Fig. 00 — The Data Economy Units: trillions of tokens (T)

The Data Beneath
the Waterline

What large language models have already eaten, the far larger private mass they cannot touch — and how an HTTP-native payment rail (x402) could turn that mass, including the data estates of dead companies, into a market.

01

What the models have eaten

Linear scale · 300T = full width
GPT-3 · 2020
0.3T Llama 2 · 2023
1.8T Llama 3 · 2024
15T Frontier · 2025–26
14–36T Public stock
All usable public human text
≈300T

Frontier training runs already consume a double-digit share of everything usable on the open web. On current trends, the public stock is fully utilized between 2026 and 2032. Then the only place left to grow is below the waterline.

02

The iceberg

Cross-section · not to scale
WATERLINE — WHAT CRAWLERS CAN REACH TRAINED INTO MODELS 15–36T tokens PUBLIC TEXT STOCK ≈300T tokens SOCIAL POSTS ~140T (Facebook alone) PRIVATE EMAIL ~400T (Gmail alone) PRIVATE CODE 10–30T ever written ENTERPRISE RECORDS CRM, logs, tickets, docs DEAD COMPANIES archives in cold storage ALL PRIVATE TEXT ≈ 2,000T TOKENS — ROUGHLY 7× THE ENTIRE PUBLIC WEB
03

x402 — a price tag on every byte

HTTP 402 · stablecoin settlement

The private mass stays submerged because there is no way to sell it in small pieces. x402 revives the HTTP 402 “Payment Required” status code: any endpoint becomes a paywall a machine can pay, per request, down to $0.001 — no accounts, no contracts. A business prices its archive; an end user prices their own exhaust.

1
Request
GET /archive/records

An agent or buyer asks for a slice of the dataset.

2
402 Required
price: 0.02 USDC

The server answers with structured payment terms.

3
Sign & pay
signed authorization

The client retries with a signed stablecoin payment; a facilitator settles it on-chain.

4
200 OK + data
seconds, no invoice

Data flows one way, money the other — a market the size of a request.

per-request pricing micropayments to $0.001 no accounts or API keys machine-to-machine works for businesses and individuals
04

The vulture — dead companies as energy

Recycling corporate data estates

Speculative model

When a company dies, its data doesn’t. Decades of CRM records, support transcripts, telemetry and internal documents sit in cold storage as an asset of the estate. In an ecosystem, nothing that dense goes to waste: the vulture converts a carcass back into energy. With x402 as the rail, a liquidator lists the archive behind paid endpoints and the estate is metabolized, token by token, into training data.

When a company winds down, its data remains an asset of the estate: CRM records, support transcripts, telemetry, internal documents. With x402 as the rail, a liquidator can list the archive behind paid endpoints and sell access per request, converting stored records into recoverable value for creditors — and into training data for models.

Stage 1
The carcass

Bankruptcy filed. Product is dead; the archive — years of records — survives on disk as estate property.

Stage 2
The listing

The liquidator audits, de-identifies where law requires, and exposes the archive behind x402 endpoints, priced per record.

Stage 3
The feeding

Labs and data agents pay in stablecoin, request by request. No negotiation, no data-room — the market is the protocol.

Stage 4
The energy

Proceeds flow to creditors; the records flow into training corpora. The company is gone — its data keeps working.

Precedent

Customer databases already trade in bankruptcy — RadioShack’s 2015 estate sold millions of customer records under court supervision. What’s missing is a rail that makes such sales granular, auditable and priced per use rather than one opaque bulk transfer.

The constraint

Privacy law doesn’t die with the company. Consent, purpose limits and deletion rights follow the records — the vulture model only works on data an estate may lawfully sell, or on consented sale by the users themselves.