Welcome to Partner Perspectives, PubMatic’s series highlighting the industry leaders shaping the future of programmatic advertising. In each blog, we showcase our most innovative partners and customers delivering exceptional results across the industry.
In this edition, we’re featuring Peyman Nilforoush, Co-Founder and CEO of InPowered, an innovator in containerization who has built decisioning products that bring advanced control and efficiency to the supply path—including solutions that work alongside PubMatic’s own containerization solution, Decision Fabric.
Can you explain how containerization enables buyers to make smarter tradeoff decisions across impression types in real-time—e.g., “should I bid aggressively on this impression or wait for a higher-value one”—in ways that weren’t possible before?
The biggest change containerization enables is actually what happens before a bid request ever reaches the DSP. Exchanges see an enormous volume of programmatic opportunities, far more than any individual DSP can ingest because of QPS and infrastructure constraints. As a result, bid requests have to be filtered and throttled before they reach the DSP.
The problem is that without advertiser-specific intelligence at that point in the supply path, the exchange doesn’t necessarily know which impressions are most valuable for a particular advertiser. An impression that looks ordinary based on broad supply signals could actually have a very high probability of driving a conversion, store visit, completed view, brand lift or another outcome for a specific advertiser. If that request gets filtered out, the advertiser’s DSP and bidding algorithm never even get the opportunity to evaluate it.
Containerization changes that. We can deploy a custom model trained around an advertiser’s specific outcomes directly into the exchange, where it can evaluate the much broader stream of available bid requests before throttling occurs.
Instead of asking the supply path simply to send more bid requests, we’re making it smarter about which requests to send. The model identifies the opportunities predicted to be most valuable to that advertiser and prioritizes those for delivery to their DSP. The DSP still controls the final bidding and buying decision. Containerization simply gives it a better opportunity set to make that decision against.
Containerization makes it possible to put that proprietary AI securely inside exchange infrastructure. So we’re effectively bringing the intelligence to the auction instead of bringing every auction to the intelligence.
The ANA’s Programmatic Media Supply Chain Transparency Study shows working media efficiency at as low as 36 cents on the dollar. From your perspective working inside containerized infrastructure, what are the root causes of this leakage when decisioning happens downstream (in DSPs/cloud), and how does containerization fundamentally change the math?
One of the less visible inefficiencies in programmatic isn’t necessarily a fee—it’s lost opportunity. A DSP can only optimize against the bid requests it receives. But exchanges see significantly more opportunities than DSPs can practically ingest. Because of QPS constraints, some filtering and throttling inevitably occurs upstream.
When advertiser-specific decisioning exists only inside the DSP, that intelligence arrives too late to influence which opportunities make it through that upstream filtering. The advertiser may have an incredibly sophisticated bidding algorithm, but it can’t bid on an impression it never sees.
Containerization moves a portion of that intelligence upstream. A custom advertiser model can evaluate opportunities directly within the exchange and predict which are most likely to drive the advertiser’s desired outcomes. Those opportunities can then be prioritized for delivery to the DSP.
That changes the efficiency equation. Rather than using limited DSP QPS to process a relatively generic sample of available supply, buyers can use that same capacity to receive a higher concentration of opportunities selected specifically for their objectives.
To us, that’s one of the biggest opportunities in containerization: make every request sent downstream more valuable rather than simply sending more requests downstream.
Containerization seems to allow smaller, more specialized decisioning models to be effective. Does this democratize advanced targeting that previously required massive scale and infrastructure investment?
Yes, but I think the bigger breakthrough is moving everything pre bid with outcome-based decisioning using sell-side signals at exchange scale. Containerized AI can sit directly on the sell side, access signals available there, and combine those signals with outcome measurement. That means we can train specialized AI models to predict very specific outcomes—for example, brand lift with Kantar, store visits with PlaceIQ, or unique household reach with Innovid—and then apply those models at decision-time.
But there’s an important technical distinction: the size and access of the container matter.
If your decisioning layer only sees a small, throttled sample of the exchange, the AI can only choose the best impression from that limited universe. The real opportunity is when the container can process 100% of the exchange’s QPS as opposed to the legacy DSP infrastructure that’s throttled.
AI is only as good as the opportunities it gets to evaluate.
When latency is removed from the decisioning chain and buyers have access to cleaner signals at decision-time, campaign performance becomes more predictable. How does this change the future of how buyers plan and deploy measurement frameworks, optimization models, etc.?
I think the biggest change is that measurement stops being primarily a reporting function and increasingly becomes an input into media decisioning. Advertisers already spend significant amounts measuring the outcomes that matter to them—conversions, sales, revenue, brand lift, attention, store visits, incremental reach and other business results. Historically, much of that information has been used to understand what happened after a campaign or inform the next campaign.
AI allows us to turn those measured outcomes directly into optimization models. We can train advertiser-specific models on historical media and outcome data to learn which opportunities are most likely to produce the advertiser’s desired result. Those models can then be deployed into containerized infrastructure and apply what they’ve learned to future media opportunities in real time.
That creates a continuous feedback loop: measure → train → deploy → activate → learn → improve. Over time, we believe advertisers will increasingly have their own AI models trained on their own outcomes and continuously improved as new measurement data comes back.
The result is a shift from measurement telling advertisers what happened to measurement actively informing what they should buy next.
Measurement tells the AI what worked. Sell-side signals help it predict what will work next. Full-QPS containerized decisioning gives it the opportunity set to act on that prediction.
Stay tuned for more Partner Perspectives as we highlight the buyers, publishers, and platforms who are driving growth and innovation with PubMatic.
