AI, Infrastructure, and the Next Investment Cycle
David GeorgeSarah WangAlex ImmermanSantiago Rodriguez
The AI investment cycle is enormous, but the panel argues that earnings—not multiple expansion—are driving the market. Since the referenced presentation came out almost four years ago, the market has risen 90%, or 17% annualized; stocks recently gained roughly 20% as multiples fell roughly 20%; and the S&P 500 trades below 20 times earnings while growing about 15%. “We’re definitely in a hot period,” but this looks unlike 2000 or 2021, when valuation and growth detached.
Hyperscaler spending is becoming an industrial cycle whose scale keeps outrunning forecasts. Alphabet, Amazon, Meta, Microsoft, and Oracle are investing about $780 billion in 2026, up from $416 billion in 2025, with expectations of more than $1 trillion annually from 2027; combined cloud backlog at Microsoft, Google, and Amazon is about $1.7 trillion. Compute demand is a “model buster”: some inputs are unavailable until 2028, yet available GPUs still command attractive prices and returns.
AI’s buildout is merging with a much broader “new age of atoms.” Global infrastructure needs are estimated at $90 trillion through 2040, spanning chips, power, cooling, factories, water, roads, transit, and skilled labor. The counterintuitive grid argument is that shared investment can benefit households: one cited study found residential electricity rates fell 40 basis points for every 10% increase in data-center capacity.
Enterprise AI has broad experimentation but remarkably little institutionalized adoption. Some 69% of S&P 500 companies report live deployments and 30% report quantifiable impact, yet only 2% track an AI metric over time. That gap—“moving from these individual deployments to really deeply influential recurring workflows”—is the application-layer opportunity, because models know much about the world but little about a specific company.
Power users and measurable business outcomes show how deep adoption might eventually run. The top 1% of AI spenders spend roughly eight times the top-10% level; inside AI-native portfolio companies, leading users spend $7,500-$9,000 monthly versus $200-$400 for median users. Chime reports more than 10% annual reductions in cost to serve for four years, Shopify reports that the share of merchants reaching five orders within 15 days grew 8% after Sidekick launched, and ServiceNow has surpassed $1 billion of AI ACV.
Falling inference costs are making multi-step agents economically viable and potentially more profitable for application companies. Agent-token usage on OpenRouter has grown 14-fold, Hebia’s financial-chat workloads became 10 times cheaper, Databricks’ router solved more problems at 35% lower cost than the strongest standalone model, and EliseAI’s fine-tuned smaller model cost 60% less. The investable unit is increasingly “the cost of getting the customer’s job done,” not allegiance to one model.
Value is bifurcating across software, consumer discovery, and private markets rather than disappearing uniformly. Only 30% of public software companies grow at least 20%, while fewer than five exceed 30%; cybersecurity, observability, and vertical AI are holding up better than horizontal applications. Meanwhile, six private champions total roughly $2.4 trillion in valuation, AI-related companies represent 86% of U.S. venture activity, and agents may relocate advertising and marketplace profits even if lower friction expands total consumption.
1. Technology has become the economy’s dominant investment cycle
The panel’s opening frame is “tech is the everything cycle”: high-tech equipment, software, and R&D now represent roughly 55% of U.S. capital spending, technology is nearly 40% of U.S. stock-market value, and eight of the world’s ten most valuable companies are American technology businesses.
George estimates that model companies have collectively raised more than $350 billion, while the AI buildout has surpassed historical railroad investment as a share of GDP. The group thinks today’s cumulative figures could look “20x higher” five to ten years from now as accessible software creation multiplies demand for compute, chips, power, and construction.
The panel treats this as the next chapter of “software is eating the world.” Previous software waves empowered a limited population of builders; AI lets far more people create automations, including long-running processes that continue working overnight.
The panel also tracks OpenAI and Anthropic’s combined annualized revenue, which has risen with unusual speed and surpassed estimates of net-new revenue added by the best software companies.
2. Strong earnings distinguish the rally from a valuation bubble
The panel raises the unavoidable question—“That’s all great, but is it a bubble?”—then points to the divergence between prices and valuations: stocks are up about 20% while trading multiples are down about 20%, meaning earnings rather than multiple expansion account for the appreciation.
The comparison with the dot-com boom is categorical: today’s largest companies are high-quality businesses trading at nothing like the roughly 100-times-P/E valuations once attached to major internet stocks. Some highly cyclical memory companies now trade at only six or seven times forward earnings.
The skeptical case remains worth preserving. Since the referenced presentation came out almost four years ago, a 90% market gain and 17% annualized return naturally invite mean-reversion concerns; the panel’s rebuttal is that a sub-20-times market growing earnings around 15% is fundamentally different from 2000 or 2021.
3. Compute demand keeps breaking every capex forecast
The five largest hyperscalers’ projected capex rises from $416 billion in 2025 to about $780 billion in 2026, with expectations exceeding $1 trillion annually from 2027. Forecasts repeatedly assumed spending would flatten, but each supposed ceiling became near-term reality as installed capacity was absorbed.
The mechanism has changed from occasional prompts to parallel, long-running agents that write code, search, use tools, and retry tasks autonomously. Because AI sits atop internet, cloud, and mobile distribution, new capacity can immediately serve billions of potential users; the panel calls the resulting demand a “model buster.”
The panel’s sharpest anecdote is that massive compute commitments criticized as reckless a year earlier came to look prescient—yet a provider still had to pause new subscriptions on its Pro plans. Across the supply chain, some materials and products cannot be secured until 2028.
Microsoft, Google, and Amazon together have about $1.7 trillion of cloud backlog. Free cash flow is temporarily depressed by construction and chip purchases, but forecasts show recovery beginning in 2028; longer-than-expected GPU and TPU lives underpin the “J-curve” between upfront expenditure and later cash generation.
4. The AI boom is becoming a physical-infrastructure boom
Every hyperscaler capital outlay becomes “someone else’s order books”: chips, power, cooling, construction, financing, materials, and skilled labor. The operating challenge differs from software’s build-once, sell-infinitely model because teams must manage vendors, capacity, financing, and uncertain demand simultaneously.
The broader frame is a “new age of atoms,” with global infrastructure needs estimated at $90 trillion through 2040. Examples include an Anderoll manufacturing facility spanning 87 football fields, Whimo depot expansion, and SpaceX’s cited $100 billion Louisiana investment—evidence that betting on technology increasingly means building factories and physical systems. As David frames it, “the factory is the product,” and factory execution can become a competitive advantage.
One counterintuitive claim is that data centers need not raise household power bills. Because poles, substations, and wires are shared fixed costs, a large stable customer can spread them across more electricity; a recent U.S. study found that each 10% increase in data-center capacity was associated with a 40-basis-point decline in residential rates.
5. Enterprise deployment is widespread, but real diffusion has barely started
The adoption funnel falls rapidly from 69% of S&P 500 companies reporting live deployments, to 30% reporting quantifiable impact, to just 2% tracking an AI metric over time. Most enterprises the panel encounters remain centered on Microsoft Copilot, leaving substantial room for recurring, workflow-level adoption.
The application layer bridges a contextual gap: models may know the world while knowing little about a specific company. Revolute, despite its sophisticated engineering organization, still partnered with ElevenLabs to connect voice capabilities securely to accounts and banking workflows—the work required to turn a model into a reliable service.
Usage is already radically uneven. The top 1% of AI spenders spend roughly eight times the top-decile level and almost as much as the 2%-10% group combined; within AI-native companies, monthly spend can range from $200-$400 for median users to $7,500-$9,000 for leaders.
Public-company evidence shows both margin and revenue effects. Chime’s cost to serve has fallen more than 10% annually for four years, Shopify reports that the share of merchants reaching five orders within 15 days grew 8% after Sidekick launched, and ServiceNow reports more than $1 billion of AI ACV alongside a ninefold increase in agentic deployments.
6. Cheaper agents turn reliability into an economic design problem
Agents require many model calls because tasks involve multiple steps, checks, and retries; the panel cites 14-fold growth in agent-token usage on OpenRouter. Caching helped Hebia make its financial-chat workload 10 times cheaper, allowing the same budget to support substantially more work.
Lower costs let an agent “try, check its work, try again,” especially where reliability previously made deployment uneconomic. The panel argues that another one or two orders of magnitude of cost reduction could unlock use cases that are currently difficult even to imagine, particularly when paired with lower latency.
Databricks’ smart router performed better than the strongest individual model while solving more problems at 35% lower cost. Fine-tuning produced similar gains for EliseAI: a smaller model became 60% cheaper and sufficiently faster for live audio.
The application-company insight is to optimize for completed work rather than defaulting to the most expensive model at every step. Routing and fine-tuning can decouple customer value from inference cost, creating a plausible path to better margins even as usage intensifies.
7. Consumer agents may expand commerce while redistributing its profits
Only slightly more than 2% of U.S. households pay for AI subscriptions, tiny beside Amazon Prime’s more than 200 million households or Netflix’s roughly 70 million. Yet paid AI products show unusually strong “smiling” retention curves as improving products bring users back more frequently.
Josh Elman’s counterpoint is that consumer AI is what people use in daily life, not necessarily what they personally expense; low household payment may therefore understate adoption. Traditional screen-time measures also miss agents doing deep research or completing tasks in the background.
Agentic discovery creates a strategic split between Amazon and Instacart. Amazon has less incremental demand to gain and more customer-relationship economics to defend, including an advertising business above $70 billion; grocery’s lower online penetration gives Instacart more potential orders to capture from reduced friction.
The disagreement stays unresolved. Agents might erode marketplace advertising and move those dollars elsewhere, but they might also create trips, dinners, and purchases that consumers previously abandoned. As the panel concludes, “both could be true”: individual profit pools can shrink while economy-wide consumption and productivity rise.
8. Software and private markets are separating winners from passengers
Roughly 75% of the public software sample is profitable, but only 30% grows at least 20%; fewer than five companies exceed 30% growth. AI budgets can displace incremental SaaS projects, so incumbents must pair durable retention—around 98% gross dollar retention—with new AI products that reaccelerate revenue.
The proposed offensive target is 10% or more of revenue-growth acceleration over 12-18 months. Cybersecurity and observability benefit as agents access more systems and create monitoring needs, while vertical companies such as Harvey, Abridge, and EliseAI are growing faster than historical precedents in their industries.
Stripe’s private-company data nevertheless suggests a SaaS “renaissance,” with growth accelerating into 2026 across young and mature businesses. Six private companies—Anthropic, OpenAI, Data Brick, Stripe, Wayland, and Revolute—carry about $2.4 trillion of combined valuation, versus $1.7 trillion for a decade of IPOs excluding SpaceX.
Staying private can let founder-led companies take longer-duration bets; the IPO is framed as another financing event. Employee tenders provide liquidity, but Carta data shows only 58% participation, while current secondary transactions are occurring at roughly the last-round price rather than the meaningful discounts seen in prior years.
AI-related companies now represent 86% of U.S. venture deal activity, up from 65% in 2025, but the label spans applications, chips, power, services, defense, and infrastructure. The panel’s next-cycle bets include robotics potentially exceeding LLMs, autonomous networks expanding miles traveled by at least an order of magnitude, AI-enabled biology and personal health, deeper enterprise diffusion beyond coding, and American Dynamism, where newer vendors still account for less than 5% of military spending.
Full transcript
1. Power users spend 20x the median
Eight of the top 10 valued companies in the world are U.S. tech companies. Since JBT came out almost 4 years ago, the market’s up 90%, which is 17% annualized. The natural instinct is, well, that’s got to come down. We’re definitely in a hot period. This puts us in a new age of atoms. Global infrastructure investment needs are estimated at $90 trillion through 2040. This goes way beyond AI and data centers. It includes power, water, roads, and transit. Live deployments at S&P 500 companies, that’s at 69%. Now, if you go to the ultimate barometer, which is a metric tracked over time, that’s actually only at 2%. AI is generating major revenue and savings. On the other hand, adoption is still extremely early. Today’s opportunity is so much around taking these capabilities and harnessing them to build reliable services. That’s all great, but is it a bubble? Um, um um.
2. Is it a bubble?
Welcome back to the A16Z podcast. I’m David George. I’m here with my colleagues Sarah Wang, Alex Immerman, and Santiago Rodriguez.
Today we’re walking through 25 key slides from our latest State of Markets presentation: the earnings behind the market’s rise, the scale of the AI buildout, the evidence of growing adoption, and what this cycle means for hardware, software, and the next generation of private companies. We’ll explain what the charts show and discuss what we’re seeing inside businesses along the way, so you can follow along whether you’re watching or listening. Sarah, Alex, Santi, thanks for joining me.
Of course.
The day this podcast goes live, we will be releasing our State of Markets presentation. And so this is a now yearly tradition from our growth team where we synthesize the biggest trends in tech, AI, infra, and markets. Today we’re just picking out a subset of interesting slides and having a discussion about them. I would direct you to look at the whole version, which is filled with a lot of nuggets. The areas that we’re going to discuss today are macro—where we’ll talk about capex, data centers, and accelerating demand at a high level—and then on-the-ground takeaways, where we’ll discuss models, apps, and some vertical deep dives.
Technology is driving an economy-wide investment boom, with rising earnings supporting market gains and AI demand pushing infrastructure spending far beyond earlier forecasts. Hyperscalers are investing all of their near-term operating cash flow to build out this capacity, as demand continues to outstrip supply in almost every case that we see. They are channeling this investment into chips, power, cooling, construction, skilled labor, and much more.
This expansion coincides with a broader need to modernize physical infrastructure, creating opportunities across industries and potentially lowering shared costs for businesses and households. So with that, let’s jump in. First slide: “Tech is the everything cycle.”
Some of this may be a little bit obvious, but the numbers are somewhat striking at this point, right? High-tech equipment, software, and R&D now account for roughly 55% of U.S. capital spending, which is just a staggering number. Tech is driving the investment cycle across all areas of the economy, from software and models to power, construction, and industrial capacity.
Tech is almost 40% of the aggregate value of the whole stock market in the U.S., and 8 of the top 10 most valuable companies in the world are U.S. tech companies. So this is a broad story. You’re seeing it in capex. This is heavily covered in the data center buildout. It’s obviously heavily covered in the amount of capital that’s been going into funding model development, with model companies together raising, I think, over $350 billion at this point.
And just to put it into very deep historical context, because this is probably the analogy that we’ve seen the most of, this buildout just surpassed railroads as a percentage of GDP. So, exciting times, massive buildout. We all happen to think that if you fast-forward 5 to 7, maybe 10 years from now, we’ll be looking at these numbers and they’ll probably be 20 times higher cumulatively.
Yeah, this does feel like the next chapter of Mark’s “software is eating the world.” Over the last 15 years, software has transformed all these industries, but only a small fraction of the population could build it. None of us could. But today, look at us: We all have a handful of automations running every night.
And so if you think about the software demand, that means a lot more compute, chips, power, and construction. It’s no surprise to see the majority of investment now going into tech.
One of the questions that we get all the time from various audiences is, “Okay, that’s all great, but is it a bubble?” The market has reached new highs, and at the same time that it has reached new highs, the trading multiples of the market are actually down. Stocks are up about 20%, while multiples are down about 20%.
What that means is that performance is driven by fundamental earnings, not increased multiples. The S&P 500 earnings multiple is below 20 times. If you just start with that, these are, in most cases, very high-quality businesses. It’s nothing like the dot-com boom in that way, where some of the highest-market-cap companies in the world had their massive stock run-ups based on increases in their trading multiples and would trade, in many cases, for around 100 times P/E. That’s not what’s happening here.
In contrast, some of the memory companies, which again are very cyclical, are trading for, call it, 6 or 7 times forward earnings. So, very different.
And I think there’s an important double-click here: Since JVD came out almost 4 years ago, the market’s up 90%, which is 17% annualized. Anytime there’s been 17% annualized growth over 4 years, the natural instinct is, “Well, that’s got to come down,” right? We’re definitely in a hot period.
But when you compare that to, as you said, the market trading for below 20 times earnings and growing 15%, this definitely feels a little different from either the 2021 period or the 2000 period, when multiples and growth were not really going together.
3. The $780B hyperscaler CapEx race
Yeah, totally agreed. I mentioned the scale of the capex buildout. In the context of the railroads, if you just look at the hyperscalers—Alphabet, Amazon, Meta, Microsoft, and Oracle—their capex in 2026 is about $780 billion. That’s up from $416 billion in 2025, and all expectations point to them spending over $1 trillion annually from 2027.
The pattern recognition is that each of these computing platforms has supported a much larger population of users and uses. In this case, this buildout can happen so quickly and demand can still outstrip supply because the number of users is driven by the fact that there’s already existing distribution.
This trend is built on top of the internet, cloud computing, and mobile phones. It can immediately reach billions of users, in contrast to previous technology cycles.
4. Stripe's Renaissance data
Yeah, and I think it’s an important point to double-click on, just because we’ve moved well beyond this model of occasional queries, right? If you think about the rise of agents, you have parallel tasks and long-running tasks. Alex, you mentioned this previously, right? You have tasks going into the evening and during the day, when you’re doing other things.
And so if you think about agents autonomously writing code, searching, and carrying out these tasks, it just gives these platforms a much larger compute requirement than ever thought before.
Yeah, absolutely. Totally agree. This is another slide that just points to what is happening on the capex side of things. Successive forecasts for the 5 largest hyperscalers’ capex have moved sharply higher in pretty short succession. Spending that once looked like a ceiling—a number of quarters or years out—has become near-term as the demand for compute keeps expanding.
Yeah, look, if you look at the chart here, I think at any point in time, the natural instinct is just to say that the investment will flatline from here. Yet for the last 4 years, as an economy, we keep underestimating the strength of the trend.
As Sarah said, with model development, usage of our installed capacity keeps being taken up. So right now, the numbers that David pointed to are the current estimates.
Yeah. I mean, maybe to use DG’s language, I think we’d all call the demand for compute a model buster at this point. One anecdote that’s telling, and that we’ve all experienced, is that Sam Alman and Sarah Frier got a lot of flak a year or so ago for their massive compute commitments. They were being reckless and aggressive, and I think at this point everyone would say they were incredibly prudent with that decision.
Even so, 2 weeks ago, we all saw that they had to pause new subscriptions on their Pro plans. Showing up with a $2,000 service? No, no thanks. It’s pretty amazing—the insatiable demand that we’re experiencing here.
Yeah, absolutely. Pretty much everyone we talk to at every stage of the supply chain is telling us the same thing—some version of the same thing: Demand outstrips supply. There are certain elements in the data center supply chain where you can’t get access to materials or products until 2028, and so this has not softened.
At the same time, you can look to the hyperscalers and see some evidence of high-quality businesses on the demand side that you can hang your hat on. Microsoft, Google, and Amazon have about $1.7 trillion of combined cloud backlog together.
Those customer commitments are building rapidly while the platforms invest heavily in the capacity to serve them. While free cash flow is depressed during this buildout, consensus forecasts show a recovery from 2028 and substantial growth thereafter.
You could look to Amazon’s latest earnings call, where they did a really good job of explaining this sort of J-curve dynamic, where the useful life of GPUs—or TPUs—is actually pretty long. You have to build out the shell, the data center, which takes a certain amount of time. You have to buy the chips, but those will have a very useful economic life for a long period of time. It’s been longer than I think any of us expected.
Yeah. And I know some of the hyperscalers have frankly gotten dinged for raising debt for capex, but similar to the model lab dynamic, I think the ones who have blinked and been less aggressive have regretted it. I know on the podcast recently that you did with Gavin, Microsoft came up, but I think this is an issue across the board.
Yeah. The other dynamic that’s been spoken about a lot is pricing on the spot markets for existing GPUs, which is just another signal that any GPU that you can bring online is being priced at an attractive rate, where the hyperscalers are earning an attractive return. Yeah, it’s a key point.
This is one of the things that we talk about all the time: each one of these successive waves just creates a tremendous amount of user or consumer surplus. What’s actually happening right now, at least as far as we can tell, is that consumers and users get a tremendous amount of value out of this. That’s why they’re using it so much.
The vendors who are serving those users are making very good money, and then you go all the way down every level of the stack to the chips, where you have to pay much higher than you did 12, 18, or 24 months ago to get access to them. Yet you could still make very high margins and create a tremendous amount of surplus for the users.
As we talk about this capex for the hyperscalers, we should think of it as someone else’s order books. These big platforms, with historically the largest profits, are pouring their cash back into AI infrastructure. That’s putting pressure on their free cash flow in the short term, as we just talked about, but it’s been a boon for chip orders, power, and construction.
That’s why this broader technology boom has become an industrial boom, and we’ve been spending more of our time looking at businesses serving this industrial boom and everywhere along the data center supply chain. I think it’s interesting because traditionally our world was more about monetizing existing IP—build once and then sell infinitely—but with these businesses, there are many more complexities that the companies need to manage, like financing, managing vendor relationships, predicting capacity, and predicting demand.
5. The new age of atoms
We’ve seen that there are some teams that excel at that and some teams that have struggled or taken a pause. It’s a little bit of a different expertise than we typically spend our time with. This puts us in a new age of atoms. Big numbers on this page: global infrastructure investment needs are estimated at $90 trillion through 2040.
Importantly, though, this goes way beyond AI and data centers. It includes power, water, roads, and transit. We’re seeing this across our portfolio. You can see this at Anderoll with their massive manufacturing facility; the scale of it is 87 football fields. Whimo is aggressively expanding its depots. And, of course, SpaceX, with its $100 billion investment in Louisiana, is betting on the future today.
From our perspective, investing in physical infrastructure isn’t exactly what it was in the past. As Santi said, it means building factories, expanding infrastructure, and creating more skilled jobs across the country at large.
And I mean, similar to what we just said on data centers, I think investing in physical infrastructure is very different from investing in software. Elon coined the term “the factory is the product” many years ago. It’s no surprise that a lot of the great founders that we’ve backed have come out of Tesla or SpaceX, because they learned that executing on the factory ends up being a massive competitive advantage as they scale.
There are many myths that we hear all the time about data centers: They’re draining all of America’s water; rich people don’t want to live near them; and, one of our favorites, my electricity bill is going to go through the roof. It may seem counterintuitive, but a data center can help lower your electricity rate.
A recent U.S. study showed that for every 10% increase in data center capacity, residential rates went down by 40 basis points. You should think about the power grid as a shared, fixed-cost space: poles, wires, and substations. A large, stable customer like a data center can help spread those costs across more units of electricity. Simply put, more demand across a shared system is a positive.
Yeah. I totally agree with what you’re saying. Investment in shared infrastructure can also improve the economics for the households and businesses connected to it.
Yeah. And I don’t think that point gets enough media attention. Dina Powell went on a podcast recently from Meta and talked about the Louisiana site that they did, where they worked with the community to lower electricity costs. So there are real examples of that today. It’s not just lip service.
Totally agreed.
So now we’re going to talk about the trends that we’re seeing at the model and application layer. On the one hand, as we’ve sort of previewed, AI is generating major revenue and savings. On the other hand, adoption is still extremely early. And then, on top of this, the trend is very much that costs are plummeting.
We previewed agents already, but agents are actually economical now for a much wider range of work. And, of course, this is changing even as we speak, right? New innovations like Jev by newer labs like Typesafe are taking this to a greater extreme. You’re starting to see things like 2 orders of magnitude in cost differences really impact the number of use cases.
It’s sort of classic Jevans paradox, which is why it’s very aptly named. All in all, this is an awesome setup for both the model layer and the app layer. These improvements are creating real opportunities for both growth and profitability among software companies.
That’s not to say that everyone will win, but it’s no surprise that AI is attracting venture dollars into a very much expanding range of industries. We’re also seeing more companies reach enormous scale in the private markets.
This is not a new chart, but it is one of my favorite charts: the combined scale of revenue for OpenAI and Anthropic. This is one that we’ve had in every GP off-site, as far as I can remember, for the last few years. We have to keep updating it every month because it gets out of date that quickly.
The point here—you can see it in the numbers—is that OpenAI and Anthropic’s combined annualized revenue has climbed to an extraordinary degree. We love to show this relative to the greatest software companies in history, and you can see on the right that, if you look at the estimates for revenue added for the best software companies ever built, the estimates for net new revenue for the leading labs have well surpassed that.
It’s a really striking combination of not only scale but also speed to get there.
But we’re still pretty early. Relative to other platform shifts, the interesting thing about the AI shift is we’re not only early in terms of the percentage of the population or percentage of users that use AI, but still very early on in the share-of-wallet gains within those users.
It’s a little bit different than, in 2008, when we would model the percentage of people with an iPhone. Here we need to think about the percentage of people who will use AI products and then to what extent they’re using the products.
Yeah, I think that point that we’re still so early is such an important one, and there’s a bunch of metrics in here that I think elucidate that. One is, if you think about live deployments at S&P 500 companies, that’s at 69%. If you’re AI-pilled, as we all are, maybe that’s something close to what we would all expect.
But then, if you move toward quantifiable impact, which is probably a good metric of how far along they are in their deployment, that’s 30%. Now, if you go to the ultimate barometer, which is a metric tracked over time, that’s actually only at 2%.
So we believe that AI is delivering results, but there is a huge amount of room to deepen its use inside the organization and track those results over time. Moving from these individual deployments that you’re seeing to deeply influential, recurring workflows is the next stage, frankly.
I would just say that most of the enterprises that we speak with, anecdotally, their exposure to AI is still mostly with Microsoft Copilot, which just shows you how far they have to go.
Yeah. This gap between what the models can do and how they’re being used is what makes me so excited about the application layer. Sarah, you did this great conversation with Ali Godsy at data bricks with Martin, and he talked about how the AI can know a lot about the world but know very little about your company.
And we're seeing this in our application-layer companies, where they're bridging this gap. One example that has stuck out to me over the last few months is Revolute. They have a really sophisticated engineering team, but they still had to partner with 11 Labs and take ElevenLabs’ best-in-class voice model and harness it to connect securely with customer accounts and banking workflows. So when I, as a customer, call, I can have my problem resolved smoothly, efficiently, and securely. Today’s opportunity is so much around taking these capabilities and harnessing them to build reliable services.
So, in addition to the fact that enterprise adoption is still quite early for AI, the other really remarkable trend that we wanted to put some stats to is that the power users are totally pulling away. You can think of AI spending rising generally across companies, but the power users—as we talked about, the most intensive users—are spending much, much more. We looked at some Yepit data for this, and if you look at median AI vendor spending, the top 1% is roughly 8 times the level of the top 10%.
What’s impressive about that is it’s almost as much as the 2% to 10% combined—basically, the 1%. So you just have these power users that are clearly early adopters.
Yeah, absolutely. Anecdotally, even inside our own portfolio companies—which you could say are pretty much fully AI-native, or AI-pilled—you’re seeing the top users spend anywhere from 7 and a half to $9,000 a month. The median users, again, for the most AI-native companies are probably closer to the cost of a monthly subscription—maybe $200, or you’re stacking 2 subscriptions on top, so $200 to $400. So that’s over 20 times the spend if you think about median versus top users.
Yeah. We talk to our own portfolio companies, and one of the questions that we discuss is how much adoption they have and how they measure that. If you just look at the percentage of money that they’re spending on AI tools compared to headcount, as an example, relatively forward-leaning, large Fortune 500-type companies are probably today at around 1%.
The most forward-leaning AI-built companies in our portfolio can be as high as 10%. All of these are just questions of how early we are into diffusion and how deep that diffusion will go.
Another example we’ve seen within our portfolio: many enterprises are really excited when they buy Cursor or Cognition for the first time, but the reality is that’s just the beginning of the journey. The runway from simply procuring and trying these tools to full adoption is massive.
And, Sarah, you touched on this early, but we’re starting to see quantifiable case studies. We’ve highlighted a couple here of public companies that actually report on the metrics where they’ve seen meaningful improvements with AI. Two that we call out: one on the cost side, for example, Chime has reported that they’ve reduced their cost to serve by over 10% a year for the last 4 years, compounded. We’re talking about almost a 50% reduction in cost to serve, supported by one of our portfolio companies, Decagon. But, in general, there are many initiatives to lower the cost to serve.
On the revenue side, an interesting one that I found was Shopify. They launched their AI Sidekick, and this AI Sidekick helps merchants get up to speed much faster. The percentage of customers that reach 5 orders within 15 days after onboarding has grown 8%. That’s a metric that Shopify tracks: once you reach 5 orders, you’re going to stick around and retain on Shopify. So it’s been a meaningful tailwind to their business as well.
Yeah, absolutely. To your point on the cost to serve coming down, if you think about the advantages that lower cost to serve gives you, you just have more room to compete. You have better pricing, maybe broader service, and you can reinvest in growth. So I think the benefits are reaching not only the customers but also the margins, and then they rotate that back in. I think the other thing that’s been really fascinating to see is that, of course, they vary, so you can’t throw everything into one bucket, but the incumbents have done a pretty nice job monetizing this as well. ServiceNow is a good example: they’ve reported more than $1 billion in AI ACV and actually a 9× increase in agentic deployments. So you’re seeing it across the stack, from incumbents to newer companies.
One of the interesting things that we’ve talked about a lot—and, again, some of our most forward-leaning companies, like Stripe, have discussed with us—is where they’re actually putting their incremental AI investment dollars. Are they putting it toward things like building new products for customers that could drive higher revenue, or are they putting it toward optimization or efficiency gains on the cost side? This is sort of a litmus test for us: where are the founders or CEOs of these companies seeing the most opportunity? The opportunity to drive revenue growth is unbounded to the upside.
The opportunity for cost improvement is different: you could take that and reinvest it, but that’s a latent opportunity that will continue to exist. If you think that the best and highest use of your dollars today is to optimize your cost structure, what does that say about the revenue opportunity for you?
Even within cost optimizations, there are different flavors. One is lowering your cost to serve, which allows you to reinvest and makes you a better business. I think 6 months ago our industry was really focused on rebuilding systems of record internally, and if your engineers are focused on rebuilding a system of record to save a couple thousand or a couple hundred thousand dollars, I expect there to be better uses of those resources. Yeah, totally agree.
So Sarah referenced ServiceNow seeing massive agentic usage. Agents are here. They’re performing tasks. Tasks require multiple steps, which require multiple model calls. That helps explain the 14× growth in agent token usage on OpenRouter. Those steps, though, are becoming more affordable. We’re seeing caching, so the system can reuse background information instead of processing it from scratch each time. I’ve seen Hebia take advantage of this. They’ve seen their financial chat workloads become 10× cheaper to run. The same budget that they had before can now support more work.
Yeah, I think your broader point is just that lower costs make it practical now for an agent to try, check its work, and try again, and open up a ton of tasks where maybe reasoning or tool use would have been prohibitively expensive. I think this is especially important in cases where reliability is perhaps the de facto reason you would or would not use an agent. Again, we talked about type safety previously, but imagine what happens when you go 1 order of magnitude cheaper, 2 orders of magnitude cheaper. I think the number of use cases that open up is actually unimaginable.
Yeah, totally. And with a much greater improvement in latency.
Yeah, absolutely.
Companies are definitely thinking more and more today about how to optimize latency, or performance at large, along with cost. Take Databricks as an example. They’re leveraging routing, choosing the appropriate model for each specific task to improve performance and cost. Their smart router performed better: it solves more problems at 35% lower cost than the strongest individual model. Another approach we’re seeing a lot of right now is fine-tuning. We’ve seen this with Harvey. We’ve seen this with EliseAI. In the case of EliseAI, they fine-tuned a smaller model, so it became much more affordable—60% cheaper—but also had way lower latency. So live use cases from an audio perspective became tenable. These engineering gains are making AI more useful but also more affordable.
If you roll that up and think about it, for an application business, the relevant unit for them is really the cost of getting the customer’s job done. If you can decouple that—you get the customer’s job done, you charge for that, and you can use better routing to make the product more economical—that’s pretty magical. Now you’re not using the most expensive model for each step, but you are getting the job done, and I think we’re going to see more margin improvements in the best app companies.
It’s a nice segue. We’ve been talking mostly about usage inside of the enterprise, but it’s important to talk about consumer, given that’s the place where we’ve had some of the largest outcomes and where we spend a lot of our time. Now, we think it’s still pretty early. There was a recent survey that showed that only just over 2% of U.S. households actually have a paying subscription for AI. Maybe subscription is not going to be the best place to monetize the majority of households, but it has been so far. And the interesting thing about subscription revenue is it has some of the best retention we’ve ever seen in consumer.
Alex likes to talk about smile curves, and it's really, really rare to see a retention profile that not only flatlines but actually smiles as the product improves and people come back to it. But that's what we've seen with AI, which demonstrates the value that the subscription is able to deliver to the person who pays.
By the way, just on subscription, what struck me about these numbers, especially if you look at the bottom-right chart, is how small they are. Think about Amazon Prime: that's over 200 million households. Netflix is a great one.
70 million households, right? I had to look it up. I thought, is that correct? I totally agree that it's just getting started.
Yeah. Our partner Josh Elman does have a definition I like that would be counter to that a little bit: consumer AI is what I use for my daily life, not necessarily what I expense. In his view, he would say it's maybe not super surprising that 97% of households that are using AI aren't yet paying for it. You are right, Santi: I do like smiling retention curves. We do like to see users coming back more and more.
Another common characteristic of large consumer platforms is time spent. If you were to look at Facebook, Instagram, TikTok, and Snap, they all get 30 to 60 minutes per day from their active users. And while the best AI assistants—and there are many up-and-coming ones right now—all aspire to be my go-to application, I hope that they're going to be incredibly persistent, always on, and super proactive, such that they may not be the place where I spend the most time in the future.
Yeah, it's going to make our job harder when we can't look at engagement via outside-in data because the agents are working in the background instead of us just looking at screen time on a social media application. But it is important. I remember when the newest models came out in December, even in our workloads, we would set off a deep-research task or an agent task and then drive up to the city or something—and that's not driving a Tesla.
That's not screen time that would have been tracked in prior generations of consumer products.
6. Amazon blocks Muse, Instacart opens up
Yeah, totally agreed.
It's early, but it is a really new and exciting time in consumer. Muse and Instinct are no doubt the new kids on the block, but they've grown really quickly. Those, along with chat GBT, are taking more and more of my queries away from traditional search. But it's a really dynamic time, and if you are an existing consumer discovery platform or an existing marketplace, you really need to be thinking about your chess moves right now.
There are a couple of questions I'd be asking. First, how much incremental demand can I get from an AI agent? How many more orders can they bring me? Second, how much of my business—how much of my profit pool—comes from owning the customer relationship and discovery?
Last week, there was a lot of news where Amazon said, "No, thank you," to Manus, but Instacart said, "Yes, please." If you think about Amazon, AWS, and Instacart, and look at their advertising revenue, it outpaces all of their operating profit. Advertising revenue and owning the customer relationship are incredibly important.
But for Amazon, how many incremental new orders or, certainly, customers am I going to get from connecting to Muse? Not that many. Whereas if you look at Instacart, online penetration of grocery is still relatively early, so there are a lot more orders to go get.
The optimistic possibility here is that there's going to be a lot more orders coming from these applications. Historically, I've had to click all these different ways through to process an order. If I offload that to an agent, all of a sudden there are no clicks and, hopefully, more GMV. The trip I once wanted to book but didn't gets booked. The order that we all wanted to place for dinner tonight happens. I'm optimistic there's going to be a lot more GMV, and that'll make up for some of this lost advertising revenue.
Yeah, it's sort of this open question: How does it get compensated for? To your point, Amazon has an over-$70 billion advertising business that's extremely high-margin flow-through, and that's totally predicated on the fact that consumers go to the website and click on the ads. If they don't do that anymore, what happens?
Today, Meta and Google famously monetize probably better than any of the internet platforms. They each make, call it, $200-plus per user in the US and the developed world. In the case of Meta, it's sort of an entertainment application. We'll see what happens with that; it's probably a little bit safer.
In the case of Google, it's funny: the whole talk track around Google two years ago, when all of this just happened, was, "Oh my gosh, what's going to happen to Google's search business?" It turns out it's been really resilient. Part of the reason it's been really resilient is that their very high-monetizing ad terms were somewhat safe, because they were things like, "I need to buy insurance," or "Find me a hotel in this city," or things like that that you can very directly monetize, but that AI was not yet capable of taking action for you on your behalf. If that changes, that could be a very different dynamic.
The other thing just to add: I think the narrative last week was a little bit like negative-sum, where it was very negative—these marketplaces are going to be impacted. I do think the positive-sum view of this is, number one, maybe that $70 billion spent on advertising on Amazon just finds a different channel and goes elsewhere. Maybe it's spent directly on the agents through a different form factor, or it just finds better ways to target people.
And, two, I think with the lower friction, we might just see more consumption. People might buy more, and that's just a positive flywheel that drives economic growth. I think we're a little bit locked into this, "Oh, this is bad for profit pools," but it actually might just be good.
Both could be true, by the way.
Yeah.
Yeah, yeah.
It could be that there are profit pools, but it could be additive overall to the economy. I think we all would be disappointed if we looked back 3 years from now and there wasn't a meaningful productivity improvement at the overall macroeconomic level, which would drive consumer spending higher in the case of advertising. But it might just be that profit pools disappear from certain places and get reallocated.
The immediate reaction has been net-market-cap positive across the ecosystem. The Meta gains have far exceeded any of the losses from the marketplaces.
7. The SaaS bifurcation
Yeah, it's a great point.
Yeah, going back to our discussion on software and moving maybe more to the public side, I think the big trend that jumps out at you from these charts is that the mix has really shifted toward slower-growing, more profitable companies. Roughly 75% of the public software sample here is profitable, and only 30% is growing at 20% or more.
Which is a staggering number: only 30% of public software companies are growing at 20% or more.
And not only 20%—I mean, that's 30%. If you draw the line at 30%, as we've discussed, the absolute number of companies is less than 5, when in the private markets basically every company we see and spend time with is growing at well in excess of 30%.
Yeah. And look, this is somewhat obvious, actually, if you talk to IT managers, CIOs, et cetera. If you look at the growth they're experiencing and what they spend on AI—
The easiest place for those dollars to come from is not spending incremental new dollars on new SaaS software projects.
8. Only 2% of enterprise AI is truly tracked
Yeah, exactly. I think we're investors in some of the SaaS software companies, including in the public markets, and the conversations that we're having with them and the founders are very focused on: How do I take my existing distribution, which in many cases is very, very strong—locked-in customers who, to your point earlier, Santi, wouldn't go anywhere—and apply really interesting new AI products to them where I can drive my revenue growth higher?
I think in order to dispel the greatest fears around the SaaS apocalypse, we just need to see a period of time where the software companies continue to post things like 98% gross dollar retention, which was always the sticking point for why they were so attractive as investments. Continue to drive efficiency, like we've talked about, but most importantly, drive revenue growth acceleration, which shows that they're safe from a defensive standpoint, but that the offensive investments they're making are actually going to drive their businesses to be better.
I thought your point in the blog post that you wrote a couple of months ago on the two paths was really interesting as a rule of thumb. Maybe share more on the finer point of the percentages.
Yeah. The more I talked to founders after it, the more I felt like almost everyone we talked to was like, "Yeah, I'm trying to drive revenue growth higher." And it's clearly, I think, the predominant path for the types of companies that we're investors in and that we've backed over the years.
Again, I think maybe relative to a year ago, when the fear was, “Okay, everyone’s going to vibe-code their software systems,” that is clearly not what’s happening in the market. But the onus is high on delivering revenue growth acceleration, and so we had said, “Let’s target 10% plus acceleration,” which is a high number. But with a magical product, given the budgets that are available to AI, it’s seemingly doable. We’ll see. I think there are a few of them that we are close to where we’ll see that over the next 12 to 18 months.
Yeah, absolutely. And, as with all things, you can’t lump every company into the software bucket. Talk about multiples coming down and growth declining as well. In fact, software has been very differentiated. If you look at this chart here, you can see that cybersecurity and observability have really stood out, and vertical software as well has generally held up much better than the horizontal applications.
I think the framework for how we think about what’s held up, or even gone up nicely, versus what’s in more secular decline, if you will, is that you really think about how AI changes the customer’s need for that product. And I think the cybersecurity risk associated with AI has been well publicized. But as you think about more software and agents creating new security and monitoring needs, this is creating greater demand, honestly, for the incumbents in these markets. You can see the callout on CrowdStrike on the right-hand side.
And then, of course, on the flip side, applications are facing different degrees of workflow change. If you think about it, one of our top CEOs, Ali Ghodsi, likes to talk about this chopping block of AI: What’s first on the chopping block? On the flip side, what is actually needed for AI? You pair those dynamics together, and I think it explains a lot of what we’re seeing in the public markets.
Yeah, I think a lot of this is pretty intuitive with what we’re seeing on the private-market side. So, as you alluded to, Sarah, on the security front, agents are accessing more and more systems and taking more actions. The OpenAI–Hugging Face incident was an eye-opening event for many, so security is paramount.
As we look at the vertical AI companies, these are some of the fastest-growing companies we’re seeing in the private markets. Harvey, Abridge, and EliseAI are growing faster than any of the precedents ever have in their industries, and that’s a reflection of customers recognizing it’s not just the model, which we talked about earlier, but how you orchestrate that and how you build the application around it. Vertical-specific workflows command these needs.
And I mean, we spend a lot of time looking at public markets, less so for investments, but we underwrite exit multiples, so we track that pretty closely. If we were to recap the year, what’s happened? In the first couple months, there was the SaaS apocalypse. We wrote a blog post about what kind of businesses would do well.
If you fast-forward to today, the software index is actually back to where it started the year, but it’s really bifurcated into some companies that are deemed AI losers and some companies that are deemed winners. The public market sometimes simplifies things a bit too much, but to your point, the AI winners are not just the ones that can improve their cost structures. They’re the ones that are actually accelerating and capturing some of these net-new dollars up for grabs. If you’re not really capturing them, then you’re probably not riding the wave.
Yep.
Yeah. We wanted to showcase some private-company operating data that gives us a view of what customers are actually doing. There’s a little bit of a narrative violation in Stripe’s SaaS customer data, which actually shows growth accelerating into 2026 across both young and mature businesses.
Stripe calls it “the renaissance.” Oh, the renaissance. I actually quite like the renaissance.
I like “renaissance.” Yeah.
It’s very good. It’s very good.
So, if you talk—we’ve been talking about public markets a bit. Sarah talked about private-market SaaS acceleration. We get a lot of questions around companies staying private for longer.
If you actually look at the top 6 companies today—Anthropic, OpenAI, Data Brick, Stripe, Wayland, and Revolute—by last-round valuation, they add up to about 2.4 trillion. This is more than the combined market cap of IPOs we’ve seen in the last 10 years, excluding SpaceX, which adds up to 1.7 trillion.
Just the amount of activity that happens within our market—again, 6 companies, 2.4 trillion—is almost as big as the Russell 2000, which is a big index in the public markets where there are hundreds of public managers that spend most of their time. So it’s increasingly exciting to spend time in these late-stage champions that can keep investing in growth, more so than maybe maximizing short-term profit.
But David, you can talk a little bit about how we advise companies on the IPO and when to stay private versus public.
Yeah, look, I mean, the IPO—especially for founder-led companies—one of the things that we’ve talked about, and that I’ve written about, is that the founder is the asset class at this point. And so the bet that we make, and part of the reason that it’s been a benefit for some of these companies to remain private, is they can many times take bigger swings in the private markets that have longer-duration paybacks.
Zuck and Elon are obvious exceptions to the rule in the public markets. But the way you see it manifest in the numbers in companies like Databricks or Stripe is massive, massive new bets in new product areas, and you can see revenue acceleration that happens as a result. You can do that in the public markets, but it’s going to catch greater scrutiny.
Obviously, Meta’s stock price, at the nadir, got below $100 a share when people were very skeptical about their investments in AR/VR. For us, the way we talk about it with our founders is just: the IPO is another financing event. What do you get out of being a public company relative to what do you get out of being a private company?
But clearly, you can reach big scale in the private markets. There are some other companies that have talked about running themselves like they were a public company, with a high focus on efficiency and basically every metric being tracked. There are also some companies that benefit from public disclosure of financials and earning customer trust, both at the enterprise level and on a consumer level. I mean, Navan went public, and we’ve seen a reacceleration just on the basis of some of the larger enterprises actually trusting them more because they’re a public business.
Yeah. And then, of course, there are examples where capital needs over time could be very, very large, and so the capital pools in the private markets are very large and can serve the needs of many of these companies. But at some point, the capital needs may even get too big for the private markets.
That’s the public markets. In the private markets, one of the things that we spend a lot of time in, too, is secondaries. So there are 2 dynamics at play. One is that we see an increasing number of companies holding tenders for their employees. If they choose to stay private, they’re allowing employees to get liquidity outside of the public markets.
An interesting data point shown here is that participation in tenders, according to Carta, has only been 58%. This is employees choosing not to take liquidity because they have so much conviction in the performance of their business. It’s the opposite of what I think is sometimes the narrative of employees cashing out. What we’re seeing is actually employees choosing not to cash out because they believe in their company so much.
Imagine San Francisco home prices if that was 100%.
Exactly.
It’d be hard. Exactly. Yeah. The tender offers that happen—we’re obviously participating in these and leading these pretty frequently, so we think they’re a good thing. They serve 2 purposes.
One, for employees and prospective employees, it is sometimes hard to compete with the liquidity appeal of public markets—RSUs that hit your account on a net-of-tax basis every quarter. So it is a weapon of competition for the private markets that want to compete for or retain their employees with the public-market companies.
The other side of it is that we’re advocates of more frequent resetting of your valuation, and so that keeps your private stock price fresh for a number of reasons. One, it’s easier to talk about that with employees and prospective employees. But 2, in the event that you want to do M&A, as many of our companies have done, you have fresh currency that you can use.
Yeah. No, maybe just the other dynamic: I think people talk about the secondary discount quite a bit because that’s a dynamic that we’ve seen over the last 5 years. Maybe coming out of 2021 and through 2022, 2023, and 2024, the median discount to the last-round price in secondary markets was meaningful because there was maybe a round that was priced too high and was out of date.
But actually, what you're seeing today is that when there are transactions in secondary markets, the discount to the last-round price is basically zero. Yeah, which just speaks to these valuations at which companies have raised: they're fresh, and there are always net-new investors willing to pay that same price, which wasn't the case for the last couple of years.
Yeah, great point. So, this is a fun slide that says, “Everything is AI computer.” AI-related companies account for 86% of U.S. VC deal activity in this 2026 snapshot, up from 65% in 2025. So, obviously, AI has become the dominant destination for venture dollars.
Yeah. But beneath that AI label, the opportunity has broadened a lot, right? So, enterprise apps, consumer apps, but also services, semiconductors, power, and defense. A lot of the attention goes to OpenAI, space AI, and Anthropic, but beyond that, our opportunity set has never been deeper and broader.
9. Robotics, autonomy, bio: what's next
Last, we thought it would be fun to talk about some of the areas that we're very excited about right now. Obviously, we touched on some of the consumer work that will now be done by agents. You could call them long-running agents, autonomous agents, or heavy-use consumer agents. If you take the perspective that you could have consumer agents do all the tasks that you wouldn't want to do, and give you things that you otherwise wouldn't have spent time or money on, it's very appealing, and I think that the distribution of this could happen pretty quickly.
Robotics is an area where we're spending a lot of time and attention. We happen to think that it could be even larger than LLMs, but probably 3 to 5 years earlier than LLMs. So, we think over the next 5 years this is going to be a massive area of investment and excitement.
Autonomy is here. Self-driving works. It's a very exciting time. If you just take a step back and talk about the auto industry and transportation industry, this is one of the biggest industries in the world that we probably don't talk about as much because we spend so much time talking about AI right now.
If you look at miles traveled by Ubers and Lyfts, it represents about 1% of miles traveled in the U.S. I think with full networks of autonomous-driving cars that are 14 times safer than human drivers, we expect that to expand by at least an order of magnitude in the coming years. Plus, there are 17 million new cars sold per year in the U.S. And I think over the next 10 years, those will all be autonomous.
AI × bio is a super exciting area. Obviously, the folks in the labs are talking about this. But new drug discovery, solving some of the most debilitating illnesses or diseases in the world, is a promise that we all are very hopeful we'll see a lot of progress on over the next 10 years.
Personal health is another one that I'm very excited about for myself, as a health maxer. But there's not been a great place or avenue to take all of the information about yourself, put it somewhere, and get very hyperpersonalized advice.
Lastly, diffusion into the enterprise beyond coding. This is one of the areas that, Sarah, you were talking about. We're so early in the actual diffusion of the technology into the enterprise that I think it's going to be super, super exciting.
And lastly, there's another area that is not inside the AI bucket, which is what we call American Dynamism: the retooling of the entire American Dynamism stack. We've been large investors in this area for a while. It's still just a small fraction—less than 5%—of overall dollars spent in the military on newer vendors like Andre, Seronic, or Castellian, and we expect that to grow dramatically as needs change.
So many areas of excitement. Obviously, we've been very active in this AI space, and we're optimistic about the effect that it's going to have on the overall economy. The buildout is massive, but we think it's going to be massively productivity-enhancing in the U.S. So, it's a blast to hang out with you guys. Thank you.