Skip to main content

AI Data Players

Who operates at each layer of the AI data value chain — and what position does each player fill?

This page is the star for the AI data player ecosystem. It maps the pattern by two lenses.

The layer lens names who runs each stage of the value chain, centralized and DePIN. The role lens names the counterparty each player becomes per transaction. The hat changes; the player remains.

How To Read​

Use this map when you scope an AI data build, size a market, or place a bet. Read it in three passes:

  • Find the layer your workflow touches — collection, connectivity, storage, compute, or application.
  • Name the player you would buy from, and the counterparty role they fill for you.
  • Check the trend signal and the risk before you commit.

Ecosystem Trend​

Across the value chain, DePIN and centralized incumbents run in parallel:

  • Collection — Centralized: Scale AI, Appen. DePIN: GEODNET, Hivemapper, WeatherXM, Grass. Trend: DePIN growing fastest.
  • Connectivity — Centralized: AWS, Cloudflare. DePIN: Helium, WiFi protocols. Trend: Helium mobile scaling.
  • Storage — Centralized: AWS S3, GCP, Azure. DePIN: Filecoin, Arweave, IPFS. Trend: cost parity for cold storage.
  • Compute — Centralized: NVIDIA, AWS, GCP. DePIN: io.net, Render, Akash, Gensyn. Trend: GPU demand outstrips supply.
  • Application — Centralized: OpenAI, Google, Anthropic. DePIN: open-weight models, DePIN inference. Trend: both growing.

Collection Layer​

Scale AI​

The centralized benchmark — the reference point for data infrastructure valuation.

  • Valuation: $29B (2025, post-Meta investment)
  • Revenue: ~$870M (2024), targeting $1.5B ARR
  • Founder: Alexander Wang (youngest self-made billionaire)
  • Core product: human-verified AI training data
  • Customers: OpenAI, Meta, US Department of Defense
  • Why it matters: Scale AI proved curated, labeled data is worth billions. DePIN protocols are building the same function with distributed infrastructure and token incentives.

GEODNET​

Global RTK positioning network with centimeter precision.

  • Stations: 20,000+ across 148 countries
  • Precision: centimeter-level RTK corrections
  • Use cases: autonomous vehicles, precision agriculture, surveying
  • Token: GEOD
  • Thesis: positioning data is foundational for every mobile AI system — self-driving cars, drones, robots. GEODNET builds the base layer. See GEODNET deep dive.

Hivemapper​

Decentralized street-level mapping.

  • Coverage: 30%+ of the global road network mapped
  • Contributors: dashcam operators worldwide
  • Output: fresh map data for AI, logistics, insurance
  • Token: HONEY
  • Thesis: Google and Apple control mapping data. Hivemapper builds a community-owned alternative that updates in real time.

WeatherXM​

Decentralized weather data network.

  • Stations: 7,000+ weather stations
  • Data: hyperlocal temperature, humidity, pressure, wind
  • Buyers: agriculture, insurance, energy companies
  • Token: WXM

Grass​

Decentralized web data collection for AI training.

  • Network: 3M+ daily active users, 8.5M monthly nodes
  • Function: web scraping and data collection at scale
  • Output: training data for LLMs and AI models
  • Token: GRASS

Connectivity Layer​

Helium​

The original DePIN protocol — connectivity as community infrastructure.

  • Hotspots: 115K+ active (2M daily mobile users)
  • Revenue: $18.3M ARR (Sep 2025 record)
  • Network: IoT (LoRaWAN) + 5G mobile
  • Token: HNT, MOBILE, IOT
  • See Telecom Players for detailed Helium analysis.

Storage Layer​

Filecoin​

Incentivized decentralized storage built on IPFS.

  • Storage capacity: 3.3 EiB committed (32% utilized)
  • Status: Onchain Cloud launched Jan 2026
  • Best for: large dataset archival, data sovereignty
  • Token: FIL

Arweave​

Permanent, immutable storage.

  • Model: pay once, store forever
  • Best for: data provenance records, attestations
  • Token: AR

Compute Layer​

io.net​

Distributed GPU compute marketplace.

  • GPUs: 327K verified (10K+ active nodes)
  • Revenue: $20M ARR (Oct 2025)
  • Best for: ML training, batch inference
  • Token: IO

Render Network​

Decentralized GPU rendering and AI compute.

  • Nodes: 5,600 (22M frames rendered in 2025)
  • Model: burn-mint equilibrium
  • Best for: graphics rendering, AI inference
  • Token: RENDER (Solana)

Akash Network​

Decentralized cloud compute with reverse-auction pricing.

  • Model: reverse auction for compute
  • Cost: 50-85% cheaper than AWS
  • Best for: general cloud workloads, containers
  • Token: AKT

Application Layer​

Open-Weight Models​

  • Llama (Meta) — largest open-weight model family
  • Mistral (Mistral AI) — European open-weight leader
  • DeepSeek (DeepSeek) — cost-efficient training breakthrough
  • Gemma (Google) — Google's open offering

Data Marketplaces​

  • Ocean Protocol — data tokenization and trading (established)
  • Streamr — real-time data streaming marketplace (growing)
  • Numerai — crowdsourced predictions market (established)

The Roles​

Every player wears multiple hats. A social platform is at once a data producer, a buyer of ML infrastructure, and a regulated entity. The position changes per transaction; the player remains. The AI data community has four sides.

Buyers​

The value-generators the industry exists to serve.

  • Frontier AI lab — buys massive pre-training corpus, rights-cleared web data, and human preference data (RLHF). Needs copyright liability closed and data freshness kept.
  • Enterprise fine-tuning team — buys domain-specific curated datasets, annotation, and evaluation sets. Fights siloed, unlabelled internal data.
  • AI product company — buys retrieval corpus, embedding pipeline, and real-time feeds for RAG. Needs freshness, commercial licensing, and low infra cost.
  • Regulated-domain AI team (health, legal, finance) — buys highly regulated data with full provenance and a consent chain. Needs HIPAA/GDPR compliance and de-identification.
  • DePIN data buyer — buys sensor and IoT streams with on-chain provenance and edge payment. Needs quality assurance and anti-spoofing.

Providers​

The organisations that produce and prepare AI data.

  • Data broker / aggregator (Refinitiv, IRI, Acxiom) — licensed, structured vertical datasets. Compete on coverage, data-rights stack, and API delivery.
  • Web-crawl operator (Common Crawl, Bright Data) — raw internet-scale text, HTML, multimodal. Compete on coverage, freshness, and permissioned commercial access.
  • Human labelling company (Scale AI, Appen, Surge AI) — human-annotated data, RLHF preference data, red-teaming. Compete on quality, turnaround, and domain-expert annotators.
  • Synthetic data generator (Mostly AI, Gretel) — privacy-safe synthetic tabular, text, image data. Compete on fidelity, privacy guarantee, and cost.
  • Publisher / news org / author — rights-cleared long-form text with high information density. Compete on copyright ownership and licensing terms.
  • DePIN data network (Hivemapper, WeatherXM, DIMO) — sensor-generated real-world data with on-chain provenance and token incentives. Win on data no central operator can collect.

Infrastructure​

The platforms and tooling the industry runs on.

  • Object storage (S3, GCS, Azure Blob) — scalable storage for raw and processed datasets.
  • Vector database (Pinecone, Weaviate, pgvector) — embedding storage and semantic search for RAG.
  • Data pipeline / ETL (dbt, Airbyte, Databricks) — transformation, movement, orchestration from source to training.
  • Data labelling platform (Labelbox, Roboflow, CVAT) — annotation workflow, quality review, active learning.
  • Data catalogue / lineage (Collibra, OpenMetadata, DataHub) — discovery, provenance, compliance tracking.
  • Evaluation / benchmark platform (EleutherAI, BIG-Bench, HELM) — standardised evaluation suites and leaderboards.

Boundary​

Sets the rules the other three sides operate inside.

  • Data-privacy authority (ICO, CNIL, DPC, FTC) — GDPR/CCPA enforcement: consent, purpose limitation, data-subject rights.
  • Copyright office / court (USCO, CJEU) — fair-use and text-and-data-mining adjudication for training data (NY Times v OpenAI, Getty v Stability AI).
  • AI Act authority (EU AI Act, national regulators) — training-data transparency obligations and prohibited data categories.
  • Data-sovereignty regulator (PIPL, DPDP, PDPA) — localisation, cross-border transfer limits, consent requirements.
  • DePIN / on-chain governance (DAO + token holders) — community-governed access, quality standards, and reward distribution.

The Five Archetypes​

The fractal pattern names five archetypes that appear at every layer. AI data is no exception.

  • Dreamer — the DePIN founder who believes token-incentivised sensor networks beat central collection; the lab visionary who says the next frontier is gated by data quality, not compute.
  • Realist — the enterprise lead who knows the GDPR consent chain before signing; the curator who says 80% of training time is data cleaning.
  • Engineer — the pipeline architect who scales 10TB to 10PB without a schema change; the labelling builder who cuts unit cost 70% with AI pre-annotation.
  • Coach — the team lead who builds annotation QA culture; the DePIN community manager who keeps contributors contributing through early coverage.
  • Philosopher — the researcher asking whether models reflect human knowledge or only internet posting; the ethicist auditing whether gig annotators are paid fairly.

When Engineer and Dreamer dominate without a Philosopher, the training data inherits its producers' biases — and the model ships before anyone asked the question.

Human vs AI Split​

Players hold positions, and each position has a human-vs-AI split that is shifting:

  • Data annotation worker — AI pre-labels; humans review edge cases and preference ranking. Direction: volume annotation AI-dominated; residual is ambiguous cases and red-teaming.
  • Data engineer — AI generates transformation code and catches schema drift. Direction: humans move to architecture and cross-system data contracts.
  • ML data curator — AI flags drift, duplicates, and quality. Direction: humans stay irreplaceable for compositional strategy and novel domains.
  • Data-rights specialist — AI tracks consent chains and flags conflicts. Direction: humans required for novel copyright and multi-jurisdiction negotiation.
  • Benchmark designer — AI generates adversarial variants. Direction: humans own conceptual benchmark design.

Competitive Dynamics​

Centralized incumbents and DePIN networks win on different dimensions:

  • Speed — centralized iterates faster; DePIN deploys faster at scale.
  • Quality control — centralized enforces it more easily; DePIN uses cryptographic verification.
  • Cost structure — centralized is fixed (data centers); DePIN is variable (community devices).
  • Geographic reach — centralized is limited by capex; DePIN is unlimited by incentive.
  • Data sovereignty — centralized is corporate-controlled; DePIN is user-controlled.

The winning model likely combines both. Centralized-style verification runs over DePIN collection. Hyperscaler compute handles frontier training while distributed GPU handles inference. Proprietary models hold the cutting edge while open-weight models handle distribution.

Investment Thesis​

Conviction and risk by layer:

  • Collection — High conviction. Position: GEODNET, Hivemapper. Risk: device saturation.
  • Connectivity — Medium conviction. Position: Helium via the telecom thesis. Risk: token sustainability.
  • Storage — Medium conviction. Position: Filecoin for enterprise. Risk: cloud price wars.
  • Compute — High conviction. Position: io.net, Render. Risk: hyperscaler response.
  • Application — Watch. Position: open-weight ecosystem. Risk: winner unclear.

Failure Modes​

Read the ecosystem wrong and the bet fails in named ways. Check these signals:

  • Backing a DePIN network with no unique data a central operator cannot collect — the moat is imaginary.
  • Treating a labelling provider as safe when synthetic data reaches human quality for the task.
  • Ignoring the boundary layer — a single enforcement action can reshape provider economics in months.
  • Confusing a player with a position — the same platform is producer, buyer, and regulated entity at once.

Changes my mind: if centralized incumbents capture every layer and DePIN networks fail to hold any durable position, the parallel-ecosystem thesis is wrong and this page is wrong.

Context​

Next question: which player type — collection network, labelling platform, or compute provider — holds the most defensible position as AI training commoditizes?

Questions​

Which counterparty's perspective is most invisible in this industry — and what routing signal gets missed as a result?

  • At what scale do io.net or Render become a credible alternative to AWS for training frontier models — and what is the missing piece?
  • How does the landscape change when AI agents autonomously source, clean, and route training data without human annotation?
  • If synthetic data reaches human-labelled quality for most tasks, which labelling players become redundant — and which new risks emerge?
  • Which existing player is most likely to acquire a DePIN data network rather than build one — and what would trigger it?