TensorCash TSC
Blocks are mined by running open-source LLMs — the verification and settlement layer for global AI inference, post-quantum from block zero.
About TensorCash
sources: official materials · verified 2026-10-13TensorCash replaces hash-grinding with useful work: miners run open-weight language models to answer questions, and the same run competes for the next block. The random number that selects each output token is derived from a hash of the latest block, the prompt and the text so far — so the model's word choices are tied to a public, unpredictable number anyone can replay. Every 256-token window is hashed against the block being built and compared with the difficulty target, Bitcoin-style: one window, one lottery ticket. Verifiers replay proofs against the open weights; blocks with failed proofs count for zero work (a miner running a cheaper fp8 setup was caught and nullified in July 2026). Each block header carries two targets — the network target and one adjusted per model, so switching to a lighter model buys no easier blocks — and Proof v4 adds a verifier-set intelligence multiplier (1x-12x) tied to window entropy. The chain itself is a Bitcoin Core fork with post-quantum ML-DSA addresses from genesis, plus a built-in financial layer: native assets with versioned legal documents, contracts-for-difference, COMPUTE options on network difficulty, and a DEX.
How it works
- Network
- Bitcoin Core fork (bcore) + forked vLLM / llama.cpp proof recorders + gnark; financial layer built in (native assets, CFDs, difficulty options, DEX)
- Consensus
- Proof-of-inference (useful PoW) · ~10-min blocks · ASERT per-block retargeting · dual targets (network + model-adjusted) · intelligence multiplier 1x-12x (Proof v4)
- Mining algorithm
- Proof-of-inference — LLM sampling seeded by hash(latest block + prompt); 256-token windows as lottery tickets
- Launched
- 2026-06
- Hardware
- GPU
Insights / Thesis
pow.family researchQuick start
tested 2026-10-17~30 min · ~$0.35 to test on a rentalMine TSC with vLLM on a rented NVIDIA GPU (Docker)
miner: vLLM (TensorCash fork) via pool worker- 1
docker run --rm hello-world && docker compose version
↳ Check Docker + Compose work
- 2
docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu22.04 nvidia-smi
↳ Confirm Docker sees the GPU
- 3
# pick a pool: LuckyPool (lpminer, stratum), Bob Labs (FearMiner, WSS) or Aria (Aria Energy worker, WSS)
↳ Broker mode: your box runs the model, the pool handles the rest. >=12 GB VRAM per card advised
- 4
# vLLM version by GPU: A100/RTX30/40 -> vLLM 0.10.0 · H100/H200 -> 0.19.0 · B200/RTX50 -> 0.19.0 built from source
↳ Full commands in the official inference-providers guide
Pitfalls we hit
- Useful-PoW is real here: verification already nullified a cheating miner's blocks (fp8 setup, July 2026), and the original closed-source miner ecosystem was opened up by third parties (suprnova's stratum spec + supr-meow-tsc, Aug 2026).
- Verification is real and adversarially tested: a miner submitting fp8 work as bf16 was caught by full re-run and had 2 blocks scored zero-work — the first documented usefulness-fraud conviction on a PoUW chain.
Tokenomics — token & supply
- Ticker
- TSC
- Maximum supply
- 21,184,153.03530240 TSC after 48 epochs (no premine, no sale, no investor round)
- Emission
- 715 TSC/block at start, multiplied by 3/5 each epoch; epoch lengths double (715 blocks ≈ 5 days first, up to 732,160 blocks). Issued 4,324,066 TSC at block 31,512 (2026-10-17) — matches epoch math to 0.001%.
- Reward now
- 55.5984 TSC (epoch 6, since block 22,165)
- Halving
- Epoch 7 at block 45,045 → 33.35904 TSC per block (x3/5 steps, not halvings)
- Unlock
- Block rewards locked 100 blocks (Bitcoin-style coinbase maturity); pools pay after maturity
- Premine
- 0%
Analytics
luckypool-stats + tscscan · harvester · just nowPools
pool compare tool → P1| pool | fee | scheme | min payout | merge | stratum |
|---|---|---|---|---|---|
| LuckyPool live tracked | 3% | PPLNS / SOLO · self-set min payout from 0.1 TSC | — | — | stratum/tcp + web — servers on four continents, pick closest |
| Bob Labs | 3% | PPLNS · min payout 0.1 TSC · pays per block after ~100 confirmations · serves paid chat (1 TSC/M output tokens, input at 1/4, 50k free credits) | — | — | wss://tensorcash.boblabs.eu/v1/ws |
| Aria (PulseMining) | 1% | PPLNS · min payout 1 TSC · Germany · +1% Aria worker fee · new wallets validated via PulseMining Discord | — | — | wss://pool.ariabrain.com/tsc-pool/ws |
Hardware & miners
full hardware db (971 devices) → P1Exchanges & liquidity
| venue | pair | kyc | note |
|---|---|---|---|
| SafeTrade | TSC/USDT | light kyc | Only venue. Price at capture from the LuckyPool page: $4.489 (0.0000549912 BTC, BTC $81,633). |
Reference material
cited on this page · links verified 2026-10-13Who to watch
sensors, not shills — verify before actingRisks
Development quiet on GitHub: all 8 org repos last pushed 2026-08-02, while the chain itself keeps producing blocks and the site lists Q4 roadmap items — releases may have moved to git.tensorcash.org; watch for divergence between site claims and shipped code.
Thin security, and concentrating: ~$36k/day issuance on a ~184 KPoI/s network where the last 100 blocks came from just 6 addresses — the top two carrying 81% of work (scan.tensorcash.org, 2026-10-20; five miners, top share 71.7% at the 17 Oct capture). Enforcement (model-adjusted targets, entropy multiplier, verifier nullification) defends the work's integrity, not the hashrate distribution.
Pool activity is concentrated and shallow: Luckypool carried 0 miners at capture (last pool block 1 month prior) vs the Aug 'Decentralisation Day' claim of ~850 addresses across 5 pools; SafeTrade is the only exchange — thin book risk on any size.
Useful-PoW is real here: verification already nullified a cheating miner's blocks (fp8 setup, July 2026), and the original closed-source miner ecosystem was opened up by third parties (suprnova's stratum spec + supr-meow-tsc, Aug 2026).
Top-100 addresses hold 38.39% of supply across 20,778 funded addresses — a concentration signature from the fast early epochs (~3.7x target); not yet a conclusion about who is behind it, but a standing watch item.
The model registry is operator-gated: deposits, FLOPs/token audits, and challenge adjudication concentrate trust in the protocol operators — the honest attack surface is faking the measurement, not winning a vote.
Verification is real and adversarially tested: a miner submitting fp8 work as bf16 was caught by full re-run and had 2 blocks scored zero-work — the first documented usefulness-fraud conviction on a PoUW chain.
tokenomics
- emission
- 715 TSC/block at start, multiplied by 3/5 each epoch; epoch lengths double (715 blocks ≈ 5 days first, up to 732,160 blocks). Issued 4,324,066 TSC at block 31,512 (2026-10-17) — matches epoch math to 0.001%.
- reward now
- 55.5984 TSC (epoch 6, since block 22,165)
- halving
- Epoch 7 at block 45,045 → 33.35904 TSC per block (x3/5 steps, not halvings)
- unlock
- Block rewards locked 100 blocks (Bitcoin-style coinbase maturity); pools pay after maturity
Don't confuse with
$TSC — cashtag cohort ticker: In X PoW-meta threads $TSC consistently means TensorCash; no other active coin claims the ticker on the tracked venues.
catalysts
Not financial advice. Verify before deploying capital.