Most AI-crypto tokens have to answer one question sooner or later: who pays for this besides the token itself? Bittensor is an unusually clear test case. Its network pays participants in newly issued tokens to produce AI services. This week brought two developments that speak directly to whether real customers are showing up. An independent tracker estimated that revenue-generating subnets have reached roughly $70 million in annualized revenue. [7] And code for “Gamma,” a proposed on-chain way for customers to pay subnets, has been merged but has not shipped to mainnet. [4]

This overview explains how Bittensor works, where TAO fits, how emissions are set, and what those developments do and don’t show.

What Bittensor does

Bittensor describes itself as an open network where independent subnets produce digital commodities like compute, inference, storage and prediction, and the chain pays participants in TAO in proportion to the value they contribute. [1]

Each subnet is effectively its own marketplace with its own rules:

  • Miners produce the commodity, such as running inference or renting out GPUs.
  • Validators score the miners’ work.
  • Subnet creators define the incentive mechanism, meaning what counts as good work.
  • Stakers back validators with TAO or with the subnet’s own token. [1]

The network can hold up to 128 subnets. When that limit is reached, registering a new subnet prunes the non-immune subnet with the lowest moving-average token price. New subnets are immune for roughly six months. [11] That makes subnet slots competitive. A subnet that can’t attract value can eventually be replaced.

TAO and alpha: two layers of tokens

Since the “dynamic TAO” (dTAO) upgrade in February 2025, every subnet has its own token, called alpha, alongside TAO. [1] Each subnet has a pool pairing TAO with its alpha, and the alpha’s price comes from that pool. [1]

TAO’s roles, per Bittensor’s documentation:

  • It’s the token the chain mints each block and splits across subnets. [1]
  • It’s the staking asset on the “root” network, whose stakers earn a baseline share of subnet emissions. [1]
  • It’s used for fees and registration costs, which are recycled back into issuance. [1]

Alpha is the subnet-level asset. Miners, validators and owners are paid in their subnet’s alpha. [1]

Supply and emissions

TAO has a hard cap of 21 million, the same number as Bitcoin. The base emission was 1 TAO per block, with a block every 12 seconds. [1] Halvings are triggered by total issuance thresholds rather than block counts: emission halves each time issuance crosses the midpoint of the remaining supply (10.5M, then 15.75M, and so on). [1]

The first halving happened in December 2025. Current emission is 0.5 TAO per block, or roughly 3,600 TAO per day. [1] There’s one subtlety. Recycled TAO, including registration costs and transaction fees, is subtracted from total issuance and can be re-emitted, which pushes future halvings further out. [1]

Each subnet’s alpha also has its own 21 million cap and follows the same halving curve, starting from that subnet’s launch. [1]

How TAO emission is divided among subnets. Each block’s TAO emission is split according to each subnet’s moving-average alpha price, which approximates demand. An “emission gate” sharply reduces the share of weak subnets (by default it centers on the 32nd-highest share). Subnets that direct miner rewards back to their owner have their share scaled down. [1] In short, the market’s pricing of each subnet’s token largely decides where new TAO goes.

How a subnet divides its alpha. The historical default split is 18% to the subnet owner, 41% to miners and 41% to validators and their stakers. [1] Since runtime v481 (October 2026), owners can change that split within chain-enforced limits. [5] This week, the computer-vision subnet Score (SN44) used that control to cut its owner share to 12% and give 44% each to miners and validators. [9]

Who pays, and why

This is the central question. Today most of what miners and validators earn comes from emissions, meaning newly minted tokens, not from customers.

That’s starting to change, at least on paper. On November 4, SubConnect, which tracks subnet revenue, estimated annualized revenue across revenue-generating subnets at about $70 million, up from about $50 million a month earlier and about $35 million when its index launched. [7] [8] According to Crypto Briefing’s summary of the report, compute and inference subnets make up roughly 78–82% of verified revenue, led by Lium (SN51), Targon (SN4) and Engy (SN53). [7]

Three caveats keep this in proportion:

  1. These are estimates. They are built from off-chain payment records, not audited on-chain totals. [7]
  2. Run rate isn’t revenue earned. An annualized figure extrapolates a snapshot. SubConnect itself says the hard part is sustaining growth while improving margins and reducing dependence on emissions. [8]
  3. Emissions are still much larger. Crypto Briefing’s summary of the SubConnect report estimates network emissions at more than $300 million a year, with external revenue covering only about 9–23% of that. [7] SubConnect’s own public update confirms the $70M revenue estimate but does not itself restate the emissions-coverage range in that post. [8]

Some subnets already use outside revenue to buy back their own alpha. Crypto Briefing reports that 14 subnets were doing so when SubConnect’s index launched, and that the number has grown since. [7]

What Gamma would change

Right now, subnets that sell services each run their own off-chain account system, where customers prepay and have to trust the subnet’s billing. [3] Gamma proposes a standard on-chain alternative.

How it works, per Bittensor’s documentation: [2]

  • A customer mints a subnet’s Gamma by burning that subnet’s alpha, either bought with TAO in the same transaction or taken from existing stake.
  • One Gamma is minted for each US dollar the burned alpha is worth, based on a TAO/USD price oracle set by root.
  • Gamma sits in a deposit keyed to a hotkey, which works like an API key. The subnet’s designated spender charges the deposit as it serves requests. Customers can set a daily spend cap.
  • Gamma can’t be transferred to another coldkey or converted back into alpha or TAO.

To resist manipulation, the burned alpha is valued at the lower of the spot price and a slow-moving “gamma price” that rises toward spot gradually, closing half the gap roughly every two hours. Pumping a token’s price briefly doesn’t buy extra credits. [2] [3]

Why it matters for token economics. Minting with TAO buys the subnet’s alpha and burns it, which shrinks alpha supply. Taostats’ explainer says the value of the burned alpha accrues to the subnet’s alpha holders, while the subnet’s operating costs are still paid from emissions. [3] Put simply, Gamma would turn customer payments into on-chain alpha burns, and it would make paid usage visible on-chain. Taostats calls Gamma charges “the closest on-chain signal of a subnet’s paid usage.” [3]

Status: not live. As of November 5, 2026, Bittensor mainnet is still on runtime 486. [4] [5] Gamma code has been merged into the Subtensor tree, but it is not in a published mainnet release yet (Bittensor Daily still lists it under “merged, not in a release yet”; the cut v487 PR was only a version bump and had not been released). [4] [5] Even after a runtime that includes Gamma ships, Gamma stays off until root enables it network-wide and each subnet owner turns it on for their subnet — both flags default to false. [2] [3] Optional quantum-resistant wallet types are tied to the same upcoming release line, with activation not yet confirmed. [6]

Competitors and alternatives

Bittensor competes on two fronts:

  • Centralized AI providers and clouds. Its leading revenue subnets sell compute and inference, the same products hyperscalers and inference APIs sell. If GPU and inference pricing gets more competitive, that hits Bittensor’s main revenue category. [7]
  • Other decentralized compute and AI networks. On compute, these include Akash, io.net, Render and Golem, which mostly run direct buyer-to-provider marketplaces rather than an emissions-weighted subnet economy. Bittensor’s model is different: it pays for output quality as judged by validators, not just rented hardware. Whether that’s an advantage depends on the subnet.

Risks

  • Emissions dependence. Outside revenue is estimated at a fraction of emissions. [7] If token prices fall, rewards in dollar terms fall too, which can push miners out.
  • Revenue concentration. Most estimated revenue comes from one category: compute and inference. [7]
  • Data quality. Revenue figures are third-party estimates from off-chain records. [7]
  • Gamma-specific risks. If a subnet is removed, its Gamma is destroyed and holders get nothing. Gamma gives no claim on anyone. The TAO/USD oracle is a trusted component. And the chain records charges but can’t verify the service was delivered. [3]
  • Governance and centralization. Root controls key switches, such as enabling Gamma and setting the oracle. [2] [3]
  • Subnet churn. Low-priced subnets can be pruned once the 128-slot limit is hit. [11]

What to watch

  • Whether a mainnet release that actually includes Gamma ships (beyond the empty v487 version bump), and when root and individual subnets actually enable it. [4] [5] [3]
  • On-chain Gamma mint and charge volumes once live. They would be the first verifiable read on paid usage. [3]
  • SubConnect’s next monthly update, especially the ratio of revenue to emissions. [7] [8]
  • More subnets changing their emission splits, as SN44 did. [9]

Bottom line

Bittensor’s design is clear: emissions bootstrap supply, and the market decides which subnets get them. The open question is whether customer payments can grow into a meaningful counterweight to those emissions. This week’s revenue estimate is a real data point, but it’s still an estimate and still small next to issuance. Gamma, if enabled, would make that question measurable on-chain, which may be its most useful feature.

Risk note: This article is for information only and is not financial advice. TAO and subnet alpha tokens are volatile and can lose value. Proposed protocol changes may change or never activate. Do your own research.

Sources