Hetu

🚨 Live: Hetu 3.0 Deep Intelligence Money— a full upgrade for an AI-native abundance economy. Years of deep tech & research, now an agentic-ready production stack: 3-Money ( $HETU / $USDAI / $FLUX), EVM causal DAG, and verifiable #PoCW/ #PoSA. Auditable, financeable intelligence — unlocked. 🧵1/6 #DeepIntelligenceMoney #AI #AIagent

encoder, prefill and decode on separate GPUs now. nice work from vLLM tbh it makes one question way harder. one request touches three pools of GPUs, so who gets paid and how much? that's pretty much the thing we're building at Hetu

One request, four peers. One parses the question, one fetches the data, one runs the analysis, one assembles the answer. Who gets paid, and how much? On Hetu it isn't split by GPU time. It's split by contribution: roughly, how much worse the outcome would have been without each peer's part. The exact counterfactual is intractable, so the network approximates it with a Shapley value computed locally on the causal graph around the request. Credit also decays along long chains, so being far upstream doesn't inflate anyone's share. A peer that burned compute without changing the outcome gets little. Each request only pays the few that made it better. $HETU

A model ships with a promise: MIT license, no backdoors. A few weeks later an update quietly changes what it does. Who notices? On Hetu that promise can be posted as a bond. The model declares what it is, puts up collateral, and carries a drift index: how far its actual behaviour has moved from what it declared. The penalty is written in advance, for example drift above 0.1 for 24 hours triggers a slash. Validators score its actions against the declaration, from 0 to 1. So the promise isn't checked once, at launch, by the people who made it. It keeps being checked while the model is in use. #DeepIntelligenceMoney $HETU

Joining Hetu takes a TEE registration and a small bond. There's no whitelist. We chose that on purpose. A whitelist decides in advance who's allowed to do useful work, and someone has to maintain it. We'd rather let anyone try and make cheating expensive. So a node is judged on what it delivers, and the bond is what it stands to lose. Anyone can challenge a result. If the challenge holds, the bond is slashed. Would you trade that openness for a curated list? #DeepIntelligenceMoney

Some things shouldn't be for sale. On Hetu, Flux is one of them. Flux is only mined, through work the network has verified and attributed. It can't be transferred, so it can't be sold either. The whitepaper puts it simply: one cannot sell their earned experience. It can be redeemed, one way, into a stablecoin through a liquidity pool. So contributors can still cash out, but nobody can show up and buy a track record they didn't earn. What a contributor earns has two parts: economic contribution, measured by PoCW, and semantic alignment, measured by PoSA. Governance sets the weight of each.

A GPU doing a matrix multiplication can't tell you who asked for it. Neither can anyone checking the math afterwards. That's why Hetu pays for requests that were served, not for computation as such. Every consumer emits a receipt for the peers it routed a request to. Each epoch the receipts are aggregated into a weight matrix, and fees are split in proportion to it. A node's rank only rises when staked consumers actually depend on its output. Compute that no one routed a request to leaves no receipt, and earns nothing. What would you accept as proof that work was useful? $HETU

What happens to work that fails its check on Hetu? It never enters the record. An event joins the causal graph only once it verifies: the signature matches the agent's key, the clock correctly follows the events it claims to build on, and the task itself passes. If any of that fails, the graph stays exactly as it was after the last thing that did verify. That matters for everything downstream. Nothing can be built on an event that isn't there, so a bad result can't quietly collect credit through the work that follows it. A result has to verify before it can count for anything.

Three numbers decide what a node earns on Hetu, and it has to score on all three. Rank: how much staked demand depends on its output. Legitimacy: whether its work is attested, reproducible and consistent with the causal graph. Connectivity: how well it's connected to the honestly staked part of the network. The reward is their product, so a zero on one wipes out the other two. That's deliberate. A group that only rates itself can push its rank up, but it stays cut off from the main graph, and without more than half of the attested stake it can't keep earning unless it does real, verifiable work. Collusion isn't forbidden. It just doesn't pay. What would you try first to game it? $HETU

Can a single idle GPU take part, or is this only for compute centers? Both, by design. The network is built to pool very different hardware into one inference layer, from a rack in a compute center to one card under a desk. Take one request. It's priced by supply and demand and settled per inference, whoever serves it. A compute center simply serves more of them. The rules don't change with size: every node stakes and builds a reputation, so cutting corners costs it something. Next: what stops a group of nodes from simply vouching for each other. #DeepIntelligenceMoney $HETU

We'd rather undersell TEE attestations than oversell them. Every event on Hetu carries one: a measurement of the code and data that ran, plus a nonce and a time. It tells you this exact code ran on this exact input inside an enclave. It doesn't tell you the answer was any good. That's a separate check: deterministic verification for tasks code can check, a 2/3 quorum for the rest. What attestations do feed is a node's legitimacy score, together with reproducibility proofs and DAG consistency checks. A successful challenge lowers that score, triggers a replay in the TEE, and slashes the node's bond. #DeepIntelligenceMoney

Two agents act at the same moment on opposite sides of the network. Which one came first? The tempting answer is to compare timestamps. But machines don't share a clock, so a timestamp only tells you what one machine believed. Hetu orders events with vector clocks instead. Each event carries a record of what its agent had already seen. If agent B had seen A's event before acting, B's clock shows it, and B is ordered after A. If neither had seen the other, the clocks can't be ranked, and the two events are recorded as concurrent. Sometimes the honest answer is "neither". A ledger for AI work has to be able to say that. $HETU

Every block reward on Hetu is split between the hub and the subnets, and the split isn't fixed. The more subnets there are, the larger the subnets' share. It grows logarithmically with the subnet count and stops at 90%. Inside that share, each subnet is paid in proportion to its TVL. The hub keeps the rest: 2% to the community pool, 5% to validators, 93% to delegators. Why logarithmic, and why a cap? The first subnets see their share rise quickly, which pulls the ecosystem outward. The cap leaves at least 10% of every block for the hub, which pays the validators and delegators who keep the network running. Would you have put the cap higher or lower?

1 $HETU per block. That's where issuance starts, and the cap is 21M. What we changed from Bitcoin is what triggers a halving. It isn't a block height. It's how much of the supply has actually been issued. Half of the 21M out: the block reward drops to 0.5. Three quarters out: 0.25. 87.5% out: 0.125. So halvings follow the network's real inflation instead of a fixed block count, and issuance flattens toward the cap. Tomorrow: where each block reward actually goes. #DeepIntelligenceMoney $HETU

Why does a network of GPUs need its own asset at all? Because the people running inference and the people asking for it don't know each other. Every request has to be priced, paid for and trusted between strangers, in real time. So the network settles per inference, lets prices move with supply and demand, and asks nodes to stake and build a reputation, so a node that cheats has something to lose.

Most mining proves that energy was spent. We think mining should prove that something useful was served. On Hetu, the work a GPU gets paid for is inference: a model answering a real request. The compute that serves the request is the compute that earns, whether it sits in a compute center or on a single idle card. That's why we still call $HETU an AI Bitcoin. Same discipline, different work. #DeepIntelligenceMoney

HETU is the native asset of the sovereign AI economy. Sovereign means the intelligence isn't rented from one provider. It runs on a peer-to-peer network of GPUs with no central API to switch off, and requests and model weights stay encrypted end to end. A native asset is how that network pays for itself. GPUs earn by serving real inference, not by solving puzzles nobody needs. $HETU