Kelp Market · Investor brief · 2026

Everyone is selling data.
Nobody can prove it.

The public web is used up — every vendor resells the same scraped photos. The real supply is the 80%+ of the world's ~13.6 trillion photos that were never published online. We give labs access to that unseen inventory, and every asset ships with cryptographic proof of where it came from and who agreed to the license. A court can check it, an insurer can check it. That's what we built.

See a live quote Why now

ed25519 · SEAL records · PRNU sensor fingerprinting · IPFS + Stripe rails

Specimen Unit economics ● Live
list price$0.50 / asset
hosting<0.1% of list
mintfractions of a cent
creator share50% · per asset
no GPUs no server trust

Context

This is already a real market

Jan 2026 real money

Cloudflare bought Human Native AI

Cash exit for a training-data marketplace. An infrastructure company decided this category is plumbing, not a niche.

May 2026 real money

Wirestock raised $23M

Creator-sourced training data, $40M claimed run-rate. Supply that can document itself is where the money is going.

Sep 2026 real lawsuits

Stability is on trial

First US jury to hear an AI training case. Every asset a lab cannot trace becomes evidence against the lab.

how training data is sold today

  • browse a catalog
  • negotiate a contract
  • receive a bulk export
  • attribution = a row in the vendor's database

The vendor vouches for itself. Ask a lab how they know an asset was licensed and the honest answer is: their database says so.

what courts and regulators now demand

  • per-asset provenance evidence
  • license chains verifiable by a third party
  • takedown and filtering traceable to source
  • machine-readable disclosure (EU AI Act Art. 50)

None of that comes out of a database row. It has to be bound to the asset itself, and that is exactly what we sell.

What we built

A verification rail,
not another catalog

We're not trying to out-catalog anyone; Wirestock and friends already won that race. Partners and creators send assets through our minting API instead. We check the capture, fingerprint the camera sensor, sign the license chain, anchor it on IPFS. The proof travels with the file. You never have to take our word for anything.

sensor tier camera-original capture, challenge-verified gear, PRNU fingerprint
attestation tier partner back-catalog imports — documentary license chain, labeled as such

Labs buy through the same rail. An agent calls our MCP server, gets a quote in seconds, pays, and downloads a bundle where every image carries its own signed manifest.

And what labs buy is inventory, not another copy of the scraped web. The published internet is fungible — every vendor resells the same photos. Most assets moving through our rail were never published anywhere online. Unseen, and documented: that is the part nobody else can sell.

mcp · kelp-market — real run, 2026-09-10
$ estimate_training_set { type: "image", quantity: 20, keyword: "surfing" }
quote 3335b6e3 · 8 assets matched · $0.50/asset · total $4.00 · expires 1h
  sample manifest:
    asset:bafybeiaygcyu5s6… → license: kelp-training ✓
    asset:bafybeic3z6mhrp3… → license: kelp-training ✓
$ # every asset: signed manifest + SEAL record, verifiable without us
$ confirm_training_set { jobQuoteId: "3335b6e3…" }
job queued → bundle.tar.gz · LICENSE.md + manifest.json + per-asset proofs

Unit economics

The money is in the proof, not the storage

hosting

<0.1%of list price

we host everything. Storage is a rounding error even with a big catalog.

mint

fractions of a cent/asset · once

plain CPUs do the fingerprinting. No GPUs.

list price

$0.50/asset

live quote, settled via Stripe Checkout

Hosting is on us. Storage cost grows in step with revenue, not ahead of it. The valuable thing we make is the proof, and only we can mint it.

Why now

The lawsuits are doing the selling

Nov 2025

Getty lost most of its UK case against Stability. But the judgment turned on what Stability couldn't document about its own training runs. Paperwork decides these cases.

Aug 2026

EU AI Act Article 50 went into force. Training on data you can't document now has a compliance price attached.

Sep 2026

Stability is in front of a jury. Runway and Luma picked up new complaints this year. Disney, Warner Bros. Discovery, and NBCUniversal sued MiniMax, which had to disclose the suit in its Hong Kong IPO filing.

Now

The public web is used up. Labs have scraped everything the internet ever published — and 80%+ of the world's ~13.6 trillion photos were never online. New supply now comes from private archives, and someone has to verify it. That is the rail.

Next

However the verdicts land, per-asset provenance becomes either a defense exhibit or a settlement ledger. That budget already exists — today it pays lawyers to reconstruct what the data was.

working today

  • MCP server, livequote, payment, and dataset delivery work end to end
  • Signed proofsed25519 manifests, one fail-closed verifier, public key in DNS
  • Sensor fingerprintingproofs bound to the physical camera that took the shot
  • Creator dashboardgear verification, uploads, payouts — supply side works

next

  • Open the quote APIany lab estimates and buys without talking to us
  • Partner catalogspartners send data through the minting API, we verify and host it
  • Two proof tierssensor and attestation tiers, kept strictly separate

The conversation

Pre-seed, small by design

We're raising against a working product, not a roadmap. The rail is live. The money goes into opening the quote API, onboarding the first data partners, and pushing creator supply while the lawsuits are still in the news.

Start the conversation Back to the product

Figures on this page are verifiable: market events are public record; the quote above is a real run against the live system.