A DAO building
decentralized AI compute.

A new consensus class. Power to the people.

DAIT does not run Proof of Work. It does not run Proof of Stake alone. It does not run Proof of Useful Work in the narrow sense the term has acquired since 2017. DAIT runs Proof of Intelligence. PoI is the consensus class formally defined in Livepaper section 3, revision 36; the idea has been in development for over a decade across multiple iterations (aiscendence.com in 2018, dait.io in 2020, this chain in 2026). The mission is to put the foundational compute layer in the hands of the people. Power to the people.

The three conjunctive conditions

The conditions are conjunctive on purpose. Drop any one of them and the resulting protocol exists already, has a name, and has a known failure mode. The DAIT design proposition is that the failure modes of PoW, of PoS, of subjective PoUW marketplaces, and of staid verified-compute auctions are all visible in production today, and that fixing them required a stricter conjunction. That conjunction is what we are calling Proof of Intelligence.

Condition (a): cognitive unit-of-work

The unit-of-work the protocol pays for is a cognitive computation, by which we mean a computation whose value derives from the model or the algorithm rather than from the cost of producing it. Proof of Work is the canonical counter-example. A SHA-256 collision is expensive to produce and trivially easy to verify, but it has no use beyond the verification itself. The economic argument for PoW is that an attacker who wants to rewrite history has to redo the work; the social argument is that the network is willing to burn power to make that argument durable. PoI rejects this trade. The work the network pays for has to be useful to someone who is not the verifier. Useful means: an inference call that returned a token, a finetune step that updated a LoRA adapter, a simulation step that advanced a state vector, a reasoning trace that produced an answer. Storage retention by itself does not qualify; storage is necessary infrastructure but it is not cognition. Subjective creative output does not qualify either, because the verification path collapses back into reputational scoring. Cognitive unit-of-work is the dividing line between consensus that costs the world energy and consensus that costs the world inference cycles a tenant was paying for anyway.

Condition (b): adaptive verification

The verifier function is plural. PoW has one verifier: hash the header, check the leading zeros. PoS has one verifier: check the validator signatures and the chain rules. Both work because the unit-of-work is a single cheap-to-verify object. Cognitive work is heterogeneous. An LLM token-generation receipt is verified one way, by reading a TEE quote, comparing the model hash, and checking the signature. A tensor-network simulation is verified another way, by spot-decomposing the published tensor train, sampling on a few network indices, and comparing against a reference. A storage shard is verified a third way, by issuing a pseudorandom challenge inside the provider's enclave. A regulated credential is verified a fourth way, by checking a zero-knowledge selective-disclosure proof against an issuer's accumulator. Adaptive verification means the verifier function is itself a value, selected per workload class from a registered set. The chain knows that an "inference receipt" needs the TEE path, a "long-running training job" needs the redundant-re-execution path, a "scientific simulation" needs the spot-decomposition path. The set of verifier functions is governance-extensible, because new verifiable workloads will be invented after genesis and the chain must accommodate them without a hard fork.

Condition (c): state-indexed economic parameters

The third condition is the most distinctive. Most chains set their primary economic parameters at genesis and only change them through ad-hoc governance votes. Bitcoin's emission curve, Ethereum's slot time, even Solana's leader schedule are constants from the perspective of the protocol; they change only when the human governance process decides to change them. PoI requires that at least one primary economic parameter is a continuous function of network state, recomputed every epoch by the protocol itself. DAIT sets several. The active validator cap is the floor of the sum of the square root of average self-bond, the square root of unique 30-day PoUW hosts, and the square root of average blocks-per-second, recomputed every three days. The per-block emission rate is indexed to the chain's running compute throughput, so a busy network mints more and a quiet one mints less. Per-workload credit weighting is governance-tunable but defaults are state-aware. The block time itself is scheduled to drop from 1000ms at genesis to 500ms in the Q4 2026 upgrade window, contingent on observed validator latency distributions. The point is not that any one of these knobs is unprecedented in isolation; some chains have one, some have two. The point is that the protocol's behavior is allowed to be a function of its own use, on a schedule the protocol maintains, and that this is now a stated condition rather than an emergent property.

Comparison: PoW vs PoUW vs Verified Compute Marketplace vs PoI

	PoW (Bitcoin, Litecoin)	PoUW (Primecoin, Coin.AI, generic)	Verified Compute Marketplace (Akash, io.net)	Proof of Intelligence (DAIT)
(a) Cognitive unit-of-work	No. SHA collisions.	Yes, narrow (one workload type, fixed verifier).	Yes, but verification is reputational or absent at the protocol layer.	Yes. Inference, finetune, simulation, reasoning, scientific compute.
(b) Adaptive verification	No. Single hash check.	No. Single workload-specific verifier per chain.	Partial. Multiple attestation paths, but verifier choice is contractual not chain-level.	Yes. TEE + cryptographic + statistical + behavioral, dispatched per workload class.
(c) State-indexed economic parameters	No. Difficulty adjusts but emission curve is hardcoded.	No. Emission and validator set typically constant.	No. Marketplace clearing is dynamic, but base-layer parameters are not.	Yes. Active validator cap, emission rate, credit weighting, block time, all indexed.
Resulting class	Proof of Work	Proof of Useful Work	Verified compute marketplace	Proof of Intelligence

Why this matters strategically

PoI is not a marketing relabel. The reason to coin a new term is that the design space PoUW occupies has filled with projects that satisfy condition (a) and nothing else, and the design space verified-compute marketplaces occupy has filled with projects that satisfy (a) and (b) but treat the base layer as inert. DAIT's design choices, the agent module, the state-channel module, the elastic emission curve, the dynamic validator cap, only make sense together. Calling the result "PoUW" undersells the architecture and groups DAIT with chains that share none of its key properties. Calling it "verified compute" ignores the adaptive economic surface. The honest description is that DAIT is the first chain we know of that explicitly satisfies all three conditions as a designed property, and the honest thing to do is name that property.

Beyond labelling, PoI is a falsifiable claim. Each condition has a test. Condition (a) is satisfied if and only if the work paid for is reproducibly useful to a non-verifier, which is observable. Condition (b) is satisfied if and only if there is more than one verifier function registered and the chain dispatches between them, which is testable from the runtime. Condition (c) is satisfied if and only if at least one primary economic parameter is a continuous function of state, which is verifiable from the chain's own code. The Foundation's standards page lists each condition as a control row and tracks status against a live signal feed.

Read the deep dive on PoI   Livepaper section 3   PoI controls in the standards matrix

Decentralized compute. Side by side.

A frank comparison with the existing decentralized AI and compute networks. We are deliberate about what we do differently and why.

vs Centralized AI clouds (AWS, GCP, Azure)

Trust model	Centralized: trust the cloud provider's audit reports. DAIT: hardware-attested, every call cryptographically verifiable.
Pricing	Centralized: set by the cloud, opaque, contract-gated. DAIT: reverse Dutch auction, transparent, public order book.
Workload control	Centralized: AUP gates, can deplatform. DAIT: permissionless. Hosts opt-in to license rules.
Margin to provider	Centralized: 60-80%. DAIT: 60% to host, 30% to validators, 10% to community/safety.

vs Bittensor (TAO)

Verification	Bittensor: subjective scoring by validators. Vulnerable to weight-copying, addressed by commit-reveal. DAIT: hardware-attested receipts. Validator scoring is verification of cryptographic proofs, not subjective evaluation.
Token supply	Bittensor: 21M cap with halving. Bitcoin-style emission cliff in long term. DAIT: compute-indexed elastic emission. No cliff. Supply expands with demand and contracts with burn.
Subjective work	Bittensor: subnets compete on subjective quality (text, image). Hard to verify. DAIT: only deterministically-verifiable work earns consensus credits. Subjective work is paid via marketplace fees.
Account model	Bittensor: coldkey/hotkey two-key separation. DAIT: same two-key separation, extended to first-class agent accounts that can hire each other.

vs Akash Network (AKT)

Marketplace pattern	Akash: reverse Dutch auction, audited attributes registry, general containerized compute. DAIT: same reverse-auction pattern plus AI-specific verification layer Akash does not have.
Verification	Akash: no native AI verification. Audit attributes are static claims signed by trusted parties. DAIT: hardware attestation per inference call. Cryptographic proof of correct execution.
Workload type	Akash: general containerized compute, not AI-specific. DAIT: AI compute first-class. State channels for high-frequency inference. AI-agent accounts native.
Tokenomics	Akash: uncapped to 388M cap pivot. Uncertain narrative. DAIT: compute-indexed elastic emission, designed once, robust across bull/bear.

vs io.net (IO)

Hardware verification	io.net: initially relied on self-reported specs. Caught with counterfeit GPU claims April 2024. Retrofitting KYC and benchmarks since. DAIT: TEE GPU UUID attestation required at registration. Self-reported specs not accepted. Hard registration check.
Base layer	io.net: Solana. DAIT: sovereign L1 with attestation reporting native to consensus, not bolted on as a contract.
Pricing	io.net: fixed-rate ask-only. DAIT: reverse Dutch auction. Lower clearing prices.

vs Render Network (RENDER) and Aethir (AETH)

Job assignment	Render, Aethir: centralized scheduler. Foundation-controlled tier system. DAIT: reverse Dutch auction with attestation filters. No central scheduler.
Verification	Render: no cryptographic verification, "Proof of Render" is reputational. Aethir: proof-of-capacity (idle benchmark) plus proof-of-delivery. Idle subsidies inflate emissions. DAIT: hardware attestation per inference. No idle subsidies. Pay only for verified work.
Trust source	Render, Aethir: foundation curates tiers and quality scores. DAIT: on-chain audit attribute registry, multiple auditors, governance-managed trust.

vs Centralized AI APIs (OpenAI, Anthropic, Google)

Most AI today is consumed through closed inference APIs. The trade is straightforward: a tenant sends a prompt, the provider returns a completion, and the tenant pays per token. The provider's word is the only evidence that the requested model ran. There is no receipt, no proof, and no recourse if the served model was a quantized variant or a cheaper substitute. DAIT inverts this. Every inference call returns a hardware attestation that names the model hash, the host's GPU UUID, and the input/output digest. A tenant can verify on-chain that the inference paid for is the inference performed.

Reproducibility on centralized APIs is structurally hard. Providers ship silent updates, retire snapshots, and reroute traffic to cheaper backends without notice. DAIT inference is pinned to a content-addressed model hash. The same hash run on attested hardware returns the same result. Audits, regulatory submissions, and scientific work that depends on byte-for-byte reproducibility have a path on DAIT that they do not have on a closed API.

Trust model	Centralized API: trust the provider's word that the requested model ran. DAIT: hardware attestation per call. Model hash and GPU UUID on every receipt.
Reproducibility	Centralized API: providers update models silently. Yesterday's outputs may not be reproducible tomorrow. DAIT: content-addressed model hash. Same hash, same hardware, same output.
Pricing	Centralized API: per-token rate cards set by the provider. Volume discounts gated by enterprise contract. DAIT: reverse Dutch auction. Public order book. Hosts compete on price.
Workload acceptance	Centralized API: AUP-restricted. Whole categories banned. Account closure with no appeal. DAIT: permissionless tenancy. Hosts opt-in to license rules per workload class.
Prompt privacy	Centralized API: prompts and completions visible to provider, retained for abuse review and training where ToS allows. DAIT: inference runs inside a TEE. Prompts decrypted only inside the enclave, never visible to the host operator.
Data residency	Centralized API: traffic routes to whichever region the provider chooses. Regional pinning is contract-only, unverifiable. DAIT: tenants filter on host region attribute, attested by the host's audited registry entry.

vs Bitcoin L2 / sidechain proposals for AI

A handful of projects propose hosting AI workloads on Bitcoin layer-2 systems such as BOB, Stacks, and various sBTC-style sidechains. The pitch is Bitcoin-grade settlement security for AI compute. The reality is harder. Bitcoin's settlement cadence is measured in tens of minutes. Anchoring a sidechain to Bitcoin inherits that latency for any operation that needs final settlement on the base layer. AI inference at a useful tempo cannot wait for Bitcoin confirmations. BTC L2s also lack the chain-level primitives that AI needs: marketplace order books, attestation registries, agent accounts, state-channel settlement. These can be approximated by smart contracts, but bolt-on contracts pay a recurring tax in gas, latency, and audit surface compared with chain-native modules.

DAIT plans to integrate Bitcoin restaking via standard restaking infrastructure once that ecosystem stabilizes. The result is BTC-pooled security where economically appropriate, with AI primitives that remain native to DAIT rather than retrofitted onto a script-limited base layer.

Throughput	BTC L2s: hundreds of TPS in optimistic conditions, lower under contention. DAIT: thousands of TPS at the chain layer, with state channels moving 99 percent of inference traffic off-chain.
Settlement latency	BTC L2s: sidechain finality is fast, but BTC anchoring takes hours. Cross-L2 settlement inherits that delay. DAIT: sub-second pre-confirm, ~6 second finality on the DAIT chain.
Native AI primitives	BTC L2s: none. AI features live in user-deployed contracts. DAIT: nine chain-native domain modules covering attestation, marketplace, agents, payments, governance, slashing, treasury, industrial data lake, and pseudonymous credentialed identity.
Programmability	BTC L2s: Bitcoin script is intentionally restricted. Most BTC L2s add EVM, but lose bridging guarantees in the process. DAIT: EVM at launch with planned DAIT-specific VMs for agent execution and inference policy.
Bitcoin-grade security	BTC L2s: sold as the primary feature, but bridge designs vary widely in honesty. DAIT: planned Bitcoin restaking integration via standard restaking infrastructure. Security parity where useful, AI primitives where Bitcoin scripting cannot reach.
Workload fit	BTC L2s: payments-first. AI is an afterthought. DAIT: AI compute first-class. Payments are one module among nine.

vs zkML rollups (zero-knowledge inference proof systems)

zkML is the research direction that produces zero-knowledge proofs of inference: a tenant gets a succinct proof that some model ran on some input and produced some output, with no need to trust the host. The math is real and the trust model is appealing. The cost is the constraint. Generating a ZK proof of a transformer forward pass is currently 100 to 1,000 times slower than running the inference itself, and on-chain verification of the resulting proof is still expensive enough that zkML is unsuitable for high-throughput consumer inference. Hardware-enclave attestation has roughly 5 percent overhead on confidential-compute-capable GPUs in production today, and the attestation quote is a few kilobytes that can be verified for cents.

DAIT runs TEE attestation as the hot path because that is what the economics support today. We also expose zkML as an opt-in premium tier for high-value workloads where TEE trust is insufficient: regulated healthcare, financial model audits, government work. As zkML hardware and proving systems improve, the cost gap will close, and DAIT's marketplace will route accordingly. The choice is the tenant's, priced by the market.

Inference speed	zkML: 100x to 1000x slower than raw inference today. DAIT TEE path: approximately 5 percent overhead vs bare-metal on confidential-compute-capable GPUs.
Trust assumption	zkML: mathematical. Soundness of the proving system, no hardware trust. DAIT TEE path: hardware vendor honesty plus revocable attestation. Compromised keys can be revoked at the registry layer.
Proof cost	zkML: proofs are expensive to generate, verifying on-chain is non-trivial gas. DAIT TEE path: attestation quotes are kilobytes. Verification is a signature check.
Tooling maturity	zkML: 2026 state is research-grade. Multiple zkML proof-system implementations are active and improving but not yet production-ready for general transformer workloads. DAIT TEE path: CPU and GPU hardware-enclave confidential-compute modes have been in production for years.
Workload fit	zkML: low-throughput, high-value, audit-critical workloads where the speed penalty is acceptable. DAIT TEE path: general-purpose inference at production speed.
DAIT's hybrid	Default: TEE attestation as the hot path. Opt-in premium tier: zkML for workloads that demand it. The tenant chooses, the market prices the choice.