Boris Abaev is CEO of Bidease, an AdTech startup on a mission to bring fully transparent, value-driven advertising to the mobile world.getty​For over a decade, cloud infrastructure has been a go-to for tech companies, and for a good reason: AWS, Google Cloud and Azure made scaling accessible, lowered capital requirements and shortened the time to market. According to Synergy Research Group, the global cloud infrastructure services market reached more than $300 billion in 2024, growing by roughly 20% year over year.For most software businesses, cloud is still the right call, but for some industries—programmatic advertising in particular—things get more complicated. I have spent over a decade building and running a mobile DSP, and today I want to share one of the most underrated infrastructure decisions in AdTech: when (and why) to consider owning your stack instead of renting it.Cloudy With A Chance Of PainWhen we were building Bidease in the late 2010s, the cloud was the only sensible option. Capital was expensive, and deployment had to be fast. We needed to ship a DSP, not run a data center. I would make the same call again today.The problem lies with the nature of programmatic advertising. Every bid request must be answered within 20 milliseconds. Every element we add—fraud filtering, supply-side scoring, ML model outputs—makes maintaining that latency more expensive. To preserve our margin, any increase in infrastructure costs ultimately means higher prices for customers.In my experience, the infrastructure costs of cloud-hosted DSPs typically account for around 40% of their revenue. This is not about specific companies; it is an inherent feature of processing billions of bid requests a day on rented compute. Those costs either get passed on to advertisers and publishers or eat into the company's margin.From Constraints Comes A Different ArchitectureBidease was built primarily for emerging markets like Southeast Asia, Latin America and parts of the Middle East, where unit economics are tighter than in the West. A cloud-based DSP simply did not work there: a campaign that needed to perform at $2 CPI could not survive a 40% infrastructure tax on top of media costs.Around 2019-2020, we decided to migrate from the cloud to our own servers in the U.S., Europe and Asia, which was neither quick nor easy. It took us years and an engineering team unlike anything a cloud-native company would build. Today, our costs run at under 20% of revenue, with sub-20 ms bid response times and pricing flexibility we never had in the cloud.What looked like a limitation at the time turned out to be what pushed us toward a better architecture. I do not think we would have made the move otherwise.​What Rented Compute Means In PracticeHere is how I see renting compute in a compute-intensive business:​• On owned infrastructure, our engineers can optimize directly against the constraints that matter: 20 ms budget, CPM and latency between continents. The cost structure is something we control, not something we inherit.• On cloud infrastructure, even the best engineers work on top of someone else's tech decisions. The vendor's architectural choices become your ceiling, which is fine until your business model needs to push past it.​When Should I Leave The Cloud?​This argument gets oversimplified in AdTech conversations all the time, so let me be careful: Owned infrastructure is not always better.​For most early-stage AdTech companies, cloud is the right starting point—and often the right long-term choice. The real question is whether you have a business where computing is the product.​From my experience, a few symptoms suggest you may need to consider owned infrastructure:• Your computing costs are growing significantly faster than your revenue.• Your latency requirements are non-negotiable and approaching the cloud limits.• You are operating in the markets where margin compression is structural, not cyclical. • Your cloud bill is starting to rival your engineering payroll.If none of these apply, you are probably better off staying in the cloud and reinvesting the capital elsewhere.​Similar Trends Observed In AI​I am writing about this now because the AI industry seems to be going through a similar realization. Frontier AI labs are building their own data centers because they need a level of speed, control and efficiency that rented infrastructure cannot provide.The constraint is not just financial. It is architectural—you cannot optimize what you do not control. According to Goldman Sachs Research, hyperscaler and AI-related infrastructure spending is projected to exceed $1 trillion over the next several years. This is not a marketing decision; it follows the same logic that pushed a handful of AdTech companies off the cloud a decade ago, just at a much larger scale.​Things I Learned the Hard Way If you are a tech leader facing this dilemma, here are a few practical observations:• Do Not Move Off The Cloud Reactively: We did not move because of a single bad quarterly bill; we moved because the cost curve was heading in a direction our business model would not survive.• Owned infrastructure Is A Multi-Year Investment, Not A Quarter-Long Project: The cost is real, and operational expertise takes years to build—we underestimated this repeatedly.• Hybrid Is Fine: We still use cloud services for parts of our stack where elasticity matters more than cost. Owning your infrastructure is not a religious commitment.• Do Not Make This Decision Based On Industry Consensus: Cloud was the consensus when we moved off it; the companies that followed it are now writing the largest checks to catch up.Looking back, the best infrastructure decisions are often the ones forced by economic reality before the industry catches up. We did not set out to build a competitive advantage; we were trying to make our business work in markets where the conventional approach failed. The advantages we gained were a side effect, not the plan.The AI industry is now facing a similar choice, and I suspect the conclusion will be the same: When compute is the product, the economics eventually demand that you own it.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?