Ben Lorica and Evangelos Simoudis on Rogue AI Agents, the AGI Bubble, and China’s Robotaxi Push.
Subscribe: Apple • Spotify • Overcast • Pocket Casts • YouTube • AntennaPod • Podcast Addict • Amazon • RSS.
Ben Lorica talks with Evangelos Simoudis about three forces reshaping AI: why recent “rogue agent” incidents may be more about permissions, identity, and governance than runaway intelligence; how the belief that AGI is near is helping sustain enormous investment in AI models and data centers despite unanswered questions about margins and ROI; and whether Chinese robotaxi companies can challenge Waymo globally by following a playbook similar to China’s rise in solar and batteries. They also discuss what enterprises should do differently as AI systems become more autonomous and how the economics of inference, training, and autonomous vehicles may ultimately determine which technologies scale.
Interview highlights – key sections from the video version:
Related content:
- A video version of this conversation is available on our YouTube channel.
- Previous conversations with Evangelos Simoudis can be found here.
- Apollo bar chart: In AI, the 41% Depends on the -59%
- [Evangelos Simoudis’s dashboard]: The AI GPS: A Dashboard for the Enterprise CFO
- Specialized AI Is Getting Easier to Build
- Open Models Will Absorb Most of the AI Spend
Support our work by subscribing to our newsletter📩
Transcript
Below is a polished and edited transcript.
Ben Lorica. All right. So today we’re back with my friend Evangelos Simoudis of Synapse Partners. His blog is at corporateinnovation.co. We are recording this on the morning of August 7, 2026. Go to Data Exchange Media for the references that we’ll cite. And also, if you’re on YouTube, hit the subscribe button. If you’re listening to this as a podcast, subscribe wherever you get your podcasts.
So, topic number one is all these headlines about dangerous AI agents. Specifically, I think it was precipitated by an OpenAI agent ending up causing some havoc at Hugging Face, and then subsequently something similar happened with Anthropic. Another agent kind of went, I guess the word they use is “rogue.” When you read the media narratives, the implication is that the models have gotten so powerful that we have to be really careful about them.
But really, if you read between the lines, what seems to have happened is that people have focused on the wrong source of risk. It seems like you had agents that were given an objective, and agents will just burn through compute to try to meet these objectives. More importantly, I guess the permissions and network security were not set properly.
In other words, if you give someone a task and give them an endless amount of compute to do that task, but you don’t lock down the network and you give them permissions to do anything they want, they will cause havoc. So I think my take, Evangelos, is that this is less a story about, “Whoa, how powerful these models and agents are,” and more about how important it is for enterprises and AI teams to actually take governance, security, access control, and network security more seriously. At least that’s my assessment of the situation. I think if you were to give one of these Chinese models the same kind of breathing room, they would be capable of causing a lot of havoc as well.
Evangelos Simoudis. I’ll point out a couple of things. First of all, Meta also indicated that one of their agents did something similar. So we have now three examples of agents that were being tested and somehow broke loose.
To me, the first thing this points out is that we really don’t yet understand the behaviors—what we are building or what kinds of behaviors what we’re building can start exhibiting. And the more complex the models become, the harder it will be for their developers to analyze those behaviors.
The second point is that it is now starting to become clearer that these agents can be used for offensive operations, in addition to just convenience- and productivity-related operations, as we’ve been doing and talking about to date, both at the enterprise and consumer level. That requires, I think, a lot deeper thinking than we have been doing, or at least than we have been talking about publicly.
I’ll point out a couple of things. First of all, we have a Chinese and a U.S. difference in approaching this topic. Recently, China held a conference where their president openly started talking about security relating to agents. I think we’ll see what happens when Trump and Xi Jinping meet late in September, and whether this is going to come up. The U.S. government to date has not put in place the kind of guardrails that would prohibit this kind of approach.
And the last thing I wanted to say, and then I’d like to hear your view, Ben, is that recently I was rereading the AI 2027 papers, which have been updated for people who have not followed the progress. I will say that scenarios they described in their original paper have already materialized. In relation to what we were just discussing about these agents jumping the sandbox, I think that creates interesting dilemmas.
We can envision—and I won’t talk about employment and all that, which those papers talk about—but I can envision situations where we have cyber-offensive operations. I’m reminded of a dinner that we participated in about 15 years ago, where Leon Panetta talked about, for the first time—at least, the first time I heard it—how we talk about cybersecurity in a defensive way, and we really need to be talking about cybersecurity offensively.
I think these agents are starting to point much more clearly to what offensive cybersecurity could mean and what it can trigger. Imagine if we start having agents stealing secrets, creating havoc. Anyway, to me, this is a lot more serious and requires a lot more serious attention by enterprises, consumers, and governments than it has received publicly.
Ben Lorica. First of all, cybersecurity is always an offense-defense capability. I think people in cybersecurity have long appreciated this.
Secondly, again, what I want to point out is that too much of the coverage being written was focused on models. But it’s really about permissions and governance and identity.
So, a couple of things for enterprise AI leaders who are starting to think about agents. First of all, I think you should adopt the attitude that autonomy should be earned rather than granted outright. Think about it: if you have a junior engineer, you’re not going to give that junior engineer permissions to touch mission-critical systems. Same thing with agents that you’ve basically just incubated and created.
What you should do—and this is kind of an exercise—is write down: here’s an agent that we’re going to use. These are the systems that it can touch. These are the permissions that we’ve granted it for each of these systems. Is it read-only? Is it read-write? These are debates worth having.
I think companies might already be overspending on certain models when maybe they should be spending more on boring things like identity, access control, network security, and things like this. So my point in bringing this topic up is that it’s fine to talk about the models and the capabilities of the models, but I think the reality is more mundane.
Evangelos Simoudis. Yeah, actually, Ben, you bring up a very good point because, in addition to the guidance that you’re providing, I would say that what this also underlines is that AI—and the adoption of AI by a corporation—is not only a technology issue. It’s a much broader issue.
I’ve been claiming for a while that it should be a catalyst for transformation. And in the process of that transformation, companies should review all of these guardrails and guidelines that you started talking about.
But the issue is not only a technology issue, particularly for the enterprise. I think the sooner enterprise leaders understand this, the better off they’re going to be.
In the same way that we were talking in past conversations about the need for corporations to use open models to reduce their costs, we’ve also been talking a lot about the need for corporations to own a lot of their IP, as opposed to outsourcing everything. But as they establish these models—whether they build them from scratch, build associated agents, or fine-tune existing open-source models—they really need to be thinking about it at the enterprise level and how it impacts various organizations, as opposed to treating it as just a technology issue.
Ben Lorica. Yeah, and this is also kind of tricky, Evangelos, especially if you’re designing and deploying an agent that, let’s say, does not mimic a human workflow or process.
For example, let’s say you’re designing an agent that onboards a customer or approves a medical diagnostic procedure or something like that. I’m just making it up. If it’s something that a human is doing now, I think it’s easier for you to list out: okay, these are the systems this agent is probably going to touch because this is what the human will touch. And these are the permissions that the agent needs to have because this is what the human has.
But if it’s an open-ended agent that is doing something more exploratory, and maybe not something that you’ve assigned to a human, you have to be, I think, even more vigilant in terms of writing down, like I said, all the systems this thing will touch, the permissions, and so on.
Evangelos Simoudis. And obviously, as part of this transformation, you also have to think about—and again, we see this with the corporations we work with—do I use AI to automate an existing workflow in the way that the workflow has been codified, or do I use this as an opportunity to rethink my workflow and make it much more AI-centric, AI-first?
That, of course, takes you into unknown territory because if you’re automating something, or if you’re infusing AI into something that is already codified and you understand it well, it is easier to create those guardrails. It is much harder when you are reimagining the workflow, and that’s where the extreme vigilance needs to come in.
Again, that’s why I think this is much more than a technology issue. It has implications for customer satisfaction, brand perception, and many other characteristics that impact corporations—valuation, market capitalization. So it’s very important.
Ben Lorica. And just to close the loop, bringing this back to the cybersecurity example: in cybersecurity, you have these people who are called penetration testers, and there’s also another group of people who are good at finding zero-day exploits.
The common thing about these two groups is that they are relentless people. They almost have OCD. They can look at something, and you and I will give up in an hour. They can take weeks until they find that one little crack in your armor.
Well, think about this situation now. You have agents that can just burn through compute nonstop. So this is the situation as far as cyber offense.
All right, which brings me to topic number two, which ties nicely with topic number one.
Listeners, as you can glean from topic number one, my objection to the media coverage was that there was too much focus on, “Oh, the models are so amazing,” and that it’s because the models are so powerful that this happened. But it turned out, at least in my estimation, that the explanation is more mundane.
Now, that doesn’t take away from the fact that a lot of people here in the Bay Area, where Evangelos and I sit—especially in the frontier AI labs—are what we call AGI-pilled. People are really focused on the idea that AGI is right around the corner, and we want to be the people who build the AGI capability.
For the most part, they have three rationales, or three stories, that they tend to tell themselves.
AGI is inevitable. So the race is not to stop it, but rather to steer it. If it’s inevitable, it might as well be us. We should build it.
The second story they tell themselves is that the ROI is so amazing—the expected value. Once we get to AGI, we can cure diseases, extend life, maybe solve global warming. So that’s why we should do it.
And then the third rationale is more like, “Hey, this is something I’ve been reading about since I was a kid. This is a great science experiment. I need to know if it can be done, so I’m going to do it.”
So those are the three motivations.
But the main reason I bring it up here is the observation that I think it’s partly marketing. Partly, it’s: we’re close to AGI, and that’s why you should continue to shovel money toward this industry and into these AI data centers that have questionable debt financing.
But I think a lot of people in these labs are truly AGI-pilled. They really believe that with the current approach of scaling, mainly built around transformers and LLMs, they will get there. What’s your reaction to this observation?
Evangelos Simoudis. I think the first thing that you need to consider is something that you had mentioned offline, which is: do we believe that our current approaches are going to get us there? And if so, what will it take, and how long should we wait?
There’s no question that in the 40-plus years that I’ve been involved with AI, the progress that we’ve made in the last five is unbelievable. It has come at a cost, and it continues to come at a significant cost.
Again, I’ll bring the enterprise perspective. The enterprise leaders I talk to, whether they’re CEOs or senior executives, are really not paying attention to AGI or not-AGI. They are paying attention to AI. They are paying attention to: what is the benefit that I will get from that, and how much will it cost me to get to that benefit?
And by the way, we’ve already seen, AGI or not AGI, certain retractions from original goals. Because even though you have increasingly intelligent systems, you may not be getting, as an enterprise, to the objectives that you’ve set for yourself.
So I think that what we are seeing here, whether it is the discussion of AGI or other expressions of intelligence—robotaxis being one of them—we live in a bubble, or whatever you want to call it. We live in a unique ecosystem, which is not necessarily shared across the world.
Ben Lorica. But the reason I bring up AGI is that I think AGI is an important element in why we’re in a bubble. Because basically part of it is: we’re so close. If we don’t do it, China will get there. So you have to keep shoveling money into these AI data centers because, to get there, we think we can get there with scale.
Now, a couple of caveats here. There are some neo-labs that disagree that the current approaches are sufficient, and they are going down different paths. Whatever you might call them, some of them are under the umbrella of world models. There are people who disagree. They’ve left these frontier labs. They’ve set up their own research labs.
Secondly, within these frontier labs—and not only the frontier labs—there’s another set of neo-labs that also adhere to this, which is much more about doubling down. The current approach is correct, and what we’re going to do is build recursive superintelligence. In other words, we’re going to build the capability for the AI to build the next version of itself.
But even within that community, there are neo-labs that, while they’re working on recursive capability, are also tweaking the underlying approach.
Set aside all of the talk about recursive superintelligence and these neo-labs. The whole point is that the AGI discussion is very, very essential to how these companies justify both their valuations and why we find ourselves somewhat in a financial bubble—or really in a financial bubble.
So, Evangelos, do you think that if you remove the discussion of AGI, what happens to valuations and the bubble?
Evangelos Simoudis. Okay, so I actually think that we have three narratives among the new labs, whether you agree or not—
Ben Lorica. Let’s focus on AGI. Yeah, yeah, yeah.
Evangelos Simoudis. But I think what’s important here is that all three are driving high valuations and very large financing rounds. There’s the AGI narrative. There’s the superintelligence narrative, which I think is different from AGI. Superintelligence meaning I become superintelligent in a specific task or in a very limited set of tasks. That would be the Claude example. And then you have the world-model labs, which deal far more with embodied AI as opposed to—
Ben Lorica. But my point with the world labs is that world labs is an umbrella term. Inside world labs is a subset of people that are actually also chasing the same thing.
Evangelos Simoudis. Okay. Anyway, to me, the companies that are developing world models are focusing primarily on embodied AI, like I said—
Ben Lorica. It’s an umbrella term. Inside the umbrella term, there’s a subset, right?
Evangelos Simoudis. Yeah. So look, I think that, again, let’s go to the investment side now. I would say that in the same way that you have researchers who are what you call AGI-pilled, I think there are investors who are the same way. They’re thinking that somehow we will reach a state of, I don’t know, world domination or whatever you want to call it, for which AGI is important, and they continue to write very large checks.
Ben Lorica. Yeah, that’s one of the rationales, right? The expected value is so high.
Evangelos Simoudis. Exactly. And frankly, I have been wondering for a long time: let’s assume that we reach AGI automatically. Why is it winner-take-all?
Why wouldn’t, for example, when the first U.S. lab claims that we have reached AGI, a Chinese lab also do the same? Or maybe even a second U.S. lab does?
Again, there are dynamics here that I think we’re sweeping under the rug while continuing to write bigger and bigger checks.
To me, what is interesting is that very recently we’re starting to see investors question what the return on their investment will be under those models, and when specific companies that they fund—whether it is a neo-lab or a data center company—will be able to, A, become self-sustaining, and, B, produce margins that are worth funding.
An investor, a venture investor, does not fund companies that promise to have 3% margins. You need something to make that investment provide significant returns, and low margins are not going to do that.
Right now, again, we’re seeing this happen with SpaceX and xAI, how they’re pivoting to become a neocloud because they realize that Grok is not being used in the way that they anticipated.
I think we will see that with Google, or maybe we have already started seeing it. This week, we had the leadership changes. SemiAnalysis has written a very interesting memo about what is happening with GCP, with Google Cloud, and where they are directing their investment.
Last week, we had the “Situational Awareness” debacle. I think we’re starting to see certain issues here that you and I believe have the same root.
Ben Lorica. Yeah, and I think we talked offline about my frustration, which is that we’re not even being given a narrative.
Basically, we’re investing in these AI data centers at such high rates—hundreds of billions of dollars, approaching trillions of dollars—because we believe what? We believe this is going to be a low-margin, high-scale business? Or is it because it’s going to be a premium business?
Maybe it’s both. Maybe OpenAI takes the consumer low-margin business and Anthropic takes the premium high-margin business.
But even if you believe either one of those, first of all, they have a lot of competition now with the open-weight models. And secondly, they’re not even telling us. We don’t even know what the narrative is and whether it’s going to be low-margin or premium.
Evangelos Simoudis. And Ben, I think that, again, speaking from what I would say is the majority of the people I talk to, when there’s a discussion about AI data centers, there isn’t even a good understanding of data centers for training models versus data centers for inference.
The model for data centers for AI inference may prove to be similar to what we had in the internet era with CDNs—companies like Akamai—because you do need to lower the cost of inference and you need to bring the inference closer to the user.
With training, that’s a whole different story. And I think, by the way—
Ben Lorica. We don’t have any visibility at all. For example, if Anthropic is charging $6,000 an hour for an hour of inference for a specific kind of model, how much is that?
Evangelos Simoudis. Exactly. No one knows. We don’t have that.
And I think CFOs—again, I’ll go to the enterprise—are starting to ask that question. That’s why I recently wrote about this dashboard. They’re asking: how much am I paying for this inference? Why am I paying that amount? Am I getting the ROI that I expect? Is that inference being used by the right projects? Not by the right people, but by the right projects.
We don’t have that visibility at the enterprise level, and we also do not have visibility into how much it costs us to train these models and for what benefit.
Everything today is a black box, and we keep talking about data centers in a very generic way. We should stop doing that. I think that’s the wrong way to approach it, both from an investment perspective, from the perspective of a consumer of AI—whether it is an enterprise or an individual—and even from a government perspective.
Look at local governments striking agreements without, in my opinion, really understanding anything more than the taxes that they will get and maybe the employment that they will gain from having a data center in their territory.
Ben Lorica. By the way, a lot of the AI data center buildout is being driven by Anthropic and OpenAI as the main customers of these data centers. But I just shared with you before we went online a new bar chart from Apollo.
Speaker 1. Right.
Ben Lorica. You know, it’s not good. The operating margins—the closer you are to the user, you go down to minus 59% operating margins if you’re the model or the application. And then only Nvidia seems to win in this scenario.
Evangelos Simoudis. And that will not continue forever. We’re already seeing China taking steps to dampen that trajectory.
Ben Lorica. By the way, the difference now, Evangelos, between the dot-com bubble that we both lived through and now is that, first of all, we have a generation of people who weren’t even born then who are now investors.
Secondly, we have social media and meme stocks. We didn’t have that back then. So the number of people driving up these prices is just off the charts.
Anyway—
Evangelos Simoudis. It is.
Ben Lorica. Let me pivot to a topic that I’ve been thinking about and that Evangelos has way more expertise in than I do, which is robotaxis.
I just came across some stats that got me thinking that we are, for the most part, in the West asleep at the wheel as far as robotaxis are concerned. Especially for those of us who have access to Waymo, we think the U.S. is far ahead in terms of robotaxis.
But the question these stats surface is going to be interesting.
Here’s a couple of stats. One, Baidu is much closer to Waymo on ride volume than people realize. I think the fact that we don’t see Baidu in the U.S. doesn’t mean they’re not in the West. They’re starting to talk to countries outside of the U.S.
Evangelos Simoudis. They’re actually testing. They’re more than talking.
Ben Lorica. Yeah. And secondly, Baidu is not the only player. There’s Pony.ai. There’s a bunch of these Chinese companies.
So then the question for me, Evangelos—and this is a much more speculative and doomsday question—is: can China do to robotaxis what it did to solar and batteries?
Evangelos Simoudis. The short answer is yes. Let me give a few more details to the setup.
Ben Lorica. Before you do that, here’s a quick question for listeners who are not familiar with the state of play as far as robotaxis.
In the West, we have Waymo, and Waymo is all over, at least in the Bay Area. But what people don’t realize is that Waymo also still relies on human operators to oversee its fleet.
So two questions. One is: what’s the state of the art of robotaxis in China? Is it basically comparable to Waymo? And how widely deployed is it in China?
Evangelos Simoudis. The total fleet of robotaxis in China is very similar, comparable to the total fleet of robotaxis in the U.S.
In the U.S., I would say the bulk of the fleet is Waymo. Behind that we have Zoox, which is an Amazon-owned company. Zoox has a little over 100 cars. Waymo has over 3,500 and maybe close to 4,000. So there’s a big difference.
Ben Lorica. That’s nationwide?
Evangelos Simoudis. Well, Waymo operates in about 10 markets, either by itself or in combination with Uber.
Zoox operates in Las Vegas. In fact, they just got permission to start charging for rides in Las Vegas. So far, the rides have been free. And they are going to get permission to operate in San Francisco. They’ve been testing in San Francisco very extensively, but they’ll also be able to charge for it.
The player that is coming very strongly behind that is Tesla with its own robotaxi. They have been operating in Austin.
Ben Lorica. Would you ride that, camera-only?
Evangelos Simoudis. I wouldn’t, and I’ve said that. But again, to give credit where credit is due, Tesla’s FSD system has made tremendous strides. They continue to improve it. The people who have been using it in their own private cars now swear by it. There are a lot of proponents.
We’ll see where it goes. But I did want to note that Tesla has very aggressive plans for both U.S. expansion of its robotaxi fleet as well as European expansion.
Europe is important because FSD has been approved in certain European countries like the Netherlands, Belgium, Estonia, and a few others, and there are more pending.
Now, going back to Waymo: Waymo has been focusing on the U.S., but it has also been testing in Tokyo and London.
Ben Lorica. London, yeah.
Evangelos Simoudis. The big difference between the Chinese companies and Waymo—just to stay on Waymo for a minute—is that China has been thinking a lot more globally than the U.S.
For example, you mentioned Pony.ai. Pony.ai is working in Singapore with ComfortDelGro. They’re working in the Middle East. Same thing, by the way, as Apollo.
Ben Lorica. Just to clarify, for both Pony and Baidu, is the experience and the implementation similar to Waymo?
Evangelos Simoudis. Yeah. These are multi-sensor, end-to-end AI systems that are—
Ben Lorica. So it’s kind of the same playbook. They have to go to a city, drive the hell out of the city for a while, map it, and they’re still using human operators.
Evangelos Simoudis. Yeah. Even though, again, we have been moving from relying on detailed maps toward what’s called mapless implementations, but again—
Ben Lorica. They’re on par with Waymo?
Evangelos Simoudis. Yeah.
Ben Lorica. Both of them?
Evangelos Simoudis. I would say so. In cities like Shanghai and Shenzhen, you see them broadly.
Ben Lorica. And for our listeners—and this might be a naive question, Evangelos—but it does make sense that the more miles you drive, the better your system gets, right? There is a benefit to the companies that really go out there.
Evangelos Simoudis. There is a benefit, but I will also say something that listeners of this podcast know: it’s not only the miles, but the type of miles.
If you’re driving the same route day in and day out, you’re accumulating miles, but you’re not accumulating experience, unless there are a lot of things happening on that route.
So it’s important.
And this is, by the way, something that Tesla is banking on, because they have collected so much data through their privately owned vehicles that they are using now to train their robotaxis.
Whereas in the case of Waymo, they are relying a lot more on simulation. That’s what we were talking about earlier with world models. World models are very important for that type of training.
Ben Lorica. So then, going back to the overarching question: can China do to robotaxis what it did to solar and batteries? You said yes. Explain why.
Evangelos Simoudis. Yeah. I think China has taken two approaches that are quite different from Waymo.
They are putting a premium on global expansion, and they are willing to collaborate much more aggressively than Waymo with local operators who are willing to install Chinese robotaxi technology on vehicles that operate in their territory.
They’ve done that, and they’re testing this in Europe. As I said, they’re testing it in Singapore. They’re testing it in the Middle East, in Dubai.
Ben Lorica. Why isn’t Waymo doing that?
Evangelos Simoudis. I think Waymo has taken the approach of: let’s nail down the U.S. The U.S. is a big enough market. We’re going to move more cautiously.
And they’ve done that throughout their operating experience. Even in the U.S., they’re a hell of a lot more cautious than Tesla is, for example, in terms of how long they’ve been testing.
Ben Lorica. But given the fact that now we’re seeing the European Union, in particular, start to push back on Chinese exports, does that mean that Chinese robotaxis will start encountering friction in Europe, particularly in the EU?
Evangelos Simoudis. Someone asked me the other day what strategies Europe should follow with regard to China.
Today, Europe takes a very narrow view. The restrictions they are imposing are not sector-wide; they are on very narrow slices of a sector.
So China says, “Well, you don’t have to use Chinese vehicles in order to use our technology. We’ll give you the technology. You can use it on any vehicle you want.”
From that perspective, you can think of Volkswagen.
Ben Lorica. So you mean that’s just driver assist? Because for robotaxis you have to install the sensors.
Evangelos Simoudis. Yeah. I’m talking about taking the stack. China says, “Take our stack, take our technology, and install it in whatever vehicle you want.”
So now you’re not talking about a Chinese vehicle. It could be a Volkswagen vehicle.
Ben Lorica. What about—I thought the Europeans were getting paranoid about AI sovereignty, models getting turned off?
Evangelos Simoudis. Well, we’ll see where that goes.
I don’t think that, for the time being—again, different interests, right? You may have one industry embracing the technology while another industry is vilifying the technology.
When we talk about robotaxis, when we talk about companies offering mobility services—and this is outside the automakers—remember, several European automakers, from Renault to Mercedes to BMW to Volkswagen, had all created mobility services divisions several years ago. All of these have now been abandoned.
So the companies that are driving mobility services in Europe are not the OEMs. They’re not the German OEMs.
That gives, I think, an opening for Chinese players to come in and say, “You’re a taxi company. You’re a mobility company. You’re an Uber competitor, or you’re Uber yourself. You can select the vehicle. We’ll give you our technology, and we’ll license it. We’ll find a business model that works for both of us.”
And I think they’ve done that in the Middle East, as I said. They’re doing it in Slovakia. They’re doing it in other parts of Europe as well. London is another place where these tests are taking place.
Ben Lorica. One observation here might be that one of the reasons Waymo is hesitant to do this is because, one, it’s expensive to just go everywhere all at once.
And it could be the case—again, going back to maybe a poor analogy to solar—that you flood the market, you’re subsidized by the Chinese government, and then at some point you dominate. Is that what’s happening here?
Evangelos Simoudis. I cannot say unequivocally that that’s what’s happening here. But I will agree with you that it has the potential to happen here.
And I think Waymo, maybe even Uber—they have reason to pay attention because Uber is also investing in robotaxis. They have this partnership with Nuro and Lucid to create Uber-branded vehicles.
Ben Lorica. How far behind are those technologies relative to Waymo?
Evangelos Simoudis. I would say at least a couple of years. The Lucid Gravity vehicles with the Nuro stack are around the Bay Area.
Ben Lorica. Waymo, in many ways, is the champion of the U.S. So if Europe wants to diversify away from China, the natural provider would be Waymo. But Waymo won’t enter their markets because it’s expensive. So they will have to somehow lure Waymo.
In the meantime, the Chinese are willing to enter their market without any subsidies from the EU, right?
Evangelos Simoudis. Yes.
Actually, I want to believe—without having any proprietary knowledge on this—that if the London experiment goes well, what does that mean? It means if this testing period, followed by a paying-customer period, starts to go well—which could mean another year, by the way—then we will see Waymo becoming more aggressive.
Remember, by the way, that companies like Waymo, starting with Waymo, do not have to worry only about China.
If you look at what’s happening here in the U.S., both in Boston and in Washington, D.C., there is very fierce opposition to Waymo operating in those cities, and the main argument is job loss.
Ben Lorica. I would imagine the same thing in London, right?
Evangelos Simoudis. Exactly. So that’s my point.
Ben Lorica. On the other hand, Evangelos, we’ve talked about this before in previous episodes. It’s not even clear that Waymo, which is already widely used in the Bay Area, isn’t losing money on each ride.
Evangelos Simoudis. I would say that Waymo, first of all, is growing very nicely. I think we have to admit that, and that’s a wonderful story to tell by itself. The number of rides is growing.
What are the margins? I think you need to look at it from a couple of perspectives.
The rides continue to be more expensive on a per-mile basis than Uber or Lyft or any of those. I think they are very close to being break-even. And if that continues—unless we have a recession or something like that—I think they will be able to start showing a profit in the near future.
Ben Lorica. So the reason they’re able to turn a profit would be because it’s in the Bay Area and they have the scale. But does that—
Evangelos Simoudis. No, no. I think they’re scaling in other markets.
Ben Lorica. No, no, no. What I mean is: does that mean they then become more confident in terms of overseas success?
Evangelos Simoudis. Yeah. Look, think of the steps that they’re taking in order to improve their margins.
They’re bringing in vehicles. They have the agreement with Geely for the Zeekr vehicle, which we’ve already started to see in the Bay Area. And they have the agreement with Hyundai for the Ioniq vehicles.
Why these are important is because the vehicles will come in a form that will make the installation of the Waymo stack, the Waymo infrastructure, a lot faster and cheaper.
It is still costing—your Waymo vehicle, the Jaguar, is still over $100,000 per copy before it hits the road, before you take everything else into account. The Zeekr vehicle, followed by the Ioniq vehicle, is going to bring that cost down significantly.
The second thing they’re doing is that they’re now on version six of their stack, and you can be assured that there are many optimizations in this version that improve performance while decreasing costs.
So they’re working very hard. And the increased performance also impacts something that you said earlier. It impacts how many teleoperators you need to have per number of vehicles.
Again, that’s why I’m saying that over the next several months, we will see technology improvements in Waymo’s fleet that have a direct impact on the financial performance of the company.
And if we continue to see the kind of adoption that we’re seeing in the markets where they operate, I think that would be very fortuitous for them.
The big question—one last comment—the big question is what is going to happen with Tesla’s robotaxi.
Of the companies that we mentioned as competitors in the U.S. market, the primary fear should come from Tesla because Tesla has the ability to manufacture robotaxis very quickly.
They already operate through FSD in every one of the markets that they will want to enter. So the question then becomes: will they get permission to operate in those markets?
And what we’re seeing from Texas is something that Musk and Tesla have been working toward: getting not city-by-city permission, like we’re seeing here in California, but statewide permission so that they can deploy in Austin and Dallas and Houston and San Antonio and El Paso without having to go city by city, by getting blanket permission.
Ben Lorica. All right. So one last question. Quick prediction from you: two years from now, European robotaxis—will we see more U.S. or more Chinese?
Evangelos Simoudis. I think we’ll see more Chinese, frankly, but with European cars. That would be my prediction.
Ben Lorica. All right, and with that, thank you, Evangelos. And again, consult the episode notes on Data Exchange Media to get links to many of the things we talked about.