Bittensor

Founded 2019 by Jacob Steeves (Const) and Ala Shaabana. TAO mainnet 2021. The only subnet-based decentralized AI network at meaningful commercial scale and the single most-watched token in the decentralized AI sector.

What Makes It Architecturally Unique

Bittensor is the only project in the decentralized AI space where the economic primitive is recursive: each subnet is a separate AI task market with its own Alpha token (AMM-priced), miners compete for subnet-specific rewards, validators score them, and base-layer TAO emits to subnets via the Dynamic TAO (dTAO) mechanism deployed February 2025.

Yuma Consensus (v3) is the underlying validation math: validator weight vectors are clipped, bonded, EMA-smoothed, and emission-distributed based on stake-weighted agreement. Subnet architecture lets anyone launch a new AI-task market with custom reward logic.

Scale (April 2026)

• TAO market cap: ~$2.37B (9.6M circulating × $238.91, post-Covenant crash)
• Subnet Alpha tokens: collectively ~$1.12B (~27% of TAO's own cap)
• Top 10 subnets capture 56.46% of TAO emissions
• SN64 (Chutes, serverless AI inference) alone takes 14.39%
• Daily TAO emissions halved to 3,600/day in December 2025 (from 7,200)
• ~70% of supply staked
• 64 active subnets with dTAO Alpha token markets

The Covenant AI Governance Crisis (April 9–14, 2026)

On April 9, 2026, the Covenant AI team exited Bittensor publicly, dumping 37K TAO ($10.2M) and publishing a letter accusing co-founder Jacob Steeves of "decentralization theatre" — unilateral emission suspensions, infrastructure deprecations, and moderation overrides inconsistent with Bittensor's permissionless framing.

Covenant had been Subnet 3's flagship project, having trained Covenant-72B on 1.1T tokens across 70+ contributors on commodity GPUs — the largest decentralized LLM pretraining run on record, with a 67.1 MMLU score. The same training run had been Chamath Palihapitiya's talking point when he introduced Bittensor to Jensen Huang on the All-In Podcast 20 days earlier.

Impact:

• TAO fell 23% ($332 → $254), wiping ~$900M of market cap
• TAO continued sliding to $238.91 by April 14
• Capital rotated to NEAR, ICP, FET
• The exit operationalized the empirical critique from the arXiv governance paper (see below)

The Empirical Critique

A peer-reviewed on-chain analysis of all 64 subnets published on arXiv (July 2025) documented:

• Stake Gini coefficient: 0.9825 — extreme concentration
• Top 1% holds median 90% of stake
• More than half of subnets are 51%-attackable by <1% of wallets
• Stake-to-reward correlation: 0.50–0.80
• Performance-to-reward correlation: 0.10–0.30
• TAO rewards capital, not AI quality by a 3–8x margin

The Covenant exit and this paper tell the same story from different angles. The governance crisis was empirically predictable.

The Jensen Huang Endorsement (March 20, 2026)

On the All-In Podcast, NVIDIA CEO Jensen Huang called Bittensor "a modern version of folding@home" after Chamath Palihapitiya walked him through the Covenant-72B training run. TAO rallied 17% that day past $300, closed up ~90% for March.

Grayscale had already filed S-1 (December 2025) for a spot TAO ETF branded GTAO. The Huang quote accelerated ETF conversations and institutional inflows before the Covenant crisis unwound them.

Leadership Transition

Jacob Steeves (Const) stepped down as Opentensor CEO in February 2026. This precedes the Covenant crisis and may have contributed to the governance vacuum Covenant alleges. The interim leadership's handling of the crisis will be a key signal for whether Bittensor's governance model is repairable or structurally broken.

Top Subnets (April 2026)

Per CoinGecko Learn: τemplar, Chutes (SN64, serverless inference, 14.39% emissions), Targon, Affine, Lium, Ridges AI, PTN, Score, Iota, Hippius.

Chutes alone — a serverless AI inference subnet — capturing ~14% of total Bittensor emissions is the commercial proof that subnet-level concentration is extreme.

Positioning vs. the Rest of the Landscape

• vs. Render / io.net / Akash — those are pure GPU marketplaces. Bittensor is a task-specific intelligence market. Commercial buyers comparing "pay for GPUs" vs. "pay for subnet outputs" are solving different problems.
• vs. Prime Intellect / Nous Psyche — those run single-purpose training swarms. Bittensor lets anyone launch a subnet for any AI task. Bittensor's generality is a strength technically and a weakness commercially (harder to pitch a buyer).
• vs. EigenLayer-based AI AVSs — Bittensor is an independent L1 with native-token bootstrapping. EigenAI and peers inherit Ethereum security via restaking (~$11B). Structural fork in the decentralized AI security model.