Whitepaper · v0.1 (draft)

Abakos: A Proof-of-Useful-Work L1 with Enforced Usefulness

Compute that pays for itself.

Abstract. Abakos is a Layer-1 blockchain — a fork of Bitcoin's btcd — that replaces SHA-256 hashing with a Proof-of-Useful-Work (PoUW) function based on matrix multiplication (GEMM), the core operation of AI. Unlike prior PoUW attempts, Abakos makes usefulness an enforced economic property: a miner earns the full block reward only for work bound to a paid compute job settled through an on-chain marketplace with stablecoin escrow. A consumer app, Abakos Chat, channels mainstream demand into this loop. The result is GPU compute priced below alternatives, because the same watt-hour earns a mining subsidy and produces paid AI output (2-for-1).

Contents  ·  1 Background · 2 Principle · 3 Architecture · 4 Chat · 5 Economics · 6 Security · 7 GTM · 8 Roadmap · 9 Risks · 10 References

1. Background & motivation

AI progress is bottlenecked by GPU compute, which is expensive and concentrated in a few hyperscalers. Meanwhile, Proof-of-Work blockchains expend large amounts of energy to produce only network security. PoUW has long promised to merge the two — to make the energy that secures a chain also produce something valuable.

The leading deployed attempt, Pearl (PRL), runs a matmul-based PoUW. But its protocol verifies only that a matrix multiplication was performed correctly — not that it served any real AI workload. Empirically (arXiv 2606.04819), Pearl's network represents on the order of 320,000 GPU-equivalents and ~112 MW, yet produces essentially zero useful AI: miners grind random matrices because that is cheaper than serving real inference. Statistical "is-this-a-real-model" checks are trivially defeated by adversarial sampling.

The core tension (verifiability vs. usefulness): tasks that are easy to verify tend not to be useful, and useful tasks are hard to verify cheaply.

2. Design principle: enforce usefulness economically

Abakos does not try to cryptographically detect "real" AI matrices (a losing game). Instead it makes useful work the dominant economic strategy:

reward_block = base_subsidy * (FLOOR + (1 - FLOOR) * useful_ratio)
FLOOR = 0.25
useful_ratio ∈ [0,1] = fraction of the block's matmul work bound to PAID escrow jobs

A miner serving paid jobs earns the stablecoin payment and the full block reward on the same compute; a miner grinding randomness forfeits 75% of the subsidy. Usefulness becomes the profit-maximizing choice.

3. Architecture

3.1 Consensus & PoUW

Base layer: Nakamoto consensus (longest chain), forked from btcd (ISC). Work function: tiled, noised matrix multiplication (NoisyGEMM) on GPU, with a Blake3 jackpot under the difficulty target and a Plonky2 zero-knowledge proof for cheap on-chain verification (building on the open cuPOW construction, eprint 2025/685). INT8 first; FP8/BF16 next (standard-compatible, vLLM/SGLang drop-in); multi-vendor (NVIDIA + AMD) from day one.

3.2 Model registry

On-chain registry binding a model_id to a weights_hash, precision, and min_vram. PoUW proofs bind their input matrices to a registered weights hash; jobs are routed to providers with sufficient VRAM.

3.3 Compute marketplace + escrow (the differentiator)

3.4 PoUW ↔ job binding

The block header commits to served job_ids, their input Merkle roots, and escrow references; the verifier checks the escrows are funded and that the PoW inputs match the committed job inputs. The hardest open problem is binding "this jackpot came from these paid inputs" cryptographically; if a perfect binding is too costly, security falls back to the economic binding (a funded escrow must exist for useful_ratio > 0), which already inverts the economics against empty mining.

3.5 Model sizes & provider tiers (open-weights only)

Closed frontier models (Claude Opus, GPT, Gemini) have secret weights and cannot be run by any decentralized network; Abakos serves strong open-weight models (Llama, Qwen, DeepSeek, Mistral, Gemma). Models that fit on one GPU are served independently (data-parallel) — ideal for a decentralized network. Models too large for one GPU are hosted by datacenter/neocloud providers with local multi-GPU boxes; model splitting happens inside one provider, never across the public internet.

4. Abakos Chat

A simple, crypto-invisible "ChatGPT for everyone": email login, pay by card or ABK (cheaper). Every message is served by the Abakos network → mainstream, paying demand flows straight into the marketplace, raising the network's useful_ratio. Early phase is hybrid (reliable partner GPUs for low latency), moving decentralized as latency stabilizes.

5. Economics & token (ABK)

6. Security & trust

The base layer inherits Bitcoin/btcd's battle-tested consensus; the novel parts (PoUW kernel, reward-split, escrow, ZK verification) are the risk surface and must be audited before mainnet. Launch is mainnet-first but gated: a short incentivized testnet (also a marketing vehicle) → security audit → mainnet + token. Transparency as policy: open-source node at testnet, an honest public "% useful compute" metric, and clear disclaimers.

7. Go-to-market

Supply first (neoclouds with idle GPUs + solo miners via a 1-click installer), then point demand at it: latency-tolerant batch jobs for fastest revenue, then Abakos Chat for mainstream scale, plus academic grants for credibility.

8. Roadmap

9. Risks

Verifiability gap (mitigated economically, not statistically); supply/demand bootstrap (supply first + ecosystem incentives + Abakos Chat); token-price dependence of the subsidy (mitigated by stablecoin revenue); decentralized inference latency (hybrid first); competition (Pearl, Gonka, io.net, Akash — differentiated by enforced usefulness + consumer app + price). The ABK token is subject to legal review; stablecoin compute sales are kept separate from token speculation.

10. References

← Join the waitlist  ·  Litepaper  ·  Discord  ·  info@abakos.ai

Whitepaper v0.1 (draft) — for information only; not financial advice and not an offer of securities. The ABK token structure is subject to legal review. Figures are illustrative.