Mary Meeker - Artificial Intelligence (AI) Trends Report

Live from Schenectady, New York, it's SAS Talk with the Metrics Brothers, Growth and CAC.

And I'm Growth, better known as Ray Wright, founder and CEO of Benchmarkit.

And I'm CAC, better known as Dave Kellogg, EIR, Balderson Capital, independent consultant, and the author of Kellogg.

And we go together like history History and the future, Ray.

That's an interesting combination. I mean, they do say that sometimes history is the best predictor of the future. Is that where you're headed?

And that's where I'm headed. And I'm calling upon my own experiences, you know, being involved with the introduction of e-commerce and then the first generation of marketing automation. My experience says that that experience and that history is going to be a good predictor of the future.

Okay. So, I also have some, some, you know, years under my belt here watching tech industry trends, and that includes 25 now years ago reading Mary Meeker's annual reports. I can't even remember where she was a banker at. She was at Morgan Stanley, I think it was. Yeah, back in the day. And unbeknownst perhaps to our audience, but she's back and she's still producing these kind of reports under a different banner.

Well, that is a great introduction to today's episode because Mary and her new VC firm— and it's not that new, it was founded in 2018— Bond just came out with a trends in AI report, and I just thought it was fabulous, Dave.

Yeah, this harkens back to the original Mary Meeker, who was famous in the internet bubble. There were 2 people who were famous, Henry Blodgett and her, for being kind of financial analysts riding the internet bubble. And Meeker's claim to fame was she would produce this annual data-intensive report, and I mean data-intensive, capital, capital everything, um, And she's back doing it again. This year under the Bond banner, she has released a— she and her team of merry men, I think it is, merry people, have released a 339-page report entitled quite simply Trends: Artificial Intelligence. And I think it'd be interesting to try and somehow wrestle this bear to the ground, Ray, and get some takeaways for the audience from this report. And most importantly, put this report on people's radar.

Yeah. And before we cover every one of those 339 pages, Dave, just kidding, let's hear a word from our sponsor. The Metrics Brothers is proudly sponsored by Benchmarkit, the B2B AI and software industry's largest, broadest, and most contextualized benchmarking index. Dave Kellogg and Ray Groth Wright have put the Metrics Brothers podcast on hiatus for the remainder of 2026. All episodes we have ever recorded are still available on your favorite podcasting app. You can reach Dave @Kellogg and Ray on LinkedIn @RayReich. Thank you so much for listening to the Metrics Brothers podcast.

Okay, Ray, you're on point today, so why don't you get us going with this, uh, romp through the 340-slide behemoth?

Yeah, and by the way, I mean, it's going to be really hard to do this in 20 to 25 minutes, so we're going to try to just hit some of the high points. So it is a 339-page report, as you said. covers a wide variety of topics from the adoption of large language models, specifically ChatGPT, the trends of cost, the trends of revenue in AI, the capital-intensive nature of building AI, and even some of the corporate adoption and geopolitical ramifications of AI. So there's a lot to cover here, Dave.

Okay, Ray, just so people think we're not crazy, the PDF has 340 pages, the slide numbers have 339. So the reason for this is that the first slide doesn't have a page number and the numbering starts on slide 2. When we refer to slides in this deck, we're not going to be referring to the PDF page number, but the slide number at the bottom and right-hand corner of the presentation. And again, if you want to find the presentation, just Google Mary Meeker Bond Trends Artificial Intelligence, and you'll find it.

Okay. You know me, brother CAC. I like to throw out a lot of numbers. So I'm going to try to not do that today. But the first thing I wanted to talk about was just The trend. Now, everybody knows this, but ChatGPT, and I use that as a proxy for generative AI, the fastest-growing technology product from an adoption perspective in history. I think the one I will hit is that it only took 2.5 months to hit 100 million users. It actually surpassed the number of search queries in 2 years that it took 11 years for Google to do. So what I took away, and I'm not throwing a lot of the page numbers out there, but pages 4, 5, and 20 are all good to look at this. My perspective is generative AI will be the technology that impacts society, business, and our SaaS industry more than anything we've ever seen, including the internet and mobile technologies, Dave.

So I'll tell you my conclusion from this, Ray. First, to add one to your slide list, I think the slides I'd look to in this context are 4, 5, 20, and 57. And look, my conclusion is, is that I have 2 conclusions from this. One, AI is highly adoptable. It can be adopted very quickly. That's for a whole bunch of reasons, from it being a familiar user interface to it being a scalable backend to— what else makes it adoptable? Well, I guess those are the 2 really. It's not hard to use and it's priced to get used by people. And the backends are scaling, so nobody needs to limit how many people use it. And it's just kind of familiar and easy to use. So this is a highly adoptable technology that if you look at Meeker's data, is landing on people, what, 5 times, 10 times, it all depends how you look at it, faster than any other technology in history. The other related point here to me, and I can't remember the slide number, is that ChatGPT is miles ahead of other people. I think it's slide number 296, not that we're jumping around here, Ray, but if I remembered correctly—

296.

Yeah, I did.

Wow.

On slide 296, you can see that ChatGPT has over 5 times the query volume of the next competitor, which is Meta AI, and like 20 times the query volume of somebody like Perplexity. So I think 2 things are actually happening here. One, AI is highly adoptable, as I think we all know at this point. It's being adopted at a pace that should scare all of us, far faster than the internet. And the second thing is that ChatGPT has one heck of a lead here. If you read the media, it feels more like a horse race than it is.

No.

So I totally agree with that. Next thing I wanted to highlight, and, and that is that the AI ecosystem is growing so fast compared to anything else. And I'll just give a couple examples. When I say the ecosystem, what do I mean? You know, developers' communities were critical to the Apple App Store growing, right? Just critical. Well, NVIDIA has invested in developers And that ecosystem has grown 6x in the last 7 years. And the number of AI startups— I mean, think about it. SaaS is 20 years old. We've got about 40,000 SaaS companies around the world. The number of AI startups in the last 4 years is already hitting 27,000. So that means the number of developers out there of AI-based solutions— when I say AI, I'm talking generative AI— is larger than ever. And the number of applications that use AI chips, GPUs, has grown almost 3x over the last 4 years. So what I've taken away from this, Dave, is just like the Apple Store was so critical to the iPhone success, this developer and AI ecosystem is going to accelerate the availability and adoption of LLM generative AI-based applications.

Yeah, I hear you, Ray. I think there's a couple of points here and there because you're talking about ecosystems in general and NVIDIA-specific ecosystem. I think the statement is true for both. I do. I'll comment on NVIDIA because I'm a fan and one of my best friends works there for Jensen. So once in a while I get to hear what it's like to actually work there. And amazing company, amazing leader. Yeah, I mean, I think it's a great point. And the thing I most like about them is they really took a full-stack kind of systems-level approach to AI, right? It started obviously with the GPUs and the adaptation from gaming to analytics to AI model training. But they bought Mellanox, a networking company. They focused on CUDA and the CUDA developer ecosystem. And they've just done a great job kind of building the whole stack to enable AI to take off. So really a systems-level perspective. And a huge part of that is the ecosystem.

Yeah.

And as Steve Ballmer said many, many, many years ago, Ray, developers, developers, developers, covered in sweat, jumping up and down. If you've not seen the video, go Google it. But certainly NVIDIA's got the developers, developers, developers.

That wasn't the Rolling Stones' Start Me Up?

That was.

That was the Windows 95 launch, Ray. What a memory.

Okay.

Half the audience wasn't born when Steve Ballmer did that, so they're going to have to Google it. But it's worth finding.

Exactly. Okay, the third thing, and that is, you know, Dave, you and I have seen the tech stack change dramatically over the last 20-plus years. But AI is reinventing not only the technology stack, but some of the economic models that the world depends on. Let me tell you a little bit about that. So the data that are training LLMs, it's up 260% annually. At the same time, the AI algorithmic efficiency is up 200%, and supercomputer performance is up 150%. But simultaneously, compute usage is up 360%. These are all on pages 14, 15, and 16 of the report. So what I walk away from, Dave, and my big takeaway from this section of the report is that data plus electricity will be to this AI revolution what oil was to the Industrial Revolution. We're gonna be consuming so much electricity, we're gonna need so much data that we can't even imagine it. I've heard electricity, and this wasn't in the report, the demand for electricity over the next 5 to 10 years is gonna be 10x what it is today.

So 2 thoughts here, Ray. Well, one contrarian. I do think that makes the argument for small models. And there are companies that I know of and/or work with that are trying to build small kind of application-specific models. So you don't need to spend $1 billion to build a foundation model and then build apps on it, that you could just build an application-specific model. So, because just like we did with Bitcoin and crypto, we threw compute at the problem, right? With the original proof of work in Bitcoin. And we then moved to proof of stake, I think it's called, which is far less compute intensive. I am optimistic that we won't continue forever this brute force approach. In addition, I think we may actually run out of training data at some point, which raises the whole question of synthetic data. But in any case, as Andrej Karpathy said, I think one of my favorite quotes on the AI revolution is simple. It's, quote, the AI revolution is compute bound, which it is. And that's why NVIDIA is doing so well. And that's why it's going to take so much power, to your point, Ray, to power all this.

Okay, we're gonna move on to the fourth point now. You and I talked about this. So I'm not a public policy guy. I'm not a politician, right? But as I said, I am a citizen of the world, and that one of the things I took away from this report was the geopolitical tensions and impact of AI. And you look at USA and China, we're in a techno-economic arms race, and that's what Mary and her team said on pages two, seven. And to me, this might be a little bit like the Cold War and the nuclear arms race. I think AI is going to be so important to the future of society. We're going to see not only major, I'd say, competition between world leaders, China and us right now, but also the impact of regulatory issues to try to keep stuff out of Chinese hands. I mean, look at right now what Jensen Huang said. He's going to lose $5 billion of revenue. next quarter because of the restrictions to sell to China, I think it's going to have huge geopolitical ramifications.

Yeah, I would say, Ray, that slides 7 and 8— first, interestingly for the deck, the deck is literally 340 slides of nothing but charts, uh, and those are 2 of the only slides in the deck that are 100% words. Um, so if you want to read them, you can. Uh, Ray's more comfortable delving into geopolitics than I am. I don't think I have a lot to add on this one other than, I don't know, the most interesting thing I've read on this lately is this whole AI regulation question, which is really not geopolitical, but just political in the US. And in the current budget bill on the table, they're trying to stop themselves from regulating AI for the next 10 years, which strikes me as a very strange thing to agree to do. Right. You should just look at the regulations you want to apply one by one, decide whether or not to apply them rather than applying a meta-regulation that says we can't make any regulations. But, but what do I know? So, uh, I do think, you know, I think your comments on China are probably correct, but I'm not— certainly not an expert in China.

Yeah, it was interesting. Once again, I, I don't want to go down the rabbit hole here, but, you know, um, the U.S. president and the head of the Chinese Communist Party had a conversation today, and the number one thing they talked about—

Is this a joke? Is this like 2 guys walking to a bar? Is it, or is this real?

No, no, this is true. And you think it was all about, you know, tariffs and trade policy. One of the biggest things that came away was the discussion on how we're going to share rare earth minerals. That was the centerpiece of the conversation, which—

Oh yeah, they get real important real fast. Yeah, yeah, I understand that.

Okay, next point that I walked away from, and, um, this was just how AI is transforming work. And I won't speak for you, Dave. I know you talked a lot about what's going on with enterprise adoption. But I was just amazed that AI job postings were up almost 450% in the last few years. Stanford actually came out with a report that said AI companies using AI and employees using AI are 14% more productive. I thought it was low. And most importantly, with ChatGPT 4.5, where they had people kind of evaluating whether it was a human responding to questions or ChatGPT, 73% of the time ChatGPT 4.5 responses were mistaken for human responses. And what I take away from this, Dave, is I do think the number of people whose careers and jobs are going to be impacted is much greater than anyone wants to talk about today. I think AI is going to dramatically impact Employment, unemployment, and I won't even go into universal basic income because that's another political thing. But I'm scared of how much it's going to impact early career and white-collar employees.

So, a couple of points there, Ray. We could talk for a long time. If we're not careful, people are going to be thinking they're listening to the All In podcast during Not SaaS Talk. But let's talk about this. My first response is, you know, learn to code, they said. And as it turns out, you should learn AI. It would be much better advice. And I do think the younger generation may have gotten a bum steer. One of the more interesting things I saw on Twitter, where I occasionally do go look, I don't post there anymore, was an AI fluency matrix that they use at Zapier. I don't know if it's Zapier or Zapier. But, and what they do is they have 4 different levels of competency and they define engineering, PM, support people, and marketing. for how you should be using AI. So, if, you know, 15 years ago we were telling young people learn to code, now we should clearly be saying learn AI. And I thought this matrix was a really useful tool. It's not in the report, but I found it recently and thought people would want to know about it. The other thing you said, I mean, I want to talk about slide 33 in this deck where it basically, they asked, this is kind of funny, they asked ChatGPT how people will use ChatGPT in 5 years. On slide 33. And then on slide 35, they asked ChatGPT how people will use it in 10 years. And this relates to how AI impacts both work and just life in general. So, you know, generate human-level text, code, and logic, business plans, legal analysis, create full-length films and games, right? It's doing some of this stuff already now, not full-length films that you want to watch, but certainly short videos. Understand and speak like a human, power advanced personal assistants, Operate human-like robots, run autonomous customer service and sales. That's gonna be a big one in the enterprise. Personalize entire digital lives, build and run autonomous businesses. Super interesting idea. A business— they talk about impacting employment like there is no employment. Then 9, drive autonomous discovery in science. And 10, collaborate creatively like a partner, which it already starts to do today. And this is a weird list because this is 5 years out and some of this is happening today. But then go look at the 10-year out. I won't read it to you, but it's pretty impressive. So, I do think AI, I mean, everybody says this with each wave of change, it's going to have a bigger impact than, you know, the internet's going to have a bigger impact than client-server. Well, it did. And client-server had a bigger impact than mainframes. It did. Mobile had a bigger impact than the internet. It did. And now, AI is going to have a bigger impact than any of the preceding disruptive technology changes. No question.

Yeah, I want to double-click. I think 33, slide 33, is really, really helpful. In fact, the one that gets not on here and I've been hearing is ChatGPT will also put a lot of therapists out of business because I know people who use ChatGPT as their mental health partner.

So Ray, you have a friend who uses ChatGPT that way? It's okay, Ray, you can be honest with us.

It's the same friend who calls it his research assistant.

Now, on that note though, Ray, slide 42, worth looking at. This is a fun one. I think it's kind of a reverse Turing test, which is percent of testers who mistake AI as human, right? And with ChatGPT-4.5, that rate is now close to 75%. And back with ELIZA in January of '25, it was closer to 25%. With GPT-4o, Which was not that long ago, just about a year, 5/24. Again, that rate was only 20-25%. So we're— it's really closing in on human capabilities, which affects all the stuff you just talked about. If you look at slide 41, there is the Massive Multitask Language Understanding benchmark, MMLU. And, you know, in 2019, people— the AI systems would score 30% on that. Now they're scoring 92%, which is actually slightly better than humans, which score at 89%. So, so all this stuff is happening fast, right? 2019, 5 years ago, it scored 30% relative to humans at 90, and today it's actually ahead of us.

Yeah, Dave, I, I think a lot of people forget that ChatGPT was only released to the general public at the end of November 2023. It's not even been 2 years yet that it's been out there. Pretty incredible. Okay. Let's get to something that's kind of aligned with what we talk about often, and that's the capital efficiency and profitability of AI companies, which is really secondary right now to investing in the technology, proving the models, training, and user acquisition. But OpenAI's revenue, right? So they had, I think, what, $3.7 billion of revenue last year, but over $5 billion of compute costs. So this isn't about profitability. This is about investing in user acquisition and training the models and free cash flow margins for the top 6 kind of AI investment companies, Amazon, Google, Microsoft, Facebook, their free cash flow margins. And I think this is page 174, Dave, have decreased dramatically. And it's due to the CapEx intensive nature of AI where these top 6 tech companies, they're now investing 15% of their revenue in CapEx, where just a few years ago it was only 8%. So—

I mean, this is why Jensen Huang is so rich at NVIDIA, and this is why Sam Altman wants to build his own chips. Uh, as currently run, the AI revolution is compute-intensive, compute-bound. Um, so people at the low end of the stack are making a ton of money. Personally, I don't expect that to go on forever. I think there's a number of ways to solve it. But yeah, right now, you can't play the— I think that it's become like Big Pharma. Like, you need to invest billions to get a blockbuster drug, and then you can make billions on it. But if you can only invest $100 million, don't even play the game. Don't certainly play a foundation model game.

That's a great analogy. Let's hope our federal government doesn't subsidize that. Okay, away from politics again, Dave. Last one. One of the slides that I really appreciated was when they were talking about, is it going to be SaaS incumbents or kind of native LLM application challengers who will win the application game? And they talked a lot about how Adobe and Microsoft seen dramatic increase in the use of their AI products, like Adobe Firefly has 5 times as many digital assets created in the last 12 months than it did the preceding 12 months. Salesforce has Agentforce. It grew over 2x over the past 14 months. And what I took away from this is for you to win as a legacy SaaS or application provider, you're going to have to have large user bases. You're going to have to have great data that you can use as a moat to train your models, and you're going to have to take some huge risk to look forward to where AI is taking the application business versus trying to protect your flanks and being worried about your current revenue sources.

Yeah, that was a great set of slides there, Ray. You're referring to a whole kind of collection starting on slide 218, roughly, in the deck, and it shows enterprise. And I agree with all this. Regular enterprise software companies are clearly embracing AI hard, Salesforce being the most recent example. But even Zoom, all the other ones you mentioned, Adobe, Microsoft, they're all in. By the way, the thing I think is so funny here is that copilots were all the rage a year ago, and now it's like copilots are dead, agents are alive, right? This stuff changes so fast that even the marketing fads change before companies could adjust their strategy, right? Just skip the copilot era, go straight to agents. So pretty funny. I think the other thing that was just fun in that deck is, you know, I think Meeker finds stuff that I just think is great stuff to count, like the number of mentions of AI on corporate earnings calls. She's got it. for Microsoft, Google, IBM, and Salesforce for the past 5 years, and it's gone from roughly 0 to between 60 and 70 times on an earnings call, right? Which is crazy. So clearly this stuff has got people's attention, and it's getting a lot of mindshare and a lot of product launches.

Yeah, so we're running out of time here. Is there anything else you want to make sure we covered before we kind of wrap up here, Dave?

Nope, Ray.

Wrap us up.

Okay, so I think Everyone knows, and we talked about how AI adoption velocity is faster than any technology lifecycle adoption curve I've ever seen. Thank you, Geoffrey Moore, for the Crossing the Chasm framework, because I think it shows how much quicker we're moving than ever before. I actually think that with the size of investments that it's going to require to win this generative AI game, I think there's going to be just a few winners. It's almost like cloud, right? You've got 3 cloud winners. Amazon, you have Microsoft, and we have Google Cloud, and then you have all the rest, including Oracle. I think even though you see all these big guys, you said ChatGPT is so much bigger, I think we're going to see 2 or 3 gorillas and a lot of, you know, using another Geoffrey Moore kind of chimpanzees that aren't going to be of significant size. And legacy SaaS, it's going to be under attack. Even more than enterprise software companies were experiencing in 2001 through 2010. I think legacy SaaS is going to be under attack.

All right. So, let's wrap this up finally. Just 2 thoughts there. One, on Geoffrey Moore, we need to talk about the chasm and elastic rebound theory, which is the metaphor he uses. So, I'm not sure this adoption is consistent with the chasm model. I'd want to talk about that. I do agree with the chimpanzee metaphor. I think that's fine. And I think all your comments on adoption are basically correct, as well as transformation. So maybe we do an episode one day, Ray, on the chasm. In fact, it could be a fun episode to talk about. But that's it on my end, Ray, unless there's anything else. Let's see you next week.

No, this was a fun report. And just encourage everyone to go get the Bond Merrymaker AI Trends Report and have a glass of wine and enjoy it because it's good.

Yep, it is. Thanks.

The Metrics Brothers AI, SaaS, and More is a production of the Metrics Brothers Growth and CAC and a member of the Benchmarkit Podcast Network. By accessing this podcast, you acknowledge that the Metrics Brothers make no warranty, guarantee, or representation as to the accuracy or sufficiency of the information provided or the humor content of the jokes told. Ray, information, opinions, and recommendations presented are, according to our spouses, probably wrong and provided for information purposes only. This podcast should not be considered professional or for that matter, unprofessional advice. We disclaim any and all liability for any direct, indirect, indirect, misdirect, incidental, special, ordinary, consequential, inconsequential, or other damages arising out of any use of, or God help you, reliance upon the information presented here. Ray Grothreich is based in New York City and available on X @RayReich. Dave Kellogg is based in Silicon Valley and available on X @DaveKellogg. Kellogg. Schedectity, which is French for unspellable, is not our actual production location. Thanks for listening.