Algorand (ALGO) sustainability report

Algorand is present on the following networks: Algorand.

Algorand reaches agreement through a pure proof-of-stake protocol in which every unit of the network's native asset held in an online account carries the same weight, and in which no bonding, no delegation to a fixed validator set, and no minimum hardware commitment is required to take part. Instead of electing a known roster of block producers, the protocol uses cryptographic sortition. At each step of each round, every participating account evaluates a verifiable random function locally, using a seed derived from the chain itself together with its own private participation key. The output tells that account privately whether it has been drawn and with what weight, and it carries a proof that anyone else can check once the account speaks. Because the draw happens in secret and becomes visible only when a selected account broadcasts, an adversary has nobody to attack, bribe or censor in advance.

A round proceeds in three steps. A small set of accounts is drawn to propose a block, and the proposal carrying the strongest sortition credential prevails. A soft-vote committee, sampled afresh, converges on a single proposal. A certify-vote committee, sampled afresh again, votes to commit that block to the ledger. Every step draws a new committee, so subverting the members of one step buys nothing for the next. Once a block is certified it is final; the chain does not fork in normal operation, and no confirmation depth or challenge window applies. Safety holds while honest accounts control more than two thirds of the online weight, and the design is deliberately biased toward never producing two conflicting ledgers, so a severe partition stalls the chain rather than splitting it.

Participation is opt-in and non-custodial. A holder registers short-lived participation keys against an account and keeps a node online; the balance backing that account never moves and is never bonded. Nodes exchange messages over a peer-to-peer gossip layer. The protocol itself changes by node-runner voting, in which produced blocks count as ballots and a large supermajority is needed to adopt a new consensus version, followed by a fixed cooldown before activation. The version adopted in August 2026 added native post-quantum account signatures alongside the existing signature scheme.

Algorand is present on the following networks: Algorand.

Under the incentive arrangement that took effect in January 2025, the account that proposes a block is paid the moment its block is certified, and the payment is credited directly to the account balance rather than appearing as a separate transfer. The payout combines two sources: a share of the fees the protocol has collected in its fee sink, and a bonus drawn from a reserve funded by the network's foundation. The bonus begins at a fixed amount per block and decays by a small fraction every million blocks, so the subsidized component tapers away over time and fee revenue is intended to become the durable source of reward.

Eligibility is open in one respect and bounded in another. Any account may register participation keys and vote in consensus, but to earn proposer payouts an account must also signal that intent by paying a raised key-registration fee and must hold a balance inside a defined band, with a floor in the tens of thousands of units of the native asset and a ceiling that discourages very large single concentrations. Holders below the floor reach the same rewards through pooled or custodial arrangements. There is no bonding, no lock-up and no unbonding queue: the balance that backs consensus stays liquid and spendable throughout.

The network does not slash. Misconduct and neglect are answered by removal rather than confiscation. An account that stops answering for its share of the work is marked absent and suspended from consensus within minutes, stops earning, and must pay the registration fee again before it can resume; a heartbeat mechanism and a safeguard for accounts whose weight rises abruptly limit false suspensions. The consequence of failure is therefore forgone reward and a small re-entry cost, not loss of principal.

Users pay a minimum fee of one thousandth of a unit of the native asset per transaction. The consensus version activated in 2026 replaced what had been an essentially flat charge with one that prices what a transaction carries, adding a per-byte surcharge above a size threshold and charging several times the minimum where a post-quantum signature must be verified; under congestion the per-byte rate rises further. Fees can be pooled across an atomic group so one transaction covers another, and every account must retain a minimum balance that grows with the assets and applications it holds.

Algorand is present on the following networks: Algorand.

The energy figure is assembled from the machines that actually run the network, rather than inferred from a mining market, which is the appropriate treatment for a pure proof-of-stake chain where no computational race takes place and producing a block costs nothing beyond keeping a node running. The starting point is the size and composition of the node population. The relay nodes that carry traffic and the participation nodes that hold registered keys and vote are counted from publicly visible network data, from peer discovery across the gossip layer, and from operator disclosures, and that count is handled as a plausible range rather than a single certain number.

A representative hardware profile is then assigned to that population. The published requirements for running the node software, covering processor class, memory, storage and network throughput, indicate the kind of machine an operator would realistically deploy, and the electrical draw of such machines is taken from controlled bench measurement of comparable equipment at load and at rest rather than from manufacturer nameplate ratings. Annual consumption is the aggregate across the estimated node set with idle draw included, because a participation node draws power continuously whether or not sortition selects it in any given round. Where a specific asset issued on the network is being reported rather than the network as a whole, a share of the network total is attributed to that asset from observed on-chain transfer volumes.

The limits of the approach deserve stating plainly. The node count and the hardware mix are estimates built from public observation and stated software requirements, not metered readings taken from the machines themselves. Virtualized and cloud-hosted nodes cannot be cleanly separated from dedicated hardware, and an operator running several logical nodes on one physical host is not always distinguishable from several operators. Where evidence is thin, the assumption selected is the one more likely to overstate consumption than to understate it, so the result should be read as a conservative upper estimate rather than a precise measurement. Figures are revised as observation of the network improves and as consensus changes alter what a node is required to store, verify and transmit.

Algorand is present on the following networks: Algorand.

Deriving a renewable share begins with where the machines are, because electricity is not the same commodity in every place. Node locations are inferred from publicly observable network data, including the addresses peers advertise to one another, autonomous system and hosting-provider registrations, and operator disclosures, and are resolved to a country or, where the evidence supports it, to a sub-national region. Hosted infrastructure complicates this: an advertised address identifies a data center rather than an owner, and since it is the data center that draws the electricity, hosting location is used in preference to any inferred nationality of the operator.

Coverage is never complete. Participation nodes that accept no inbound connections, and nodes sitting behind relays or proxies, are not directly observable. Where the geographic spread cannot be established from the network's own traffic, the distribution of a network with a comparable operating profile is used as a stand-in, chosen for similar participation requirements, similar reasons to run a node and similar hosting patterns, on the reasoning that operators facing similar conditions make similar siting choices. That substitution is a recognized source of uncertainty and is the main reason the renewable share is less robust than the consumption estimate it rests on.

The resulting location distribution is weighted by estimated consumption and matched to regional electricity statistics, so each portion of the network's power draw is assigned the generation mix of the grid that supplies it. The renewable proportion is the consumption-weighted average of those regional mixes, not a simple count of nodes by country. Grid statistics are annual averages and do not capture hourly variation, so a node drawing power overnight in a solar-heavy region is treated identically to one drawing at midday.

Energy intensity is reported as a marginal quantity: the additional electricity associated with one further transaction, given the current node set and the throughput the network is carrying. On a network whose nodes run continuously and whose consumption barely responds to load, that marginal quantity is small and falls as throughput rises, which is a property of the accounting convention rather than evidence of an efficiency gain. Regional generation mix is taken from Share of electricity generated by renewables, compiled by Our World in Data from Ember and the Energy Institute's Statistical Review of World Energy.

Algorand is present on the following networks: Algorand.

Emissions are derived from the consumption estimate rather than measured, by pairing each geographically attributed portion of the network's electricity use with the carbon intensity of the grid that supplies it. The location inference used for the generation mix applies here unchanged: addresses observed on the network and hosting-provider registrations place consumption in a country or region, gaps are filled from the distribution of a structurally similar network, and the outcome is a consumption-weighted spread across grids rather than a headcount of nodes by country.

The split between scopes governs how the figures should be read. Scope 1 covers emissions from sources the operators of the infrastructure control directly, which in practice means fuel burned on site, primarily in backup generators. For a network whose nodes are ordinary servers in data centers and offices rather than dedicated industrial plant, direct combustion attributable to the network is negligible and the reported scope 1 figure is effectively nil. Scope 2 covers the indirect emissions embodied in the electricity those machines purchase, and carries essentially the entire footprint. It is computed on a location basis, using the average carbon intensity of the supplying grid, rather than on a market basis that would credit renewable energy certificates or power purchase agreements held by individual operators, because contractual instruments of that kind cannot be observed from network data and are not reflected here.

Several things sit outside the boundary. Emissions embodied in manufacturing, shipping and disposing of the hardware are excluded, as is the electricity consumed by wallets, indexers, block explorers and other services built on top of the network. Grid carbon intensities are annual averages, so short-term shifts in the generation mix are not captured, and every uncertainty affecting the location estimate propagates into the emissions figure.

Greenhouse gas intensity is expressed as a marginal quantity, the additional emissions associated with one further transaction at the current node set and throughput, and it moves inversely with activity because the underlying consumption scarcely responds to load. Carbon intensity values are taken from Carbon intensity of electricity generation, compiled by Our World in Data from Ember and the Energy Institute's Statistical Review of World Energy and made available under the CC BY 4.0 license.