Madhive powers precision ad campaigns with Google Cloud TPUs and SADA Services

Advancing adtech with scalable, cost-effective AI

In the competitive world of adtech, abundant data often fails to translate into actionable insights. Madhive, the unified operating system for local media, is bridging the gap by embedding models and AI and redefining how advertisers connect with audiences. “From a business perspective, we deal with large, complex datasets,” says Aaron Brown, Chief Scientist at Madhive. “Our secret sauce lies in creating a canonical signals dataset to drive meaningful customer connections.”

Madhive’s technology embeds petabytes of behavioral data into high-dimensional spaces, enabling precise audience segmentation that transcends traditional demographic buckets. By combining diverse data—such as ads and CRM inputs—into a unified “semantic space,” Madhive uncovers actionable patterns and predicts customer behavior.

“These high-dimensional vectors let us reason, group, and generate insights with mathematical precision and contextual depth,” explains Brown. “Instead of assuming every consumer reading about a car brand is automatically in the market to purchase a car, we dive deeper into behavioral signals to discern intent and make more accurate predictions. That user reading about cars could just be a hobbyist—passionate about cars but with no intention to buy in the near future. We’ve been super successful in translating these determinations into business outcomes because after we’ve narrowed in on who your customers are, our unique and completely vertically integrated bidder can execute to deliver to each and every one of them. It’s a perfect product for our advertisers who need to efficiently find high-fit customers that are ready to act.”

Business challenge 

Struggling to keep pace with GPUs

As Madhive’s data demands grew, their traditional GPU-based systems struggled to keep pace. Processing petabytes of real-time behavioral data became prohibitively expensive and required extensive engineering efforts.

“The first big challenge was scalability,” says Brown. “GPUs require spreading workloads across many machines, which adds unnecessary complexity and engineering time.”

Beyond scalability, limited throughput hindered the development of Madhive’s advanced AI models. “If you have a workload that keeps the hardware running hot continuously, like our real-time behavioral data streams, the costs associated with maintaining high throughput are substantial,” explains Brown.

To overcome these challenges, Madhive needed a scalable solution capable of efficiently handling massive volumes of data without bottlenecks.

Solution 

Powering a solution with Google Cloud TPUs

Google Cloud Tensor Processing Units (TPUs), emerged as a potentially transformative option for Madhive. TPUs, designed specifically for machine learning workloads, promised cost savings as well as unmatched throughput for embedding and processing massive datasets.

“For us, the real advantage of migrating to TPUs wasn’t just cost, but also throughput,” says Brown. “TPUs enable real-time processing, which is critical. Behavioral signals come in constantly, and we need to act on fresh data. For more fleeting behaviors, we layer additional models to ensure our predictions remain relevant and therefore valuable.”

With TPUs, scaling Madhive’s advertising model becomes much simpler. “You can almost think of it as double-clicking to add more TPUs, and it just works,” explains Brown. “That’s the promise of Google Cloud. It’s a refreshing experience.”

Optimizing machine learning workloads with SADA

To maximize the potential of Google Cloud TPUs, Madhive collaborated with SADA, An Insight company and a multiple-time Google Cloud Partner of the Year. SADA brought deep expertise in optimizing machine learning workloads while reducing costs and streamlining integrations.

“We have amazing engineering talent at Madhive, and we could probably do all of this machine learning and TPU work ourselves—but thankfully, we don’t have to,” Brown says. “SADA’s machine learning expertise allows us to focus on our core innovations. If I need a gut check, I can message our SADA account lead, and his team handles the rest.”

In the close collaboration with Madhive, SADA provides a concierge level of service for routine operations and edge cases that may crop up. “Our SADA engineering contacts will say, ‘Don’t worry, we’ll run some tests and talk to the Google Cloud TPU product managers.’ They give us the white-glove treatment,” says Brown. “They’ll tell us if what we’re doing with TPUs will work or not—they’ll give you straightforward and technically reasoned answers. That level of engineering collaboration is phenomenal.”

Impact

Driving customer success with AI-powered insights

As a result of collaborating with SADA on implementing their adtech workloads on Google Cloud TPU-powered infrastructure, Madhive’s platform has been able to empower local advertisers. They now compete with the precision and sophistication typically reserved for corporations with large data science budgets. By combining fresh, high-quality data with contextual insights, Madhive enables small and local advertisers to create campaigns that resonate on an emotional level.

“Madhive’s relationship with SADA has been instrumental to our success,” says Brown. “This collaboration has allowed Madhive to focus on our core mission of innovating adtech.”

SADA helped Madhive optimize their machine learning infrastructure for cost and scalability by providing expertise on TPU architecture and implementation. This enabled Madhive to leverage the full potential of Google Cloud’s advanced hardware.

“A huge differentiator with SADA is their ability to help us scale seamlessly while we focus on delivering value to our clients,” says Brown. “Their machine learning expertise has allowed us to build on our success without getting bogged down in operational challenges.”

Additionally, the integration of Google Cloud TPUs has unlocked substantial efficiencies for Madhive. The platform now provides real-time insights, supports new products, and drives cost-effectiveness.

“What excites me most is the storytelling potential of this data,” says Brown. “By zooming into specific regions or platforms, clients can refine their strategies in real time. For example, if you’re targeting racing simulator enthusiasts in Maryland, our system might highlight untapped audiences in Washington, D.C., or recommend unexpected ad channels like TBS over Sky Sports. It’s about turning intuition into data-driven action.”

Innovating at the forefront thanks to SADA

As Madhive continues to innovate, their relationship with SADA ensures that they remain at the forefront of the industry. “We’re just scratching the surface with TPUs,” says Brown. “There’s so much more Madhive and SADA can build on top of this. The efficiencies we’re unlocking in adtech are just the beginning. It’s about delivering campaigns that not only perform well but also create meaningful connections with audiences.”

In addition, real-time data processing allows Madhive to act on fresh behavioral signals—critical in an industry where timing is everything. “For example, if someone reads a car magazine eight seconds ago, our system reflects that in real-time,” says Brown. “This level of freshness and precision sets us apart.”

Shifting the adtech paradigm

This fractal-like view of data informs not just customer ad campaigns but also creative strategy and future product development. 

“This isn’t incremental progress; it’s a paradigm shift,” says Brown. “For example, say you’re a maker of organic, fair-trade dog biscuits. Using our semantic space, we can identify and refine audiences that align with your brand. Our system might suggest yoga enthusiasts as a key demographic—a connection you might not have uncovered on your own.”

Through SADA’s close collaboration with Madhive on Google Cloud TPU optimization, they were able to:

• Fine-tune cost structures
• Scale AI workloads efficiently using TPUs
• Unlock substantial cost savings while maintaining high throughput
• Enhance the performance of embedding models critical to audience targeting