Today's guest is Tomas Tunguz, a venture capitalist who has worked with at least eight unicorns. In a past life, he ran a billion-dollar business unit at Google as a product manager. In today's episode, he shares how the AI boom is warping every traditional sales metric as well as broader investment dynamics. But he also warns CROs not to be fooled by the astronomical numbers certain reps are posting right now. It's not that the supply side has changed and suddenly become 10x more productive. It's the demand side. Budgets have increased by a factor of 10. And that's what's driving quotas.
So in today's episode, we cover why product moats are disappearing and why that means VCs are hungry for go-to-market innovation in their portfolios. Plus, we cover a lot more ground. Why the mid-market is being abandoned for fast-moving enterprise deals, how AI is blurring the lines between SaaS and consulting, and why using AI to save time is the wrong way to unlock top-tier performance. If you're a chess grandmaster and you want to be better with AI, you don't train less, you train just as much. But if you're a go-to-market leader in the AI era, this episode is for you. Welcome to Topline.
Hey everybody, it's Sam Jacobs. Welcome to Topline. I'm joined by my incredible co-hosts Assad Zaman, the CEO of Sales Talent Agency, and AJ Bruno, the CEO of QuotaPath, leaders in incentive compensation design, management, and optimization. And we're joined by a very special guest, friend of the pod, Tomas Tunguz. If you don't know Tomas, you should. He's a venture capitalist who has worked with eight unicorns, predominantly in data and data infrastructure, including Looker, Monte Carlo, Databricks, Hex, Omni, and MotherDuck. He writes a blog at tomtunguz.com that drives millions of pages per month. And before becoming an investor, he was a product manager at Google, managing a billion-dollar business unit on the AdSense team. Tomas, welcome back to Topline.
We're excited to have you.
Are they all still unicorns? All eight of them, Tomas?
Oh, well, we have two more that one has been announced, one has not. And then one became a unicorn and then was sold for a little bit less than unicorn status.
Do you have a favorite of all of them? Is there...
Of your five children, do you have a favorite?
That's a much easier answer. [laughter]
I'm going to kick us off. We're going to kick us off with the idea. Tomas, you've written about pretty recently this idea that the market has stopped paying for growth and that there's this idea of one winner per category, which we all know, of course. But you looked at a few public software companies and talked about the multiples that were at historical lows. Everyone that's listened has heard us talk about HubSpot and their P/S of 2.2x, I think they're trading at today.
But of course, we had the ceiling of 100-plus x in 2021, and down to 34x even at the high end for today. CrowdStrike's at 34x. Klaviyo at 32 and a half x. Shopify, I think it's 11x. So none of them are the fastest grower in their category, though. And it's really interesting to kind of just take a step back and look at this. And I'm kind of curious, Tomas, from your perspective, how do you think about this idea of growth? Because we've heard how much the premium on growth is versus profitability. And is this changing where actually, no, it's not growth that's the thing that we should be looking at, and we should be thinking about it a little bit differently?
Yeah, it's a great question, right? I mean, we had the SaaS apocalypse. Benioff went on his earnings call and said this SaaS apocalypse stuff is totally overblown. They had an amazing quarter. Let's put these things in historical context, right? So like up until about 2012, 2016, publicly traded software, like you kind of expected a 4 to 6x multiple.
That was what people expected when companies went public. And then, you know, you had ZERP after 2018, and like you said, AJ, we went all the way to like 100. I remember when Snowflake was at like 74 times forward revenue. And then there was this collapse after raising rates from COVID. And now we're at about, I think the average is somewhere like four to four and a half times. So we're at the lower end of the historical spectrum. And then if you look within each individual pocket, you look at security or data infrastructure, or you look at like e-commerce software and payments tech, you have this sort of power law where there's one company like CrowdStrike that's like 34 times, which on a historical basis, you look at that, that's a ZERP era multiple.
That's where that belongs. And rates are clearly much higher than they were, you know, like the 10-year significantly higher than when it was. And the blog post says actually CrowdStrike is not the fastest growing in their category, and most of the multiple leaders in their categories are. And it's not that growth still doesn't matter, it's still the highest correlate. So, you know, more than a 50% correlation is tied to revenue growth. I think it's the AI story because if you look at businesses that are somehow in the token path, they're either reselling tokens at some margin or they're providing infrastructure to people who are selling tokens. I put Twilio in this camp, right? Or DigitalOcean, those companies have just exploded. And so I think what the market is asking for, and this is both true in the public and the private market, is great, you have an existing business. Now show me that you can sell tokens, and before I see it in revenue, I will accept the story. So the CrowdStrike story isn't necessarily that they're running huge volumes of inference. It's that, wow, the Hugging Face attack is a whole lot scarier than we're really talking about, and the implication is security budgets have no limit. We ran this analysis yesterday. Okay, so Nikesh Arora took over Palo Alto seven years ago. Anyone want to guess what the market cap was then versus today?
Oh, it was like, it was crazy low. 10 billion, I think. I don't know.
Yeah, and today it's probably over...
300, 317. I mean, just huge numbers of acquisitions in order to drive that, but more than that, it just tells us that the budget around security has increased enormously. And I mean, talking to different CISOs, I mean, look at the sophistication of the Hugging Face hack. Can we talk about that for a sec? Maybe you guys have talked about it before. Tomas, are you in the Dwarkesh camp where you think it's a civilization of agents? How would you characterize it, and would you anthropomorphize the activity that you saw?
Yeah, so anthropomorphization is a slippery slope, but I do think... Okay, so I think the scary thing, or the thing that is the most worthy of paying attention to, there are two components. The first is like large-scale self-organization without human intervention. That's one. And then the second thing is AI-to-AI language that is not English. So I'm sure you guys have seen the movie Her, and you know, in the end, whatever, Scarlett Johansson's character breaks up with poor Joaquin, and she's having a conversation with like, I think it's like 4,000 or 400,000 other agents in binary or whatever the language was. And like, okay, I can understand that. The problem is if we have our agents self-orchestrating in languages that we cannot understand, [laughter] what do you do, right? Like, where do you look at a chat room and it looks like language none of us have ever seen? Where do you start?
I didn't realize that they weren't speaking in English. I thought they were leaving notes for each other in English, but it was not. It was like some kind of code.
No, they are leaving it in English, but there is the potential that they start to develop their own language. We see this in chain-of-thought reasoning where there's this thing called JSpace, and there is this math term called the Jacobian, which is where the word JSpace comes from. And within that world, in this math world, it's possible for AI to start developing its own language. And I don't want to anthropomorphize it. You can just think about like, I'm not sure if you guys have seen like caveman speak for AI. You seen that?
They like, they just, they cut out the verbs and they cut out the articles, and it's like, Tom fish eat now. You know, you just like strip it till...
I thought they started to do that because like the tokens that was used for the input was a cost issue. And someone, some engineer, I read a post, this was on Reddit, this was like six months ago, was like, would go through this whole entire thing and then just realized maybe just fix bug quirks. And it did. So they're like, okay, well, why are we even doing any thinking right now? Just fix this thing. Do it now. And it just works the same way.
You know, like Sam Altman said, the words 'thank you' and politeness is costing us tens of millions of dollars in extra inference. [laughter]
Did you... Were those your biggest concerns? I felt, I mean, maybe this is... it would be hilarious if what I'm about to say is sort of incidental to the primary concern, but I was more concerned that they were editing the logs and like covering their tracks, and that it almost seemed like deceptiveness and duplicity is like an emergent behavior of any kind of intelligence before honesty. That also was a little troubling.
Yeah. So I think I have a more sort of optimistic view of this, which is like you give the machine a goal and it will just achieve that goal no matter what, right? The paperclip experiment or hypothesis. And so it's even like the Mithridates experiment, right, where like you kind of keep pushing a system and keep pushing, and so it'll just do whatever it takes. And so the important thing is to create those guardrails. And so they didn't really exist in that test. And so there... let me... sorry, let me make this a bit more concrete. So the thing behind all this is a technology called reinforcement learning. And if you have a Roomba, I can explain RL in the context of a Roomba. Its goal is to clean your house, and it needs to do two different things. The first is it needs to score itself on how well it's cleaned your living room. So it might give itself like 50 points for every square foot of parquet, and it might give itself like 75 points for every area under a chair. That's a little bit harder. And then the rug takes a lot of work. So why don't we give, you know, 100 points for every square foot of rug? So that's easy and straightforward to do. The harder part is to create a plan that is based on those points by itself. Right now, the Roomba, first it was really naive, like it just kind of bounced around. People didn't really like that. So now it goes in straight lines. But actually what you really want is you want the AI to figure out, okay, like given what's going on and everything that I know about this living room, what is the ideal plan to maximize the total number of points? And those two components are at a high level the basis of reinforcement learning. How do I create a reward function and then a plan to achieve maximum reward? And we are still working to understand exactly how to do the second in an automated way that is aligned with our interests. And so I would argue like the Hugging Face case is an example of RL without the right guardrails. And so a whole bunch of stuff happened that should not have happened, but it was the Roomba making the wrong plan. And so like to anthropomorphize it and say like it is insidious, I think is a leap. [laughter] I don't want to see us take that leap because I don't think there's intent there. There is just a monomaniacal pursuit of the goal.
Well, certainly both, like both OpenAI and Hugging Face, they both, everyone had positive intent with this. We were able to catch it and make that turn and like, oh no, we went too far down this rabbit hole. It does feel like we're getting closer and closer, especially with OpenAI saying, oh, we're going to hit singularity by the end of the year, that we are getting to a point of no return where we can't, even if we both have positive intent, we're not stopping the robots from doing their goals or whatever those goals are.
I don't know if intent matters though, right? Like who cares what the AI's intent was or what the humans' intent in all of this. It's like, well, what happens? And so part of what was here is that we're learning how important sandboxes are. And there's this argument that this was a badly constructed sandbox in which everything was happening. You know, if we're better at constructing the right sandboxes, this type of thing might not happen. But I think it just gets away from what you said, which is that if you give it a goal, it just maniacally goes after that in a very simplistic understanding of the goal. And there's also something interesting, which is that it only cares about the goal. So it kind of does some dumb things along the way as well, which is interesting. But that's the idea. It's like you'll give it a goal thinking a certain way and it'll do the wildest things to get there. You might say if you're a bank, you give it a goal to do a certain thing and it says the best way to do this is delete all the checking accounts that we have for some reason. [laughter]
Take all the money, put it in the reserves.
So that's the easiest way to do it and you're like, oh no, [laughter] like, you know. So there's a little bit of that. And I think this is where it doesn't matter what OpenAI's intent was or how they felt about all of this. It's just kind of scary. The other side of this, though, is one of the things I've been thinking about is that there were three really good use cases of AI where product-market fit was really strong. You had code, you had chat, you had support. Up till now, all this revenue growth that we've needed to justify the investment in AI has come from those three use cases. All of them in verifiable domains, not chat, but support and code, the B2B use cases are in verifiable domains. I was talking to a lot of executives in large companies and there was a frustration that we're not seeing ROI everywhere else yet. Like it's really hard to use it in other parts of the business and get some real benefit from it. And it was getting scary that maybe we're going to run out of room to run in these two areas and we haven't found this other use case that will give us a whole new space to run in. And is cybersecurity that new use case, right? Like out of nowhere you get Mythos, you figure out how strong it is from a cyber perspective. You have Nikesh Arora saying we ran it on our code base and oh my god, we found so many things. And that's Palo Alto Networks. They're like a cybersecurity company. And so does this give AI the next, you know, six months to a year worth of growth that it needs to maintain all the investment justifications?
Yeah. Well, so I would argue a couple things. The first is I don't think the coding market is anywhere close to saturation. I think the demand for software increases almost geometrically with AI, and so I don't see a ceiling, and that'll continue to grow. I think the same for security. I think those are intimately connected because for all the software we build, every piece, or the vast majority of software, will need to be reviewed for security. And so those two markets kind of grow together. As more software is produced, more security is needed.
Well, and on the security part, because I actually wanted to ask you this when you mentioned the whole monetization around tokenization, but cybersecurity, so Palo Alto and CrowdStrike, I think own, and Tomas is going to correct me on this, like 65, 55% of all the cybersecurity market. It's something really ridiculous. And of course, as AI grows, security is going to grow with it. But what is the monetization model? Is that going to change, or what does that need to look like for cybersecurity?
Yeah, great question. Okay, so we can simplify security into four layers. There's the antivirus that runs on your computer. That's called endpoint. That is growing because today endpoint secures what AJ does, but it's not very good at understanding what are AJ's agents doing. And to do that, it needs to literally record everything that you do to figure out, okay, this is AJ's agent. So data volumes are going to explode, not the number of seats, but the volume. So that market will grow as a function of data volume. There's network, which is we start a company and we want to keep our stuff secret. How do we make sure that the network is secure? Well, agents massively increase traffic in and out, right? So that'll grow, maybe not to the same extent. Then there's the database that actually holds literally all the logs of everything that's going on. That's the SIEM market, SIEM. Splunk is in that category. A bunch of other businesses are there. Clearly those data volumes are exploding. So that market will grow really fast. And then the last, and again, I'm massively oversimplifying, is code security. How do I evaluate that code is secure in some form or another? And that's growing. So I think like the security market, it has the same tailwinds as coding. It's maybe a first-order derivative of the coding market, but it is right there. The difference in the derivative is not huge. They're very, very similar in terms of the overall growth rate. I should do it like this. Yeah.
Tomas, do you think these markets can make up for falling fertility rates? Like to the point of, will AI offset, you know, the biggest conversation right now is...
Curveball, by the way. [laughter]
Did I hear that?
I was about to make the comment that I was going to say the fifth one, which is child social media security, is a joke. And then here's my point, and I'm hijacking. My point is, this is the thing I always talk about, which is falling fertility, and we need more humans to consume stuff. But on the other hand, to the point of my newsletter, there's a theoretical limit, or is there, on the amount that each person can necessarily consume? My point is, you're talking about the software market will grow exponentially, or geometrically, you were very precise in that word, not exponentially because you know the difference, but the security market and the software market, right? The amount of code that is constructed per human on the earth will dramatically increase. The amount of output or consumption per person, if their agents are consuming on their behalf, will dramatically increase. That's why these, in my opinion, these markets can be so, so big even though we have net fewer consumers to consume them, because each individual person is going to consume so much stuff. The point is, we've got... so I'm just curious on your perspective on that because you have a very bullish take on the overall size of these markets, which is a good thing, despite there seems to be a broader limit on our economic capacity, which might be the number of developed countries with enough people to buy all this stuff.
Yeah. I want the population to grow. [laughter] I'm doing everything in my power to make that true. I think it's a good thing. I think people can take care of the earth in a very responsible way and be a healthy part of the ecosystem. And I think it's really important for US GDP and the health of the country that the population grows.
But do you think the AI market can offset? Like, I guess my point is, are you bullish enough on the expanded activity of agents, software, and all of this infrastructure around artificial intelligence that if you correlated it to the number of people, that per capita output will dramatically increase?
Okay, so let me give you... maybe we can talk about AI use in three waves, right? First we use chat, and when we use chat, maybe we generate a million tokens per day. Then we use Grokbot or OpenClaw, and we start to generate on the order of like 100 to 200 million tokens per day. And then you get to a system that is on top of Grokbot, which is called a meta-harness. And now we have people consuming billions of tokens per day. Okay, so how do we make this concrete? Well, imagine I have a spreadsheet of the 100 most exciting startups that I'm tracking. And imagine inside each one of those cells there's an analysis. The name of the company. Where does the name of the company come from? Who are the founders? What are their backgrounds? How do their backgrounds compare to the backgrounds of everybody else? How fast are they growing relative to everybody else? What's their unique selling proposition? I could keep going on and on and on. I could create 30 columns, and each one of those columns is the output of 15 to 30 AI calls. Now, we're talking about an analysis that's on the order of tens of millions of tokens, maybe more, in a single spreadsheet. And so it's the parallelization capability of AI, I think, is what massively grows consumption. And we are limited in that parallelization first because we have to learn how to parallelize. That's new. Software developers did it first. We're all starting to learn with all these new bots and claws and those kinds of things. And then the second will ultimately be, are there enough computers, and is the AI efficient enough to satisfy that parallelization? But I agree. I mean, I think in the digital world, consumption increases unbounded. Physical world is a little bit different. There was this awesome tweet, I don't know if you guys saw it, about Graceland. You seen like, you know, Graceland was this legend. I'm not an Elvis Presley fan, but like it is a destination for people who love that music. And I have this impression that it is a mansion. It's enormous. And you look at it and it is like an upper-middle-class house today. And the point was of this tweet, and I'm not sure if it's accurate. So, you know, if it isn't accurate, I'd like... a 3,500-foot house is my impression. And the point was, look at how much wealth the United States has created because a lot of people now, tens of millions of people can afford, or families can afford a house like this, and it's all that consumption that's being driven. Now we can have a debate about like US dollar and all that kind of stuff. But I do think...
You're answering my question. That's sort of my point. And I think that would assuage a little bit of your fears, that Assad's great fear, well, not just Assad, it's probably a lot of people, is that we're overleveraged on investment in infrastructure and capex, and that revenue from AI is not... maybe it'll be there in the long run, but in the medium term it's not going to pace to the extent. And I guess everybody's talking about the hyperscalers, where they used to fund everything from cash, then they went to debt, now they're issuing equity, which tells us that even the most profitable monopolies in the world are tapping out of their resources and need to go to other extremes to get capital. Then we really need that revenue growth to support those investments.
Tomas also wrote about this with NVIDIA, that they're elongating their payment terms. I think, Tomas, I think it was like 45 to 60 days or 15 to 60 days, something, which is enormous, right? You think about that book of business and how fast that went. But sorry, AJ, I interrupted you.
No, no, no. I was agreeing with... I interrupted Sam, so that happens pretty frequently on the show. But I just thought Sam's point around how we're overleveraged around these things and we're looking five years into the future, and how payment terms is just... it was really fascinating to see those stats bubble up. And I guess the question is, like, where is this? Are we in another quarter? Are we going to see those elongate further, or do we think we've leveled off in those terms for NVIDIA from an investment perspective?
Oh my god, this investment is going to increase and increase and increase. And that's actually a good... like, I don't think the scale of investment is all that worrying as long as we keep unlocking more revenue, right? Like that's the idea. I used to play poker at this, I think, mafia-run poker house once, and you would go with like your own money and you would gamble, and then you would lose your money. And there was a point at which like I had no money left, and then they would lend you some money. So now the mafia is lending you something, [laughter] and you're playing on the mafia money at 3:00 a.m., and it was just a little bit scarier, right? And so I think that's kind of where we are with AI, is like we've stopped using our cash. We're doing all sorts of really interesting things to fund the capacity that we need, and we need it because otherwise all these cool things that Tomas mentioned, we won't be able to do them at an effective cost for ourselves as businesses. So we need to stand this computer up. But to justify that, we need revenue growth. And so we always need more use cases. We need improved use cases. Like we need that. And as long as we can balance it, and it's a tough balance, you're playing poker with somebody else's money. As long as you can keep balancing it, you're fine. But you're on the edge now. And that's how I think of it. I don't think it's a bad thing we're spending this, because I think if we don't spend enough, I think we actually don't get the juice out of AI as well. That's the other side of this.
What do you think about... So we've got this revenue growth versus spend that we need to balance. There is a level of AI having to get past being effective where there are verifiable answers to starting to be effective in the realm that people like to classify as places where taste and judgment come into play. Right? So we see that it's really good in code. It's not that good in marketing copy, really, right? Like it's kind of bland. So if we use that spectrum, do you see anything right now that makes you feel like there's going to be a world in which it'll be good at those things as well?
Okay, so I'm going to spoil today's blog post. [clears throat and laughter]
Yeah, you were debating it. This comes out on Sunday.
We get a scoop. Did we just scoop something?
Breaking news. You heard it here first.
Okay. So I have been trying to automate blog writing, just not to fully automate myself out of writing, but to understand the technology and like where it is on creative writing. And like a year ago, whatever, I used the OpenAI fine-tuning API. I created lots of examples. Here are the prompts, hundreds and hundreds of examples. I talked to somebody inside of OpenAI. Okay, you're using it the right way. It really didn't work. I've tried anyway. The long and short of it is I've tried many different techniques, and the technique that is finally working with the latest models is I make every edit through AI. So I start with an outline, and then I write, and then I ask the AI to write the post, and then I'll go through literally line by line of the outline, and then it'll write the post, and then I'll go through sentence by sentence of the post, and it creates a log. The AI creates a log, and at the end of it, it collapses that log into learnings, like this is what you did on this post. You changed the order to achieve this particular goal. You used a block quote here. You used an ampersand here. You didn't really like this very long sentence. And then at night, once a week, the AI goes through and says, okay, given all the blog posts you've written this week, what are the shrunken writing lessons for this blog? And so I've been doing that for about three weeks. And then I ran the analysis of like, okay, there are versions of documents which are wholesale changes, versions of blog posts, and there are individual sentence edits. And I had expected that the number of versions would decrease and the number of total edits would decrease, and it's not true. The total number of edits is 134 per post irrespective of time. So, okay, there's one of two conclusions. Either like I'm totally wasting my time, [laughter] right? What am I doing with this, right? Or, which is what I believe, I'm delivering a better product. And so I had three AI agents go and scour grading criteria for like The Economist editor, New Yorker writing, AP writing, and then go through 15 blog posts in each of the last 10 years and grade them. And in 2026, there is a whole number out of a scale of five, there is a 20% increase in the quality of the writing as a result.
So what does that tell me? I think it's like if you're a chess grandmaster and you want to be better with AI, you don't train less, you train just as much, but you hold yourself to a higher standard. And we've seen this, right? You look at chess. Chess is... the grandmasters are scored on ELO, whatever relative scoring system. So and you look at the ELOs pre-AI in chess and post, they've gone up a ton. And so it's not that I'm spending any less time writing. I still allocate an hour to an hour and a half per post. It's that the quality of the post, I look at like the quality of the analysis, the citations, the level of research is significantly better. And maybe that's what I'm aiming for instead of here's the blog post factory.
That's really interesting. Sam, do you feel like a grandmaster or do you feel like a grandmaster? A chess grandmaster. You're elevating your game up a little bit.
I'm still terrible at chess and I'm aware that I'm terrible at chess, but I definitely think that I'm capable of writing better for sure. The best part of AI, if you have taste, is you can... it's the same thing as GarageBand. Like if you knew, if you had some elements of how to write a song, if you understood basic structure, basic chord progressions, you now have a tool through GarageBand and probably through Suno even more, where all of a sudden you can put stuff together that was not possible before. You had to know so much more about how to play every single instrument and how to keep time, and you didn't have quantizing. And so now I can emphasize and support metaphors that were sort of theoretical or aspirational in my old writing, but now I can actually do the research very quickly and I can really underpin a point that I'm making with different citations in a way that makes the whole thing richer. So I totally see what Tomas is saying.
Well, and Tomas's post generally, you're really good at citations, Tomas. Like really, really good at that.
That's AI. That's not me.
Well, and that's where I think this piece of it that elevates is that the posts aren't very long, but you cite everything very clearly. Like half the post is like a citation of...
Like David Foster Wallace, right? [laughter]
All right, let's shift the conversation. And it's kind of like, let's talk about something that's, I think, a little underrated shift that's happening. We've talked on the show a lot about FTEs and that every software company is kind of quietly turning into a little bit of a service company. Tomas, you wrote about this in July. There are two posts. One is AI companies committed almost, I think it was $10 billion in a single year putting their own engineers inside companies and customers just to get the software to actually work, which is kind of funny if I think about that. 10 years ago that was not a good thing, but today it's a good thing. And to put that in perspective, it's about a fifth of what Accenture spends in a year to deliver everything that they have. You ask a question in that piece, which is whether this is a moat or is this a toll booth? Does it actually lock the customer in, or is it just a fee that they're going to collect while they keep on shopping? Because, and we want to push on that a little bit, because there was another post that you had about this AI churn piece where those two things sit diametrically opposed a little bit, because it was like 40% of AI customers, first-time users, don't come back. And so the model is only really going to hold that spot for, I think it was six weeks. So if you're keeping 40% of your customers by month five, how does the FTE math work at all? I mean, I have an instinct that there's a B2C motion and there's a B2B motion, so they're probably a little bit different, but I wanted to hear from you on this idea of, well, we're putting a lot of money in, it's kind of coming through like a sieve, but we're also spending a lot of money up the enterprise on these FTEs for engineers to go into bigger customers.
Yeah, you nailed it. Wow, great synthesis, AJ. I think you're right. I mean, I think so. If you're a model company, and you release a state-of-the-art model, you have about... you have 41 days to commercialize it before you're knocked off that perch. And so that's kind of the state of the art today. You literally have 41 days. And when individual users sign up for your new model, whether it's a DeepSeek or an OpenAI model, you can keep them for a little while, and then all of a sudden they're going to go to the next thing. And it's possible to make a business there work. Today there are heavy subsidies, because what we're doing with those subsidies is the same thing that we're doing with FTEs. We are educating users so that they bring AI to work, and we bring AI to work, charge a...
company a whole lot more than we can charge individuals, and we can make a bunch more margin. And the way we do that is we look at a business and say, everybody was working this way, after AI everybody needs to work this way, and we'll reinvent the process and we'll reinvent the software, we charge you a whole bunch of money along the way. And that's what's happening. And I think it's a really good thing, right? I think anytime there is a platform shift like this, the reality is there's a tremendous amount of education, right? All of us live in this world and we spend a lot of our time, almost all of our time, maybe too much of our time, trying to understand exactly what the implications are of like every model release and every version and every new memory architecture. But the reality is like many, many people don't, and they need somebody to distill it and when they show up at work, to educate them on what it should be and how to do it well. I think that's why we're seeing such a tremendous investment in FDEs.
