Gonka plans AI ASICs, but offers no shipping date yet

By Sophie Lindqvist

Gonka co-creator Daniil Liberman says open models now account for 80% of global AI usage, even as the company moves from distributed GPU inference software toward specialized AI-chip development.

The announcement supplies no independent support for that usage estimate. It also names no chip architecture, process node, wafer size, capacity target or manufacturing partner. Gonka has set no shipping date or performance benchmark. The hardware remains a development goal, not an announced product.

What does Gonka already operate?

Gonka operates a coordination layer that pools computing capacity from independent hardware providers. Developers reach that shared network through an API compatible with the OpenAI format. This lets smaller operators combine their hardware instead of competing alone against large cloud groups.

Liberman and his brother David, the company's other co-creator, have built technology companies, including work in augmented reality at Snap. He said Gonka started with inference because developers can use it directly in commercial applications. Distributed training remains possible, but it carries different technical requirements.

The company's argument rests on a distinction between accessible software and concentrated hardware. Open-model users can inspect and adapt models under license terms. They still need accelerators, memory, power and facilities to run those models at scale.

Liberman called that dependency “GPU feudalism.” “You can build a business on somebody else's infrastructure, but somebody else still sets the terms,” he said. He warns that providers can change prices, limit access or withdraw services after businesses depend on them.

He cited historical usage figures as evidence that shared protocols can attract physical infrastructure and paying demand for a durable AI-compute market, at least at the inference layer. The announcement published no supporting figures.

Does open-model access change the hardware bottleneck?

No, not on the evidence presented. Liberman said open models have widened access at the software layer while chips, memory, power and facilities remain concentrated among a small group of infrastructure providers. His 80% figure describes model usage, not the distribution of physical computing assets.

Coordination software cannot remove concentration in semiconductor supply, chip manufacturing or energy access. Even thousands of network participants may depend on the same narrow set of suppliers and constrained industrial routes. Gonka therefore does not remove hardware concentration merely by adding another coordination layer.

That limitation brought application-specific integrated circuits, or ASICs, into the discussion. Liberman said specialized hardware could improve inference economics. He also required engineering work and real-world measurement before those claims become credible.

“Open models are an enormous achievement,” he said. “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.”

What would Gonka's accelerator need to prove?

Liberman compared the effort with Bitcoin mining, where economic incentives supported repeated generations of specialized chips. AI presents a harder test, he said, because models and memory requirements change while inference workloads place varying pressures on hardware. Processing speed alone would not make a useful accelerator.

His stated priorities are memory, data movement and software integration. The announcement does not identify a target memory configuration, benchmark or hardware partner. Nor does it explain how a prospective chip would differ from the GPUs now serving the network.

That omission matters because Gonka's value proposition rests on network participation. A chip must meet the distributed network's operating requirements before Gonka can justify manufacturing it. The company has not yet shown whether that design work satisfies its own requirements.

When does the hardware arrive?

Gonka's immediate work remains improving reliability in distributed GPU inference and connecting more applications to existing capacity. Liberman described specialized AI chips as the next development area, not a finished product. He said Gonka would work with people who can build and deploy the hardware, but named no participants.

Before moving to manufacturing, financing and deployment with infrastructure operators, he said, the design needs economic validation, software integration and proof that it can meet the network's operating requirements. No tape-out date, foundry process or production volume appears in the announcement.

His broader aim is to let more operators own productive capacity and participate in the AI economy. “Who controls it will shape who gets to build,” Liberman said. Gonka's challenge is commercial as much as technical: an alternative must offer dependable capacity and acceptable pricing, not merely another software layer. Until the company publishes capacity, pricing and validation data, its ASIC plan remains an unproven alternative to concentrated accelerator supply.