Compute & Inference.
Crypto rails for the GPU shortage plus the markets that serve models on them — real wedge today, structural question for 2027.
Thesis
Why this sector, why now.
This is the merged compute-and-inference stack: the GPU networks (Render, io.net, Aethir, Gensyn) that aggregate hardware, and the inference markets (Bittensor, Allora) that sell model serving on top. The supply thesis is that pooling consumer + small-data-center GPUs creates a durable cost edge over hyperscalers; the demand thesis is that model serving fragments away from centralized providers toward specialized subnets. I track both halves because they rise and fall together — a network with idle GPUs and no inference demand is a token, not a business. The counter-case is brutal: hyperscaler capacity catches up and centralized inference keeps getting cheaper every quarter. Pick projects with real institutional customers, not Discord-bot rendering jobs or emissions-funded validators.
Signals I track
What would move my read.
- 01
Average GPU utilization across networks (proxy for real demand)
- 02
A single non-crypto-native customer adopting decentralized inference at scale
- 03
Tokenomics surviving low-utilization periods without emissions propping revenue
Kill shot
What would kill the thesis
Hyperscaler GPU prices fall 30%+ and centralized inference keeps cost-per-token dropping faster than decentralized supply can match — leaving only the privacy / geography / censorship long tail.
Coverage
Projects on the radar.
Bittensor
ConvictionDecentralized machine learning network — subnet model — TAO token
Deep read →
io.net
ConvictionDecentralized GPU network — aggregator model — institutional traction
Deep read →
Render Network
ConvictionVeteran decentralized rendering — pivoted to AI compute — Solana migration done
Deep read →
Aethir
ConvictionEnterprise-grade decentralized GPU cloud (NVIDIA H100/H200/B200) for AI and gaming; $9M Pre-A backed by HashKey, Animoca and Framework Ventures.
Deep read →
Hyperbolic
ConvictionOpen-access GPU marketplace and serverless inference layer aggregating idle compute for affordable model hosting.
Deep read →
Nous Research
ConvictionOpen-source AI lab running the Solana-based Psyche network for decentralized model training, off a $50M Paradigm Series A at a $1B valuation.
Deep read →
Prime Intellect
ConvictionDecentralized AI training protocol pooling globally-distributed compute to produce open frontier models (the INTELLECT series).
Deep read →
Akash Network
WatchingCosmos-based decentralized cloud — earliest DePIN cloud — slower growth
Deep read →
Allora Network
WatchingSelf-improving inference network — collective intelligence — newer than Bittensor
Deep read →
Gensyn
WatchingDecentralized ML compute protocol stitching idle GPUs into a verifiable training network, off a $43M a16z-led Series A.
Deep read →
NetMind.AI
WatchingDecentralized GPU compute network (NetMind Power) pooling idle GPUs for AI/ML, NMT token; led by serial AI founder 'Kai' (ex-ProtagoLabs).
Deep read →
Ritual
WatchingA sovereign execution layer ('Ritual Chain') for running AI models inside smart contracts, built by two ex-Polychain partners on a $25M Archetype-led round.
Deep read →
Dynex
SkepticalFair-launched neuromorphic-computing PoUW blockchain (DNX) with an anonymous team and no venture funding.
Deep read →
NodeAI
SkepticalSmall Ethereum-based GPU/AI-node rental and revenue-sharing platform ($GPU); no disclosed institutional funding, founder attribution unverified.
Deep read →
Going deeper
Bespoke Compute & Inference dive for your fund.
Draft thesis · editorial voice in progress, edits land continuously