaker 04 It's just like the nature, the nature of what drive these kinds of individuals. They want to be there when the AGI happens. Whatever that means, right? Like technology history teaches you that you can sustain lack of profits for a very, very long time, as long as you give. Investors in the market, other reasons to believe why they should suspend that disbelief for another year, another year, another year. And for the moment, I think that's still the case.
19:38 Speaker 01 I asked him about this too, especially considering Menlo Ventures invested in at least two separate financing rounds for Anthropic over the last several years. He sees the ROI on AI arriving alongside more superior chips and models, even if it's difficult to see exactly how it will all play out.
19:56 Speaker 01 When you think about the return on investment that a firm like yours needs to have, where do you think it's going to come from? Do you think it's going to come from the models or from the applications that are built on top of the models or something else?
20:08 Speaker 02 I think it's going to come from both for sure. I mean, both the foundation model layer and the application layers, not for every single company, obviously, but for for certain ones are doing quite well, right? And they're they're selling a lot of software and seeing a lot of usage and obviously a lot of API usage and they're generating real revenue. I mean, this is this is not a sort of blip in time or a blip on the radar where, you know, we're going to get some false positive results. Like this is this is real. And, you know, you can see it in the board meetings that my I attend or my partners attend and the results of their companies are returning. So this is this is very real. And I think you'll see it from both layers.
20:50 Speaker 01 And everyone broadly knows that companies like OpenAI, Anthropic are, yes, they're making money, they're growing their revenues, but they're also burning way more than they're bringing in. And maybe that will invert at some point.
21:02 Speaker 02 It's really hard to put a pin on when that's going to happen. I'm, I feel extremely bullish on the idea that the cost to train models over time is going to go down, right? There'll be, you know, better silicon, there'll be new ways to train models, there'll be optimizations that, you know, that I'm probably not even aware of that are being cooked up right now and being produced in literature that will come out soon. So I'm confident that that will find a path to have cheaper inference and training over time. On the on the application layer, I think it's already there. Like, you know, most of these folks are just calling in using API calls into the foundation models anyways. And so the cogs and OpEx for for them sort of to produce these these software solutions is pretty, pretty small.
21:49 Speaker 01 We need to take another quick break. We'll be right back.
22:00 Speaker 03 This episode is brought to you by Subaru. You know that place your mind always wanders to? The lake? The mountains? The homes of the people you love most? The completely redesigned 2026 Subaru Outback is built to help you get there. With legendary capabilities, standard symmetrical all wheel drive, and the confidence to go further. The 2026 Subaru Outback. Love never looked so good. Visit Subaru.com/Outback to learn more.
22:28 Speaker 01 We're back, talking about AI spending and what it means for companies to start pouring money into these products to transform how they do work. Before the break, we talked about what it would look like for all these companies to start showing the kind of growth and profitability to justify these massive AI investments. But how exactly does the AI industry go from a collective total of about 15 billion in revenue to each major AI company doing tens or even hundreds of billions in revenue a year? Investors like Tim and Nathan see a major opportunity in AI agents. Right now, companies are mostly paying for access to LLMs attached to a text box. You type in a question or a command, maybe you feed it some data or code to work with, and it spits something out in return.
23:10 Speaker 01 But what if these LLMs could go off on their own and complete tasks for you? That's the dream being pursued by Anthropic, OpenAI, and scores of their competitors. So far, though, only Anthropic has publicly put out something trending in that direction, with the launch in October of a beta version of its Claude model that can move a digital mouse cursor, click buttons, and type text all on its own. It's too early to tell right now, but companies might pay a whole lot more for these kinds of agents than they do for the AI models of today.
23:40 Speaker 02 The software is just getting quite good, and the move to agentic workloads is getting quite strong. And the capabilities of what an agent can do, again, I'm not going to say that it replaces what a human being can do. And frankly, people should not be writing agentic workloads and architecting them in the way that a human does. It turns out that that doesn't really work well, is my understanding. But when you do deploy these things, they're quite productive and it certainly accelerates workflow, like I said.
24:10 Speaker 01 Yeah, you all are very bullish on agents in the report and I feel like it's the topic de jour. It's the thing I hear constantly, agents, agents, agents. I don't think people have seen a lot of them practically in the world yet, but Anthropic is an early mover here through their API. You can already do some kind of basic agent-like stuff on your desktop. What are agents going to do, especially as it relates to like the enterprise? I feel like we'll see some big consumer swings from Google and OpenAI in the coming months, but when they get deployed into companies, what happens?
24:42 Speaker 02 Well, I think it's important to understand what an agent really is, not to sound overly reductive or anything, but what it really is, an orchestrated series of steps to execute. We talk about it a lot. Really what that is, is an architectural shift more than, you know, some kind of revolution, I would say necessarily that's going to change the way software is written or replace humans. Like, yeah, we talk about it a lot, but that's because we spend a lot of time meeting new companies or pitching to us and even our existing companies and disarticulating their architectures and really understanding how they're built. And what we're seeing is that, I talked about in the report a little bit, some of the RAG architectures that were pretty extensive in 2023 have been replaced by agentic-style
25:27 Speaker 02 architectures. And it's in the report as well. You can actually see how the primary architectural approach and application construction has changed in 2024. A lot of it went away from prompt engineering and almost, I think, more than halved, I want to say, and went into agents. So really what it is, is a mindset shift in terms of how people are building software more than anything else.
25:49 Speaker 01 Does it mean that a company's backend infrastructure becomes increasingly automated? That fewer people are needed to maintain it?
25:56 Speaker 02 Yeah, I'm not going to go so far as to say that. Again, I think of it more as how software is built more than anything else, as opposed to replacing people or specific tasks. It's more about, going back to that security example that I mentioned a moment ago, having agents that can go off and in parallel execute a series of tasks that a human would have to do serially, right? Like that kind of, that's where the throughput acceleration really comes from, is the automation, but the parallelization of the efforts as well.
26:27 Speaker 01 On the enterprise use of agents, I guess, just to stay on this for a second, it seems profound, but at the same time, like, when do you think customers will start to interact with agents? Not just like a company using it for its code generation.
26:43 Speaker 02 Oh, it's almost assuredly happening already, particularly in customer support issues right now. I mean, you are more than likely, if you're on a website trying to return a pair of shoes that you bought from some site online, you know, talking to software that was constructed as an agent, giving it the order number, you know, the address, they're giving you the shipping label to print out. Like, I can guarantee that most of your audience has already interacted with agents, whether they realize it or not.
27:14 Speaker 01 When I talked to Nathan about this, he was confident that the rise of AI agents will translate to companies spending much more on AI. As you'll hear us discuss, it's going to take considerable time to work out the cost structure here, though, as running an AI agent will need to eventually be cheaper than the cost of a human doing the same task.
27:33 Speaker 04 We spent the last 10 years doing robotic process automation and task mining. Companies like UiPath, Solonius are probably the biggest names here. They made billions of dollars going to enterprises diagnosing the various click streams and software tools that individuals would work on and finance, HR, et cetera, and trying to figure out which ones of those processes showed the most repeatability, is like the cow path. They would write scripts to automate those processes. And so now, like, the promise of this sort of software agent, general purpose agent idea, is to do much of the same thing, but give that capability to everybody and have software systems that can automate processes without them having to be so codified.
28:17 Speaker 04 Where there is maybe a bit of nuance, a bit of reasoning nuance, like these kind of RPA agents could not reason if there was some ambiguity or a change in the data structure, or something you really had to think about that couldn't be encoded as an equation or something deterministic. If you do follow the, this idea that complex reasoning would require more thinking or more thinking time, however much thinking time there is, is just a coin operated like purchasing system on Nvidia GPUs or other, other compute systems. So the more complex reasoning, the more time spent, the more compute cost. And then I think it's a question of, hey, is whatever task you're trying to use an agent for, is that some total of the seconds and minutes it spends thinking and completing a task with
29:02 Speaker 04 the success ratio and failure ratio cheaper than what a human can do? Or are there like interesting like handoffs that you can have between an agent and a team of humans, depending on the complexity of the task? Perhaps like a task might involve some like real world manipulation, like the, hopefully the agent can't jump out of the computer and like go to the world.
29:24 Speaker 01 I mean, I guess an interesting thought exercise there is if you are a travel AI agent, would you charge someone for successfully creating and booking a trip, right? You probably would look at what a human would cost to do that and at least better undercut it. But like you said, the cost of these reasoning models is very high. People can experience this with OpenAI's O1 in ChatGPT, which is their reasoning model. It takes a lot of time to run and think. And that's like you said, expensive, that's all being powered by Nvidia. Even though like you may be trying to undercut what human labor would cost, your costs to run a product like this may be going up drastically. So what does that do to the companies and the, and the balance sheets of these companies?
30:10 Speaker 04 In principle, the dogma is you'll need less people to do the same job. I think that's probably true. I do still think that there are tasks where you still want to have a third party do them, even if you could technically audit yourself as a company with O1 because it was better than some top four auditor. Like, no one who actually asks you for an audit is gonna take that seriously. They're still gonna want you to buy from a top four auditor. So I think if we can disconnect the kind of pieces of work that companies and people purchase because somebody else did it and put their stamp of approval on it from like pieces of work that they can totally do themselves and this is acceptable, then I think like
30:52 Speaker 04 agents will do like far better on the secondary option versus the first unless like the auditors like use it themselves and then hopefully like increase their margin and turn around their product faster.
31:03 Speaker 01 Are you seeing at the early stage teams being smaller than they would have been before already because of AI?
31:09 Speaker 04 It's hard to say if it's because of AI because we did go through this growth spurt, probably more than the spurt during the 2021 where teams just became enormous. So in the overhang of that and the reductions in force that were conducted subsequent to the pandemic, entrepreneurs already at that point were saying, I don't want to have to deal with big teams. I much prefer to have smaller teams, not too many offices, more senior people. And now with generative AI and technically speaking, you could probably do more with less, so it's a bit hard to disconnect those two. But I don't think people are hiring more because they have AI tools.
31:50 Speaker 01 I'd like to thank Nathan and Tim for joining me on the show and thank you for listening. I hope you enjoyed it. If you have thoughts about this episode or what you'd like to hear more of, you can send a message to decoder@theverge.com. The team really does read every email. Neelai will also be answering some reader questions for the last show of the year. So if there's anything you want to ask, make sure you get that in soon. We also have a TikTok. Check it out at at DecoderPod. It's a lot of fun. If you like Decoder, please share it with your friends and subscribe wherever you get your podcasts. Decoder is a production of The Verge and is part of the Vox Media Podcast Network. Our producers are Kate Cox and Nick Stat and our editor is Callie Wright. Our supervising producer is Liam James and the Decoder music is by Breakmaster Cylinder. See you next time.
32:39 Speaker 05 This week, it pays extra to be a MyLowe's Rewards member. Through October 14, members get $10 MyLowe's money when you spend $200 or more on an eligible purchase during MyLowe's Money Days. Also using your MyLowe's Rewards credit card? Even better, that's an additional $20 MyLowe's money. Use it to save on your next Lowe's haul. Exclusions, restrictions, and more terms apply. See Lowe's.com/MyLowe'sMoneyDays for details. MyLowe's money expires 30 days after issuance. Subject to change. Credit offer subject to credit approval.
Previous Rewind: Bluesky CEO Jay Graber on the future of federated social media Next Microsoft AI CEO Mustafa Suleyman says conversational AI is the next web browser Page updated October 26, 2026 · transcript and summary are machine-generated