Gonka

Position

Gonka's thesis: Bitcoin's security model (PoW, permissionless entry, no stake required) but with the work being useful AI computation instead of hash lotteries. Where Bittensor pays stakers ~60% of rewards and Akash/io.net are marketplaces, Gonka directs 100% of rewards to GPU providers for verified inference work:

The Bitfury investment (Dec 2025) is the signal that matters: a Bitcoin-mining incumbent concluding that GPU-PoW-for-inference is the closest living relative of its own business model. CEO Val Vavilov: the model "channels hardware power directly toward productive AI workloads, which aligns perfectly with Bitfury's mission to support open, resilient, and high-impact digital infrastructure."

Architecture

Sprint — Proof-of-Work 2.0

• Miners ("Hosts") connect GPUs (H100, A100, Blackwell; eventually ASICs) and compete in a ~10-minute Sprint cycle on a 2.3B-parameter transformer — the competitive proof phase.
• Outside the Sprint window, the same hardware serves paid workloads: inference (Kimi K2.6, MiniMax M2.7, DeepSeek V4 Flash are the flagship models) and training.
• Verification: results checked by validators via cryptographic proofs of compute; the L1 records balances, transactions, and the cryptographic artifacts proving Hosts performed AI work correctly. All computation happens off-chain.
• Whitepaper goal: near-100% of network compute toward training and inference instead of consensus busywork — against >90% of GPU cycles consumed by consensus in typical PoS networks (per Bitfury's announcement framing).

Governance — Proof-of-Compute-weighted voting

• Voting power ∝ verified compute contribution (PoC).
• 20% of PoC-derived voting weight activates automatically; the remaining 80% unlocks only by locking GNK as collateral — influence requires real compute work + economic skin in the game.
• Inference pricing is miner-votable on-chain: the launch rate has held flat at $0.000314 per million tokens (pricepertoken), an order of magnitude below mainstream API pricing. Heavy Kimi-model volume loading GPUs is expected to force the pricing vote.

Token (GNK)

• Payments for compute, validator staking, governance.
• OTC only (SafeTrade; HeX orderbook DEX) — no tier-1 listing yet. Market cap ~$80M class (early stage).

Measured Output — the DeAI Dashboard numbers

The reason Gonka earns a vault note now is independent measurement:

• 72.34B tokens/day across 626 GPUs — #2 among all tracked entities (behind only Bittensor's Chutes at 425B), #1 among standalone networks, roughly matching io.net's 71.45B on a fraction of the silicon (io.net's B200 index: 402.20 vs Gonka's 626 tracked GPUs).
• Institutional backing alongside a CertiK audit; Bitfury's $50M (Dec 2025) plus Coatue.
• ~6,000 H100-equivalent GPUs aggregated in the first months post-mainnet (mainnet: August 2025).
• Dashboard readings are volatile (~10% week-over-week swings) — appropriate for a network still in rewrite.

Trade-offs

Strengths

• The useful-PoW thesis, live. Nobody else has Bitcoin-style permissionless security and measured inference output in one system. No stake-to-participate means GPU owners enter with hardware alone — the lowest-capital permissionless entry in the sector.
• Radical price transparency. Flat, governance-set $0.000314/M tokens is test-level pricing for production inference; pricepertoken tracks it independently.
• Reward integrity. 100% of rewards to compute providers (vs ~40% Bittensor) removes the staker-over-miner extraction that Bittensor critics highlight.
• Independently measured. Throughput claims are checkable against the DeAI Dashboard, not just marketing.

Limitations

• Pre-production stability. A January 2026 flaw in the original Sprint design let Hosts overstate hardware capacity; the fix grew into a ground-up rewrite that is well advanced but unfinished. Track record is the thinnest of any top-5 tracked network (mainnet Aug 2025).
• Dynamic pricing unbuilt. The whitepaper's dynamic pricing model remains unimplemented; flat-rate pricing is sustainable only while test-level demand subsidizes it.
• Liquidity and access. OTC-only GNK; no tier-1 exchange; no US/EU retail on-ramp.
• Verification depth. Validator-checked cryptographic proofs of compute are weaker than optimistic-verification + fraud proofs (Gensyn) or TEE attestation (Phala) for workloads where correctness of reasoning (not just execution) matters.
• Model narrowness. A handful of flagship open models vs 128+ specialized Bittensor subnets.

Attack surface

• Capacity fraud — the January class of bugs: hosts claiming hardware they don't operate. The Sprint rewrite is the systemic fix; watch its audit trail.
• Consensus-period capture — a coalition dominating Sprint cycles could gate paid-work allocation; PoC-weighted governance concentrates influence in large farms.
• Underpriced inference race — if the pricing vote keeps rates near $0.000314/M, host economics may only work for subsidization-phase actors; the network needs the dynamic-pricing mechanism to survive demand maturation.
• Single-verification path — no fraud-proof game or TEE redundancy yet; a validator-collusion event has no secondary check.

Use cases

Cheapest measured inference in the sector

Flat $0.000314/M tokens for flagship open models (Kimi K2.6-class) — for bulk, non-sensitive, throughput-bound workloads where correctness risk is tolerable.

Useful-PoW mining

Bitcoin-familiar operators (Bitfury's lineage) repurposing GPU fleets: hardware-only entry, no token purchase required, rewards proportional to verified useful work.

Sovereignty angle

Permissionless, stake-free GPU provision with governance set by compute contribution rather than capital — structurally the closest thing to "Bitcoin for AI compute." The censorship-resistance depends on the rewrite landing and validator decentralization actually materializing; treat as thesis-in-progress.

Evidence at a glance

Primary sources

• Gonka — gonka.ai; FAQ — gonka.ai/docs/FAQ.
• GitHub — github.com/gonka-ai.
• Bitfury investment — VentureBeat (Dec 2025).
• DeAI Dashboard coverage — Crypto Daily via Binance Square (Aug 14, 2026).