As artificial intelligence reshapes advertising, much of the industry’s attention has focused on AI agents, creative automation and media buying.
PubMatic believes the bigger opportunity is to rethink the infrastructure that powers advertising—and where intelligence should live.
In a conversation with tipsheet, Co-Founder and CEO Rajeev Goel discusses the company’s long-term AI strategy, why intelligence belongs closer to the supply path, and how autonomous agents, protocol-based buying and AI-native infrastructure could reshape the economics of the open internet, potentially redefining the role of today’s sell-side platforms (SSPs).
Topics covered include:
- PubMatic’s long-term vision for AI-native advertising
- What “agentic advertising” means
- Why AI should transform advertising rather than simply automate workflows
- How AgenticOS is performing in early deployments
- Why PubMatic believes intelligence belongs in the supply path
- The role of AdCP, Decision Fabric and containers in an agentic ecosystem
- When protocol-based media buying reaches scale
- Why the sell side gains strategic advantages in an AI-native world
- How PubMatic is evolving beyond its traditional SSP roots
- The company’s outlook for conversational advertising
- AI economics and token costs
- PubMatic’s future positioning
Scroll down for the interview, which has been lightly edited for clarity.
AI product vision
tipsheet: Over the past year, PubMatic has made a series of announcements that seem to tell a larger story than simply introducing individual AI products. Was there a long-term roadmap behind those announcements, or did the strategy evolve alongside the technology?
Rajeev Goel: It’s a bit of both.
Like many companies, we began using generative AI internally with our engineering team. That’s not surprising because engineers tend to be the earliest adopters. We reached a very high degree of proficiency, and earlier this year we shared that more than 85% of our software is now written by AI.
At the same time, PubMatic has been using machine learning for 15 to 20 years. Real-time bidding was built on machine learning, so AI has always been part of our business. What we’re talking about now is the next phase.
As we saw the results in engineering, we started asking how AI could improve the way customers interact with our platform.
Unlike many companies, we looked beyond simply making existing workflows faster. We asked two broader questions.
First, how can AI help reimagine the advertising supply chain and eliminate much of the complexity that exists today? Second, how can AI unlock performance advertising for the open internet?
I believe the open internet will ultimately survive as a performance medium. The industry has been trying to solve that challenge for years, myself included, and I think AI is the ultimate unlock.
That’s the real power of AI. It isn’t incremental productivity. It’s the ability to transform the industry. That vision is what led us to prioritize generative AI and why you’ve seen us move so aggressively in this area.
Agentic media buying
tipsheet: Everyone is using the word “agentic” today, but PubMatic has been talking about it for some time. How do you define agentic advertising?
Rajeev Goel: To me, agentic means autonomous AI systems executing the advertising value chain. There are three important elements.
First, it has to use AI.
Second, it has to be autonomous. You give the system an objective or desired outcome, but it isn’t simply following a script. It reaches decision points where it determines whether to do A or B on its own.
Third, it has to be applied to advertising. Whether it’s discovery, planning, activation, measurement or optimization, AI is executing those functions across the value chain.
tipsheet: Much of the industry’s excitement around agents centers on autonomous media buying and selling, where AI systems interact directly with one another. PubMatic has been one of the earliest companies pursuing that vision. How is it progressing?
Rajeev Goel: It’s progressing very well. Before I answer that, let me add one point because I think it’s useful to contrast agentic AI with another term that’s become popular.
A lot of people are talking about orchestration. You’ve written about it, and others have as well. To me, orchestration is simply a fancy word for using AI to improve existing workflows. I think that’s lazy and unimaginative.
It’s like standing at the beginning of the Industrial Revolution and saying we should train horses to run faster so they can compete with steam engines and automobiles.
AI is going to enable an entirely different reality. We need a vision for what that reality looks like and then build toward it as we learn.
Regarding agentic media buying today, we’re seeing a great deal of learning and experimentation. Many companies are running pilots with us, and a growing number are now beginning to scale after seeing the results.
(Goel pointed to recent AgenticOS case studies with Level Agency and Havas.)
More broadly, there’s tremendous curiosity across the ecosystem. Everyone is trying to discover new ways of working together, remove unnecessary complexity and allow employees to focus on higher-value work.
Those motivations are driving rapid AI adoption across advertising.
AI changes the supply chain
tipsheet: How do you balance that agentic future with PubMatic’s core SSP business?
Rajeev Goel: Our business has evolved considerably. This year marks nearly 20 years since PubMatic was founded. We’re well known as a sell-side platform, but that’s only one of several products we offer today.
We have Activate, which enables direct buying within the SSP. We have AgenticOS, which simplifies advertising for both buyers and publishers. We have Connect, our data platform with more than 300 data partners. We also have an intelligence layer that combines our own models and data with those of our partners.
With AgenticOS, we believe we can do much more at the point where consumer attention, publisher content and consumer consent all come together.
That’s an incredibly powerful place to make AI-driven decisions, which is why we’ve been so aggressive in opening our platform to advertisers, agencies and DSP partners.
tipsheet: It still sounds like an SSP strategy, but one increasingly powered by autonomous agents.
Rajeev Goel: Yes. Let me give you an example.
As buyers begin executing campaigns agentically, they quickly move beyond the traditional assumption that every transaction requires both a DSP and an SSP.
Once AI becomes autonomous, the advertiser no longer needs to worry about which technical capabilities are being used. The AI determines the optimal way to achieve the desired outcome.
That’s when you begin to realize how much of today’s industry is locked into existing user interfaces.
Trading teams and ad operations teams become comfortable with a particular interface, and that creates a kind of invisible lock-in around incumbent platforms.
Once AI removes that dependency, the interface itself becomes far less important.
Maybe the buttons are arranged differently. Maybe there are no buttons at all because the AI is executing everything.
Once you break free from that UI lock-in, an entirely new set of possibilities opens up for how advertising gets executed.
Agentic advertising protocols
tipsheet: Thinking about protocols, we’ve spoken before about Ad Context Protocol (AdCP). What momentum are you seeing, and how is the governance effort progressing?
Rajeev Goel: It’s progressing very well and continuing to move at a rapid pace. We’re a co-founder and launch partner of AdCP, and we’ve been deeply involved at the board level.
One of the things AdCP and AgenticAdvertising.org do well is remain tightly focused and move with a high degree of speed and agility.
People often ask how that relates to organizations like the IAB Tech Lab, but I don’t think that’s the right comparison. There are multiple standards bodies serving different purposes.
The Tech Lab is the backbone for how hundreds of billions of dollars move through the programmatic ecosystem today. Given that responsibility, you would expect it to move differently because so much infrastructure and commerce depends on those standards.
AdCP is solving a different problem. It makes sense for it to move faster and experiment more aggressively.
For the foreseeable future, I think the industry benefits from having different governance models operating at different speeds. It’s been exciting to watch AdCP continue bringing new capabilities to market.
tipsheet: How do you think about the relationship between AdCP and the Agentic Real Time Framework (ARTF)?
Rajeev Goel: I think they’re different use cases.
To that end, we launched Decision Fabric in June because we wanted to open up many of the capabilities we’ve been building with Activate and AgenticOS to the broader ecosystem.
With Decision Fabric, partners can run their own containers inside our infrastructure. They can bring their own intelligence and models while also tapping into PubMatic’s intelligence.
I think that has the potential to fundamentally change the traditional buy-side versus sell-side divide. We’re well positioned for that future.
When I think about what a DSP could look like several years from now, it may be very different from today’s model.
Instead of one general-purpose DSP, you might have specialized DSPs optimized for specific objectives. One might excel at driving store visits. Another could specialize in financial services. Another could optimize for consumer packaged goods.
An advertiser could use a portfolio of specialized DSPs, all operating within Decision Fabric.
PubMatic provides the infrastructure. We provide access to more than 300 data partners, our own AI models, publisher inventory and AI agents. Partners bring their own models into that environment.
It becomes a much lower-cost and lower-overhead way for DSPs to operate, and that’s why we’re seeing significant interest from newer DSPs today.
Protocol timeline
tipsheet: When do you think protocol-based buying reaches meaningful scale? Whether it’s AdCP or ARTF, when does this become mainstream?
Rajeev Goel: I’ve made a prediction that by the end of 2028, about 25% of media buying in our industry will be agentic.
By the end of 2030, I believe that number reaches 50%.
That’s specifically around agentic buying aligned with AdCP.
On the container side, I don’t yet have the same level of visibility. We’ve launched our initial partners including inPowered AI, MiQ, Chalice AI and SWYM[dot]AI. We’re still working through some of the foundational scaling challenges with those partners.
Today, it isn’t yet a completely repeatable motion. On the agentic side, however, we already see that repeatability. I think containers will reach that point fairly soon, and once they do, I expect adoption to accelerate quickly.
The walled gardens
tipsheet: If the walled gardens build powerful agentic infrastructure of their own, doesn’t that create significant competitive pressure? How do you build a durable moat when companies like Google and Meta are investing so heavily?
Rajeev Goel: I don’t think those companies are primarily focused on monetizing the open internet.
They’re focused on monetizing the media they own.
If you look at Google’s financial results over the past several years, Search continues to grow, YouTube continues to grow, while the Network business, which represents the open internet, has declined.
The open internet itself isn’t declining.
To me, that suggests a strategic shift toward owned media rather than the open internet. I think that creates opportunity for companies like PubMatic.
Over the last five years we’ve been building several important advantages.
- First, we work with more than 2,000 premium publishers representing roughly the top 100,000 websites, mobile apps and streaming services across the open internet.
- Second, we have Activate, which enables direct buying inside the SSP.
- Third, we have AgenticOS, with more than 20 AI agents and additional capabilities on the way.
- Fourth, we have our intelligence layer, combining our own models with more than 300 data partners.
- Finally, we own and operate our infrastructure. Through partnerships like NVIDIA, we’ve invested in the GPUs, CPUs and infrastructure required to build AI at scale.
Taken together, I don’t think anyone else has that same combination of capabilities operating globally.
Beyond that, I think the sell side is the natural place for AI decision-making for three reasons.
First, we see all available inventory. We’re not constrained by traffic shaping or query-per-second limitations that prevent DSPs from seeing the entire opportunity. Second, we have richer, consented data because it hasn’t yet been filtered or packaged before reaching a buying platform. Third, we simply have more time to make decisions.
As the open internet shifts toward performance advertising, advertisers need much richer AI models than traditional brand campaigns required.
Those richer models need additional time to execute. By running inside containers on the sell side, we have more time to execute sophisticated models than exists within the traditional DSP-SSP workflow.
Changes with OpenWrap
tipsheet: Last week PubMatic announced its partnership with Playwire, making Playwire the recommended monetization partner for publishers transitioning from OpenWrap Web. Why?
Rajeev Goel: OpenWrap is our wrapper solution for header bidding, and it exists in a few different forms.
OpenWrap Web serves browser-based monetization, while OpenWrap SDK supports mobile apps and connected TV.
As we evaluated where we wanted to invest our engineering and product resources, it became clear that our greatest opportunities are in mobile apps and CTV.
Playwire is a terrific partner and is well positioned to continue innovating in browser monetization. That allows us to focus our resources where we see the fastest industry growth.
Mobile apps and connected TV are growing much faster, so that’s where we want to continue investing our engineering, product and go-to-market efforts.
This was really a strategic decision about focus and partnering with a company that can continue serving publishers well in that environment.
tipsheet: Is that the same thinking behind Activate? A broader effort to deepen relationships with buyers?
Rajeev Goel: Going deeper with buyers has absolutely been part of our strategy. In many ways, that effort began years ago with supply path optimization. We’ve continued expanding that vision ever since.
The question we’ve consistently asked is how we can innovate for buyers so they achieve better performance. When buyers perform better, publishers generate greater yield. Those two objectives reinforce one another.
New advertising surfaces
tipsheet: Consumers are spending more time interacting with AI assistants and chatbots. Is conversational advertising an opportunity for PubMatic?
Rajeev Goel: Absolutely. We think about it in two broad categories: commerce environments and more general-purpose AI environments.
Consumers are increasingly shopping, researching and discovering products through these experiences. That creates opportunities to deliver highly contextual advertising based on what consumers are trying to accomplish at that moment. Of course, it’s critically important to respect privacy and consumer expectations.
But it’s clear that these environments are becoming an increasingly important part of how consumers learn, discover and make purchasing decisions. You’ll hear much more from us on this topic in the future because we believe it’s a significant opportunity.
Managing AI economics
tipsheet: AI has introduced a new cost structure for every technology company. How do you think about token costs and AI infrastructure costs inside PubMatic?
Rajeev Goel: We really think about two categories of cost.
One is token costs. The other is the capital expenditure required to process those workloads.
There’s been a lot of discussion over the past several months about enterprises trying to optimize token costs. My reaction is simple: companies optimize every expense they have.
We look at travel costs, office expenses. We look at every line item. Of course we’re going to optimize token costs as well. What matters is whether those investments increase the productivity of our teams.
If AI enables our employees to build better products and deliver more value to customers, then there is a clear return on that investment.
The infrastructure side is different because we own and operate our own systems. That means we have to think carefully about GPUs, CPUs and the capital required to support AI workloads. Whether you own that infrastructure yourself or consume it through a cloud provider, compute has become a meaningful input cost.
Our responsibility is to make sure we’re extracting as much value as possible from every dollar we invest in compute.
In five years
tipsheet: Five years from now, do you think people will still describe PubMatic as a sell-side platform?
Rajeev Goel: I don’t think so. In fact, I think it’s probably a mistake to describe us that way today. Being a sell-side platform is certainly our history and our foundation, but today it’s only one part of what we do.
We also have Activate, AgenticOS, Connect and Commerce Media Solutions. Some of those products have been in market for several years.
Our focus is becoming an AI-native platform that delivers performance advertising. AI is built into how we operate internally. It’s built into how customers interact with our platform. And it’s built into the products we deliver.
Whether people ultimately call that an SSP or something else isn’t especially important to us.
What’s important is building an AI-native platform that helps advertisers and publishers achieve better performance.