All right. Fantastic. I will go ahead and kick us off here on stage. Thomas Seifert, CFO of Cloudflare. Really grateful to you joining us multiple years in a row. Thank you
Yep
for taking the time.
I'm aging here.
Thomas, I love that I get to ask you this question because you're going to give me the most conservative possible response.
The most thoughtful.
Most thoughtful, most balanced. We have a debate with investors on this concept of the inference economy.
Yeah.
Which is the idea that, look, we're so early in agents and agentic capabilities that the best part of whatever product cycle is tied to agentic inference is still ahead of us.
Yep.
Now, you're already putting up 36% growth at the very beginning of this product cycle. If we were to extrapolate out and say, "Look, the best part of Cloudflare's growth opportunity tied to Workers, tied to SASE, tied to Act 4 is still ahead of you," that would suggest directionally that the best part of your growth story is also still ahead of you. Talk to us a little bit about the framing there. Why would that not be true, or what would we need to see to feel even better that it is going to be true?
One of our founders, Michelle, we make fun of her, because every year when we have kickoff for the year and talk about the targets for what is to come and how we approach the year, she says we're only getting started. Now this is our what, eighth year as a public company. So the beginning of the year, she said it the eighth time. I truly think that what you just said is true. When we think about our life in acts, Act 1, Act 2, Act 3, Workers AI, we'll talk about it, Act 4, content protection, how we monetize it, we think that that opportunity is going to dwarf what is already behind us. Last year, we were sitting here, and agentic AI was still far away.
The step-up of the LLMs only came in November, and we were primarily concerned about convincing you all that our sales execution was on track, that productivity was coming, that we felt comfortable investing in sales capacity, and that was driving the early steps of acceleration and growth in a quite significant way. Now I think we are what you described in this paradigm shift of what an agentic AI world is going to do to us. We find ourselves at a point where the network, the architecture, the biggest competitive moat of Cloudflare, the network itself, seem to be designed for what is going to come, and support what agentic AI workflows are looking for. Super- low latency, secure sandboxes to execute code, and so on. What you look for are the signs that signal the early innings.
Because of the architecture of the network and the benefits, developers are flocking to Cloudflare. That is an early sign that I think is signaling this dramatic shift. Developer count was up 2 million developers quarter-over-quarter alone. You see that the acceleration of revenue is now happening in what we call variable revenue. So consumption-based revenue. Another big shift, I think, in the agentic world is that we move and disrupt SaaS from a seat-based metric into highly variable consumption-driven revenue. That is where we see acceleration. We're at a point now where pretty much all the large names of AI-native companies, whether it's Lovable, Figma, Anthropic, OpenAI, are building on Workers on top of the network. One of the stats we gave at Investor Day was that we think about 80% of the significant AI enterprise infrastructure is a customer of Cloudflare.
All the signs you want to see. There's one trend that I think is signaling what is going to come when we talk about Act 4. There was this vertical that we always were highly underrepresented. It was content creators. Media was always a vertical that was revenue weak at Cloudflare because there was not a lot of value for us just moving bytes from one place to another. They are now shifting towards Cloudflare to start to protect and make their content that they create scarce in order to generate leverage. I think there are so many structural signs that signal what is coming, and that is, to be very honest, what gave us the confidence at Investor Day to raise our North Star. We said for many years now 40% was our North Star. We shifted that to 50%.
We'll exit this year with significant momentum in that direction. I think there's clearly a world, and that is what we also said, beyond 2027, where we will live in a world beyond the Rule of 50, and that will be the result of acceleration Act 3 and Act 4.
Let me pick up on the network architecture comment that you just made. Many investors in the room will spend time with the hyperscalers, and the hyperscalers will invest hundreds of billions
Yep
in their CapEx guidance to be able to support training clusters, but also inference clusters, and there's some debate around mix between training and inference that we'll leave aside for now.
Yeah.
The interesting thing about Cloudflare is you are not investing hundreds of billions of dollars, and yet you have this product cycle and compute tied to inference at the edge. Maybe square that circle for us.
Well, a big part I think is already what you mentioned. When the CapEx wave started, we said training is not our world, at least not at this point in time, not for the super large models. Things might change as models get smaller, and maybe run even more on-premise on enterprises, on open models. Our world is inference. Inference is an interesting world, right? When we talk about AI, it's in the larger standard GPU-driven discussion. There was a very interesting part at Investor Day in Rita's presentation on how important CPU capacity is in order to orchestrate agentic AI workflows. Where does the data come from? Instruction, what needs to get computed, what API needs to get access. That is our world. That is what Cloudflare's network is today already.
A reason why our gross margins used to be so high, and still are high, is because we really understand well how workloads get scheduled. That allows us really to take that strength into this AI world of inference, where we think GPU is important, but the demand curve is really bending when it comes to CPU and memory capacity, and that's where we excel.
In this world of inference, I think an important advantage for us is that we do not have to provision ahead of demand. Two years ago when we started work on Workers AI, there were so many questions. When will we see revenue? What is the pricing? We always said it's not what's important. What is important is to get as many developers on the platform. We want to understand the variety of inference workloads because that is what we need to optimize.
We need to understand what hardware stack we have to put in place and learn how to schedule AI workloads in a way that gets us to a utilization that is similar to our Act 1 and Act 2 products. That work is now paying off, right? We have made tremendous progress on how we schedule inference workloads on CPU capacity. It doesn't mean we don't need GPU capacity.
We deploy GPU capacity pretty much in all of our PoPs in all of our data centers. But that is the strength we built from based on the fact that this investment can follow a demand signal, and we do not need to provision ahead of it. That doesn't mean that our CapEx will stay low. CapEx can go up, but it will go up in conjunction with me telling you that revenue is going to come up too.
We are quite comfortable that the CapEx efficiency is going to hold for a while.
You said something interesting there, which is the training opportunity may change as models get smaller. Not quite the training opportunity
Yeah
but there is a scenario where the industry evolves, and if we're training smaller models, maybe Cloudflare has a role in that. Maybe just elaborate a little on what you mean.
I don't want to go too far there and get ahead of my skis, but
Very fair
What I think is that how you optimize. Smaller models and open models will put pricing pressure on tokens.
Absolutely.
How you cope with that from a revenue perspective, in order to get throughput up so you optimize revenue in a different way, I think is going to be really interesting, and I think this is our playground. It will be another vector that will play to our strengths, whether that is inference or early training adoption of open models to enterprise-specific use cases on enterprise-specific data, maybe even on-premise in order to make sure that data leakage is contained.
We are actually hearing more about this idea that enterprises could bring more compute on-premise. Maybe pick up the thread there. What do you think the opportunity is for Cloudflare, whether it is on the sovereign SASE side or more on the on-premise compute side?
Well, we were just debating on centralized large data centers versus a highly decentralized network. That is what we are. We drove data sovereignty and data privacy products from a very different perspective before AI was even a topic. Those products are now the basis of making sure that either critical infrastructure customers, sovereign nation customers, federal business in friendly parts of the world can contain their data privacy and data sovereignty risk. We are built for that. You literally can wrap something around a byte of data that you are a citizen of Zurich, and your credit card can only move between those two cantons, and can only be in a data center of that security level. That is what the network was built for, and that is why we think this will be another growth vector that will play well for us.
Absolutely. Okay. I want to ask you about Act 4. If I think about some of the products that you've announced around identity payments gateway, one of the comments that Matthew's made is, look, one of the reasons this has become important to Cloudflare is because the existing payments networks don't give you the throughput that
Yeah
you would need to support agentic transactions
Yeah
at scale. I would otherwise have been skeptical about Cloudflare's ability to build all of these pieces of infrastructure, but for the fact that you've done bold things with your architecture before, and you have Stephanie Cohen driving some of this for you. Talk to us a little bit about what the building blocks are to be able to actually realize the vision on agentic payments, agentic transactions, pay per crawl, Act 4.
Let's start with pay per crawl. This is how it all started when we saw that crawl rates and links were completely diverging, and content creators lost an opportunity to monetize the content that they created. These were the early innings of Act 4, and we started with established product like bot mitigation of our tool in order to give content creators a tool to make their content scarce and put them in a position to negotiate for people paying for crawling their content. This is how it all got started. There are two sides of the situation. Meanwhile, we are working on both sides.
You probably saw our research announcement with OpenAI to index what we have. Today, LLMs crawl literally every page that is important every minute because they do not know what changed, and you do not want to miss that an important football player dropped out of the game two minutes before it starts. This costs significant money, bandwidth that ultimately is paid by their content creators. More than 20% of the web is behind us. We can signal what changes. We know what changes. We know what has refreshed, and that is valuable knowledge for the LLMs that they are ready to pay for because it contains the complexity of what they need to do. It reduces the cost, and it makes a friendlier ecosystem. Significant progress. Now we talk about how we monetize transactions.
When 20% of the web is or more is behind on our network, when all these agentic AI agents and workflows go through our network, we are in the flow of what commercial transactions are going to come, and you want to be a part of it. We announced a couple of products that are first step in that direction. We announced what we call Monetization Gateway, so content providers can literally sell what they have. We introduced Cloudflare Wallets, which allows you and your agents to have autonomous transactions. We announced identity tools, and there is more to come. The big lift, to be honest, is still in front of us because Matthew, I think, gave some stats on the last earnings call we had of how many 10,000 transactions per second a significant payment provider today is able to handle.
That is literally in the 10,000s of transactions. We think we need to be in the millions or tens of millions of transactions in small value, high volume on our network. We are quite excited about that. There is significant effort going in that direction together with partners. Literally, every credit card, every large fintech company is part of this effort. We think that is why it all accumulates to Act 4 being probably the biggest opportunity that is still in front of us, and it is going to dwarf whatever we have done so far.
Say more about the heavy lifting to realize the vision. Do you already have the ability to do the throughput at the hundreds of thousands and millions?
We are not at the millions yet.
Okay.
This is work in progress. There's significant engineering effort going in that direction, yeah.
What about from a partnership and an ecosystem standpoint?
All the partnerships are announced. There's Visa, there's Mastercard, there's American Express, there's PayPal, there's Coinbase. It's a large set of partners that need to come together, but we'll provide the infrastructure in order to enable that. Yeah.
When is the right time for us to get excited about seeing it in the numbers?
Well, remember, it's about building a durable business model, right? We pace our time in light of the opportunity. We cannot take this lightly. Processing millions of transactions per second will be a responsibility and obligation. We want to make sure that the tech is ready for it. There's still so much earlier stuff that needs to happen that gets us excited when we talk about Act 3. I think that acceleration is coming, and we're in a good track.
This is a little bit of a framework monetization question for you. When I think payments, I think 275 basis points as a take rate.
Yeah.
This is a little bit interesting because I guess you could do it as a take rate model. You could do it as a per transaction model. What are some of the things going through your head?
This is where you pay people to give you good advice. We are working with specialists in this arena, even down to academics, figuring out what the best model is. I think there's no one size fits all, so there will be a differentiated approach. I think we, like so many things when we start, we look at it from a first- principles perspective, and we really need to start from scratch in order to come up with a model that is fit for purpose.
Very fair. Okay. Let me ask you a little bit about the growing piece of your business that is tied to pool of funds and
Yep
consumption revenue. One of the comments that you made this quarter was customers consumed more intra-quarter, which is where some of the upside came from relative
Yes
to the guide. How do you think about, there's a lot of ease of use in pool of funds that reduces friction. At the same time, if more of your revenue is coming from consumption revenue, it creates some variability from quarter to quarter.
Yes.
How do you adopt that into your forecasting?
Well, it's a really good topic. Literally, a large cross-functional team is meeting tomorrow for two days to really rethink most of our processes from scratch. Forecasting is one. We have to tie way deeper into the infrastructure to understand what drives consumption. A SaaS model was a demand-limited model, right? You have the time, you figure out how much you get. In this world, we are not demand-restricted, it seems, right? It's what product and what supply do we have at the right point in time, how you forecast that, what are the early indicators. You have to go far below the billing metrics that indicate or could signal consumption. In a SaaS model, you can ask your customer and say, "How many people do you intend to hire next quarter?
How many more seats are you going to occupy on this SaaS product?" Our customers don't know where demand is going. You have to really rethink what indicators you take. For this year at least, we talked openly at our earnings call about it. In this world, guidance becomes more difficult. If you want to keep the range, you have to shift the midpoint, right?
Yeah.
Because we want to surprise you all to the upside and not to the downside. This approach only goes so far, right? Because the numbers become so big and the growth rates become so steep. We have to figure that out. But we have an excellent data science and finance team. We'll get this under control.
What do you think was unique about 2Q? It sounded like something changed
Yeah
in 2Q in terms of the way customers were consuming.
Yeah.
Where they started consuming more relative to whatever your baseline forecast was.
Yeah.
What do you think was unique about 2Q, and why wouldn't the same thing happen in 3Q?
That it goes faster, or?
That customers consume more than your baseline forecast.
That would be a good thing, right? I am not going to exclude that.
Sure. Why do you think it was 2Q specifically?
If the upside comes from variable consumption, it is more agentic AI traffic. It is more requests. It is faster adoption of products across pool of funds. If it is a pool of funds contract, people burn through their commitments way faster. We renew mid-quarter. That is the dynamic around that. You asked for structural signs of us participating. I think this is one of the structural signs where we see that this flywheel is starting to spin, and it is probably going to spin faster and not slower moving forward.
Is there more heavy lifting to do on the go-to-market for enterprise?
Heavier lifting, I don't think is the right word, but we see that serving agentic AI customers is different. Let me put it this way. In terms of the personas, in terms of the sales cycles, they are way shorter. The deal sizes are way bigger, but it needs a significant support in terms of forward-deployed engineers that can nerd out on a super technical level. I don't think this is a major lift, but it is an evolution of the go-to-market model, but it caters to us, right? Having forward-deployed engineers, people that are able to talk technology with the customers, that has been our strength. Yes, that evolution is happening, and it is happening pretty fast.
Let me ask you a little bit about the core businesses with CDN and SASE.
Yeah.
One of the hypotheses that we're pulling at is if agentic traffic goes up, core CDN has historically been. Or network security and CDN inversely, forward proxy, reverse proxy, fundamentally have been a function of network traffic. I guess the question for you is, have you started to see a change in network activity? You have the Cloudflare Radar where you show percentage of agentic traffic versus human traffic, but it's more a question on, does it create an operate cycle in CDN? There's so much more to talk about on the security side, too. Is there going to be an underlying operate cycle in the core Act 1, Act 2 businesses?
Yes. But for more differentiated reasons. I think we see an uptake for sure in Act 1 because verticals that have not been on the network are moving to the network, right? Share in the media vertical and publisher vertical has gone up quite a bit. For sure about this. I think you'll see an uptick in SASE-driven products. All enterprises start to deploy AI like there's no tomorrow, and then we wake up and figure out, we need to control what data they access, what data shouldn't leak. There's a whole slate of products that comes with this. That's what I meant. An agentic AI request might be lower in gross margin, high growth rate, but it comes with a significant opportunity to wrap other value-added products and services around it.
We see that in Act 2, so there's an acceleration from the baseline, not to the level of Act 3, but you can clearly see that.
Maybe stay on this point. Why would agentic traffic be lower gross margin? What's the attach rate on the product side that's higher?
Well, we are quite good at making progress on how we schedule, but the utilization of the throughput on a given CPU or GPU in terms of how many tokens you can generate within a given set of compute capacity is different, right? We are getting that up, so gross margin is improving, but that is the reason why agentic AI work gross margin is lower. It comes with really low cost to book. We make it up on the operating margin side. The workload needs, it needs to be secure. It needs sandboxes, AI Gateway products that all come along with it that make for an upsell or a cross-sell opportunity. It all happens on a platform. If we are lucky, the customer is on a pool of funds contract, so the friction to adopt is really, really low.
That's why every request. I remember I always say that when I teach an incoming class at Cloudflare, I said one of the things that really stood out in my interview many, many years ago with Matthew was, and there was no Act 2, there was no Act 3. Thomas, when we think about pricing and customers as the most important aim is to never discourage a byte of data from moving through the network. This stood for me, and this drives a lot of the decisions, how we price, how we discount, how we think about product is never discourage a byte of data from moving through the network because we will find opportunity to wrap value around that byte of data.
In a world of agentic AI, where the request, the AI, agentic AI or agent-initiated requests are going through the roof are multiples of human-initiated requests. That opportunity, of course, multiplies, too.
Let me ask you a question on competition. At various points through Cloudflare's history, I think back to the early CDN days where Akamai was part of the conversation. Then I think through some of the wins that you've announced recently on the security side, where Zscaler had been the dominant SASE player, all the way through Act 3, where you've won hyperscaler displacements more recently. Maybe just take a step back from all of this. How do you see competition evolving in the spaces you compete holistically rather than us creating a map of saying, "Well, here are each of the products, and here's who competes in each lane"?
We have formidable competitors out there, so you have to acknowledge this. I think there is a strength to the business model that is really hard to attack, and that makes it easier, not easy, but easier to defend. There is this architecture of the network that is just so superior, and our principle and primitive to keep every product on every server in every region is just a flywheel in itself, right? We started with five products, I think, when we went public. We are now getting to 70. The products live off each other. The reason why SASE products have high gross margin is because they live off the infrastructure that Act 1 has built. Act 4 will live off the infrastructure that Act 3 has built and therefore has these significant margins.
You are now in a world where you have so many independent growth vectors driving the business, that there are so many ways you can go in order to evolve the business model. I think this makes this very unique, and that is why over the years now, the competitive position and strength has increased, in some areas dramatically increased compared to competition. That flywheel of Act 4 will make that mode even higher.
I want to end here with a conversation that we started with Matthew right after the riff on how do you operate a company in 2026 such that it is different from perhaps the SASE era or the 2015 era? I would love to hear how have things changed reflecting inward on Cloudflare. You have always been an innovative company. I think of Birthday Week, for example. Share with us a little bit about what has changed behind the scenes to the extent that some of that innovation has been amplified.
When we announced our employee reduction in the second quarter, there are many big question marks of why we would do that, how could inefficiencies have built up that much? It took us until Investor Day and people got a glimpse into what we call Cloudflare OS, product that again, we developed ourselves that we are now starting to sell, and the harness and AI infrastructure we built ourselves for people to understand how we think about productivity and how the company would evolve. We spent with 80 people tomorrow and the day after tomorrow to really rethink every process in the finance stack from an AI-first perspective. We deeply believe in that. There were question marks when Matthew last year, at the end of last year, went out and said, "Everybody stops hiring fresh outs and grads, but we will hire 1,111." That is a product.
That's why the strange number, and that is what we did. I was at the Austin office the other day, and there's 350 interns working on the craziest projects, none biased by any legacy or any process. They never worked in a billing process that works from completely free to reimagine how that process could look like in an AI-first world, or how demand generation could look like in that world. A lot of these interns will get job offers. So there is an approach to this where you just give credit to Matthew and Michelle for rethinking things like that from a very different perspective. There were many our executive team who said, "Really? Now we hire 1,111 interns. How do we do that?" Now we look back and say, "We'll do that next year again." It was exactly the right thing to do. Yeah.
Hey, fantastic. Thomas, thank you for your time. Please join me in thanking Thomas.
Thank you.