Gonka eyes AI chips to tackle compute concentration

Gonka has outlined plans to expand from distributed GPU inference software into specialised AI chip development. Co-Creator Daniil Liberman set out the proposal as part of the company's broader push to widen access to AI compute infrastructure.

Liberman said hardware and supply chains underpinning artificial intelligence are becoming increasingly concentrated, even as open models have widened access at the software layer. Developers may be able to inspect and adapt models, he argued, but still depend on a small group of infrastructure providers for the chips, memory, power and facilities needed to run them at scale.

That concern sits at the centre of Gonka's current model. The company operates a coordination layer that lets independent hardware providers contribute computing capacity to a shared AI network, while developers access it through an API compatible with the OpenAI format.

Gonka began with inference rather than training because inference is the service developers can use directly in commercial applications, Liberman said. Distributed training remains a possible path, but it brings different technical demands.

"My brother David and I are co-creators of Gonka. We have spent years building technology companies, including in augmented reality at Snap. Today, we are focused on a question that goes underneath the applications: who will own the infrastructure on which artificial intelligence runs? Because if AI becomes a basic tool for almost every business, access to compute becomes access to the economy. That is why our work is moving from protocols toward chips.

Think about what happens when your company starts depending on AI every day. Your software engineers use it. Your customer support uses it. Your products begin to depend on it. Now imagine that the provider changes the price, limits access, or decides that your application can no longer use its service. How much freedom does your business actually have?

We have described this as a kind of GPU feudalism. You can build a business on somebody else's infrastructure, but somebody else still sets the terms. Our concern is that this dependency becomes much deeper as AI becomes more capable. We need an alternative that gives more people a meaningful ability to participate in the infrastructure itself," said Daniil Liberman, Co-Creator at Gonka.

Open models

Open models have changed the software side of the market by allowing users to inspect, adapt and run models themselves, subject to licensing terms, Liberman said. But access to model weights does not resolve the underlying question of compute ownership.

He said 80% of global usage now comes from open models, but argued that a gap remains between software openness and infrastructure concentration. In his view, AI will remain difficult to access on equal terms unless the economics and availability of compute open up alongside the models themselves.

Gonka's answer is to let smaller infrastructure operators work together rather than compete alone against large cloud groups. Its network is designed to connect independent providers and turn their hardware into a shared pool for AI inference workloads.

Historical usage figures suggest such a shared protocol can attract physical infrastructure and meet paying demand, Liberman said. He framed that as evidence that software coordination can support a durable market for AI compute, at least at the inference layer.

Hardware bottleneck

The next problem, he said, is that coordination software cannot remove concentration in chip supply, manufacturing or energy access. Even if thousands of operators join a network, they may still depend on the same narrow set of semiconductor suppliers and constrained industrial routes.

That has led Gonka to examine application-specific integrated circuits, or ASICs, for AI workloads. Liberman said specialised hardware could improve the economics of inference, but added that any such claim would need to be validated through engineering work and real-world measurement rather than broad promises.

"Open models are an enormous achievement. They give people the ability to inspect a model, adapt it, and run it themselves, subject to its licence. Right now, 80% of global usage comes from open models. But having the weights does not give you the hardware to run that model at scale. You still need accelerators, memory, power, and a place to operate them.

So we can have openness at one layer and concentration underneath it. A developer can download the model and still depend on the same small set of infrastructure providers. That is the gap we started with. If we want AI to remain accessible, we need to connect the openness of the model to the economics and availability of the compute," said Liberman.

He compared the effort with Bitcoin mining hardware, where economic incentives supported repeated generations of specialised chips. But AI presents a different engineering challenge, he said, because models and memory demands change, and inference workloads place varying pressures on hardware.

For that reason, a useful AI accelerator would need more than processing speed, Liberman said. He pointed to memory, data movement and software integration as core requirements if new chips are to run relevant models efficiently on the network.

Next steps

Gonka's immediate focus remains on improving reliability in distributed GPU inference and connecting more applications to existing capacity, Liberman said. He described specialised AI chips as the next area of development, not a finished product.

He also made clear that the company is not giving a shipping date or performance target. Instead, the process will require economic validation, software integration and proof that new hardware can meet the network's operating requirements before moving to manufacturing, financing and deployment with infrastructure operators, he said.

"We believe AI compute is becoming a foundational economic resource. Who controls it will shape who gets to build. Gonka is our contribution to the coordination layer. Specialised hardware is the next part of the problem we want to work on, together with people who can build and deploy it. The future we want is one in which more people can own productive capacity and participate in the AI economy," said Liberman.