727 FinanceAuthor: Z Potentials
The gap between what models can do and how they are actually used is exactly what makes me so excited about the application layer.
Among the top 1% of enterprises by AI tool spending, median spending is roughly eight times that of the top 10%.
For founder-led companies, the founder is the asset class.
This article is compiled from the official a16z podcast episode "AI, Infrastructure, and the Next Investment Cycle." David George, Sarah Wang, Alex Immerman, and Santiago "Santi" Rodriguez are four partners at a16z's Growth Fund. In this episode, using the 25 key charts from the annual State of Markets report as a guide, they discuss the earnings behind the market rally, the hyperscalers' compute arms race, the extremely low adoption rates at the model layer and application layer, and the latest observations on consumer, the public software market, IPOs, and the secondary market. The episode was published on September 30, 2026.
David: Welcome back to the a16z podcast. I'm David George, and joining me today are my colleagues Sarah Wang, Alex Immerman, and Santiago Rodriguez. In this episode, we'll walk through the 25 key charts from the latest State of Markets report: the earnings behind the market rally, the scale of the AI buildout, the evidence of adoption growth, and what this cycle means for hardware, software, and the next generation of private companies. We'll explain what's behind the charts and also talk about what we're seeing on the front lines of enterprises. Sarah, Alex, Santi, thanks for coming.
On the day this podcast goes live, we'll also be publishing the full State of Markets report. This is our growth team's annual tradition—bringing together the biggest trends across technology, AI, infrastructure, and the broader market into one document, and today we'll pick out some of the interesting charts to discuss. If you want the full picture, be sure to read the complete version—there's a lot of substance in it. The areas we'll cover today include macro—CapEx, data centers, and accelerating demand; and front-line observations—models, applications, and a deep dive into several verticals. Technology is driving an investment boom across the entire economy: earnings growth is underpinning the market rally, and AI demand is pushing infrastructure spending far beyond previous forecasts. Hyperscalers are pouring nearly all of their near-term operating cash flow into capacity buildout, and demand continues to exceed supply in almost every case we see. That money flows into chips, power, cooling, buildings, and skilled labor, and more. This expansion coincides with the broader need to transform physical infrastructure, creating opportunities across industries.
Alright, let's dive in. Page one: technology is "the everything cycle." This may be somewhat obvious, but the numbers on this page look quite striking today—high-tech equipment, software, and R&D now account for about 55% of U.S. capital expenditure, which is a staggering figure. Technology drives investment cycles across every sector of the economy, from software and models to power, construction, and industrial capacity. Tech stocks account for about 40% of total U.S. stock market capitalization, and eight of the top ten companies by global market cap are U.S. tech companies. This is a grand narrative—you can see it in capex, you can see it in data center construction, and you can see it in the capital flowing to model companies—which together have raised more than $350 billion. Put in a longer historical context, because this is the most-cited analogy: the scale of this buildout as a share of GDP has already surpassed the railroads of their era. We all have a sense that if you fast-forward five to seven, or even ten years and look back at these numbers, the cumulative total will likely be more than 20 times what it is now. So this really does feel like the next chapter of Marc Andreessen's "software is eating the world."
Alex: Over the past 15 years, software has transformed all these industries, but at the time only a small fraction of people could build software—none of us could. And today, look at us: every one of us has several automated tasks running every night. If you extrapolate that logic to software demand, it means far more compute, chips, power, and construction. So it's no surprise that most investment is now concentrated in technology. We often get asked the same question by listeners: this is all impressive, but is it a bubble? The market has hit new highs, but the market's current trading multiples have actually come down—stocks are up about 20%, but multiples are down about 20%. That means the gains are driven by fundamental earnings, not multiple expansion. The S&P 500's earnings multiple is below 20x, and most of the companies that make up the S&P 500 are very high-quality businesses. This is completely different from the dot-com bubble—back then, the highest-market-cap companies in the world saw their stock prices surge on multiple expansion, with many stocks trading at 100x earnings. That is not the case at all now. By contrast, some memory companies—which are inherently cyclical—trade at six to seven times forward earnings. So the difference is stark.
Santi: There's an important addition here: since GPT came on the scene nearly four years ago, the market is up 90%, or 17% annualized. When you see four consecutive years of 17% annualized growth, the instinctive reaction is "this has to pull back," and we are indeed in a hot period right now. But when you put it together with what we just said—the market trading at below 20x earnings with earnings still growing at 15%—this does feel different from 2021 recently, and different from 2000, when multiples and growth weren't really in sync.
David: Completely agree. I mentioned the scale of the capex buildout earlier in the "railroad analogy." Just looking at the hyperscalers—Alphabet, Amazon, Meta, Microsoft, and Oracle—their 2026 capex is about $780 billion, versus $416 billion in 2025; all expectations point to more than $1 trillion per year starting in 2027. There's a pattern here: each generation of computing platform has supported far more users and use cases than the one before. This buildout can happen so quickly while demand still exceeds supply because the user base is driven by distribution channels that already exist. This trend builds on the internet, cloud computing, and mobile, so in theory it can reach billions of users immediately—in stark contrast to previous technology cycles.
Sarah: I think this is worth expanding on. We've gone far beyond the "ask a question occasionally" model. Think about the rise of agents: parallel tasks, long-running tasks—as you mentioned earlier, tasks keep running while you sleep at night and do other things during the day. If agents can autonomously write code, do search, and execute these tasks, then the compute these platforms need will far exceed anything we've ever imagined.
David: Completely agree. This chart shows what's happening on the capex side: capex forecasts for the five major hyperscalers have been repeatedly and significantly revised upward in a very short period of time. Spending that once looked like a ceiling—something to be reached several quarters or even years later—has now become a near-term target, because compute demand keeps ballooning. Looking at this chart, I think at any given point the instinctive reaction is "investment will plateau from here." But over the past four years, as an economy, we have consistently underestimated the strength of this trend. As Sarah said, as models evolve, utilization of installed capacity keeps getting maxed out. So the numbers David just mentioned are the actual current estimates.
Alex: To use DG's phrase, we'd all call current compute demand a "model buster." A compelling example we've all experienced: about a year ago, Sam Altman and Sarah Friar took a lot of heat for enormous compute commitments, and everyone thought they were reckless and aggressive. And now, everyone would say that decision was extremely prescient. Even so, just two weeks ago, we all saw OpenAI have to pause subscriptions to its Pro plan—a $2,000-tier service—and it was still oversubscribed. What we're experiencing is a kind of demand that is hard to satisfy.David: Everyone we talk to across every part of the supply chain tells us a different version of the same thing: demand exceeds supply. Some parts of the data center supply chain can't get materials or products until 2028. So this tightness isn't easing. At the same time, you can see evidence of high-quality business on the demand side: Microsoft, Google, and Amazon together have about $1.7 trillion in cloud backlog. These customer commitments are accumulating rapidly, and the platforms are investing heavily to serve this demand. So while free cash flow is under pressure during the current buildout period, consensus forecasts show it recovering from 2028 and then growing significantly. You can look at Amazon's most recent earnings call—they explained this J-curve dynamic well: the actual useful life of GPUs (or TPUs) is actually quite long. You first have to build the data center "shell," which takes time; then you buy chips, but the economic life of the chips is very long, and longer than any of us expected.
Sarah: Yes, I know some hyperscalers are indeed being criticized for taking on debt to fund capex, but similar to the dynamics in model labs, those who were more conservative out of hesitation later regretted it. In your recent podcast with Gavin, you talked about Microsoft, but I think this is an industry-wide issue.
Santi: Yes, another frequently discussed signal is the spot market price of existing GPUs: any GPU that can bring compute online can be rented out at attractive prices, and hyperscalers can earn considerable returns. This is the key point we often discuss: each such wave creates enormous user and consumer surplus. From what we've seen so far, consumers and users are getting tremendous value from it—which is also why they use it so heavily. The vendors serving these users are earning well. Then you go layer by layer down to the chips—the cost of acquiring chips is much higher than 12, 18, or 24 months ago, but you can still earn high gross margins and create a great deal of surplus from users.
David: We should think of hyperscaler capex as someone else's order book. These historically most profitable platforms are continuously pouring cash into AI infrastructure. As just said, this suppresses their free cash flow in the short term, but it is a huge boon for chip orders, power, and construction—which is why this broader technology boom has become an industrial boom. We are spending more and more time looking at the companies serving this industrial boom, across every link of the data center supply chain. Interestingly, traditionally our world has been more about monetizing existing IP—build once, sell infinitely; but these businesses have far more complex matters to manage: financing, managing supplier relationships, forecasting capacity, forecasting demand. We see some teams excelling at this, while others struggle or even pause. This is slightly different from the expertise we usually spend time on. This brings us to the "new era of atoms." This page has the big numbers: by 2040, global infrastructure investment demand is estimated at $90 trillion. But importantly, this goes far beyond AI and data centers, and also includes power, water, roads, and transportation. We see this everywhere in our portfolio: Anduril's massive manufacturing facility, the size of 87 football fields; Waabi aggressively expanding their transport hubs; and SpaceX's hundred-billion-dollar-scale investment in Louisiana. Betting on the future today, from our perspective, is different from the past. As Santi said, this means building factories, expanding infrastructure, and adding skilled jobs across the country.
Santi: Speaking of data centers, investing in physical infrastructure is very different from investing in software. Elon proposed years ago that "the factory is the product." It's not surprising that many of the great founders we back come from Tesla or SpaceX, because they know that, at the execution level, the factory ultimately becomes a huge competitive advantage in scaling. We often hear many myths about data centers: they will drain America's water resources; rich people don't want to live nearby; and our favorite—my electricity bill is going to skyrocket. But perhaps counterintuitively, data centers can actually help lower your electricity bill. A recent U.S. study showed that for every 10% increase in data center capacity, residential electricity prices fall by 40 basis points. You can think of it this way: the grid is a shared fixed-cost space—poles, wires, substations. A large and stable customer like a data center can spread these costs over more units of electricity. Simply put, more demand on a shared system is a good thing.
David: Yes, I completely agree with what you said. Investment in shared infrastructure can also improve the economics of the households and businesses connected to it.
Sarah: This hasn't received enough media attention. Dina Powell recently mentioned Meta's Louisiana project on a podcast, where they worked with the community to lower electricity prices. There are real examples today—this is not empty talk.
Sarah: Completely agree. Now let's talk about trends at the model layer and application layer. On the one hand, as we previewed, AI is creating considerable revenue and savings; on the other hand, adoption is still very early. On top of that, the trend is very clear: costs are plummeting. We have already previewed agents, and now agents are becoming economically viable across a wider range of work. And even as we speak, the landscape is still changing: new innovations such as Jev from new labs like TypeSafe are pushing this even further. You will see that a two-order-of-magnitude cost difference really affects the number of use cases. This is somewhat like the classic Jevons paradox, so the name is quite fitting. Overall, this is an excellent setup for both the model layer and the application layer. These improvements are creating real opportunities for software companies in both growth and profitability. It's not that everyone will win, but unsurprisingly, AI is attracting venture capital into an increasingly broad range of industries. We also see more companies reaching enormous scale in the private markets. This is not a new chart, but it is one of my favorites: the combined revenue scale of OpenAI and Anthropic. Over the past few years, we have shown this chart at every GP internal meeting and updated it every month, because it becomes outdated so quickly. The key point is visible in the numbers: the combined annualized revenue of OpenAI and Anthropic has grown to an extraordinary degree. We like to compare it with the greatest software companies in history—on the right you can see that even by estimates of the best software companies in history in terms of net new revenue, the leading labs' net new revenue has already far surpassed them. This is not just scale, but an astonishing combination of scale and speed.
Santi: But we are still early. Compared with other platform migrations, what is interesting about the AI migration is that we are not only early in the percentage of the population using AI, but also early in those users' share of wallet for AI. This is different from 2008—back then we modeled by "the percentage of people who own an iPhone"; now we need to consider both how many people will use AI products and how intensively they will use them.
Sarah: Yes, the point that "it's still early" is very important, and here is a set of metrics to support it. First, the actual deployment rate among S&P 500 companies is 69%—if you are an "AI addict" like us, that is probably close to your expectations. But when you look at quantifiable impact—which may be a good indicator of their deployment progress—it is 30%. And at the ultimate metric, one tracked over time, it is actually only 2%. So we believe AI is delivering results, but there is still enormous room to deepen usage within organizations and track those results over time. Moving from these scattered deployments to truly deep-impact, recurring workflows is frankly the next phase. I can say that for most enterprises we have spoken with, their exposure to AI is still mainly through Microsoft Copilot, which itself shows how far they still have to go.Alex: Yes, the gap between what models can do and how they're actually used is exactly what makes me so excited about the application layer. Sarah, that conversation you and Martin had with Ali Ghodsi of Databricks was great—he talked about how AI can know the world very well but know almost nothing about the company you work for. We're seeing this across our portfolio companies in the application layer—they're bridging that gap. One example that struck me over the past few months: Revolut has a very mature engineering team, but they still partnered with 11 Labs to integrate 11 Labs' best-in-class voice models into a secure connection with customer accounts and banking workflows, so that when I call as a customer, my issue gets resolved smoothly, efficiently, and securely. The opportunity today is largely about harnessing these capabilities and building reliable services.
Sarah: Beyond enterprise adoption still being early, another notable trend we wanted to illustrate with data: heavy users are pulling away across the board. You could argue AI spending is rising broadly across companies, but what we call the heaviest users are spending much more. We looked at some YipitData data: the top 1% of enterprises by AI tool spending have a median spend roughly eight times that of the top 10%.
What's striking is that the top 1% is almost equivalent to the sum of the top 2% to 10%.
Even within our own portfolio companies—which you could say are almost entirely AI-native—top users spend around $7,500 to $9,000 per month. And the median user—again, at these same most AI-native companies—spends closer to the price of a monthly subscription, around $200, maybe plus another subscription, $200 to $400. So if you look at the gap between median and top users, it's more than a 20x spending difference.
David: Yes, one question we discuss with portfolio companies is: how much adoption do they actually have, and how do you measure it. For example, look at their AI tool spending as a percentage of employee costs: for relatively forward-leaning large Fortune 500 companies, it's about 1% today.
And the most forward-leaning AI-native companies in our portfolio can be as high as 10%. So it's all about how early diffusion is and how deep it will go. We've already seen another example within the portfolio: many enterprises are very excited the first time they procure Cursor or Cognition, but the reality is that's just the beginning of the journey. From procurement and trial of these tools to full adoption, there's a long runway.
Santi: Sarah mentioned earlier that we're starting to see quantifiable case studies. We named several public companies in the report that actually disclose these metrics and are seeing material improvements from AI. On the cost side: Chime disclosed that over the past four years they've reduced their cost to serve by more than 10% annually, compounding to nearly a 50% reduction in cost to serve—supported by our portfolio company Decagon. On the revenue side, I find the Shopify example interesting: they launched an AI assistant (Sidekick) to help merchants get up to speed faster. As a result, the percentage of customers reaching five orders within 15 days of onboarding increased by 8%. Shopify tracks this metric because once merchants reach five orders, they stay and keep using Shopify. So it's also a meaningful tailwind for their business.
Sarah: Absolutely. Speaking of the advantages from lower cost to serve: you have more room to compete—better pricing, potentially broader service, ability to reinvest in growth. These benefits reach not only customers but also margins, and then capital cycles back into investment. Another very fascinating phenomenon is: of course companies differ and you can't generalize, but incumbents have also done quite well in this cycle. ServiceNow is a good example—they disclosed AI annual contract value (ACV) exceeding $1 billion, with agentic deployments growing 9x. So you see it across the entire stack, from incumbents to new companies.
David: We often discuss an interesting question—our most forward-leaning portfolio companies, like Stripe, have talked with us: where exactly are they spending incremental AI investment? Is it investing in building new products for customers to drive higher revenue? Or is it for cost-side optimization and efficiency gains? It's like our touchstone: where do these companies' founders and CEOs see the biggest opportunity. The upside opportunity from driving revenue growth is unbounded, while the opportunity on the cost side...
David: Cost-side improvements—of course you can reinvest them, but that's always just a potential opportunity that will always be there. If you think the best use of capital today is optimizing your cost structure, what does that say about your new revenue opportunities?
Santi: Even within cost optimization, there are different flavors. One is lowering cost to serve, which lets you reinvest and become a better company. I think six months ago, our whole industry was really focused on "rebuilding systems of record." If your engineers are busy rebuilding a system of record just to save a few thousand or tens of thousands of dollars, then I think those resources have better uses.
Sarah: Yes, if you add it all up, for application businesses the relevant unit is really the cost of "completing a customer task." If you can decouple that—you complete tasks for customers and charge for it, and you can make the product more economical with better routing—that's amazing, because now you're not using the most expensive model at every step, but the task still gets done. I think we'll see more margin improvement from the best application companies.
Alex: Completely agree. Sarah mentioned ServiceNow seeing massive agentic usage. Agents are already here, they're executing tasks. Tasks require multiple steps, and multiple steps mean multiple model calls—which explains the 14x growth in agent token usage on OpenRouter. But these steps are becoming cheaper. We're seeing caching applied—systems can reuse background information instead of processing from scratch each time. I saw Hebbia take full advantage of this: their financial chat workloads saw a 10x reduction in running costs. The same budget can now support more work.
Sarah: Yes, I think your broader point is: lower costs now let agents check for errors, retry, and suddenly unlock a massive number of tasks—where previously inference or tool calls might have been prohibitively expensive. This is especially important in scenarios where reliability is exactly what determines whether to use agents. Also, we talked earlier about TypeSafe—imagine when costs drop by an order of magnitude, two orders of magnitude, the number of use cases unlocked is truly hard to imagine.
David: Plus latency improving dramatically. Absolutely.
Alex: Yes, so companies are now thinking increasingly seriously about how to optimize latency or performance and cost at scale. Take Databricks, for example—they use routing, selecting the right model for each specific task, to improve performance and cost. Their intelligent router performs better: solving more problems at 35% lower cost than the strongest single model. We're now seeing a lot of another approach: fine-tuning—Harvey and Elise AI are examples. Elise fine-tuned a smaller model, reducing costs by 60% with much lower latency, making some real-time audio use cases feasible. These engineering improvements make AI more useful and cheaper.
David: We've mainly talked about enterprise usage, but the consumer side matters too—after all, that's where we've had some of our biggest returns and spent a lot of time. We think the consumer side is still early. A recent survey showed that only slightly more than 2% of U.S. households actually pay for an AI subscription. Maybe subscription isn't the best way to monetize most households, but so far it is. One interesting thing about subscription revenue is that it has some of the best retention we've seen in consumer. Alex likes to talk about the "smile curve"—it's really rare for a retention curve to not only flatten but actually "smile" upward as products improve and people come back. But that's what we're seeing with AI, which speaks to the value subscriptions deliver to payers.
Sarah: Speaking of subscriptions, what strikes me in these numbers—especially looking at the chart in the bottom right—is how small they are. Think about Amazon Prime, that's over 200 million households. Netflix is also a good example—70 million households. These numbers made me double-check, I thought "is this a mistake." So completely agree, this is just the beginning.Alex: Yes. Our partner Josh Elman has a saying I like, which may be a bit counterintuitive: consumer AI is something I use in my daily life, not necessarily something I expense. So in his view, it may not be surprising that 97% of households are using AI but not yet paying for it. You're right, Santi, I do like a smiling retention curve—we like to see users coming back more and more frequently. Another common feature of large consumer platforms is time spent: active users on Facebook, Instagram, TikTok, and Snap spend 30 to 60 minutes a day. And today's best AI assistants—many of which are now emerging—want to be my go-to app. I want them to be extremely persistent, always on, and super proactive, in which case, in the future, they may not be where I spend the most time. Yes, when agents work in the background and we no longer look at engagement through external data the way we look at social media screen time, our job will become harder, but this matters. I remember when the latest model was released in December, even in our own work, we would launch a deep research task or agent task and then drive into town—that doesn't show up in traditional screen-time metrics, and it's different from the engagement tracked by past generations of consumer products.
Santi: Yes, completely agree. It's still early, but the consumer side is truly a new and exciting moment.
Alex: Muse and Instinct are undoubtedly new players, but they are growing very fast. Together with ChatGPT, they are increasingly taking away my queries from traditional search.
This is a very dynamic moment, so if you are an existing consumer discovery platform or an existing marketplace, you really need to think about the next step now. I would ask a few questions. First, how much incremental demand can I get from AI agents? How many additional orders can they bring? Second, in my business, what percentage of the profit pool comes from ownership of the customer relationship and discovery? There was a lot of news last week: Amazon said "no, thank you" to Muse, but Instacart said "yes, please." Think about Amazon (and AWS) and Instacart: their advertising revenue exceeds all of their operating profit. So advertising revenue and ownership of the customer relationship are extremely important. But for Amazon, how many incremental new orders or new customers would connecting to Muse bring? Not that many. And look at Instacart: online penetration in groceries is still relatively early, and there are still many orders to win. The optimistic possibility is that more orders will come from these apps. In the past I had to click through in various ways to place my next order; if I hand this to an agent, suddenly there are no clicks, but hopefully there is more GMV. Trips I once thought about booking but didn't will get booked; the dinner order we want to place tonight will get placed. So I am optimistic that there will be more GMV, offsetting some of the lost advertising revenue.
David: But there is an open question here: how does it get compensated? Take Amazon as you mentioned—it has an advertising business of more than $70 billion, extremely high-margin streaming monetization built entirely on the premise that consumers visit the website and click on ads. What happens if they no longer do that? Today, Meta and Google are famously better at advertising than any internet platform, earning more than $200 per user in the U.S. and developed markets. Meta is basically an entertainment app, so let's see what happens—it may be relatively safe. On Google's side—interestingly, two years ago when all this first happened, the whole narrative around Google was "oh my God, what will happen to Google's search business," and it turned out to be very resilient. Part of the reason is that their high-monetization ad words are relatively safe—those words are more like "I need to buy insurance" or "find me a hotel in this city."
David: Or something like that—you can monetize it directly, but AI can't yet do these things for you. If that changes, it could be a very different dynamic.
David: Yes, I want to add one point: last week's narrative had a bit of a "negative-sum" flavor, very negative—"look, these marketplaces will be affected." But I do think the positive-sum view is: first, that $70 billion spent on Amazon ads may just move elsewhere through a different channel—perhaps spent directly on agents in different forms, or finding better ways to reach people; second, with lower friction, we may see more consumption—people may buy more, which is a positive flywheel driving economic growth. I think we are a bit trapped in the "this is bad for the profit pool" narrative, but it may actually just be good. ... Yes, both can be true at the same time.
David: The profit pool may change, but overall it is incremental to the economy. I think if we look back in three years and there has not been meaningful efficiency gains at the macro level—which would push consumer spending in advertising higher—all of us will be disappointed. But it may also be that profit pools in some places disappear and then get reallocated.
Alex: The immediate reaction was that the net market value of the entire ecosystem was positive. Meta's gain far exceeded the marketplaces' losses.
Sarah: Yes, going back to our discussion about software, perhaps more of it is shifting to public markets. The most prominent trend in these charts is that the composition has clearly shifted toward slower-growth, more profitable companies. About 75% of the public software sample here is profitable, while only 30% is growing at 20% or more.
David: That's a striking number: only 30% of public software companies are growing at more than 20%—and not just 20%, that's 30%.
Alex: If you draw the line at 30%, as we discussed, the absolute number is fewer than five. And in the private market, every company we see and spend time with is growing far above 30%.
David: And honestly it's a bit obvious: go talk to IT managers and CIOs—look at their growth and the money they spend on AI, and the easiest source of that money is no longer spending incremental dollars on new SaaS software projects.
David: We have invested in some SaaS software companies, including in the public market. Our conversations with their founders are very focused: how do I use my existing distribution—which in many cases is very strong and has customers locked in, as you said to Santi earlier, without churn—to launch truly interesting new AI products for their applications and push revenue growth even higher? I think to dispel the biggest fear around the "SaaS apocalypse," we need to see a period in which software companies continue to report numbers like 98% gross retention—which has always been the key point of their appeal as investment targets; continue to drive the efficiency gains we talked about; but most importantly, drive revenue growth acceleration. That would show both that they are safe on defense and that their offensive investments will actually make the business better.
Sarah: I think the point in that blog post you wrote a few months ago about "two paths" is very interesting as a rule of thumb. Maybe you can share a bit more detail about the specific percentages.
David: Yes, the more I write and the more I talk with founders, the more I feel that almost everyone we've talked to says, "Yes, I'm trying to push revenue growth higher." That is clearly the main path for the companies we invest in and support. This is very different from a year ago—when the fear was that "everyone will rewrite their software systems with vibe coding," and that is clearly not happening in the market. But there is a lot of pressure to deliver revenue growth acceleration. So we said: aim for acceleration above 10%—that is a high number, but with magical products and the budgets available for AI, it seems feasible. We'll see. We have a few companies already close, and we'll see results in the next 12 to 18 months.Sarah: Yes, absolutely. And as with everything, you can't put every company into the same "software" bucket and talk about multiple compression, growth deceleration. In fact, software has become highly differentiated: look at this chart—cybersecurity and observability are truly standing out, and vertical software overall is holding up much better than horizontal applications. The framework we use to think about "what's holding up, or even doing quite well, versus what's in a longer-term decline" is: really think about how AI changes customer demand for this product. The cybersecurity risks brought by AI have been well publicized; but as more software and agents create new security and monitoring needs, this actually creates greater demand for incumbents in these markets—you can see the CrowdStrike annotation on the right. Conversely, applications are facing varying degrees of workflow change: one of our best CEOs, Ali Ghodsi, likes to talk about "AI's guillotine" and what gets cut first. You can also think about it from the other side: what does AI actually need. Putting these two dynamics together, I think, explains most of what we're seeing in the public markets.
Alex: Yes, I think this is very consistent with what we're seeing in the private markets. As you, Sarah, hinted in security, agents are accessing more and more systems and taking more and more actions—the OpenAI and Hugging Face incidents were eye-opening for many people. So security is top priority. Also, look at vertical AI companies: they are the fastest-growing companies in our private markets—Harvey, Abridge, Elise—growing faster than any predecessor in their respective industries. This reflects customers recognizing that: the key isn't just the model—we talked about this earlier—but how you orchestrate it, how you build applications around it, and the workflow requirements specific to vertical domains.
Santi: We spend a lot of time looking at the public markets, not to invest, but to price exit multiples, so we track them very closely. If we sum up this year: the first few months were the "SaaS apocalypse." We wrote a blog post about what kinds of companies would perform well. Fast forward to today, the software index has actually returned to where it was at the start of the year, but it has really bifurcated into perceived AI losers and winners. Public markets sometimes oversimplify things too much. But to your point, AI winners aren't just those who can improve cost structures, but those who are truly accelerating and capturing this new money; if you're not capturing it, you may not be riding this wave.
Sarah: Conversely, we want to show some operating data from private companies to see what customers are actually doing. There's actually a "counter-narrative" finding in Stripe's SaaS customer data: growth accelerated entering 2026, for both younger and more mature businesses.
David: Stripe calls it "RenaSaaS."
Santi: Oh, "Renaissance."
David: I actually quite like the term "RenaSaaS."
Sarah: Yes, I like "RenaSaaS" too.
Santi: We talked a bit about public markets, and Sarah talked about SaaS acceleration in the private markets. We're often asked: will companies stay private longer? Look at the top six companies ranked by latest round valuation today: Anthropic, OpenAI, Databricks, Stripe, Waymo, Revolut, totaling about $2.4 trillion. That exceeds the total market value of IPOs over the past decade (excluding SpaceX), which cumulatively is about $1.7 trillion. So just the volume of activity in our market alone—again, six companies, $2.4 trillion—is almost as large as the Russell 2000 index, which is a massive public market index where hundreds of public fund managers spend most of their time. So spending time on these late-stage champions is increasingly exciting—they can continue investing in growth rather than maximizing short-term profits. But David, you can talk about how we advise companies on IPOs, and when they should stay private...
David: Yes, look, regarding IPOs, especially for founder-led companies, one thing we've been discussing and I've written about is: at this point, founders themselves are the asset class. The reason we make bets, and part of why some companies benefit from staying private, is that they can typically make bigger bets in the private markets, and those bets have longer payoff periods. Zuck and Elon are obviously exceptions in the public markets. But you can see it manifest in the numbers of companies like Databricks and Stripe: making huge new bets in new product areas, and subsequently seeing revenue acceleration. You can do these things as a public company too, but it comes under much tighter scrutiny—obviously, Meta's stock price fell below $100 at its lowest point, when there was extreme skepticism about their AR/VR investments. For us, the way we talk to founders is: an IPO is just another financing event. What do you gain by becoming a public company relative to a private one? Obviously you can reach significant scale in the private markets. There are also companies that already operate like public companies: highly focused on efficiency, essentially every metric tracked. And some companies may benefit from public financial disclosure to win customer trust at the enterprise and consumer levels—for example, after Navan went public, we saw it accelerate again, partly because some larger enterprises trusted them more because they were a public company.
David: Yes, and of course there are also examples where capital needs can become very large over time—the private market capital pool is very deep and can serve many of these companies, but at some point, capital needs may even become too large for the private markets to accommodate.
Santi: That's the public markets. Another area in the private markets where we spend a lot of time is the secondary market. There are two dynamics here. First, we're seeing more and more companies doing tenders for employees—if they choose to stay private, giving employees liquidity outside the public markets. One interesting data point here: Carta shows tender participation rates are only 58%—employees are choosing not to take liquidity because they have tremendous confidence in their business's performance. This is exactly the opposite of the sometimes-portrayed narrative of "employees cashing out"—what we're seeing is employees choosing not to cash out because they believe so much in their companies.
David: Imagine if that were 100%—what would happen to San Francisco housing prices.
David: That would be tough. Indeed. We obviously frequently participate in and lead these tender offers. We think they're good things. They serve two purposes. First, for employees and potential employees, it's sometimes hard to compete with the liquidity appeal of the public markets—RSUs vesting into your account on a net-tax basis every quarter. So this is a competitive weapon for the private markets when competing for talent or retaining employees against public companies. The other side: we advocate for more frequent valuation resets. This keeps your private stock price fresh, for several reasons: one, it's easier to talk about with employees and potential employees; two, if you want to do M&A—many of our companies have—you have fresh currency to use.
Santi: I think secondary market discounts have been discussed a lot because that's been the dynamic of the past five years: coming out of 2021, 22, 23, 24, the median secondary market discount to the latest round price was significant, because some rounds may have been priced too high or were stale. But what we're actually seeing today is that when there are secondary market transactions, the discount to the latest round price is essentially zero. This precisely shows that companies are raising at fresh prices, and there are always new investors willing to buy at the same price—which wasn't the case over the past year or two.
David: Yes, good point. This interesting slide is called "Everything is an AI computer." In this 2026 snapshot, AI-related companies account for 86% of US VC deal activity, up from 65% in 2025. Clearly, AI has become the dominant destination for venture capital.
Alex: Yes, but under the AI label, the opportunity has actually broadened enormously: enterprise applications, consumer applications, and also services, semiconductors, power.
Alex: And defense. A lot of attention is focused on OpenAI, xAI, and Anthropic, but beyond that, our opportunity set has never been as deep and as broad as it is now.David: Finally, we thought it would be fun to talk about some of the areas we're very excited about right now. Obviously, we touched on some of the consumer work that will be done by agents—you could call them long-running agents, autonomous agents, heavy-use consumer. If you think about it this way: a consumer agent can do all the tasks you don't want to do and give you things you wouldn't have spent time or money on—that's very compelling, and I think its diffusion will be fast.
Robotics is an area we've spent a lot of time and attention on, and we happen to think it could even be bigger than LLMs—but it may arrive three to five years earlier. So over the next five years, this will be a huge area of investment and excitement.
Autonomous driving is already here, driverless is already feasible, and this is a very exciting moment. Take a step back and look at the auto and transportation industry: it's one of the largest industries in the world, but we may not talk about it much because we spend so much time talking about AI now. If you look at Uber and Lyft's miles traveled, it's only about 1% of total U.S. miles traveled. I think that with an autonomous driving network that is 14 times safer than human drivers, that number will expand by at least an order of magnitude in the next few years. In addition, the U.S. sells 17 million new cars each year, and I think in the next decade they will all be autonomous.
David: AI times biology: this is a super exciting area, and people in labs are obviously talking about it—new drug discovery, solving some of the world's most debilitating diseases. We all hold enormous hope for this promise, and we'll see a lot of progress in the next decade.
David: Personal health is another area I'm personally very excited about—as a health maximalist, I've always felt there's no good place to put all your information about yourself and get highly personalized advice.
David: Finally, diffusion into the enterprise and beyond coding: this is what you just talked about, Sarah—the actual diffusion of technology in the enterprise is still so early, and I think this will be super, super exciting.
David: Finally, there's another area that isn't in the AI bucket, which we call American Dynamism—the re-arming of the entire "American Dynamism" technology stack. We've invested in this area for a long time and are big investors. It's still a small portion, less than 5% of total defense spending. Newer suppliers like Anduril, Saronic, and Castelion—we expect it to grow substantially as demand changes.
David: There are so many exciting areas. Obviously we're very active in AI, and we're optimistic about its impact on the overall economy—the scale of buildout is enormous, but we think it will be a massive efficiency gain for the United States. It was a pleasure recording with you. Thank you.