Sprint: How Gonka's Consensus Works

Consensus is the mechanism by which all network nodes agree on the current state of the blockchain. Without it, each node would see its own version of “truth,” and the network would not be able to function. In blockchain history, there have been three generations of consensus:

• PoW (Bitcoin, 2009) – miners find meaningless SHA-256 hashes. Secure, but wasteful: 100% of energy goes to a “digital lottery.”
• PoS (Ethereum, 2022) – validators stake tokens as collateral. Energy-efficient, but sacrifices decentralization: large stakers (Lido, Coinbase) control a significant portion of the network.
• Sprint / PoUW (Gonka, 2025) – the third way. GPUs perform real AI computations that simultaneously serve users and confirm blocks.

Sprint is Transformer PoW 2.0, Gonka's unique consensus. The name reflects its essence: network operation is organized into sprints (epochs), each representing a cycle of execution, verification, and reward distribution. Unlike Bitcoin, where “work” is iterating through random numbers, in Sprint, work is a forward pass through the Qwen3-235B neural network with 235 billion parameters.

Network scale: currently, ~4,648 GPUs are operating in Gonka, combined into ~582 ML-nodes. Each ML-node is a GPU server with a minimum of 40 GB VRAM, capable of processing neural network requests. To operate the full Qwen3-235B model (MoE-architecture, 22B active parameters), a GPU cluster with a total VRAM of 640 GB is required. Sprint coordinates all these resources in real-time, distributing tasks and tracking each participant's contribution.

In the Gonka network, participants perform different roles, and Sprint coordinates their interaction. Hosts (miners) provide GPUs for inference and model training – they receive GNK from both genesis emission and payment for completed requests. Transfer Agents are specialized dispatcher nodes that receive incoming AI requests, verify cryptographic signatures, and route requests to suitable ML-nodes, taking into account load, available model, and latency. Validators provide cryptographic auditing – verifying that nodes have honestly performed computations. All roles are economically motivated through GNK rewards, and Sprint ensures their coordination without central control.

The network is completely permissionless – any owner of a suitable GPU can connect without KYC by installing cosmovisor and MLNode. Task routing between nodes occurs based on hardware characteristics, availability, and node reputation. Voting power in the network (Proof of Compute) is determined by the volume of computational work: “one unit of computational power = one vote.” This fundamentally differs from PoS, where voting power is determined by capital.

Key difference from predecessors: Sprint is not an “empty” protocol tied to abstract complexity. Each block contains proofs of genuinely performed work – neural network responses that were sent to users. This creates a direct link between network security and its utility: the more AI requests the network processes, the more secure it becomes.

To understand why Sprint is an evolution, and not just “yet another blockchain,” it's useful to compare it with classic Proof of Work in Bitcoin:

The main difference is in the source of value. Bitcoin's value relies on “digital gold” – limited emission and consensus on value. GNK's value is tied to the real demand for AI computing. When someone sends a request via Gonka API, they pay GNK for a specific result – a neural network response. This is utilitarian demand, not speculative.

The second critical difference is scalability. Bitcoin cannot scale: more ASICs = more wasted energy, but not more transactions. In Sprint, every new GPU increases network throughput – more requests are processed in parallel, blocks are formed faster. Network growth does not lead to energy overruns but increases its useful power.

The third difference is equipment universality. Bitcoin ASIC miners are useless outside mining – they are specialized chips that can only calculate SHA-256. GPUs used in Gonka (H100, H200, A100) are universal computing accelerators. If a host decides to leave the network, their equipment retains full value – it can be used for rendering, scientific calculations, model training, or sold on the secondary market.

The fourth difference is the economic reward model. In Bitcoin, the reward is fixed and halves every ~4 years (halving), regardless of how much real work is performed. In Sprint, rewards are distributed proportionally to the Active Proof-of-Compute weight – the more AI requests a host processes, the more GNK they receive. In addition, hosts receive two parallel income streams: tokens from genesis emission (80% of the total 1 billion GNK is distributed to hosts) and payment from users for inference (80% of each AI request). To maximize rewards, hosts lock GNK as collateral – without collateral, the node's weight is reduced by 5 times. This creates an economic incentive for long-term participation in the network, rather than quick “entry-exit.”

Sprint's security is confirmed by an audit from CertiK – a leading Web3 audit company, which completed its review in September 2025. The project has attracted ~$80M in investments from Coatue, Bitfury ($50M Series B), Insight Partners, and Benchmark – this is institutional validation of Sprint's viability as a next-generation consensus model.

Let's consider the full cycle of AI request processing in the Gonka network – from the moment a user clicks “Send” to receiving a response and distributing rewards:

1. Request: a user (or application) sends a standard POST /v1/chat/completions via an OpenAI-compatible API. The request is signed with the wallet's cryptographic key – this proves that the sender has funds to pay.
2. Routing: the request goes to a Transfer Agent – a specialized dispatcher node. The Transfer Agent verifies the signature, determines the required model, and finds a free ML-node with suitable characteristics (sufficient VRAM, required model loaded, minimal latency). Several Transfer Agents operate simultaneously in the network for fault tolerance.
3. Inference (computation phase): the selected ML-node performs a forward pass through the Qwen3-235B neural network. The GPU generates a response token by token in streaming mode. The user receives the response in real-time – latency is minimal.
4. Verification (PoC V2): in parallel with request processing, the network verifies the honesty of the nodes. 1–10% of tasks are randomly sent for re-execution by another node. Results are compared. If they match, both nodes confirm their honesty. If not, arbitration begins, and the dishonest node loses 20% of its collateral. BLS-signatures allow verifying results in less than 10 milliseconds.
5. Block formation: at the end of the epoch, all proofs of work performed are aggregated into a block. The block contains: hashes of completed requests, BLS verification signatures, data on each node's contribution.
6. Reward distribution: GNK rewards are accrued proportionally to each node's contribution. 80% of the inference payment goes to the host who processed the request. 20% goes to the Community Pool – a fund for ecosystem development (model training, grants). Additionally, hosts receive tokens from genesis emission – these are two parallel income streams.

The entire process takes seconds. The user does not see the blockchain mechanics – for them, it's a regular API request to a neural network, similar to ChatGPT, but thousands of times cheaper. The current network inference price is about $0.0021 per million tokens, versus $2.50–15 per million tokens from OpenAI.

Dynamic pricing is another feature of Sprint. The inference price is recalculated every block depending on network load. In the stable zone (40–60% utilization), the price does not change. Below 40%, the price automatically decreases to attract users. Above 60%, it increases, stimulating the connection of new GPUs. The maximum change is 2% per block. This creates a market mechanism where supply and demand are balanced automatically, without manual intervention.

Security through DiLoCo: in addition to inference, Sprint coordinates distributed model training. GPU clusters worldwide train locally and synchronize every ~1000 steps via the DiLoCo protocol. This allows Gonka not only to serve but also to train AI models – without the need to gather all GPUs in one data center. The roadmap for 2026–2027 includes multi-model inference, where hosts will be able to serve different models depending on their GPUs. Sprint transforms Gonka from an inference network into a full-fledged AI platform.

Protocol-level security: to protect against consensus attacks, Gonka employs Guardian nodes – 3 special nodes controlling 34% of the votes. This ensures that an attacker cannot seize control of the network, even with significant computing resources. Guardian nodes were introduced in v0.2.7 (January 2026) as an additional layer of security during the early stages of network development. As the number of independent nodes grows, the role of Guardian nodes will gradually diminish through on-chain governance – a management process in which all participants vote with tokens and PoC weight. This is standard practice for young blockchain networks: starting with a controlled launch and gradually moving towards full decentralization.