Bittensor (TAO) sustainability report

Bittensor is present on the following networks: Bittensor.

Two distinct mechanisms operate here and conflating them produces a misleading description. The first is an ordinary chain consensus: a proof-of-stake network built on a general-purpose blockchain framework, where holders nominate validators, block production runs as a randomized slot lottery and finality is reached by a separate voting process over chains rather than individual blocks. This is what orders transactions and secures the ledger, and it resembles other networks in the same framework family.

The second mechanism is what the network exists for and has no counterpart elsewhere. The network is divided into subnets, each an independent market for a particular kind of machine-produced output. Within a subnet, participants in a producing role submit work and participants in a validating role evaluate it, scoring each producer on a normalized scale. Those scores are submitted on-chain as weights and aggregated by a consensus algorithm that weights each evaluator's opinion by the stake behind it and produces a ranking reflecting the stake-weighted median view rather than any individual judgment. Rewards are then distributed according to that ranking. The design deliberately makes it unprofitable for an evaluator to score dishonestly, because rewards accrue to those whose assessments agree with the stake-weighted consensus, and an evaluator who deviates persistently earns progressively less.

A restructuring introduced per-subnet tokens and shifted the allocation of rewards between subnets from a governance-style vote toward a market mechanism driven by where participants commit capital, so subnets compete for emissions rather than receiving an administered share.

On penalties, precision matters. Participation as an evaluator requires meeting a stake threshold and ranking among the top holders by stake weight, and failing to maintain that standing removes the permit to participate, with associated bonds deleted. Deviating from consensus reduces what an evaluator earns. This text does not describe the protocol as slashing principal in the sense used by networks that confiscate a bond for equivocation; the documented mechanisms operate through reduced emissions and loss of participation rights.

Bittensor is present on the following networks: Bittensor.

Emissions are issued per block and divided between three groups: the producers whose output was ranked highest by consensus, the evaluators who scored them, and the owner of the subnet in which the work took place. A producer ranked below the effective threshold earns nothing at all and eventually loses its registration, so the distribution is sharply concentrated rather than broadly shared, and the marginal participant is not merely underpaid but unpaid. Evaluators earn in proportion to stake and to the degree their assessments agreed with the consensus view.

Entering a subnet as a producer requires paying a registration cost that adjusts with demand for places, which prices access and limits indiscriminate registration. Holders who do not wish to operate infrastructure may commit stake to an evaluator and share in its earnings after a commission, in the manner of delegation elsewhere. Following the restructuring, committing stake to a particular subnet converts it into that subnet's own token, so the holder's return depends on that subnet's success rather than on the network as a whole, and capital flows between subnets are what determine how emissions are allocated across them. The issuance schedule includes periodic halvings that reduce the rate of new supply.

Transaction fees on the chain itself are modest and are not where the economics of participation lie. The dominant cost for a producer is not a protocol fee but the computing hardware required to generate output good enough to be ranked highly, and the dominant cost for an evaluator is the infrastructure needed to assess that output. This is the defining feature of the network's incentive structure: the protocol is a payment and scoring layer sitting on top of a large amount of off-chain computation, and the real competitive expenditure happens in hardware rather than in fees or bonded capital. Participants who cannot cover that hardware cost exit, in much the way unprofitable mining equipment is switched off elsewhere, though nothing is confiscated when they do.

Bittensor is present on the following networks: Bittensor.

The figure reported for this network covers the infrastructure that runs the chain itself, and the boundary matters more here than for most networks, so it is stated first. The chain is a proof-of-stake network of the ordinary kind, and its validator and full-node population is estimated in the ordinary way: node counts drawn from public listings and network crawlers, representative hardware inferred from the published requirements for running the software, and measured power draw for machines of that description applied across the population with idle draw weighted heavily.

What the boundary leaves out is larger than what it contains. The work this network pays for is machine learning inference and training carried out by participants in the producing role, and that work runs on graphics processing hardware whose power draw is an order of magnitude above that of a validator node and which is loaded continuously rather than intermittently, because a participant who stops computing falls in the ranking and stops earning. That consumption is real and it is not in the reported figure.

It is excluded because it cannot presently be estimated to a standard that would justify publishing it. Registered participants per subnet are readable on-chain, which bounds the count, but the hardware behind each registration is not disclosed, and it varies enormously between subnets because the computational demands of the different kinds of output being produced are not comparable. An estimate built by assuming a configuration and multiplying it across registrations would carry an error range wide enough to make the result misleading in either direction.

One consequence deserves stating. Rewards accrue only to participants ranked above a threshold, so there is a standing incentive to deploy as much capable hardware as expected earnings justify, which makes the network's true consumption respond to reward value in a manner closer to proof of work than to a conventional stake-based chain. The reported figure, resting on a validator population that does not move in that way, is stable while the quantity it omits is not. It should be read as a floor for this network rather than as a total, and it will be restated as the compute population becomes observable.

Bittensor is present on the following networks: Bittensor.

The renewable share reported here is derived from the same boundary as the consumption figure, and therefore from the chain's own validator and full-node infrastructure rather than from the computing population the network pays for. That infrastructure is distributed in the manner typical of proof-of-stake networks and is located through network observation, hosting provider address ranges and what operators choose to disclose. Each location is matched to published statistics on how electricity is generated in that region, and the renewable share is the average across regions weighted by the consumption attributed to each.

The exclusion described in the methodology section shapes this figure as much as it shapes the consumption one, and in a way that is not neutral. Concentrations of graphics processing hardware are sited according to electricity price, available grid capacity, cooling conditions and the availability of suitable premises, which is a different set of considerations from those governing where someone runs a validator. A renewable share computed across the computing population would therefore not be expected to match the one reported here, and there is no basis at present for saying in which direction it would differ. The figure describes the grid mix behind the chain's own infrastructure and should not be read as characterizing the network's activity as a whole.

Two limits apply that are common to this kind of estimate. Grid statistics are annual and regional, so seasonal and daily variation is invisible and a regional average stands in for a specific facility. Contractual renewable purchases are not counted, since what is described is the physical grid mix rather than a procurement arrangement.

Energy intensity per transaction should be read with unusual care. It is period consumption within this boundary divided by transactions settled, so it is an accounting average across the chain's infrastructure and not the energy that an additional transaction causes. Source data is processed by Our World in Data from Ember and the Energy Institute's Statistical Review of World Energy: Share of electricity generated by renewables.

Bittensor is present on the following networks: Bittensor.

Emissions are derived from the consumption estimate by applying a grid carbon intensity to each location in the distribution described for the renewable share, and summing across the chain's validator and full-node infrastructure. The boundary is therefore the same one used throughout: operational electricity for the machines that run the chain, and not the graphics processing hardware that performs the work the network pays for. Hardware manufacture is excluded as well, in common with figures of this kind.

Scope 1 covers emissions from sources under the operators' direct control, principally on-site combustion such as backup generation. Across commercial hosting this is a negligible contributor and is reported as such rather than modeled in detail. Scope 2 covers emissions embodied in purchased electricity and constitutes effectively the whole of the reported footprint.

The structural point raised in the methodology section has a direct consequence here. Because rewards flow only to participants ranked above a consensus threshold, and ranking depends on computational output, an increase in the value of rewards draws in additional hardware and raises the network's real emissions with it. That relationship between reward value and emissions is characteristic of proof-of-work networks and is unusual in a network whose consensus is stake-based. It does not appear in the reported figure, because the population it acts on sits outside the boundary. Readers comparing this network against others on the reported number are comparing validator infrastructure against validator infrastructure, which is a narrower comparison than the number alone suggests.

Greenhouse gas intensity per transaction is period emissions within this boundary divided by transactions settled, and carries the same caveat as the energy intensity: it is an accounting ratio rather than a marginal figure. Figures are restated each period, and the omission described here will be closed if the hardware behind subnet registrations becomes observable. Carbon intensity data is processed by Our World in Data from Ember and the Energy Institute's Statistical Review of World Energy, and is made available under a Creative Commons BY 4.0 license: Carbon intensity of electricity generation.