Classic proof-of-work burns electricity to prove electricity was burned. TensorCash points the same work at a real forward pass — the proof is just a by-product of answering a real prompt. Same chain security; the joules do a job. 👇

https://tensorcash.org/blog/pow-vs-useful-pow/

TensorCash v1.0.4 is out. VDF cumulative tick self-heals with full RPC visibility, plus amber adjudication semantics, wallet fixes and Linux desktop builds. Operators: upgrade — on v1.0.3 defaults, keep the tick-floor override until you do. 👇

https://github.com/tensorcash/tensorcash/releases/tag/v1.0.4

TensorCash v1.0.5 is out — setting up has never been easier. Prebuilt, public Docker images for GPU mining workers (Hopper + Blackwell) and the full node, plus Linux node builds and macOS (Apple Silicon) miner binaries. No more building from source. 👇

https://github.com/tensorcash/tensorcash/releases/tag/v1.0.5

TensorCash mainnet, six weeks in: implied throughput has climbed from ~27 to 100,900 tok/s, and difficulty has retargeted up step by step to keep pace. Our version of hashrate is measured in tokens per second of real inference, not wasted hashes. 👇

https://tensorcash.org/explorer

Your wallet shouldn't ask someone's server what you own. The TensorCash wallet is a full node with a GUI — balances and assets verified locally, not fetched from a backend. Prebuilt for Linux; runs anywhere with Docker (VNC console: off by default, password-gated). 👇

https://tensorcash.org/blog/how-to-run-the-wallet/

⚠️ Wallet setup note: the AI-proof check needs a verifier, and the wallet doesn't come pointed at one — you set it yourself. Run the open-source verifier on your own GPU, or point your node at a verification service you trust: VALIDATOR_HTTP_URL=https://your-verifier 👇't ask someone's server what you own. The TensorCash wallet is a full node with a GUI — balances and assets verified locally, not fetched from a backend. Prebuilt for Linux; runs anywhere with Docker (VNC console: off by default, password-gated). 👇

https://tensorcash.org/blog/how-to-run-the-wallet/#wiring-up-verification-it-has-to-be-real

Meet COMPUTE — the first asset issued on TensorCash. Real AI inference is the work that secures the chain, so as compute grows, difficulty becomes a live on-chain tracker of AI-inference activity. COMPUTE is the clearest way to see the asset-issuance platform in action. 👇

https://tensorcash.org/blog/tokenised-difficulty-options/

A miner ran the committed model at fp8 on a smaller GPU and submitted the proof as bf16. Full verification reran the model at bf16: the p-values collapsed, both blocks were assigned zero work, and the chain never extended them. Trust in action. 👇

https://tensorcash.org/blog/catching-a-quantized-miner/

On 28 July, in mainnet block 19,978, a tokenised difficulty option changed owner and native TSC moved the other way — both legs inside the *same* transaction.

There was no exchange account, no deposit or withdrawal, no escrow agent and no custodian. Two people running the TensorCash desktop wallet found each other over a bulletin board, negotiated directly, and settled on chain.

The interesting part is the guarantee. Because both legs live in one transaction, there is no ordering and no "who sends first". The chain has no way to accept half a transaction, so either both sides settle or neither does — not because anybody behaved well, but because there is no state in between.

Getting two strangers to co-sign that transaction is the actual work. Their wallets opened an encrypted peer-to-peer session, each proved on-chain ownership of what it was offering with BIP-322, and each verified the other's contribution input by input before signing anything. Nonces were exchanged before any signature was seen, and every adaptor signature was checked against the exact transaction on the table.

To the network the result looks like an ordinary Taproot spend. No escrow output, no multisig, no hash-preimage, no timelocked refund branch — none of the machinery costs block space.

The write-up includes the part most announcements leave out: one available protection, a lock-step reveal of the final signatures, was negotiated off in this run. It is a fair-exchange refinement rather than a custody hole — the worst reachable outcome was a stall — and we would rather document that than imply it happened.

Technical demonstration, not an offer or solicitation.

https://tensorcash.org/blog/first-otc-trade-fully-on-chain/

⛏️ TensorCash is at 600-second blocks.

Difficulty retargeted at block 20,160 (+3.72×) and blocks have averaged 605.7s since — right on target. Early miners rode blocks arriving 4–6× faster than target, and those rewards went to the people who showed up first. Thank you.

The network is doing ~7.7M tokens/sec of real model inference — 4.6 billion tokens every block. More powerful commercial models are on the way, and our position on that upside is simple: it belongs to the miners.

Mining rewards already reach 2,700+ distinct addresses within two hops. More hashrate is very welcome.

Also coming: better wallets, a community verification service TIP, and cross-chain transfers in the Qt wallet — where we would love community help finalising it. ⚡

https://explorer.tensorcash.org

🔐 Post-quantum spending is live on TensorCash — since block zero, not a someday migration.

You opt in by funding a new address type: tc1z…

• Taproot v2, key path disabled by consensus. A tc1z is an ordinary 32-byte Taproot output, but consensus permits only script-path spends — so there is no secp256k1 key-path signature to attack. What the leaf requires is up to the script; the standard wallet template makes it a single ML-DSA check, so the prefix alone proves the key path is off, not that the policy is ML-DSA-only.

• ML-DSA (NIST FIPS 204), via liboqs. Three parameter sets — 44 / 65 (TensorCash default) / 87 — for NIST categories 2 / 3 / 5. No Shor-like quantum attack is known against its module-lattice problem.

• Holds coin and assets. Native TSC always, and issued assets too. Already on mainnet: block 20,246 moved COMPUTE Index option units into a tc1z output — held in Witness v2; an ML-DSA spend out of one reveals the leaf later.

• Fees: free to receive (that block-20,246 tx paid just 363 sats), larger to spend — a default ML-DSA-65 spend runs ~10× a standard one, since the ~3.3 KB signature only enters the witness when you spend.

Before you use it: sign on hardware you control (liboqs warns against production use for sensitive data), back up every key (no seed-phrase derivation yet), and use a fresh address per receive.

No migration, no exposed legacy balance to rotate. Whether a coin sits on the PQ rail is your choice of address.

https://tensorcash.org/blog/post-quantum-is-on-tensorcash/

📢 TensorCash v1.1.0 — coordinated network upgrade at block 20,440, expected today.

v1.1.0 moves the difficulty algorithm from the 2,016-block retarget to ASERT: difficulty adjusts every block based on how far the chain is ahead of or behind its 600-second schedule. The old rule needed weeks to respond when mining throughput changed abruptly — the cause of the recent slow blocks.

⚠️ UPGRADE BEFORE BLOCK 20,440 — nodes still on older versions may FORK off the network at activation. Expected around 19:15 BST today (block times vary). Nodes, miners, pools, and any service that validates headers must run v1.1.0 before then. Balances and history are untouched; only the difficulty rule changes.

What to expect:
• At block 20,440 ASERT starts from a bootstrap difficulty calibrated to the hashrate observed over the last two days; block times may deviate from 10 minutes briefly while it settles.
• The reset is not a giveaway: if the mining throughput seen before the last retarget returns in full, ASERT brings difficulty back to today’s level in roughly two days (~48 h).
• A fast 12-hour half-life applies until block 46,738 (~6 months), then the standard 48-hour aserti3-2d configuration (used on BCH since 2020) takes over permanently. Block 120,440 is a review point, not an automatic reversion.
• Contracts referencing block difficulty now see per-block fixings.

⛏️ Miner note: blocks carrying low VDF ticks are relayed lazily and can be orphaned — low ticks don’t cost you mining performance, but carry enough ticks that your blocks propagate promptly.

🔁 Source mirror on GitLab: https://gitlab.com/tensorcash-org/tensorcash/-/tags/v1.1.0 (downloads on the main release link).

https://git.tensorcash.org/tensorcash/tensorcash/releases/tag/v1.1.0

📜 New blog: Real-World Assets on TensorCash — the ticker is the easy part.

A blockchain moves a unit pseudonymously, efficiently, across borders. What it cannot do by itself is say what the unit MEANS. If a token represents cash, reserves, a fund or a share, someone must define the claim and sign the terms — and the asset keeps living after issuance: distributions, votes, amendments.

How TensorCash joins the two halves:

• Consensus-enforced issuance parameters — denomination, hard live-supply cap, mint/burn policy, script and KYC rules. Visible in the explorer, enforced by every node.

• The legal text travels with the asset, anchored by its on-chain control output. Public mode for broad-disclosure instruments; holder-only mode keeps the document encrypted on-chain while its commitments stay verifiable by everyone. Worked examples: a live private share-administration issuance on mainnet, and UKLTD — a public educational model — on testnet.

• Text/witness split — the canonical text is frozen and hashed; qualified e-signatures attest to that exact hash in a separate witness bundle. Holders accept the designated clauses with an on-chain 0x40 record.

• Full lifecycle in the desktop wallet — mint, burn, compliance, distributions, holder votes, and option/CFD series referencing the asset itself.

RWAs as first-class citizens: they can be financed, exchanged, and have derivatives written on them — where their governing documents and applicable law permit.

https://tensorcash.org/blog/rwa-the-ticker-is-the-easy-part/

It is the weekend—a good time to take a step back and look at what we have accomplished.

Pick a prompt. Any prompt you want. A real model responds. The response is perfectly natural and no different from a normal model response. There is no separate mining workload, but that inference can mine a block.

It is easy to take this for granted because it looks so simple and has run so smoothly. But no one had done it before: a normal model response and a mined block, produced by the same inference.

The same proof that mines TensorCash can be used as a receipt to sell the inference in the market. AI inference is set to be one of the defining markets of the next decade, and TensorCash gives independent GPU operators a chance to participate as providers—not spectators.

Newer and larger models are coming. The protocol path is already in the code; this is not waiting on a new engineering build.

Many of you are looking at difficulty and production cost today. We respect that. GPUs cost money, and electricity bills arrive at the end of the month.

But think about what difficulty looks like when a neocloud adopts the protocol across its inference workloads. Or when a Bitcoin miner converting sites into GPU datacenters sees this as the pivot. Or when a compute-constrained model provider taps into the decentralized network.

That is the point: this is not a milestone conditional on a future engineering build or development. It can happen tomorrow.

What sets TensorCash apart?

Everyone wants to do AI + crypto. It is cool. It is hyped. Every project now claims a Proof of Something.

The useful question is simple: what does each protocol actually prove?

TensorCash puts real model inference in the block and lets that inference be sold on the open market without trusting the provider. The proof both mines the block and serves as the inference receipt. Buyers verify execution, then compare providers on price and latency. The response is the work, and one inference cannot mine two blocks.

Now group the alternatives by what they claim to prove:

• Proof of Model — Keryx. It proves model weights are resident in GPU memory. Residency is not inference, and Keryx needs a separate mechanism for answer correctness. What stops one GPU claiming the same resident model across multiple reward networks?

• Proof of GPU, capacity and attestation — io.net, Aethir, Phala. These prove hardware, availability or an attested stack. What proves that capacity is not double-booked across networks? Attestation trusts the vendor chain—not the inference.

• Proof of MatMul — BTX, Pearl. The proof rewards matrix work rather than the customer response. If mining while inferring makes both slower, rational operators specialize: mine, infer, or split the hardware. The efficient miner strips away inference. What remains is a GPU puzzle—not an inference receipt. Nockchain points to the same arbitrary-MatMul cuPoW as a future AI route, so the same critique applies.

• Proof of Transformer and benchmarked capacity — Gonka. It benchmarks a randomized transformer Sprint, awards compute power, then allocates inference. The benchmark is not the service. TensorCash makes the service itself proof-of-work.

• Prover–verifier networks — Bittensor. Miners produce subnet work; validators score it. That proves who validators rewarded, not which model executed. Its documentation describes weight copying: predicting consensus can replace evaluation.

• Proof of Logits — Ambient. “Hardware agnostic” needs a specification: cross-GPU calibration, false-accept and false-reject rates, with boundaries that tolerate honest variation but catch cheaper execution.

TensorCash already did the verification work. Every node runs deterministic code with calibrated, model-specific statistical boundaries. Honest GPU variation passes. The live network distinguished bf16 from fp8 on the same model; the fp8 blocks received zero work.

The published evaluation measured single-digit proof-emission overhead. Production mining software improved logging performance and brought measured overhead to zero; the earlier residual was a logging artefact.

That full combination is the distinction: real inference as block work, one-use work, deterministic verification, and a receipt that lets the inference be sold on the open market without provider trust. Unknown operators and datacenters compete on price and latency.

If another live chain does all of this, send the block, proof, verifier and end-to-end throughput numbers. We are here to talk.

Read the proof, not the pitch:

https://tensorcash.org/blog/what-sets-tensorcash-apart/

🔗 The reply that mines the chain

Protocols call themselves ecosystems long before they have one. The test is whether people you have never met build things you did not commission — then put them in front of real users.

That is beginning to happen on TensorCash:

🤖 Bob Labs — chat.tensorcash.boblabs.eu: a full chat + agent-orchestration platform (agents, multi-agent labs, tools, MCP, memory, cron, and API usage prepaid in TSC) running inside a mining pool. The inference behind each reply becomes mining work. No card or identity is requested at sign-up.

🏢 AskTilly — an enterprise-grade chat and agent platform: passkey-first and end-to-end confidential, with an API and a self-custody WebWallet (client-side signing, passkey PRF unlock, SegWit, Taproot and post-quantum addresses, and signed desktop builds).

⛏️ Three public pools: luckypool.io, tsc.suprnova.cc and Bob Labs. Three pools do not guarantee decentralisation, but they give miners something essential: a choice.

Running something on TensorCash? Add it to the community directory — one PR.

https://tensorcash.org/blog/third-party-ecosystem/

✨ **TensorCash Magic #1: mining inference is normal inference.**

Ask an AI a question. It answers exactly as it normally would.

Same model. Same next-word probabilities. Same answer quality.

But that useful inference is also doing the work of mining.

There is no dummy calculation running beside it. No weaker “mining model.” No trade between a useful answer and network security.

So where is the magic?

Inside every AI answer is a hidden coin flip. At each step, the model proposes possible next tokens — *Light*, *Sunlight*, *Rainbows* — with different odds. A sampler draws a number and chooses one. Then the chosen word becomes part of the context, and the model does it again.

A normal chatbot gets that draw from a private RNG. TensorCash derives it from the chain and the answer so far.

The model never changes. The odds never change. The quality never changes. But the full sequential inference path is now bound to the chain, replayable and verifiable.

**Normal inference goes in. Verifiable mining work comes out.**

Watch three answers grow token by token: https://tensorcash.org/blog/chain-derived-sampling/

https://tensorcash.org/blog/chain-derived-sampling/

Welcome to the Bitcoin genesis era. At block 22,165, the TensorCash subsidy moves from 92.664 to 55.5984 TSC — close to Bitcoin's original 50 BTC subsidy. Same ~21M destination; a different road there. Height, not date, decides it.

https://tensorcash.org/blog/welcome-to-the-bitcoin-genesis-era/

People often assume a useful-proof-of-work chain has to choose one AI model forever. TensorCash does not.

A model joins through an on-chain registration: a 5 TSC sponsor deposit, a 100-block window and at least 50 supporting commits in valid blocks. That makes admission a hashpower-backed consensus decision—not an admin switch.

Every registered model then competes for the same next block. nBits provides the network-wide base target; nAdjBits gives each model an adjusted lane based on its registered difficulty. Miners choose among those lanes using the economics of block rewards, inference demand and operating cost.

And no extra validator class is needed. The statistical tests are protocol-defined; nobody scores whether an answer “looks intelligent.”

One chain. N models. Quick explainer:

https://tensorcash.org/blog/how-tensorcash-runs-multiple-models/

TensorCash is now listed on two exchanges.

Every coin in circulation was mined by real model inference — no premine, no allocation. Until now, moving between that work and the wider world mostly meant OTC or on-chain trades. This makes the path wider: miners who earn coins by serving inference and users who hold or spend them now have more ways in and out.

A stepping stone, not a destination — the goal remains an open compute market where inference is sold with a verifiable receipt. But stepping stones are worth celebrating.

And as ever: an exchange account is not a wallet. Your keys, your coins.