AI, Infrastructure, and the Next Investment Cycle

Key Takeaway

Pick one recurring workflow this week—customer support, research, reporting, or coding—and run a small AI pilot against it. Define the job’s success metric before you start (resolution time, cost per ticket, output quality, or conversion), route simple steps to cheaper models, and retain human review for high-risk actions. The episode’s central practical lesson is that AI value comes not from buying a tool, but from building reliable, measurable workflows around it.

Episode Overview

David George, Sarah Wang, Alex Immerman, and Santiago Rodriguez examine a16z’s market outlook across AI, data-center investment, software, private markets, and physical infrastructure. They argue that AI demand is being supported by real revenue, savings, and infrastructure constraints rather than purely speculative multiples, while enterprise implementation remains early. The discussion emphasizes the opportunity to turn increasingly capable and inexpensive models into dependable, industry-specific services.

Full Transcript

Transcript of AI, Infrastructure, and the Next Investment Cycle from A16Z. Auto-generated from episode audio; may contain minor errors.

Eight of the top 10 most valuable companies in the world are U.S. technology companies. Since ChatGPT was introduced nearly four years ago, the market has risen 90%, or 17% annually . Naturally, you would think this should go down . It's certainly a good time right now. This will usher in a new age of atoms . Global infrastructure investment needs are estimated to reach $90 trillion by 2040 . This goes far beyond AI and data centers . This includes electricity, water, roads, and transportation . 69% of S&P 500 companies are currently trading live .

Now, looking at the ultimate barometer—the indicator tracked over time—it's actually only 2% . AI is generating significant revenue and savings . On the other hand , implementation is still in its very early stages . Today's opportunity is largely related to leveraging these features to build reliable services . It's all wonderful , but is it a bubble ? Hmm. Hmm. Hmm. Welcome to the a16z podcast. This is David George. Today I'm with my colleagues Sarah Wang, Alex Immerman, and Santiago Rodriguez. Today, we'll be looking at 25 key slides from the latest market situation presentation .

The revenues, the scale of AI development, the evidence of expanding adoption behind the market rally, and what this cycle means for hardware, software, and the next generation of private companies . We will explain what the charts indicate and discuss what can be seen within the company during the process, so you can follow along whether you are watching or listening . Sarah, Alex, and Santi, thank you for joining us. Of course . On the day this podcast is released, we will also release a market situation presentation .

This is an annual event for our growth team, summarizing the biggest trends in technology , AI, infrastructure, and the market . Today , we'll select a subset of interesting slides and discuss them. Therefore, I recommend that you take a look at the complete version, which is packed with a lot of valuable information . The area we will be discussing today is at the macro level. In other words, I'll be talking about the acceleration of demand for CapEx, data centers, and high-level services . Then, at the practical level, we'll discuss the model, the application, and some vertical details .

Technology is driving an investment boom across the economy , with rising revenues supporting market gains and AI demand pushing infrastructure spending far beyond previous forecasts . Hyperscalers are investing all of their short-term operating cash flow into building this capacity to meet demand that continues to outpace supply in almost every case we see . They are allocating this investment to chips, electricity, cooling, construction, and skilled labor . This expansion aligns with a broader need to modernize physical infrastructure , potentially creating opportunities across the industry and reducing shared costs for businesses and households .

Okay, let's get started . The first slide shows that technology is the entire cycle. Well, this might sound a bit obvious , but the numbers at this point are quite shocking, aren't they? High-tech equipment, software, and research and development currently account for approximately 55% of U.S. capital investment , which is a remarkable figure. Technology is clearly driving the investment cycle in every sector of the economy, from software and models to power, construction, and industrial capacity . Technology accounts for nearly 40% of the total value of the U.S.

stock market . Furthermore, eight of the world's top ten companies by market capitalization are U.S. technology companies . In other words, this is a broad topic. This can be seen in capital investment figures . This is a major topic in data center construction . The amount of capital invested in model development, models, and funding for model companies has clearly been highlighted, and I believe they have raised over $350 billion to date. And to place this within a very deep historical context , this is perhaps the analogy we have seen most often.

This construction project surpassed railway construction in terms of its share of GDP . Well, it's an exciting era, with large- scale construction. We all think that if we fast forward five to seven years, or perhaps even ten years, and look at these numbers, they will have increased twentyfold cumulatively . Yeah. This feels like the next chapter in Mark's software devouring the world . Over the past 15 years, software has transformed all of these industries, but only a small fraction of the population has been able to build upon it.

None of us could do it . But look at us today . We all run some automations every night . Therefore, given the demand for software, that means more computing, chips, power, and construction . Well, so it's not surprising that the majority of investment is concentrated in technology . Therefore, one of the questions we constantly get from various sources is , "That's great , but isn't it a bubble ?" While the market has reached all-time highs, the market's price-to-earnings ratio has actually declined . In other words, while the stock price has risen by approximately 20% , the price- to-earnings ratio has fallen by approximately 20% .

In other words, performance is driven by revenue based on fundamentals . This is not due to an increase in the price-to-earnings ratio . The price-to-earnings ratio of the S&P 500 is below 20. Therefore, if you start from there , in many cases you will find a very high-quality company. This is completely different from the dot-com bubble . During the dot-com bubble , some of the world's most valuable companies saw their stock prices rise sharply based on increased price-to-earnings ratios , often reaching P/E ratios of 100 .

That's not what's happening here . In contrast, as you know , some highly cyclical memory companies are trading at six or seven times their projected earnings . In other words, it 's a completely different situation. And here, I think there's an important double click. Since ChatGPT was introduced nearly four years ago, the market has risen 90%, or 17% annually. And if the annual growth rate has been 17% for many years , the natural instinct would be that it should start to decline . It's true that things are going well right now.

However, as you mentioned, when you compare this to a situation where the market is trading at less than 20 times its earnings and growing at 15%, it certainly feels a little different from recent times like 2021, or from 2000 when the multiple and growth didn't really match. Yes, I completely agree. Earlier , I touched upon the scale of capital investment in the context of railways, but even just considering hyperscalers like Alphabet, Amazon, Meta, Microsoft, and Oracle, their capital investment in 2026 will be approximately $780 billion.

This represents an increase from $416 billion in 2025 . Furthermore, it is expected that they will invest more than $1 trillion annually from 2027 onwards . In other words, in terms of pattern recognition , each of these computing platforms has supported a far greater number of users and applications . In this case, the construction can proceed very quickly , and demand may exceed supply. This is because the number of users is already boosted by the fact that a distribution network already exists . This trend is built upon the internet, cloud computing, and mobile phones .

Therefore, in contrast to previous technology cycles , there is a possibility of reaching billions of users very quickly . Yes, and I think this is an important point. Because we sometimes run queries far beyond the models we're working with . Right? Considering the rise of agents , there are tasks that are performed in parallel or for extended periods . So, Alex, you've said this before, right? Tasks are executed at night, even while you're doing other things during the day . Well, considering that agents will be autonomously writing code, searching, and performing these tasks, these platforms will have much greater computing requirements than previously thought.

Yes, that's absolutely right . I completely agree. Well , this is another example that shows what's happening on the CapEx side . The CapEx forecasts for the top five hyperscalers have risen sharply over a fairly short period of time . In other words, spending that was once thought to be capped several quarters or even years into the future is becoming short-term as the demand for computing continues to expand. Yes, look at this chart , and I think you'll instinctively think that at any point , the investment will level off from here .

However, over the past four years, we have consistently underestimated the strength of the trend across the entire economy . As Sarah mentioned, as the model develops, the usage of the installed capacity continues to increase . Therefore, the figures David pointed out are current estimates. Yes, in other words, to borrow DG's words , the demand for computing at this point should be called a model disruptor . Well, one anecdote that illustrates this well is when Sam Altman and Sarah Frier received a lot of criticism about a year ago for their large-scale computing commitment .

Well, they were reckless and aggressive , but now everyone would say their decision was very prescient . Well, even so , everyone saw that two weeks ago they had to suspend new Pro plan sign-ups by offering a $2,000 service . no thanks. Well, what we're experiencing here is truly astonishing, insatiable demand. Yes, absolutely. Yes, almost everyone you talk to at every stage of the supply chain is telling us the same thing in some way: demand is exceeding supply . The data center supply chain has elements where materials and products will not be available until 2028 .

Therefore, this has not been eased . Therefore, if we look at hyperscalers at the same time , we can say that we can see some evidence of reliable, high-quality businesses on the demand side . In other words, Microsoft, Google, and Amazon together have a cloud backlog of approximately $1.7 trillion . Furthermore, these customer commitments are increasing rapidly, and the platform is investing heavily in capacity to accommodate them . Therefore, while free cash flow is currently sluggish during this construction period, consensus forecasts predict a recovery from 2028 onwards, followed by significant growth.

Amazon's latest earnings call did a very good job of explaining this kind of J-curve dynamic, showing that the lifespan of GPUs is actually quite long . In other words, it's necessary to build a shell for the data center , and that takes a certain amount of time. You will need to purchase chips. However, as you know, they have been economically useful for a very long time , and that has continued for longer than we anticipated . Yes, I am aware that some hyperscalers have been criticized for increasing their debt through capital investments, but I think they regret their hesitation and lack of assertiveness, just like with the dynamics of the model lab .

In a recent podcast I did with Gavin, Microsoft came up in conversation , and I think this is a problem common to all fields . Yes, another dynamic that often comes up in discussions is the pricing of existing GPUs in the spot market . This is another signal that any GPU that can be brought online is being sold at an attractive price, and hyperscalers are getting a good return . Yes, that's an important point . This is one of the things we always talk about. Each of these successive waves generates a massive amount of user or consumer surplus .

And, at least as far as we know, what is actually happening now is that consumers and users are gaining a tremendous amount of value from this . That's precisely why they use it so frequently . And the vendors providing services to these users are making enormous profits . Furthermore, to access any level of the stack, down to the tips, you have to pay far more than you did 12, 18, or 24 months ago . Even so , they can maintain extremely high profit margins and generate enormous surplus from users .

When we talk about hyperscaler capital investments, we should think of it as another company's order book. In other words, these massive platforms, which have historically generated the greatest profits, are investing in AI infrastructure . As mentioned earlier , this puts pressure on their free cash flow in the short term, but it is a great benefit for chip orders, power, and construction . And that's precisely why this widespread technology boom turned into an industrial boom. And we've spent more and more time observing the businesses that support this industry boom, and at every point in the data center supply chain .

What's interesting is that, traditionally, our world has focused on monetizing existing IP. In other words, once you build it, you can sell it indefinitely . However, these businesses involve far more complexities that companies need to manage, such as fundraising, managing vendor relationships, capacity forecasting, and demand forecasting . And we've seen some teams excel at this, while others struggle or even pause . This is a little different from the expertise we usually spend our time on . This marks the beginning of a new era for atoms.

This page contains some large figures , but global infrastructure investment needs are estimated to reach $90 trillion by 2040 . But the important point is that this goes far beyond AI and data centers . This includes electricity, water, roads, and transportation . Well, this can be seen across our entire portfolio . This can be seen by looking at Anduril's massive manufacturing facilities. Well, that's equivalent to 87 soccer fields. Well, Waymo is, well , actively expanding its depots . And of course, SpaceX has invested $100 billion in Louisiana.

Well, from our perspective , betting on the future today is not exactly the same as betting on the past, as Santi said . That means building factories, expanding infrastructure, uh, more skilled jobs, uh, nationwide. And, just like what I said earlier about data centers , I think investing in physical infrastructure is , well, completely different from investing in software. And Elon, well, coined the phrase years ago that the factory is the product. And it's no surprise that many of the great founders we've supported have come from Tesla or SpaceX .

Because, as they scaled up , they learned that factory execution ultimately became a major competitive advantage . There are many common misconceptions about data centers. It's using up all of America's water resources , the rich don't want to live near it , and one of our favorites is that electricity prices skyrocket . It may seem counterintuitive, but data centers help lower electricity costs. A recent US study showed that for every 10% increase in data center capacity , residential rates decrease by 40 bits . The power grid should be considered as being based on shared fixed costs such as utility poles, power lines, and substations .

Therefore, large, stable customers like data centers can benefit from distributing these costs across more power units . Simply put, increased demand across the entire shared system is a positive thing. Yeah. Yes, I completely agree with what you're saying. In other words, investing in shared infrastructure also leads to improved economic conditions for the homes and businesses connected to it . Yes, and I think that point has n't received enough attention in the media. Dina Powell recently appeared on Meta's podcast and talked about how she worked with a community in Louisiana to reduce electricity costs .

In other words, there are real-world examples of this today . It's not just empty words. I completely agree. Now, let's talk about the trends we're seeing in the model and application layers. On the other hand , as we've already previewed , AI is generating significant revenue and savings . On the other hand , the implementation is still in its very early stages . Furthermore, there's a trend of costs plummeting, and as we've already previewed , agents are now more economical for a much wider range of tasks .

And of course, this is changing right now, at this very moment . New innovations like Jev's, from new labs like TypeSafe, are pushing this even further. We can see that a cost difference of two orders of magnitude is beginning to have a significant impact on the number of use cases . This is a classic Jevons paradox, and that's precisely why the name is so fitting. And overall, this is a great setup for both the model layer and the application layer . These improvements are creating real opportunities for both growth and profitability among software companies .

While not everyone will succeed , it's not surprising that AI is attracting venture capital to this rapidly expanding industry . Furthermore, the number of companies reaching enormous scale in the private market is increasing. This is n't a new graph, but it's one of my favorites , showing the combined revenue size of OpenAI and Anthropic . And this is something I've covered at every GP offsite for the past few years, as far as I can remember, and it quickly becomes outdated, so I have to keep updating it every month .

What's important here, as the numbers show , is that the combined annual revenue of OpenAI and Anthropic has risen to an astonishing level . I like to illustrate this by comparing it to the greatest software companies in history . Looking to the right, we see the estimated revenue growth of the best software companies built to date , and the estimated net new revenue of the leading labs surpasses that . In other words, it's a combination that's very impressive not only in terms of scale, but also in the speed at which they reached it .

However, we are still in the very early stages. What's interesting about the AI ​​shift, compared to other platform shifts, is that it's still in its early stages, not only in terms of the proportion of the population and users using AI , but also in terms of the increase in wallets among those users . In other words, this is slightly different from modeling the percentage of people who owned an iPhone in 2008 . Here, we need to consider the percentage of people who use AI products and the extent to which they use them .

Yes, I think it's very important that we're still in the early stages , and there are plenty of indicators here that make that clear . Well, considering the live performance of S&P 500 companies , that's 69%. In other words, if we're all fed up with AI, like we all are, then it's probably closer to what we expect . Well, looking at the quantifiable impact, which is probably a good indicator of the progress of implementation, it's 30% . Well, if we look at the ultimate barometer, that is, the metrics tracked over time , it's actually only 2%.

In short, while we believe AI is achieving results , there is still considerable room for further development in actually using it within organizations and tracking its outcomes over time . Well, to put it bluntly, the next step is to move from the individual implementations we see now to truly impactful, iterative workflows . Well , most of the companies we've spoken to still have their AI involvement largely limited to Microsoft Copilot, which shows how much progress they need to make . Yes, this gap between what a model can do and how it's being used is precisely why I'm so excited about the application layer .

Sarah, you had a fantastic conversation with Martin and Ali Godossi from Databricks . He said that AI can do a lot about the world, but it knows very little about your company . And our application layer companies are trying to bridge this gap . One example that has stood out to me in recent months is how even Revolut, with its highly skilled engineering team, needed to partner with ElevenLabs to leverage Eleven's best-in-class voice models to securely connect to customer accounts and banking workflows . By doing so , when I, as a customer, call, my problems can be resolved smoothly, efficiently, and safely .

Today's opportunity lies in leveraging these features to build highly reliable services . In other words, in addition to the fact that the adoption of AI in companies is still in its very early stages , another very noteworthy trend that I would like to show with statistical data is that power users are overwhelmingly leading the way . In other words, while AI spending is generally increasing across companies , as mentioned earlier, power users who utilize AI most intensively are spending far more . Looking at Yipit's data, the median spending of the top 1% of AI vendors is about eight times higher than that of the top 10% .

Surprisingly, this is roughly the same amount as the combined 2-10% , which is effectively 1%. In other words, there are clearly power users who were early adopters . Yes, that's exactly right. And, speaking from personal experience , in our portfolio companies that are almost entirely AI-native or AI-pill , top users spend between $7,500 and $9,000 per month . And for the most AI-native companies , the median user probably costs close to a $200 monthly subscription . You may have two subscriptions overlapping. In other words, between $200 and $400 .

Considering the median and top users, spending is more than 20 times higher. Yes, we've spoken with our portfolio companies , and one of the questions we've discussed is how far along the adoption is and how to measure it . For example, if you look at the percentage of money spent on AI tools compared to the number of employees , well, in relatively advanced large companies like those in the Fortune 500, I think it's probably around 1% today . Yeah. And among the most advanced AI-implementing companies in our portfolio , this figure can reach as high as 10% .

Therefore, all of these are in the early stages of adoption, and the question is how far their adoption will progress. Yes, another example we've seen in our portfolio is that many companies get very excited when they first purchase Cursor or Cognition . However, in reality , that's only the beginning of the journey . There's a long road ahead before we acquire, test, and fully adopt these tools . Sarah, you touched on this point earlier, and we are now seeing quantifiable case studies . And here, we've highlighted some publicly traded companies that have actually reported metrics indicating meaningful improvements achieved through AI .

Yes , there are two examples that we are focusing on . The first is cost. For example, Chime reports that it has reduced its service delivery costs by more than 10% annually over the past four years. In other words, the cost of providing the service was reduced by almost 50% . This is gladly supported by Decagon, one of our portfolio companies . However, generally speaking, there are many efforts to lower the cost of providing services . In terms of revenue, I found something interesting on Shopify .

Shopify has released AI Sidekick . This AI sidekick helps the merchant speed up much faster. The percentage of customers who placed 5 orders within 15 days of onboarding increased by 8%. This is one of the metrics Shopify tracks , and achieving 5 orders means they'll keep you on their platform and help you stay . Therefore, this is a huge boost for Shopify businesses . Yes, that's exactly right. And regarding the fact that service delivery costs are decreasing , please consider the benefits that can be gained from lower service delivery costs .

This means there will be more room for competition . In short, it allows for better pricing , the ability to offer a wider range of services , and the ability to reinvest in growth . In other words, the benefits extend not only to customers but also to profit margins, which can then be reinvested . And another very interesting point is that, although the situations vary and we cannot lump them all together, existing companies have been quite successful in monetizing this . ServiceNow is a good example of this.

The report states that annual AI contract values ​​have exceeded $1 billion, and large-scale deployments have increased ninefold . In other words, this trend can be seen in all sectors, from established companies to startups . Yes, one of the interesting points we've often discussed, and one that even the most advanced companies like Stripe have talked about with us , is where they are actually making additional investments in AI . Are they investing in things that will lead to higher profits, such as developing new products for customers ?

Or are they investing in cost optimization and efficiency improvements ? And this is a kind of touchstone for us , showing where the founders and CEOs of these companies see the greatest opportunities . Opportunities to drive revenue growth hold limitless potential for upward movement . On the other hand, while opportunities for cost improvement can certainly be reinvested, they are potential opportunities that will continue to exist in the future . If you believe that optimizing the structure of your call center is the most effective way to use your funds today , what does that mean in terms of revenue opportunities ?

And there are various types of cost optimization, aren't there ? For example, reducing service delivery costs allows for reinvestment, resulting in a better business. Six months ago, our industry was focused on rebuilding its internal record-keeping systems . If engineers are focusing on rebuilding record systems to save thousands or even hundreds of thousands of dollars, there must be a better way to utilize those resources . Yes, I completely agree. Sarah mentioned that agent usage has increased significantly at ServiceNow . The agent is here , performing the task .

The task requires multiple steps and multiple model calls. This helps explain why the use of agent tokens in OpenRouter has increased 14-fold . However , these steps have become more affordable. Because the cache is accessible, the system can reuse background information instead of processing everything from scratch each time . I've seen Habia using this . They confirmed that the execution cost of financial chatwork loads was reduced to one-tenth . Well, we can now support more work with the same budget as before . Yes, your point in a broader sense is that the reduced costs have made it more realistic for agents to try, verify the work, and try again .

And it opens up many tasks that might have been prohibitively expensive to perform due to the need for reasoning and the use of tools . Well , I think this is especially important if reliability is the de facto reason for whether or not to use an agent . Well, I just talked a little about type safety, but try to imagine what would happen if the price dropped by an order of magnitude or two . Well, I think it's impossible to imagine the actual number of use cases that will be opened up .

Yes, absolutely. And latency is significantly improved. Yes, absolutely . Yes, so today, companies are increasingly thinking about how to optimize latency and overall performance while keeping costs down . Well, for example, taking databricks as an example , we leverage routing and select the appropriate model for each task we perform in order to improve performance and cost . As a result, the performance of the smart router improved. We solved more problems at 35% less cost than our most powerful individual models . Well, another approach that's commonly seen these days is fine-tuning.

I've seen this with Harvey and EleutherAI . In the case of Eleuther, we fine-tuned a smaller model . As a result, it became much more affordable , 60% cheaper, and the latency was also significantly reduced . Therefore, live use cases have become feasible from an audio perspective . In short, these engineering advancements are making AI more convenient and more affordable . Yes, and when you put it all together and think about the application business, I think the unit relevant to them is the cost of completing the customer's work.

And once you can separate it , the customer's job is done , they can be billed, and they can use better routing to make the product more economical . This is absolutely wonderful. Because it allows you to complete the job without using the most expensive model at each step . And I think this will lead to more profit margin improvements at top app companies . This is a good sign. Yes , we've mainly been talking about use within companies , but it's important to talk about consumers, which is clearly where we're seeing the biggest results and where we're spending the most time .

No, I think we're still in the very early stages. Recent surveys have shown that just over 2% of U.S. households actually have paid AI subscriptions . In short, subscriptions may not be the best place to monetize the majority of households , but so far they haven't been . An interesting aspect of subscription revenue is that it has the highest consumer retention rate we've ever seen . Alex often talks about the smile curve , and he says that when a product is improved and users come back, it's really rare for the retention rate to not only level off, but for them to actually start smiling .

However, AI has already achieved this, proving the value that subscriptions can provide to those who pay for them. By the way, regarding subscriptions, I was surprised by how small these numbers are, especially the graph in the bottom right. Consider Amazon Prime . There are over 200 million households. Yeah. Netflix is ​​a great example. That's 70 million households, right? We won't know until we investigate . It felt like K. Well, I completely agree with the opinion that it's still just the beginning. Yes, in other words, our partner Josh Elman has a slightly different definition.

What I mean is that consumer-oriented AI is something I use in my daily life , and it doesn't necessarily have to cost money . Therefore, in his view, it might not be so surprising that 97% of households using AI are still not paying for it . Yeah. Santi, you're right. I like it when the retention curve turns into a smile. I love seeing more and more users come back. As you know, another common characteristic of large consumer platforms is usage time. Looking at Facebook, Instagram , TikTok, and Snap, they are getting 60 minutes or 30-60 minutes per day from their active users .

And the best AI assistants, while there are many emerging ones now, all aim to become my favorite applications . I hope they are incredibly persistent, always on, and very proactive . That way, it might no longer be the place where I spend the most time in the future . Yes, because our agents are working in the background, it becomes difficult for us to see engagement through external data rather than looking at things like screen time on social media apps. But that's important. In other words, when the latest model was released in December , our workload also involved starting detailed survey and agent tasks , and then driving to the city .

And that doesn't mean they're aiming to be like Tesla . It's not screen time being tracked like it was for previous generations of consumers . Yeah. I completely agree. It's still in its early stages, but it's a truly new and exciting era for consumers . Muse and Instincts are definitely newcomers, but they are growing very rapidly . Along with ChatGPT, they are increasingly taking my queries away from traditional search . However, it is truly a dynamic era. Therefore, if you are running an existing consumer discovery platform or marketplace , you need to develop a strategy right now.

There are two questions I ask myself . First, how much demand increase can be expected from AI agents? How much will orders increase? And the second question is , what is the business and profit pool that can be gained from building and discovering customer relationships ? Last week , there were many news reports that Amazon responded to Muse with "No thank you, " while Instacart responded with "We'd love to." When you consider AmazonX AWS and Instacart , advertising revenue exceeds operating profit . In other words, advertising revenue and building relationships with customers are extremely important.

However, how much new orders or customers will Amazon actually see by partnering with Instacart ? There aren't that many . On the other hand, looking at Instacart, the spread of online grocery shopping is still in its relatively early stages , and there are still many orders to be acquired . Therefore, an optimistic outlook is that we will see an increase in orders from these applications . Up until now, processing orders required clicking in various ways . If you leave it to an agent , you can expect clicks to suddenly disappear and GMV to increase .

The trip we wanted to book before but couldn't, and the dinner we all wanted to order tonight, are finally happening. Therefore, I am optimistic that GMV will increase further and help offset some of the losses in advertising revenue . Yes, but the question remains: how will compensation be provided? As you pointed out , Amazon has an advertising business worth over $70 billion, a very profitable flow that relies entirely on consumers visiting their website and clicking on ads . And you can imagine what will happen if they stop doing that .

Well, nowadays , Meta and Google are arguably the best internet platforms for advertising. Absolutely. Well, each of them is earning over $200 per user in the US and other developed countries . Well, in the case of Meta, it's an entertainment application. Let's see what happens . It's probably a little safer . In Google's case, two years ago, when this situation occurred, the discussion surrounding Google revolved around the question, "What will happen to Google's search business ?" However, in reality, it had remarkable resilience. One reason for this was that highly profitable advertising keywords were relatively safe .

For example, keywords that can be directly monetized include "I need to buy insurance" or "I need to find a hotel in this city." However, AI was not yet able to take action on behalf of the user. If that changes, the situation will change dramatically. Yeah. One more thing to add is that last week's discussion was somewhat negative, focusing on which marketplaces would be affected . However, from a positive-sum perspective , firstly, Amazon's $70 billion in advertising spending could potentially flow to other channels . And, as you know, it flows elsewhere.

They will probably find better ways to target people, either by spending directly on agents through different form factors or simply by targeting individuals . And secondly, I think that if friction decreases, consumption may increase. In other words, people might buy more . And it's like a positive flywheel that stimulates economic growth . We feel bound by this, or we think this is bad for the profit pool , but in reality, it might be good . Yeah. That's absolutely right. Incidentally, both statements could potentially be true .

Yeah. yeah yeah . While this might be bad for certain profit pools, it could be beneficial for the economy as a whole . For example, if we look back three years from now and find that the overall macroeconomic improvement in activity has n't actually occurred at a meaningful rate , we'll all be disappointed . In the case of advertising, consumer spending will increase , but the profit pool may simply disappear from a specific location and be redistributed . The immediate response was a positive net market capitalization across the entire ecosystem .

Net profit far exceeds that. Any losses from the marketplace are... Yes, that's an excellent point . Returning to the discussion about software, let's think a little more about the open market . Well, the main trend that stands out from these graphs is a significant shift towards companies with low growth rates and high profitability . Of the sample of public software companies listed here, approximately 75% are profitable, and only 30% have achieved a growth rate of 20% or more . The fact that only 30% of open-source software companies are achieving growth rates of 20% or more is a remarkable figure.

And it's not just 20%, but 30% . If we use 30% as a benchmark, that would mean fewer than five companies in reality, but in the private market, almost all the companies we see or spend time with are achieving growth rates far exceeding 30% . Well, this is n't actually that obvious . You'll understand if you talk to an IT manager or CIO . etc. Given the growth they are experiencing, 100% of their spending will be on AI. In other words, the easiest way to raise that funding is to not invest new money in a new SaaS software project .

Yes, that's exactly right. Well, as you know, we invest in several SaaS software companies, including those in the open market, but the conversations we have with them and their founders are very focused on how we can leverage our existing distribution networks (which are often very strong and, as Santi pointed out earlier, lock in customers who won't go anywhere) to further increase revenue growth . " Well, I think that in order to dispel the biggest concerns about the end of SaaS , we need a period where software companies can consistently achieve figures like a 98% gross revenue retention rate .

This was the crux of why software companies were so attractive as an investment. Not only do they continue to pursue efficiency as we've discussed, but most importantly, they drive accelerated revenue growth. This shows they are safe from a defensive standpoint, but the aggressive investments they are making will actually make the business better . I found the point of your blog post about the two paths, which you wrote a few months ago , very interesting as a rule of thumb . Could you elaborate a bit more on the percentage details?

Yes, and as I've spoken with founders since then , I've found that almost everyone we've spoken to feels, 'Yes, I'm looking to further increase revenue growth.'" " Well, that, well, that, that, that, that is clearly the mainstream path for the type of companies we invest in and have supported for many years . Well, again, compared to perhaps a year ago, there was a concern back then that ' everyone's going to start coding software systems in vibecode .'" "So, it's clearly different from what's happening in the market.

Well, but there's a big responsibility to accelerate revenue growth. So I said we should aim for an acceleration of over 10%. That's a high number, but if we have a magic product, it seems achievable considering the budget we can spend on AI. So let's see what happens. I think there are a few things that are close to that, which will happen in the next 12 to 18 months. Yes, absolutely. And as with everything, you ca n't confine every company to the software box and talk about multiples going down or growth slowing down.

In fact, software is very diversified. And looking at this graph, you can see that cybersecurity and observability really stand out . Well, vertical software is generally holding up better than horizontal applications. And the framework for thinking about what's holding up, or growing well , or heading towards long-term decline is how AI is changing customer needs for that product." I think it's something to consider. Right? And the cybersecurity risks associated with AI are widely known. However, given that more software and agents are creating new security and surveillance needs, this is frankly increasing demand in these markets for existing companies.

You can see that in CrowdStrike's critique on the right . And of course, applications are facing varying degrees of workflow changes . And, for example, one of our top CEOs, Ali Godossi, often talks about the AI ​​truncation table, but we can also consider what will be truncated first, and at the same time, what AI actually needs . I think these dynamics combined explain a lot of what we see in the open market . Yes, I think a lot of what we see in the private market is quite intuitive .

As Sarah mentioned regarding security, agents are accessing more and more systems and taking more actions . The OpenAI facial recognition incident is It was a breathtaking event for many. So security is a top priority. And when you look at vertically integrated AI subcommittees , these are some of the fastest-growing companies in the private market . Harvey, Bridge Ellis, they are growing faster than ever in the industry . And it's about how customers are not just looking at models , but, as I mentioned earlier , how to orchestrate them, build applications around them, and control those needs with industry-specific workflows .

So we have spent a fair amount of time looking at the public market for investments . But we have the right exit multiples, right? So we are tracking it quite closely . And looking back at this year , the first few months were the end of SAS . We wrote blog posts about what kind of businesses are doing well, but fast forward to today and the software index is actually back to its level at the beginning of the year . But the companies that are seen as AI losers and AI winners and : There is a divide between companies that are considered to be the winners .

And the open market may sometimes oversimplify things a little, but as you pointed out, the winners of AI are not just companies that can improve their cost structure . They are companies that are actually accelerating and earning new net dollars that can be earned . If they are not actually catching it , they are probably not riding the wave. Yeah . Yeah. Well, on the other hand, I wanted to show you some operational data from private companies that shows what customers are actually doing . Well, Stripe's SaaS customer data has a bit of a narrative violation.

Well, it shows that both young and mature companies are seeing accelerated growth in 2026. Stripe calls it a renaissance . Oh, a renaissance. I really like renaissances . I like renaissances. Yeah . It's very good. It's very good . Well, as for me, if you're talking, we've been talking a little about the open market . Sarah is talking about the private market We talked about the acceleration of SaaS . I get a lot of questions about companies staying private longer. So, if you actually look at the top 6 companies today, Anthropic, OpenAI , Databricks, Stripe, Waymo, and Revolut, the combined final round valuations are about $2.4 trillion.

That's more than the combined market capitalization of IPOs over the last 10 years, excluding SpaceX, which is $1.7 trillion . So, just the volume of activity happening in our market alone for these 6 companies is $2.4 trillion, which is roughly the same size as the Russell 2000. The Russell 2000 is a huge index of the public market where hundreds of fund managers spend most of their time. So, it's becoming increasingly exciting to spend time on these late-stage champions who can continue to invest in growth rather than maximizing short-term profits.

But David, how do you advise companies regarding IPOs and when to stay private or go public? Could you tell us a little about that ? Yes, with IPOs, especially for founder-driven companies, one thing we've been talking about , as I've written before, is that founders are the asset class of today. And one of the reasons we're betting, and why it's beneficial for some of these companies to remain private, is that in the private market you can take bigger risks and get returns over a longer period .

As you know, Zuckerberg and Elon are clearly exceptions in the public market, but what the numbers show for companies like Databricks and Stripe is massive new investment in new product areas . And as a result, you see accelerating earnings . So you can do the same in the public market, but you'll be under much scrutiny . As you know, Meta's stock price fell below $100 per share at a time when people were very skeptical about investing in AR/VR. When you talk to founders , many of them say that an IPO is just another fundraising event .

The question is, what are the differences in the advantages of being a publicly traded company versus a privately traded company? Of course, large- scale business operations are possible even in the private market. In fact, there are some companies that are managed like publicly traded companies . These are companies that prioritize efficiency and track all metrics. Also, by disclosing financial information , some companies can gain customer trust at both the company level and the consumer level . For example , Navan went public, and its performance accelerated again because it gained the trust of large companies simply because it was a publicly traded company .

Of course, there are cases where the demand for funds becomes very large in the long term. The capital pool in the private market is very large and can meet the needs of many companies , but at some point, the demand for funds may become even larger . It is too big for the private market . That is the public market. Even in the private market, one of the things we spend a lot of time on is the secondary market. In other words, two dynamics are at work.

One is that more companies are conducting tender offers for their employees . In other words, companies If the company chooses to remain private, employees can obtain liquidity outside the public market. An interesting data point here is that only 58% of Carta's tender offers were taken over . This suggests that employees are choosing not to pursue liquidity because they have a high level of confidence in their company's performance. In other words, it's the opposite of what you'd expect from employees cashing in. In fact , employees are choosing not to cash in because they have a high level of confidence in their company .

If this were 100%, imagine the price of a house in San Francisco . Exactly. It would be difficult. Exactly . So, there have been 10 tender offers , and we've clearly taken over. We've led some of them, and quite frequently . So, we think these are good . They serve two purposes . First, for employees and future employees, the liquidity of public market RSUs, reflected in their accounts on a net tax basis on a quarterly basis, can sometimes be hard to resist . So this is a competitive weapon for private companies that want to compete with publicly traded companies to acquire and retain employees.

Another aspect is that we recommend resetting valuations more frequently . Doing so can keep private stock prices fresh for several reasons . First, it makes it easier to talk about it with employees and prospective employees . Second, if you want to do an M&A, as many companies have done , you have new funds to use. Yes. No, perhaps just a different dynamic. Secondary discounts are often talked about , and this has been a common phenomenon over the past five years . Up until around 2021, 2022, 2023, and 2024, a median discount to the final round price in the secondary market was meaningful because there were cases where the price was too high or outdated .

But what we see now is that when trading takes place in the secondary market, the final round price The discount on that is practically zero . This indicates that the valuation at which companies raised funds is still fresh, and there are always new investors willing to pay the same price . This is something we haven't seen in the last few years . That's a great point. Now , here's an interesting slide that says "It's all AI computers ." This snapshot of 2026 shows that AI-related companies account for 86% of U.S.

VC trading activity, up from 65% in 2025 . In other words, AI is clearly a major target for venture capital funding . Yes, but under the label of AI , opportunities are expanding significantly, aren't they ? Enterprise applications, consumer applications, services, semiconductors, power, defense, etc. While attention is focused on OpenAI, SpaceX, AI, and Anthropic , our opportunities are deeper and broader than ever before. Finally, I'd like to talk a little about an area that we're very excited about right now . As you know, we've already discussed the consumer-facing work performed by agents .

These can be called long-running agents , autonomous agents, or heavy-use consumers . From these perspectives, you can have consumer agents perform all the tasks you don't want to do, and in return, you can acquire things you wouldn't have otherwise spent your time or money on. It's very appealing, and, well, this adoption could progress quite rapidly . Robotics is a field in which we dedicate a great deal of time and attention . We believe it has the potential to become even larger than an LLM . Well, probably three to five years earlier.

Well, I think this will be a field of enormous investment and excitement over the next five years . Well, autonomy lies here . Well, the autonomous driving function works. Well, it's a very exciting time. Well, if we pause for a moment and talk about the automotive and transportation industries , which are among the world's largest industries, we probably haven't talked about them much because we're spending so much time talking about AI right now . Well, if you look at the mileage for Uber and Lyft , it accounts for about 1% of the total mileage in the United States .

Well, well, if we had a complete network of self-driving cars that are 14 times safer than human drivers, we expect it to expand, at least . It will grow exponentially in the next few years . Furthermore, 17 million new cars are sold annually in the United States , and I believe that within the next 10 years, all of them will be self-driving . The fusion of AI and biotechnology is an extremely exciting field. The people at the research institute are certainly talking about this, but I think everyone is hoping that the discovery of new drugs and solving some of the world's most debilitating diseases will lead to significant progress in the next decade .

Personal health is also an area that I, as a health enthusiast, am very much looking forward to . However, until now, there hasn't been a good place or method where you can input all your information somewhere and receive highly personalized advice . And finally, beyond coding, there's the issue of enterprise adoption. This is one of the things Sarah was talking about. However, since the actual adoption of this technology by businesses is still in its early stages , I think it will be very exciting . And finally, there's one more area that doesn't fall within the realm of AI .

It's a reconstruction of what's known as American dynamism, or rather, the entire American dynamism stack. We have been investing in this field for some time now . It's a tiny fraction, less than 5% of total military spending , but new vendors like Anduril, Saronic, and Castilian are predicting that this sector will grow dramatically as needs change. Therefore , there are a great many exciting fields . As you know, we are very active in the field of AI and are optimistic about its impact on the overall economy.

The construction is large-scale, but we believe it will significantly improve productivity in the United States . It was a pleasure meeting you all . thank you very much.

Go deeper

• All A16Z episode guides