The Overclock
Story

In 2014, Kubernetes changed how software could be deployed. It was not yet ready to run globally, or to survive failure at scale. Greg Osuri and Adam Bozanich founded Overclock Labs in 2015 to close that gap and make Kubernetes production-ready at internet scale. Everything that became Akash started here.

Overclock built global Kubernetes clusters: federated deployments, fault-tolerant orchestration across continents, multi-region failover, and cross-cluster scheduling. These became the primitives a marketplace for compute would later need. We wrote the first lines of Akash code in this period. Our open-source work went into production at NVIDIA, AWS, IBM, HashiCorp, and Docker, so other teams could build on it. Code written in this period is still running.

As our clusters grew, the central control plane became the limit. When it failed, everything failed with it, and it held back performance, reliability, and growth at the same time. We concluded that compute would scale globally only if it could grow through many independent participants instead of one operator. A blockchain gave us a way to coordinate a network like that, so we built on one.

In 2018, Overclock published the Akash whitepaper and gave the model a name. The Supercloud. Control and resources, fully decoupled. Any cloud-capable hardware can supply compute while a protocol-level control plane handles pricing, scheduling, and access globally. The same paper named machine learning as the primary use case.

The team started on Ethereum and hit its scaling limits. The design needed a sovereign chain built for the purpose, and Cosmos offered one. To bootstrap a real community rather than a speculative one, incentivized testnets distributed AKT to the validators and developers who actually ran the network. Those early validators became users, operators, and long-term stakeholders. It was the beginning of a cloud its own users would own.

Akash went live as a general-purpose cloud, not a specialized one: production Docker containers and everyday workloads, running on a network with no single operator. Developers could deploy to a cloud that no one company controlled.

Storage made databases possible, and service discovery made networking possible. Applications stopped being demos and ran as production systems, on the same architecture the 2018 whitepaper set out.

The roadmap, the meetings, the funding decisions, the disagreements. All of it public, recorded, and searchable by anyone. Contributors grew from a few dozen to more than 500 worldwide, funded by a community model that did not depend on a single balance sheet.

Researchers, startups, and labs could rent NVIDIA GPUs on demand, without a long-term contract or a procurement process. At a time when GPUs were hard to get, the network gave more people a way in.

Work on the user experience made Akash easier to use, and more people used it. The roadmap moved up the stack, from supplying resources to running complete workloads. Next are autonomous agents: software that acts on people’s behalf and needs compute it can use on its own, with clear limits on what it can do and spend.

Developers and teams have opened more than 640,000 leases on Akash, each a workload placed with a provider. Anyone can check that figure on-chain.

Training still needs large, concentrated clusters. Inference doesn’t: an open model can run wherever a suitable GPU is available, so people no longer have to wait on one provider’s hardware. AkashML is our inference service on Akash. It runs open models on capacity we source across the network, and it earns by making that compute easy to use.