Bittensor — Double-Down Investment Memo (Aug 2024)

Executive Summary

• Bittensor (TAO) coordinates compute, data, bandwidth, and electricity from miners, and routes it through validators that score usefulness; users benefit from increasingly powerful AI apps.
• EV3 began building a position in Q1’23 when the network ran roughly $30M of GPUs (~GPT‑2‑class). It now connects an estimated $150M of GPUs, with performance comparable to GPT‑3.5 (or better in some domains).
• We expect a “killer app” to emerge from the Bittensor ecosystem within 18 months, creating substantial upside for TAO as apps and subnets compound.

Investment Rationale

1. Killer app within 18 months.
2. Value accrual to TAO as successful subnets/apps grow.
3. Not priced in: TAO’s success remains a non‑consensus bet despite headline valuation.

Further Reading

• Official: Website, Medium, Whitepaper
• Explorers: TaoStats, TaoPill, BittensorWiki
• Podcasts: Novelty Search, Bittensor Hub, Proof of Coverage
• Research: Messari, Delphi, Stateless, Collab+Currency

Apps & Demos Built on Bittensor

1) Can the ecosystem spawn a “killer app” in 18 months?

Subnets split TAO emissions across distinct objectives. Miners compete to deliver the highest‑ranked responses at minimal off‑chain cost; validators score and monetize. Many subnets showed 80%+ improvements within months of launch. Example focuses include data collection/labeling/storage; pre‑training/fine‑tuning/inference/evals; image/3D/protein generation; roleplay/productivity/DeFi agents; trading; fraud/disease detection.

Selected Subnets With Momentum

1. SN25 — Protein Folding (Macrocosmos): Early throughput already exceeds ROSIE’s multi‑year totals; tackles AlphaFold3 constraints (usage caps / non‑commercial restrictions).
2. SN34 — Deepfake Detection (BitMind): Hard‑mixture‑of‑experts outperforms prior SOTA on generalized deepfake benchmarks; public demo via BitMindID.
3. SN2 — zk‑Verified Inference (Inference Labs): 4k+ CPUs and 30TB+ RAM in four months; integrations signaled by Benqi and TrustMachine.
4. SN13 — Web‑scale Text Scraping (Macrocosmos): Top miners scraped >300M posts each (>1.3B combined), surpassing Grass’s first dataset (550M posts).
5. SN11 — Roleplay LLMs (Impel Labs): Beats GPT‑3.5‑Turbo on EQ benchmarks with smaller models; powers Dippy (280k+ MAUs).
6. SN39 — Edge‑grade Image Gen (Wombo): Optimizes SDXL for consumer GPUs; 26% latency improvement month‑over‑month; ties into apps with 200M+ downloads.

2) Do successful subnets imply success for TAO?

Yes, via Yuma Consensus and Dynamic TAO. Bittensor evolves from Foundation‑defined parameters → permissionless subnets (Revolution, Q4’23) → Dynamic TAO (Q4’24), where subnets launch their own tokens and TAO emissions are deposited as protocol‑owned liquidity into AMM pools, weighted by market valuations of subnet tokens. This creates perfectly competitive markets among miners, validators, and subnet creators—all must stake TAO; emissions follow a Bitcoin‑like schedule (4‑year halvings; 21M cap).

As competition intensifies, it becomes uneconomic to leave the network: validators need access to the best models (distribution), miners need access to multi‑app demand, and subnets need the infrastructure/app developer base—driving lock‑in for TAO.

3) Is success already priced in?

Despite a multi‑billion market cap/FDV and leadership versus other decentralized‑AI tokens, TAO remains non‑consensus among institutions. Critiques (subjective consensus, “not real AI,” or “memecoin”) miss the core advantage: mobilizing massive infra and developer attention on‑demand to achieve double‑digit monthly efficiency gains. We expect Bittensor apps to reach 1M users by end‑2024, 10M+ by 2025, and 100M+ by 2026 as momentum compounds.

Known challenges include: centralization at the physical layer (PoS transition expected 2025), extractive subnets, validator weight‑copying, chain halts, and limited institutional custody/staking support. We view these as solvable given the flywheel strength.