Bittensor Tokenomics: Yuma, dTAO and the 2025 Halving

Bittensor tokenomics is the only decentralized AI-inference marketplace with subnet-level tokenomics to have crossed a halving and a structural emissions shift with multi-year live data on which to test theory. The TAO mainnet has been running since 2021; the protocol has survived two major architectural shifts in the last 18 months—dTAO (February 13, 2025) and Taoflow (rolled out from November 2025, fully active by December)—and its first halving on December 13, 2025. Older compute-marketplace tokens like Akash (Sep 2020) and Render (Apr 2020) predate it, but neither has subnet-token architecture comparable to dTAO. The newer AI-agent-marketplace peers—Virtuals and ai16z (both late 2024), plus Bittensor’s various imitators—are all younger than 18 months. Olas dates to 2021 but only pivoted into AI-agent infrastructure recently. None of them has crossed a halving, dealt with subnet liquidity asymmetry at scale, or been audited by independent on-chain research after a structural emissions shift.

That makes Bittensor the only public benchmark a serious AI-marketplace designer can study. This piece is an integrated technical reference: the architecture, the math behind Yuma Consensus, the mechanics of dTAO and Taoflow, the 41 / 41 / 18 emission split, the subnet deregistration rule introduced in September 2025, what the first five months of post-halving data show, and the design lessons that hold (and don’t) for a founder building an AI-marketplace token in 2026.

For broader context on where Bittensor sits inside the AI-tokenomics landscape, see AI Agent Tokenomics; this article goes deeper into the protocol itself.

Architecture: root, subnets, miners, validators, nominators

Bittensor is structured around a root chain (Subtensor) and up to 128 subnets. Each subnet is an independent AI task—text generation, image embedding, time-series forecasting, financial signal extraction—with its own miners, validators, and (since dTAO) its own alpha token.

Bittensor architecture: root, subnets and role split.

Five roles share the protocol:

• Miners produce the AI commodity for a given subnet (an inference, an embedding, a prediction). They get paid by validators that consider their output good.
• Validators evaluate miner output and publish a weight vector—a score per miner—every epoch (360 blocks, ~72 minutes). Validators must hold enough TAO stake to qualify.
• Subnet owners define the validation task, run incentive code, and receive 18% of subnet emissions.
• Nominators (delegators) stake TAO with a validator without running infrastructure themselves; they earn a pro-rata share of that validator’s reward stream minus the validator’s commission.
• Root subnet (Subnet 0) is a special subnet without its own alpha token. Stake in the root subnet influences subnet-level weights and, before dTAO, was the sole mechanism for allocating emissions across subnets.

The economic primitive is straightforward: stake gates participation; validators score miners; the score determines emission shares; everyone gets paid in TAO (and, post-dTAO, in subnet alpha tokens).

Yuma Consensus, emission split, and halving trigger

Three numerical mechanisms govern TAO supply and reward distribution.

Block emission. The chain mints 1 TAO per block before the halving and 0.5 TAO per block after; one block lands every ~12 seconds, so daily emission was 7,200 TAO before December 13, 2025, and is 3,600 TAO today. Maximum supply is capped at 21 million TAO, mirroring Bitcoin.

Block split. Every minted block is divided across the three productive roles in a subnet:

Halving trigger. Unlike Bitcoin, Bittensor halves on emission count, not block height. The first halving fired when cumulative issuance reached 10.5 million TAO, exactly half of max supply. The protocol then drops the block reward by 50%. Because some emitted TAO is recycled (returned to the protocol via burns, slashing, and unused subnet registration fees), the halving date drifts: more recycling means a later halving, because recycled tokens don’t count toward the threshold.

So the often-repeated “every ~4 years” rule is only true under zero recycling. In practice the second halving will slip beyond 2030.

Yuma Consensus. This is where Bittensor’s design becomes interesting. Validators submit a weight vector—how much each miner deserves of the block emission. To resist a single validator (or a coordinated minority) pushing extreme weights, the protocol computes a stake-weighted median with clipping. For each miner j, the consensus weight W̄ⱼ is the highest value w such that validators controlling at least κ of total stake agree the weight should be ≥ w:

After the consensus weight is computed, each validator’s individual weight Wᵢⱼ is clipped down to W̄ⱼ: weights above the consensus are removed and earn nothing for the validator who set them, neither for the miner who received them. The two-sided punishment is intentional—it disincentivizes both over-reporting (collusion to inflate a miner) and accepting inflated weights.

This is the part most third-party summaries skip, and it explains a lot of validator behavior. If you submit weights wildly different from the consensus, you lose emission. If you submit weights identical to other validators, you might be “weight copying” and lose to the bond-trust mechanism. The optimal validator strategy is to evaluate independently but converge with the majority of trusted peers—which is hard enough that the Weight Copier working paper (Opentensor, May 2024) treats it as an unsolved free-rider problem.

import numpy as np

# Validators (rows) submit weights to miners (cols).
# Each row sums to 1 (Bittensor normalizes weights).
W = np.array([
    [0.6, 0.3, 0.1],   # Validator A
    [0.5, 0.4, 0.1],   # Validator B
    [0.5, 0.4, 0.1],   # Validator C (consensus)
    [0.4, 0.4, 0.2],   # Validator D
    [0.9, 0.05, 0.05], # Validator E — outlier, tries to inflate miner 0
])
# Normalized stake share of each validator (sums to 1).
S = np.array([0.30, 0.25, 0.20, 0.15, 0.10])
kappa = 0.51

# For each miner, find the highest weight w such that
# stake supporting (Wij >= w) >= kappa.
def consensus_weights(W, S, kappa):
    cons = np.zeros(W.shape[1])
    for j in range(W.shape[1]):
        candidates = np.sort(W[:, j])[::-1]
        for w in candidates:
            support = S[W[:, j] >= w].sum()
            if support >= kappa:
                cons[j] = w
                break
    return cons

cons = consensus_weights(W, S, kappa)
# Clip each validator's weights at the consensus.
W_clipped = np.minimum(W, cons)

print("Consensus weights:", np.round(cons, 3))
print("Validator E original [0]:", W[4, 0], "→ clipped:", W_clipped[4, 0])

Subnet economics: alpha tokens, bonding curves, deregistration

Pre-dTAO, every subnet was a pure cost center for TAO holders: it received emission, but the only way to “own” a subnet’s success was to validate it. dTAO changed that by giving every subnet its own alpha token and its own constant-product AMM against TAO.

How alpha tokens come into existence. A user stakes TAO into a subnet’s reserve pool. The pool maintains the invariant TAO_reserve · alpha_reserve = k, the same x·y=k AMM that powers Uniswap V2. The user receives alpha tokens at the current pool price; the more TAO already staked in the subnet, the higher the alpha price climbs. This is a bonding curve: there is no order book, the price emerges mechanically from reserve ratios.

Alpha emission. The chain mints alpha tokens for each subnet at twice the base rate of one alpha per block, allocated between (a) the subnet’s alpha reserve, which deepens AMM liquidity, and (b) outstanding rewards for miners, validators, and the subnet owner. Alpha follows the same halving schedule as TAO, which is why the December 2025 halving compressed subnet incentives at exactly the same rate as TAO emission.

Subnet creation. Anyone can create a subnet by paying a lock cost in TAO (auction-priced; historically in the low hundreds when slots were free, but with the 128 cap saturated since 2025 the cost runs into the low thousands of TAO—roughly $1–2M at recent prices). The cost is held by the protocol and refunded—minus emissions the subnet owner already received—if the subnet later deregisters. The proposed 256-slot expansion would compress the lock cost back down.

Subnet deregistration. The 128-subnet cap was always a hard constraint, but until September 17, 2025 there was no eviction mechanism—once full, the chain stopped accepting new subnets. The new rule changes that: when the cap is hit and a new subnet wants to register, the chain identifies the subnet with the lowest exponential moving average (EMA) of alpha price, outside its 4-month immunity period, and deregisters it. On deregistration, all alpha tokens in that subnet swap back to TAO at the final pool ratio and are distributed to alpha holders pro-rata. The subnet owner gets the lock cost back minus the cumulative emissions they have collected.

The rate-limit is at most one deregistration every two days. This produces a slow, predictable churn rather than a single cliff. As of May 2026, the chain has been at or near the 128-subnet cap for several months, so the EMA-price filter is the live mechanism deciding which AI tasks the network thinks are worth running.

These numbers are illustrative ranges synthesized from live data; for current values check taostats.io/subnets directly.

dTAO migration and Taoflow: why market-driven emissions

The most consequential design choice in Bittensor’s last 18 months is the answer to a single question: who decides which subnets get emission?

Legacy (pre-February 2025). A council of 64 root-subnet validators chose subnet weights directly. They earned reputation, paid for nothing, and faced very little accountability. The pattern that emerged is well known to anyone who has watched DAO governance: collusion, apathy, and informal cartels. A subnet that didn’t curry favor with root validators—even if its AI output was strong—couldn’t get emission.

dTAO (February 13, 2025). Root-validator allocation was replaced with a market mechanism: each subnet’s share of TAO emission became a function of its alpha token’s price (relative to TAO) in the AMM. If you believed in a subnet, you staked TAO into its pool, which raised the alpha price, which raised the subnet’s emission share. Conversely, unstaking pushed the alpha price down and reduced emission. The signal moved from “what does the root cartel like” to “where is real capital being deployed.”

dTAO worked, but it exposed a second-order asymmetry: subnets with deep pools could absorb large amounts of sell pressure without their alpha price moving much, while subnets with thin pools saw their alpha price collapse on the smallest unstake. The result was a liquidity-rich-get-richer dynamic: the subnets that got emission first kept it, even when their actual usage stagnated.

dTAO: TAO ⇄ alpha AMM and emission flow.

Taoflow (rolled out from November 2025). The protocol replaced price-based subnet share with flow-based subnet share. Instead of looking at “where is the alpha price highest right now,” the chain tracks net TAO flow into each subnet (staking inflow minus unstaking outflow) and smooths it with an exponential moving average:

Each subnet’s share of the block’s TAO emission becomes proportional to its smoothed flow F_t. By December 2025, 100% of TAO emission is allocated by flow.

Two design properties this fixes:

1. Scale invariance. A subnet that grows from a small pool by attracting genuine inflow gets credit, even if its absolute alpha price stays modest. Old large pools no longer dominate by inertia.
2. Negative-flow penalty. Subnets with sustained net outflow have F_t fall toward zero, and their emission share collapses to zero. There is no “I have a $10M pool, I deserve emission forever” defense anymore.

The trade-off: flow-based allocation is more game-able through wash staking (rapid stake/unstake cycles to fake inflow). The EMA dampens this but doesn’t eliminate it; protocol developers have publicly flagged it as an open monitoring problem.

For wider context on why market-driven mechanisms beat governance-driven ones for resource allocation, see Demand-side tokenomics.

Validator, miner, and nominator economics in numbers

The 41 / 41 / 18 split is the headline; the actual flow of money to a stake-holder is more layered.

Validators. A validator’s 41% share of block emission is divided across all validators in that subnet by stake share. If a validator controls 5% of subnet stake and the subnet captures 2% of total daily emission, that validator’s daily TAO income before commission is:

0.41 × 0.05 × 0.02 × 7,200 (or 3,600 post-halving) = ~2.95 TAO/day pre-halving, ~1.48 TAO/day post-halving.

The validator keeps a commission (take) of 0–18%, default 18%. The remaining 82%+ is paid to nominators (delegators) pro rata to their stake with that validator. So a validator with a 10K TAO self-stake plus 90K TAO from nominators, running on a subnet that captures 2% of emission, post-halving:

• Validator’s own emission: 0.10 × 1.48 = 0.148 TAO/day from self-stake
• Plus 18% commission on the 90K nominator stake’s emission: 0.18 × 0.90 × 1.48 = 0.240 TAO/day
• Total: ~0.39 TAO/day from this subnet

A validator typically operates across multiple subnets, with stake split. The economics scale linearly with stake share inside each subnet and with how many subnets the validator participates in. The full validator handbook is in the Bittensor staking docs.

Miners. Their 41% share goes through the Yuma weight clipping described above. A miner who ranks high (large W̄ⱼ across many validators) captures a large slice; a miner whose weights got clipped to zero earns nothing for that epoch. The variance is high—miner economics are closer to a tournament than a stake-yield product.

Subnet owners. Their 18% is the most predictable: it doesn’t depend on stake or Yuma, just on the subnet’s emission flow. Combined with the recovered lock cost on deregistration (minus collected emissions), the owner role is engineered as a moderately compensated curator, not a profit center.

Nominators. Their economics are the cleanest: stake X TAO with validator V, receive 0.82 × (V’s emission share) × (X / V’s total stake) per epoch. Post-dTAO advertised yields run 5–30% APY across root-subnet TAO staking and dynamic-subnet alpha staking—the closest TAO has to a “savings product” once you net out TAO/alpha price volatility and validator commission risk.

TAO emission and halving simulator

The numbers above hold under a specific set of assumptions: zero recycling, full subnet utilization, no protocol changes. In practice the supply curve drifts. The simulator below lets you tune the inputs and see how cumulative supply, halving timing, and net new tokens per year shift.

Halving aftermath: what 5 months of post-halving data show

The first halving fired on December 13, 2025. Five months later, three things stand out in the data.

Float compression continues. Roughly 70% of circulating TAO was staked before the halving, mostly with validators and in subnet alpha pools. Post-halving, the percentage hasn’t decreased—if anything, slightly more TAO is locked, because reduced inflation makes existing stake more valuable relative to outside-the-network alternatives. Effective tradeable float is in the low single millions of TAO.

Subnet incentives compressed in sync. Because alpha emission halves on the same schedule as TAO, the entire subnet incentive layer dropped by 50% overnight. There was no observable mass exodus of miners from subnets in January–February 2026, but the bar for being a profitable miner rose: hardware costs that broke even at 7,200 TAO/day total emission now require ~2× the rank inside a subnet to break even at 3,600. This is consistent with the Pine Analytics bear case framing that the long-term miner equilibrium thins out.

Institutional layer formed. Two weeks after the halving, on December 30, 2025, Grayscale filed an S-1 for a Bittensor trust. The filing is the first time a US-regulated institutional vehicle has taken TAO seriously as a single-asset product, and it changes the demand side of the equation. Pre-halving Grayscale produced a research piece framing the supply shock; the S-1 was the next step. It is the most significant non-protocol news of the post-halving period.

These three facts—tightening float, compressing miner economics, institutional demand entry—are the live experiment a designer in 2026 has access to. None of them was a sure prediction in early 2025.

Pitfalls: concentration, weight copying, sybil, halving cliff

The academic record on Bittensor is not uniformly flattering. The most-cited recent paper is Lui & Sun, “Bittensor Protocol: The Bitcoin in Decentralized AI? A Critical and Empirical Analysis” (FLock.io, June 2025). Working with on-chain data across all 64 active subnets at the time, they document four problems a designer should reckon with.

1. Stake concentration drives rewards more than quality. Lui and Sun show that the correlation between a miner’s reward and their stake-adjusted rank is much stronger than the correlation between reward and any quality proxy. The 41% miner share is technically distributed by Yuma weights, but the validator weighting itself follows stake, so a miner aligned with high-stake validators outperforms a higher-quality miner aligned with low-stake validators. The fix they propose—performance-weighted emission split, composite scoring, trust-bonus multiplier—has not been adopted in protocol.

2. Weight copying as a sustained free-rider problem. The Opentensor Weight Copier paper formalizes the issue: validators who simply copy the consensus weight vector earn the same emission as validators who do real evaluation work, but spend nothing on inference. The protocol’s defense is the bond-trust mechanism (validators bond to specific miners over time; bond holders earn extra), but it’s a partial fix—copiers still capture a meaningful slice of the validator pool.

3. Sybil pressure on subnet registration. Subnet registration costs (lock cost) discourage casual creation, but the marginal cost of an additional miner identity inside an existing subnet is low. A coordinated actor can run multiple miner identities, split the work, and capture a disproportionate slice of Yuma-weighted rewards. This is the same dynamic that breaks naive airdrop designs; for the formal game-theoretic treatment see Mechanism design.

4. Halving cliff for marginal subnets. Most subnets aren’t profitable at the median—they survive on the thin slice of emission above their operating cost. When TAO + alpha emission halves overnight, the marginal subnet’s operating margin disappears. The deregistration mechanism then accelerates churn at the bottom, which is design-intended but worth flagging: founders launching a subnet in 2026 are entering a much more constrained reward environment than founders in 2024 did.

A fifth pitfall sits at the dTAO layer specifically.

5. Liquidity asymmetry in dTAO AMMs. Constant-product AMMs are price-sensitive to trade size relative to pool depth. A subnet with a 5K TAO pool sees its alpha price move 10% on a 250 TAO unstake; a subnet with a 100K pool sees the same trade move price by 0.25%. Before Taoflow, this meant small subnets were structurally disadvantaged on emission share. After Taoflow, it means small subnets are more vulnerable to price manipulation by anyone willing to spend a few hundred TAO to suppress an alpha price.

For a methodological treatment of how to stress-test designs against these failure modes, see Agent-based modeling.

Design lessons for AI-marketplace token founders

Bittensor is not a template to copy. But it has run long enough that a handful of design choices have proven themselves robust, and a few have proven themselves problematic. A short list, in order of confidence:

Lessons that hold:

1. Stake-weighted ranking with clipping protects against minority pumping but requires a concentration counter-mechanic. Yuma’s clip-to-consensus rule works—pure pumping is expensive. The miss is on the supply side of validator stake: nothing in Yuma prevents a small set of validators from holding most of the stake and quietly converging on the same weight vector. Performance-weighted emission (Lui & Sun’s proposal) is one fix; capped voting power per validator is another.

2. Market-driven emission allocation outperforms governance-driven allocation when liquidity is sufficient. dTAO solved the cartel problem; Taoflow solved dTAO’s price-bias problem. The combined design—stake to vote, but vote with persistent flow, not snapshot price—is the most defensible piece of Bittensor’s architecture and the most copyable.

3. Subnet deregistration via EMA-price with an immunity period is a clean lifecycle primitive. Hard caps create churn cliffs; immunity periods prevent registration-griefing; EMA-based ranking ignores transient pumps. This pattern travels well to any protocol with a bounded slot count.

4. Hardcoded halving by emission count (not block height) is more flexible than Bitcoin’s version. Recycled tokens delay halving; this means treasury operations (burns, slashing) can be used to lengthen the supply ramp without forking. It’s a strict superset of Bitcoin’s design.

Lessons that don’t hold (and what to do instead):

5. Don’t depend on a root subnet for governance. Bittensor inherited this from its 2021 launch and has been trying to reduce root-subnet influence ever since. New AI-marketplace designs should bake market allocation in from day one, not retrofit it.

6. Don’t underprice subnet creation when the cap is loose. Early-Bittensor subnet lock costs were low; the lattice of low-value subnets that resulted is now being cleaned up via deregistration. A higher floor, refunded only on graceful exit, would have produced fewer ghost subnets.

7. Don’t let validator copying go undefended. The bond-trust patch helps but doesn’t solve it. New designs should consider commit-reveal weight schemes or randomized validator subsets per epoch.

For the broader framing—what AI-tokenomics archetypes exist and where Bittensor sits among them—see AI Agent Tokenomics.