Why Gonka exists
Open models still need hardware
Some AI models can be downloaded and run independently. But a large model needs a powerful server, sometimes with several expensive graphics processing units, or GPUs. Someone has to buy or rent that equipment, configure it and keep it running.
Developers therefore often pay for ready-made access through an API: an interface that lets an application send requests and receive responses. This is convenient, but the provider sets the prices, available models and access conditions.
Gonka's creators propose bringing independent operators' hardware into a shared network. Serving AI applications would then no longer require one company to own all the servers and build all the data centers. Hardware owners can supply computing power, and developers can use it through a common protocol.
In an interview with the Liberman brothers, this idea comes down to a simple problem: freely available models are of limited use if the computing power to run them remains expensive or out of reach.
What the network does
When someone asks an AI to explain a bug in code or translate an email, a model that has already been trained processes the request. This is called inference.
Gonka supplies computing power for this work. Providers, known as hosts, run models on their own hardware. The protocol sets the rules for how they interact, how computation is checked and how they are paid.
An application developer does not need to negotiate with every server owner. They connect through an available service or gateway and send requests to supported models. Developer documentation.
How it compares with centralized services
Gonka makes it possible to obtain computing power from different providers operating under shared rules. This reduces dependence on a single infrastructure owner and opens the market to operators that could not build a large cloud service on their own.
There is also a question of trust. With a conventional API, developers largely rely on the provider's promise to use the requested model. Gonka makes verification part of the network's design.
Lower cost is another aim. Pooling computing power, keeping verification economical and rewarding providers are intended to help. Actual savings depend on the model, demand and access terms: decentralization alone does not guarantee the lowest price. The economic reasoning is described in Gonka's tokenomics.
How Gonka works: the main principles
Computing power has to be demonstrated
In an open network, saying “I have a hundred powerful servers” is not enough. There needs to be a way to check what resources a participant actually supplies.
Gonka uses Proof of Compute for this purpose. Participants perform verifiable computing tasks, and the results help determine their weight in the network.
Gonka's approach is to limit the time spent on these tests and keep hardware available to process user requests. The tasks involve computation characteristic of AI workloads. The authors intend this to encourage the development of hardware that is also useful for running models. The mechanism explained on Gonka's blog.
Participants check each other's work
A provider might try to cut costs, for example by using a simpler model than the one requested. The network therefore includes checks performed by other participants.
Having the entire network recompute every answer would be expensive. Gonka uses selective verification and economic consequences for violations. The aim is to make persistent cheating unprofitable while keeping verification costs reasonable. How inference validation works.
Checking that a model ran correctly does not mean checking whether its answer is true. An honestly executed request can still produce an error made by the AI itself.
A blockchain helps participants follow shared rules
Models run on hosts' hardware. The blockchain records and coordinates participants' actions, allowing them to maintain a shared history of settlements and apply the protocol's rules.
Agreement on the state of the network is called consensus. A core principle of Gonka is to connect a participant's influence to their verified computing contribution. Specific rules may evolve, but accounting for the resources supplied is central to the project. Gonka whitepaper.
The GNK coin connects computing to the network's economy: it is used to pay for work and reward providers. Additional rewards help attract hardware while the network builds demand. For an application's end user, these settlements can stay behind the scenes.
How Gonka differs from other decentralized networks
Decentralized computing takes different forms. Akash, for example, offers a market for resources on which you can deploy your own application. Bittensor connects separate subnets with different tasks and rules for evaluating results.
Gonka focuses on organizing computation for AI: verifying providers' computing capacity, serving model requests and checking the work. For developers, this means access to a supported model without maintaining the hardware themselves.
Its distinctive approach is to combine these mechanisms in one protocol. Speed, quality and cost still need to be compared for the particular task at hand.
Who created Gonka, and how to try it
Gonka originated with Daniil and David Liberman and the engineering team at Product Science. Gleb Morgachev is among the protocol's co-creators. Independent computing providers helped launch the network. Read more in Who is behind Gonka?.
Start with the available ways to try Gonka. Before connecting your own application, check the broker monitoring page to see how reliably providers work and how quickly they respond.