The Race to Build the Inference Layer

For the last several years, the defining image of AI infrastructure has been the giant data center: enormous buildings, massive GPU clusters, and staggering amounts of power assembled to train increasingly capable models.

But training a model is only the beginning.

Once that model gets used, a different workload takes over. Every prompt, agent, voice interaction, software workflow, robot, and device running AI requires inference. Training happens periodically. Inference happens every time someone asks the model to do something.

That changes the infrastructure problem.

The next generation of AI capacity will need to be available continuously, respond quickly enough for real-time applications, and increasingly exist closer to the companies, networks, people, and devices using it.

Inferra is building for that world.

Bringing Compute Closer to Demand

Inferra is building distributed infrastructure for AI inference, with a focus on bringing real-time compute closer to where AI is actually being used. The company describes its mission as “bringing real-time inference to the metro edge” and building a local layer for real-time AI.

The idea is straightforward. As AI moves from centralized model training into products people and businesses use continuously, proximity begins to matter.

Voice systems need to respond almost immediately. Robots and other physical systems have to make decisions in real time. Agents are becoming part of business-critical workflows. Wearables, industrial systems, and connected devices create their own demands around speed, reliability, and connectivity.

Some of those workloads will continue to run perfectly well in enormous centralized data centers. Others may benefit from compute sitting much closer to the places where the applications themselves are operating.

Inferra is betting that this local layer becomes an important part of the AI infrastructure stack.

They’ve Built Infrastructure Before

A company like Inferra is unusually difficult to build because almost nothing about it is purely a software problem.

AI infrastructure ultimately runs into the physical world. Compute needs power. Facilities need sites. Hardware has to be procured and deployed. Networks have to be connected. Capacity has to arrive when customers need it, and the entire system has to work reliably once it does.

That is what made Aidin Aghamiri stand out to us.

Aidin co-founded ITRenew and spent 15 years helping build it into a major data center infrastructure company. ITRenew worked with some of the world’s largest data center operators, managing and repurposing hyperscale hardware and building infrastructure for cloud, enterprise, edge, and AI workloads. Iron Mountain acquired the company in a transaction implying a minimum enterprise value of $925 million.

More importantly, Aidin has already spent the years since ITRenew working directly on infrastructure optimized for AI inference. He isn’t entering the category because inference suddenly became an interesting market. He has been working his way toward this problem for years.

His co-founders bring complementary experience. Mohammad Javanbakht has spent more than two decades in real estate and infrastructure, the kind of background that matters when the product ultimately depends on getting physical projects built. Matthew Mickelson has spent more than 25 years building enterprise and infrastructure systems and is focused on the distributed computing layer that turns physical capacity into something customers can actually use.

Together, they cover an unusual amount of the stack: data center infrastructure, real estate and development, and the systems required to operate distributed compute.

For Inferra, those aren’t adjacent skills. They are the company.

Inference Changes the Shape of Compute

The opportunity exists because inference has different characteristics from training.

The world’s largest training jobs benefit from concentrating enormous amounts of compute in a small number of places. Inference is more fragmented. It happens whenever and wherever applications are being used, and different workloads optimize for different things.

Some will simply want the cheapest available compute. Others will care about latency, connectivity, dedicated capacity, resiliency, or where data is processed. Applications operating in the physical world may have requirements that look very different from a chatbot answering a question in a browser.

The result is unlikely to be a single architecture.

Hyperscale data centers will remain essential. But as inference becomes a larger share of AI computing, there is room for additional layers of infrastructure between those giant clusters and the people and devices using the applications.

Inferra is betting that metro-edge infrastructure becomes one of them.

If that thesis is right, the opportunity is much larger than any individual facility. AI is becoming a continuous workload embedded across software and, increasingly, the physical world. The infrastructure serving it will need to evolve alongside it.

Why We Invested

Inferra isn’t an obvious seed investment. It is capital intensive, operationally complex, and competing in a market attracting some of the largest companies and deepest pools of capital in technology.

Those are real risks.

But they are also part of why the team matters so much. Aidin has already spent much of his career building data center infrastructure at scale. Mohammad understands the physical realities of getting infrastructure developed. Matthew has spent decades building the systems that infrastructure depends on.

They are starting Inferra with a lot of scar tissue from the exact kinds of problems they’ll have to solve again.

We also think inference is early in becoming its own infrastructure market. As AI moves from something people occasionally query to something embedded throughout software and the physical world, the amount and shape of compute required to serve it will change dramatically.

Someone has to secure the sites, find the power, deploy the compute, connect the network, and build the systems that make all of that capacity usable.

It’s complicated, physical, expensive work.

Which is exactly why we think this team has a chance to matter.

Visit Inferra at https://www.inferra.ai/.