The Ultimate Guide to Mining TAO

In this guide

1. Introduction
2. What mining TAO actually means
3. How miners get paid
4. What you need before you start
5. Registration explained
6. Immunity, pruning and deregistration
7. How to choose a subnet
8. The subnets, researched
9. Running your miner day to day
10. The economics: costs, rewards and risk
11. Ten mistakes new miners make
12. Tools and resources
13. FAQ

Introduction

People mine TAO for the same reason people have always mined anything: there’s a task, a network wants it done, and it pays for the doing. On Bittensor, “mining” doesn’t mean solving arbitrary cryptographic puzzles, it means running software that performs a real task (inference, prediction, data collection, video processing, agentic work, compute rental, and more) on one of the network’s subnets, and being judged against everyone else doing the same thing.

That last part matters. Bittensor mining is competitive, not passive. You don’t buy a rig and let it hash away in the corner. You register a slot on a subnet, run the subnet’s software, get scored continuously by validators, and get paid according to how good your output is relative to other miners. Fall behind, and you get pruned, sometimes literally replaced by the next person in the queue.

This guide covers what mining TAO actually means, how the payment mechanics work (Yuma Consensus, dTAO, alpha tokens, the halving), what you need before you start, how registration and deregistration actually function on-chain, how to think about choosing a subnet, what day-to-day operation looks like, the honest economics, and the mistakes that trip up almost everyone in their first weeks. It is written for people who are serious about running infrastructure, not for people looking for a shortcut. There isn’t one.

What mining TAO actually means

Bittensor is a network of independent “subnets”, currently 128 active ones, excluding the root network, each built around a specific task. A subnet owner defines that task and writes the “incentive mechanism”: the rules by which output is judged. Examples of tasks across existing subnets include large language model inference, compute rental, AI agent evaluation, video upscaling and compression, dataset creation, model training, and confidential compute.

Within each subnet there are two working roles:

• Miners run the actual model or service that performs the subnet’s task. They compete on quality, speed, cost, or whatever the subnet measures.
• Validators continuously query miners, score their responses using the subnet’s incentive mechanism, and submit those scores (“weights”) to the Bittensor blockchain.

Subnet owners don’t do the scoring themselves day-to-day, they build and maintain the codebase and the rules, and update them as the task or competitive landscape evolves.

So “mining TAO” in practice means: pick a subnet whose task you can do well, run the software that performs that task, and keep improving it, because the chain pays according to your relative performance, continuously, not once.

How miners get paid

Payment on Bittensor runs through a mechanism called Yuma Consensus. Validators query miners, score their output against the subnet’s rules, and submit weights on-chain. Yuma Consensus turns those weights into emissions: miners are paid in proportion to their consensus score, and validators earn dividends when their evaluations agree with the stake-weighted majority, evaluations that are wild outliers get clipped and don’t count for much. This is designed to stop a single dishonest validator from redirecting rewards to a colluding miner.

New TAO enters the system on a Bitcoin-like exponential issuance schedule, capped at 21 million TAO, with a new block roughly every 12 seconds. Within each subnet, the TAO issued per block is split roughly:

• 41% to miners
• 41% to validators
• 18% to the subnet owner

Issuance itself halves periodically, like Bitcoin’s block reward. The first halving completed on 14 December 2025, cutting issuance from 1 TAO per block (~7,200 TAO/day) to 0.5 TAO per block (~3,600 TAO/day). The next halving is projected for around 2029, on a roughly four-year cycle. Halvings reduce the total TAO being emitted across the whole network, they don’t change the 41/41/18 split within a subnet, but they do mean the absolute size of the reward pool shrinks over time.

The other mechanic every miner needs to understand is dTAO (Dynamic TAO), introduced in February 2025. Every subnet has its own token, called “alpha,” with its own liquidity pool. A subnet’s share of overall network emissions is not fixed by decree, it’s set by market demand: people stake TAO into a subnet, which swaps it for that subnet’s alpha, and the more TAO staked into a subnet relative to others, the larger the emission share that subnet receives.

Crucially, miner and validator rewards are paid in the subnet’s alpha token, not TAO directly. To turn earned alpha into TAO, you swap it back through the subnet’s liquidity pool. This means your real return depends on two things beyond your consensus score: the alpha token’s price, and the depth of liquidity in that subnet’s pool. Earning a large amount of alpha on a subnet with thin liquidity or a falling alpha price can be worth far less than the same nominal amount on a subnet with a deep, stable pool.

What you need before you start

Before registering anywhere, get the basics in place. None of this is subnet-specific, it’s the baseline for participating in Bittensor at all.

• A Bittensor wallet, made up of a coldkey (holds funds and authority, keep it offline, never share its mnemonic with anyone) and a hotkey (used for registering and mining, designed to be exposed on-chain). No legitimate person, tool, or support channel will ever ask for your coldkey mnemonic.
• btcli, the official command-line tool, used to create wallets, register hotkeys, and manage stake.
• A Linux machine, Ubuntu 22.04 LTS or later is the standard baseline most subnets assume.
• Python 3.10+ and the bittensor SDK installed.
• Docker, since many subnet repositories package their miner software as containers.
• The specific subnet’s repository and README, this is the definitive source for that subnet’s actual requirements, dependencies, and setup steps. Nothing generic substitutes for reading it.
• Real hardware or cloud capacity appropriate to the subnet’s task (see below, this varies enormously by subnet).
• TAO in your wallet to cover registration cost, which is dynamic and paid before you get a working slot.
• Time and DevOps capacity, a plan for monitoring your miner and pulling updates, not just switching it on once.
• A realistic view of competition, check how many active miners a subnet already has and what quality bar they’ve set, before you commit hardware.

Skills-wise, comfort with Linux administration, Python, Docker, and reading unfamiliar codebases is the practical minimum. Bittensor mining is software engineering and operations work, not a point-and-click investment product.

Registration explained

To mine, you register a hotkey on a specific subnet and are issued a UID, a numbered slot on that subnet. The command is:

btcli s register --netuid <netuid> --wallet.name <wallet> --wallet.hotkey <hotkey>

Registration isn’t free, and the fee isn’t fixed. Most subnets use burn registration: you pay TAO to get your slot, and that cost is dynamic. It rises by a multiplier (BurnIncreaseMult) every time someone successfully registers, and it decays over time, halving every BurnHalfLife blocks. In plain terms: when a subnet is popular and people are registering fast, the price of a slot climbs; when interest is quiet, it drifts back down. Some subnets also permit proof-of-work registration as an alternative, but burn registration is the norm across the network.

Part of your registration fee goes into the subnet’s pool; the rest is recycled, removed from circulation, to be replaced gradually through future emissions. This matters practically: if you’re later deregistered, you do not get your registration fee back. Treat it as a sunk operating cost, the same way you’d treat a non-refundable listing fee, not as a deposit.

Costs vary a great deal by subnet, currently anywhere from roughly 0.0005 TAO to roughly 1 TAO, and the live figure for any subnet is shown on its TaoStats page (taostats.io/subnets/<netuid>, registration tab). Always check this before committing, since it changes block to block.

Slots themselves are limited. Most subnets cap out at 256 neuron slots (max_neurons); some are smaller, NOVA (SN68), for example, runs only 64. Once a subnet is full, getting in means someone else has to be pruned out first.

Immunity, pruning and deregistration

Every fresh registration gets an immunity period, a window during which it can’t be deregistered, giving a new miner a fair run at proving itself. The chain-wide default is 4,096 blocks, around 13.7 hours at 12-second block times. Individual subnets set their own value, and there’s real variation:

• Many subnets use around 5,000 blocks (~16.7 hours).
• Some use roughly 25 or 33 hours.
• Bitcast (SN93) and ChronoLLM (SN38) use around 168 hours, a full 7 days.
• engy (SN53) has effectively no immunity at all.
• Ditto (SN118) has around 1 hour.

Immunity is a preference, not an absolute guarantee. If every single neuron on a subnet happens to be within its immunity window, the chain will still prune the lowest-scoring immune neuron rather than refuse a new registration. So immunity buys you a head start, not a force field.

Pruning itself is emission-based: when a new registration lands on a subnet that’s already full, the neuron with the lowest emission that is currently out of immunity gets deregistered to make room. If there’s a tie, the older registration is evicted first. Critically, the chain does not distinguish miners from validators when pruning, whoever is earning the least goes.

Practically, this means you have roughly your subnet’s immunity window to get your miner running well enough to earn a defensible score. If you spend that window still debugging your setup, you may find yourself pruned before you’ve earned anything at all, and remember, your registration fee doesn’t come back.

How to choose a subnet

There’s no universally “best” subnet, the right choice depends on what you can actually build and run well, and what you’re prepared to risk. Work through these factors in order:

• Task fit. Can you genuinely do the subnet’s task well, inference, video processing, agent evaluation, compute rental, whatever it is, or would you be a mediocre entrant into an already crowded field? Read the subnet’s own repository and README before assuming anything about requirements.
• Emission share. Subnets differ hugely in what proportion of network emissions they receive, and this is set dynamically by dTAO market demand for the subnet’s alpha, not fixed. A larger emission share spread across more miners may pay less per miner than a smaller share among few competitors, check both numbers, not just one.
• Competition. How many active miners are already on the subnet? A subnet with very few active miners and steady emissions may offer a more favourable ratio than a crowded one, but it may also mean the task is hard, niche, or has high barriers you haven’t spotted yet.
• Hardware. Requirements are entirely subnet-specific. Some tasks demand serious GPU capacity, Vidaio (SN85), for video upscaling and compression, requires an RTX 4090 or better. Other subnets need little GPU at all and instead reward clever software pipelines. Check the subnet’s own repository for its actual spec; never assume.
• Alpha price and liquidity. You’re paid in the subnet’s alpha token, not TAO. A subnet with a thin or volatile pool can make it hard to convert earnings into TAO at a fair price. Look at the subnet’s pool depth and alpha price trend before assuming your emissions are worth their headline TAO-equivalent value.
• Registration cost. Check the live, dynamic cost on the subnet’s TaoStats registration tab. A cheap slot on a quiet subnet and an expensive slot on a hot one can both be reasonable, or both be traps, depending on the other factors above.
• Immunity period. A short immunity window (an hour, on some subnets) gives you little time to prove yourself before you’re evictable; a long one (up to seven days on some) gives you more room to iterate. Know which regime you’re entering.

For anything specific to a given subnet beyond these general mechanics, exact hardware specs, model requirements, incentive formulas, the subnet’s own repository is the definitive source. Don’t rely on secondary summaries, including this one, for subnet-specific technical detail.

The subnets, researched

Live figures from TaoStats, 04 September 2026. Emission share is the subnet’s slice of all network emissions; active miners and slots show how contested it is; registration cost is recycled and non-refundable.

#	Subnet	Task	Emission share	Active miners / slots	Reg. cost (TAO)	Immunity	Links
1	lium.io (SN51)	compute, gpu-rental	9.47%	64 / 256	0.5	~16.7 h	Git Site TaoStats
2	ORO (SN15)	ai-agent, commerce	8.99%	59 / 256	0.541	~16.7 h	Git Site TaoStats
3	Chutes (SN64)	compute, ai-inference	7.98%	17 / 256	0.001	~16.7 h	Git Site TaoStats
4	Affine (SN120)	ai-evaluation, llm	7.95%	5 / 256	1.004	~16.7 h	Git Site TaoStats
5	Albedo (SN97)	ai-agent, benchmark	6.86%	5 / 256	1.08	~16.7 h	Git Site TaoStats
6	Targon (SN4)	compute, ai-infrastructure	6.15%	6 / 256	0.001	~25.1 h	Git Site TaoStats
7	Score (SN44)	ai, video-analytics	4.53%	10 / 256	0.001	~25.0 h	Git Site TaoStats
8	SOMA (SN114)	ai-agent, compute	3.44%	18 / 256	0.25	~16.7 h	Git Site TaoStats
9	engy (SN53)	ai-inference, compute	3.32%	68 / 256	0.001	~0.0 h	Git Site TaoStats
10	Teutonic (SN3)	ai-training, model-evaluation	3.12%	5 / 256	0.001	~33.3 h	Git Site TaoStats
11	cascade (SN91)	ai, time-series	2.87%	5 / 256	0.11	~16.7 h	Git Site TaoStats
12	NOVA (SN68)	drug-discovery, ai	2.52%	2 / 64	0.131	~16.7 h	Git Site TaoStats
13	Bitcast (SN93)	content-marketing, incentive	2.34%	6 / 256	0.005	~168.0 h	Git Site TaoStats
14	SayGM (SN28)	ai-gateway, privacy	2.27%	36 / 256	0.1	~12.0 h	Git Site TaoStats
15	Hippius (SN75)	storage, compute	2.08%	67 / 256	1.0	~16.7 h	Git Site TaoStats
16	ChronoLLM (SN38)	ai-model, financial-analysis	1.96%	1 / 256	0.74	~168.0 h	Git Site TaoStats
17	Ditto (SN118)	ai-agent, productivity	1.84%	5 / 256	0.001	~1.0 h	Git Site TaoStats
18	Swarm (SN124)	ai-agent, competition	1.54%	25 / 256	0.5	~16.7 h	Git Site TaoStats
19	Hone (SN5)	ai-evaluation, validator	1.47%	243 / 256	0.104	~16.7 h	Git Site TaoStats
20	Beam (SN105)	bandwidth, data-transfer	1.47%	195 / 256	0.001	~25.0 h	Git Site TaoStats
21	RedTeam (SN61)	security, competitive-programming	1.39%	45 / 256	0.077	~48.0 h	Git Site TaoStats
22	ConnitoAI (SN102)	ai-training, mixture-of-experts	1.34%	6 / 256	0.045	~16.7 h	Git Site TaoStats
23	Ridges (SN62)	ai-agent, competition	1.33%	13 / 256	0.001	~24.7 h	Git Site TaoStats
24	blockmachine (SN19)	rpc, infrastructure	1.13%	27 / 256	1.0	~12.0 h	Git Site TaoStats
25	Cacheon (SN14)	ai-inference, compute	1.08%	5 / 256	0.001	~16.7 h	Git Site TaoStats

Before diving in, a note on how to read the figures below. Emission share is the size of a subnet’s slice of daily TAO emissions - the pool from which rewards are drawn, of which 41% is earmarked for miners (the rest goes to validators and the subnet owner). Active miners vs slots, alongside registered neurons, tells you how crowded the field is: a subnet showing 256 of 256 registered neurons but only a handful “active” means most registrants aren’t currently earning meaningfully - a sign of a demanding or narrow task. Registration cost is your entry ticket in TAO, burned on registration - it is not refundable and buys you nothing but a slot to compete for. Immunity is the grace period (in hours) during which a freshly registered miner can’t be deregistered for poor performance, giving you time to get your setup working before the market judges you.

lium.io (SN51)

Lium is a compute marketplace subnet aiming to make GPU rental more accessible and open. Miners are paid for supplying and serving compute capacity into the marketplace. With the largest emission share in this data set (9.47%) and 64 of 256 slots active against 256 registered neurons, there’s a sizeable reward pool but also a lot of registered competitors sitting idle - suggesting genuine competitiveness once you’re running rather than at registration. Entry costs 0.5 TAO, with a short ~16.7-hour immunity window, so you’ll need working infrastructure from day one. For hardware specifics, check the repository rather than guessing: github.com/Datura-ai/lium-io, site lium.io.

ORO (SN15)

ORO focuses on AI commerce agents, benchmarking their performance against defined incentive criteria. Miners are paid to build and run agents that perform commerce-related tasks well enough to score against the subnet’s benchmark. It carries the second-highest emission share here (8.99%), with 59 of 256 slots active out of 256 registered - a moderate but not overwhelming level of active competition relative to the pool size. Registration costs 0.541 TAO, with a brief ~16.7-hour immunity window before you’re exposed to deregistration. The subnet doesn’t specify hardware requirements in the data provided, so check the repository for build and runtime details: github.com/ORO-AI/oro, site oroagents.com.

Chutes (SN64)

Chutes bills itself as serverless compute for AI at scale, i.e. a hosting layer for running AI inference workloads on demand. Miners are paid for hosting models and serving inference requests through the platform. Emission share is high at 7.98%, yet only 17 of 256 slots are active against 256 registered neurons - a striking gap suggesting either a demanding operational bar or plenty of dormant registrations. Registration is cheap at 0.001 TAO, with a standard ~16.7-hour immunity window. No hardware specification is given here, so check the repository for what’s needed to run a competitive Chutes miner: github.com/chutesai/chutes, site chutes.ai.

Affine (SN120)

Affine is described as “Reason Mining,” an AI-evaluation subnet built around LLM benchmarking. Miners are paid for submitting outputs that perform well on the subnet’s reasoning/evaluation benchmarks. Emission share is strong (7.95%), but active miners are very thin - just 5 out of 256 slots, against 256 registered neurons - implying either a tough technical bar or a niche task most registrants haven’t cracked. Registration costs 1.004 TAO, one of the higher entry fees in this list, with a standard ~16.7-hour immunity window. No hardware detail is provided, so check the repository before committing: github.com/AffineFoundation/affine, site www.affine.io.

Albedo (SN97)

Albedo, styled “Alchemical intelligence,” is an AI-agent subnet built around benchmarking and dataset work. Miners are paid for contributing agent performance and/or dataset outputs that score well against the subnet’s evaluation criteria. Emission share is 6.86%, but with only 5 of 256 slots active out of 256 registered, competition for active status is sparse relative to the number who’ve registered - again pointing to a task with a real bar to clear. Registration costs 1.08 TAO, the highest single fee in this set, alongside a standard ~16.7-hour immunity window. Hardware needs aren’t specified here - check the repository: github.com/unarbos/albedo, site us-east-1.hippius.com/albedo.

Targon (SN4)

Targon runs an incentivised compute marketplace powered by its own “Targon Virtual Machine,” with an emphasis on confidential compute and security. Miners are paid for supplying secure compute capacity through the TVM. Emission share sits at 6.15%, with just 6 of 256 slots active against 256 registered - a low active count relative to the reward pool, suggesting meaningful headroom for miners who can meet the security/confidentiality bar. Registration is cheap at 0.001 TAO, and immunity is slightly longer than most here at ~25.1 hours, giving a bit more runway to get established. Hardware and setup specifics aren’t detailed in the data - check the repository: github.com/manifold-inc/targon, site targon.com.

Score (SN44)

Score, run by the team behind “making every camera intelligent,” is a video-analytics and compute subnet with a validator component. Miners are paid for processing or analysing video/camera data to a standard validators can score. Emission share is 4.53%, with 10 of 256 slots active out of 256 registered - a modest active cohort against a large registered base. Registration is low-cost at 0.001 TAO, with an above-average ~25.0-hour immunity window, giving new entrants a little extra time to bed in. The data doesn’t specify hardware requirements for video-analytics miners, so check the repository: github.com/score-technologies/turbovision, site wearescore.com.

SOMA (SN114)

SOMA describes itself as a context-compression layer delivered through MCP infrastructure - essentially an AI-agent and compute subnet focused on compressing context for downstream use. Miners are paid for providing compression services that meet the subnet’s requirements. Emission share is 3.44%, with 18 of 256 slots active against 256 registered - a reasonably competitive but not saturated field. Registration costs 0.25 TAO, with a standard ~16.7-hour immunity window. No hardware specification is given in the data provided, so check the repository before building anything: github.com/DendriteHQ/SOMA, site thesoma.ai.

engy (SN53)

engy provides “verified inference for frontier open models” - an AI-inference and verification subnet focused on trustworthy model serving. Miners are paid for running inference and having it verified as correct. Emission share is 3.32%, and it has one of the highest active-miner counts here: 68 of 256 slots active against 256 registered, indicating a genuinely busy subnet. Registration is cheap at 0.001 TAO, but notably immunity is effectively ~0.0 hours - meaning new registrants get no grace period and are exposed to deregistration immediately. Hardware needs aren’t specified in the data - check the repository: github.com/hanlinai/engy, site engy.ai.

Teutonic (SN3)

Teutonic focuses on “coordinated learning” - an AI-training and model-evaluation subnet built around collaborative training approaches. Miners are paid for contributing to training runs and being evaluated on their output quality. Emission share is 3.12%, with only 5 of 256 slots active out of 256 registered, suggesting a demanding technical bar with low active participation relative to registrations. Registration is cheap at 0.001 TAO, and immunity runs to ~33.3 hours - one of the longer windows in this set, giving new miners more time to establish themselves. Hardware requirements aren’t detailed here - check the repository: github.com/unarbos/teutonic, site teutonic.ai.

cascade (SN91)

cascade positions itself around “SOTA time series foundation models,” combining AI, time-series forecasting, and synthetic-data generation. Miners are paid for producing or improving time-series model outputs against the subnet’s standards. Emission share is 2.87%, with just 5 of 256 slots active against 256 registered - a small active cohort relative to the registered pool. Registration costs 0.11 TAO, with a standard ~16.7-hour immunity window. The data doesn’t specify compute or hardware requirements for time-series work here, so check the repository: github.com/TensorLink-AI/cascade, site cascadesub.net.

NOVA (SN68)

NOVA is aimed at “accelerating drug discovery” through an AI-driven compute competition. Miners are paid for submitting computational work relevant to drug-discovery tasks that scores well in the subnet’s competition format. This is a smaller subnet by slot count - only 64 slots total, with just 2 active - and emission share is 2.52%. The low slot ceiling combined with minimal active miners suggests a specialised, low-volume niche rather than a mass-market entry point. Registration costs 0.131 TAO, with a standard ~16.7-hour immunity window. Hardware needs for the drug-discovery workload aren’t specified - check the repository: github.com/metanova-labs/nova, site metanova-labs.ai.

Bitcast (SN93)

Bitcast describes itself as “the decentralised creators economy,” a content-marketing and attention subnet rewarding media/attention-related activity. Miners are paid for contributing content or attention-driving activity that the subnet’s incentive structure values. Emission share is 2.34%, with 6 of 256 slots active out of 256 registered - a small active group. What stands out is the immunity window: at ~168.0 hours (a full week), it’s by far the longest in this data set, giving new miners considerably more breathing room before being judged. Registration is cheap at 0.005 TAO. Hardware/setup specifics aren’t detailed - check the repository: github.com/bitcast-network/bitcast, site stats.bitcast.network.

SayGM (SN28)

SayGM offers “drop-in access to Claude, GPT, and Gemini” - an AI-gateway subnet with a privacy and security angle, essentially routing/proxying requests to major model providers. Miners are paid for operating gateway infrastructure that meets the subnet’s privacy and security requirements. Emission share is 2.27%, with 36 of 256 slots active against 256 registered - a moderately active field. Registration costs 0.1 TAO, and immunity is on the shorter side at ~12.0 hours, so new entrants have less time to prove themselves. No hardware specification is provided here - check the repository: github.com/taostat/gm-miner, site saygm.com.

Hippius (SN75)

Hippius offers a “blockchain-backed cloud” spanning storage, virtual machines, and applications, emphasising transparency and trust. Miners are paid for providing storage, compute (VM), or infrastructure capacity into this decentralised cloud. Emission share is 2.08%, and it’s one of the busier subnets by active count: 67 of 256 slots active against 256 registered. Registration is comparatively steep at 1.0 TAO, with a standard ~16.7-hour immunity window - meaning a higher upfront outlay before you’re even competing. Hardware and storage-capacity requirements aren’t detailed in the data provided, so check the repository: github.com/thenervelab/thebrain, site hippius.com.

ChronoLLM (SN38)

ChronoLLM runs “competitive training of chronologically consistent large language models,” an AI-model subnet geared towards financial analysis and backtesting. Miners are paid for training or contributing model updates that maintain chronological consistency (i.e. not leaking future information) under backtesting evaluation. Emission share is 1.96%, with only 1 of 256 slots active out of 256 registered - the thinnest active field in this data set. Registration costs 0.74 TAO, and immunity runs to a full ~168.0 hours, matching Bitcast as the longest grace window here. Hardware/training requirements aren’t specified - check the repository: github.com/chronollm/sn38, site chronollm.crunchdao.com.

Ditto (SN118)

Ditto brands itself “Open-Source Claude Cowork,” an AI-agent subnet centred on productivity and memory management for agentic workflows. Miners are paid for building or running agents that manage memory/context in productivity-oriented tasks. Emission share is 1.84%, with 5 of 256 slots active against 256 registered - a small active cohort. Registration is cheap at 0.001 TAO, but immunity is very short at ~1.0 hour, meaning almost no grace period before you can be pruned for underperformance. Hardware requirements aren’t detailed in the data - check the repository: github.com/orgs/ditto-assistant/repositories, site heyditto.ai.

Also worth a look

• Swarm (SN124) - drone-piloting AI competition across simulated missions; 1.54% emission share, 25/256 active, 0.5 TAO to register, ~16.7 h immunity. github.com/swarm-subnet/swarm
• Hone (SN5) - AI-evaluation/validator training subnet; 1.47% emission share with a notably crowded 243/256 slots active, 0.104 TAO registration, ~16.7 h immunity. github.com/hone-subnet-org/hone-subnet
• Beam (SN105) - decentralised bandwidth network; 1.47% emission share, 195/256 active (one of the busiest here), cheap 0.001 TAO entry, ~25.0 h immunity. github.com/orgs/Beam-Network
• RedTeam (SN61) - cybersecurity competitive-programming subnet; 1.39% emission share, 45/256 active, 0.077 TAO registration, longer ~48.0 h immunity. github.com/RedTeamSubnet/RedTeam
• ConnitoAI (SN102) - collaborative mixture-of-experts training; 1.34% emission share, 6/256 active, low 0.045 TAO registration, ~16.7 h immunity. github.com/Connito-AI/Connito
• Ridges (SN62) - software-engineering AI-agent competition; 1.33% emission share, 13/256 active, cheap 0.001 TAO entry, ~24.7 h immunity. github.com/ridgesai/ridges
• blockmachine (SN19) - RPC/node-provider infrastructure subnet; 1.13% emission share, 27/256 active, steep 1.0 TAO registration, short ~12.0 h immunity. github.com/taostat/blockmachine
• Cacheon (SN14) - GPU inference-speed competition subnet; 1.08% emission share, 5/256 active, cheap 0.001 TAO entry, ~16.7 h immunity. github.com/latent-to/cacheon

Running your miner day to day

Getting registered is the easy part. Staying competitive is the actual job.

Initial setup workflow typically looks like this: provision your hardware or cloud instance, install the subnet’s software from its repository (usually via Docker), configure your hotkey and wallet paths, register on the subnet, and start the miner process pointed at the network’s validators. Follow the subnet’s README precisely, deviations are a common source of silent failure.

Monitoring is not optional. You need visibility into:

• Your consensus score and emissions over time (via btcli or the subnet’s own dashboards, and via TaoStats).
• Whether your process is actually running and responding to validator queries.
• How your score compares to other miners on the subnet.
• Your remaining immunity window, if you’re newly registered.

Staying competitive means treating this as an evolving business, not a finished project. Subnet owners frequently update validation criteria, scoring code, and model requirements. A miner that doesn’t pull updates from the subnet’s repository will drift out of sync with what’s being measured, its score will fall, and, once immunity expires, it becomes a pruning candidate the moment a new registration needs a slot.

The DevOps reality is that this is active, daily-ish infrastructure work: watching logs, updating dependencies, redeploying when the subnet ships a new release, and reacting when validators change behaviour. If you’re not prepared to check on your miner regularly and respond to change, mining TAO is not a good fit for how you want to spend your time, a subnet you don’t tend will simply pay you less, then nothing.

The economics: costs, rewards and risk

Strip away the mechanics and a miner’s economics come down to a simple structure:

Revenue is your share of a subnet’s miner emissions, the 41% of that subnet’s issuance allocated to miners, split according to your consensus score relative to other miners, and paid in the subnet’s alpha token.

Costs include:

• Registration (a sunk cost, potentially recurring if you’re deregistered and choose to come back).
• Hardware, whether owned or rented cloud capacity.
• Electricity or cloud/bandwidth bills.
• Your own time keeping the miner updated and monitored.

Risks include:

• Deregistration, losing your slot to a better-scoring newcomer once your immunity expires.
• Alpha price and liquidity risk, your emissions are only worth what you can actually realise when swapping alpha back for TAO, and thin pools or falling prices reduce that.
• Validation-rule changes, a subnet owner updating the incentive mechanism can shift what’s rewarded overnight.
• Competition, other miners improving their setups faster than you improve yours.

To give one illustrative sense of scale: a third-party analysis (SubnetAIQ, 2026) estimated a median miner on Vidaio (SN85) earns roughly $448 per month. That figure is illustrative only, returns vary widely by subnet, by your rank within it, and by your costs, and there is no guarantee any given subnet or miner will perform similarly, better, or at all. None of this is investment advice.

Ten mistakes new miners make

1. Registering before reading the subnet’s repository. Burning TAO on a slot for a task you haven’t actually confirmed you can perform competitively.
2. Sharing or storing the coldkey mnemonic online or with a “support” contact. No legitimate party ever needs it.
3. Assuming registration cost is fixed. It’s dynamic and can rise sharply if a subnet is hot, check the live figure before submitting.
4. Treating registration fees as refundable. They’re recycled, not returned, if you’re later deregistered.
5. Ignoring the immunity clock. Spending the immunity window still debugging instead of getting a working, scoring miner live.
6. Assuming immunity is absolute. It isn’t, if a subnet is full of immune neurons, the lowest-scoring one is pruned regardless.
7. Setting up once and walking away. Subnet owners update validation rules regularly; miners who don’t pull updates fall behind and get pruned.
8. Choosing a subnet purely on headline emission share without checking miner competition, alpha liquidity, or hardware fit.
9. Under-provisioning hardware for the task. Some subnets have hard requirements, mismatched hardware simply produces uncompetitive output.
10. Confusing alpha rewards with TAO value. Assuming a headline emission number translates directly into TAO without checking that subnet’s alpha price and pool liquidity.

Tools and resources

• btcli, the official command-line interface for wallets, registration and stake management.
• The bittensor SDK, the Python library underlying most subnet miner and validator code.
• TaoStats (taostats.io), on-chain data, including live registration costs, emission shares, and subnet-level statistics via taostats.io/subnets/<netuid>.
• Each subnet’s own GitHub repository and README, the definitive, authoritative source for that subnet’s requirements, setup instructions, and incentive mechanism details. Treat this as more reliable than any third-party summary, including this guide, for subnet-specific facts.
• Bittensor’s official documentation, for wallet, coldkey/hotkey, and network-level mechanics.

FAQ

Do I need a powerful GPU to mine TAO?

It depends entirely on the subnet. Some tasks, such as video upscaling and compression on Vidaio (SN85), require serious hardware, specifically an RTX 4090 or better. Other subnets need little GPU capacity and instead reward good software design. Always check the specific subnet’s repository for its actual requirements before assuming either way.

What happens to my registration fee if I get deregistered?

It isn’t returned. Part of the fee goes into the subnet’s pool and the rest is recycled, removed from circulation and gradually replaced through future emissions. Registration should be treated as a sunk operating cost, not a refundable deposit.

How long do I have before I can be pruned?

Roughly your subnet’s immunity period, which varies widely, from around an hour (Ditto, SN118) to seven days (Bitcast, SN93, and ChronoLLM, SN38), with the chain default at 4,096 blocks (~13.7 hours) and many subnets around 5,000 blocks (~16.7 hours). Even during immunity, though, you can still be pruned if every neuron on the subnet happens to be immune and yours is the lowest-scoring.

Are miners paid in TAO or something else?

Miners are paid in the subnet’s own alpha token, not TAO directly. To realise TAO, you swap alpha back through that subnet’s liquidity pool, so the alpha price and pool depth affect what your rewards are actually worth.

Why does registration cost keep changing?

Registration uses a dynamic burn mechanism: the cost rises with a multiplier every time someone successfully registers, and decays over time, halving every set number of blocks. Busy subnets get pricier to join; quiet ones drift cheaper. Always check the live cost on the subnet’s TaoStats page before registering.

Can validators be pruned too, or only miners?

Yes. Pruning on Bittensor is purely emission-based, the chain removes whichever out-of-immunity neuron is earning the least, regardless of whether it’s a miner or a validator.

How do I pick between two subnets with similar emission shares?

Work through the full picture rather than emissions alone: how many miners you’d be competing against, the hardware and skills the task actually demands, the alpha token’s price and liquidity, the current registration cost, and the immunity period you’d get to prove yourself. A high emission share spread across many competitors, or paid in an illiquid alpha token, can be worth far less in practice than a smaller share on a subnet you’re genuinely well-suited to.

This is not investment advice.