ACT 0 — BEFORE CIVILIZATION
Fire. Stone. Speech. Trade.
For 300,000 years, humans existed in a world of continuous reality — gestures, faces, weather, hunger. There were no tokens. Only the world itself, perceived directly.
ACT 1 — THE FIRST ENCODING
Humanity learned to encode reality.
Writing turned speech into marks. Numbers turned quantity into symbols. Currency turned trust into a portable thing. Each was a discrete substitute for something continuous — the first tokens.
ACT 2 — EVERYTHING BECAME A TOKEN
Information became tokens. Value became tokens.
In 2017 a paper called 'Attention Is All You Need' redefined the substrate of machine cognition: the token. In 2015 a different paper, the Ethereum white paper, redefined the substrate of digital value: the programmable token. The substrates are the same word. That is not an accident.
00 / OPENING SEQUENCE
Civilization runs on tokens.
One word, two technologies, one substrate. The token at the bottom of a Transformer and the token at the bottom of a smart contract are not analogies — they are the same architectural move, made twice, by two different fields, within five years of each other.
Read top-to-bottom. The thesis only assembles itself by the end.
BASELINE · INFO ↔ VALUE
ChatGPT inference tokens / day (~2024)
0
B tokens
Altman, public statements; order-of-magnitude
Stablecoin annual settlement (2024)
0
T USD
Visa Onchain Analytics, adjusted
Context window — frontier model
0
M tokens
Gemini 1.5 / Claude class
On-chain stablecoin supply
0
B USD (2024)
USDT + USDC dominant; ~90%+ of float
Tokens generated by all LLM APIs (est.)
0
T/day
Industry-wide; aggressively conservative
Active onchain wallets (monthly)
0
M
Artemis aggregate, all chains
ENTER THE ABSTRACTION
●TOKEN.CIV v0.1/●INFO TOKEN ↔ VALUE TOKEN/●TWO HALVES OF THE SAME ABSTRACTION/●AI THINKS IN TOKENS/●MARKETS THINK IN TOKENS/●CIVILIZATION COMPILES TO TOKENS/●TPS · TOKENS PER SECOND/●ATTN(Q,K,V) = softmax(QKᵀ/√dₖ)V/●ERC-20 · BPE · GPT · OFT · DEX · DAG/●TOKEN.CIV v0.1/●INFO TOKEN ↔ VALUE TOKEN/●TWO HALVES OF THE SAME ABSTRACTION/●AI THINKS IN TOKENS/●MARKETS THINK IN TOKENS/●CIVILIZATION COMPILES TO TOKENS/●TPS · TOKENS PER SECOND/●ATTN(Q,K,V) = softmax(QKᵀ/√dₖ)V/●ERC-20 · BPE · GPT · OFT · DEX · DAG/●TOKEN.CIV v0.1/●INFO TOKEN ↔ VALUE TOKEN/●TWO HALVES OF THE SAME ABSTRACTION/●AI THINKS IN TOKENS/●MARKETS THINK IN TOKENS/●CIVILIZATION COMPILES TO TOKENS/●TPS · TOKENS PER SECOND/●ATTN(Q,K,V) = softmax(QKᵀ/√dₖ)V/●ERC-20 · BPE · GPT · OFT · DEX · DAG/
●01 / TOKEN AS INFORMATION
AI thinks in tokens.
Strip the chat UI and the personality away from a language model and what's left is one operation: given N discrete symbols, predict the (N+1)th. Everything called 'reasoning' falls out of doing this at sufficient scale.
PIPELINE · TEXT → INTELLIGENCE
[00]
Text
raw input
▶
[01]
Tokenize
BPE / SentencePiece
▶
[02]
Embed
ℝ⁴⁰⁹⁶ vector
▶
[03]
Attention
softmax(QKᵀ)V
▶
[04]
Feed-forward
MLP × N layers
▶
[05]
Predict
argmax / sample
▶
[06]
Token out
1 token
ANATOMY
[01]
What is a token?
A token is the atomic discrete unit a language model can read, write, or attend to. Most tokens are subwords — fragments shorter than a word, longer than a letter. 'civilization' might be 3 tokens; '文明' is 2 in many tokenizers; a Chinese character costs more tokens than an English word.
[02]
Byte-Pair Encoding (BPE)
Start with characters. Find the most common adjacent pair. Merge it into one new token. Repeat 50,000 times. The vocabulary that survives is the alphabet a model thinks in.
[03]
Embedding — a token becomes geometry
Every token is mapped into a vector — typically 1024–8192 dimensions. Distance in this space becomes meaning. The vector for 'king' minus 'man' plus 'woman' lands near 'queen'.
[04]
Attention — tokens look at each other
Each token issues a Query, every other emits Keys and Values. The match decides how much each past token influences the next prediction. softmax(QKᵀ/√dₖ)V is the line of math that built ChatGPT.
[05]
Context window — the model's short-term memory
GPT-2 (2019): 1,024 tokens. GPT-3 (2020): 2,048. GPT-4 (2023): 32K. Claude 4 / Gemini 1.5 (2024–25): 1–2M. The window decides how much of reality a model can hold in its head at once.
[06]
An LLM is a token-prediction machine
Strip everything away — the chatbot UI, the personality, the supposed reasoning — and what remains is one task: given N tokens, guess the (N+1)th. Intelligence, surprisingly, falls out of doing this at scale.
BOTTOM LINE
“An LLM is a token-prediction machine. Reasoning emerges as a side effect.”
●02 / TOKEN AS VALUE
Markets think in tokens.
A token here is the inverse of the AI token. Instead of compressing meaning into a discrete symbol, it lifts ownership into a discrete symbol. The symbol carries rules: who can hold it, how it transfers, what it does when it moves. Money learned to compute.
PIPELINE · ASSET → PROGRAMMABLE VALUE
[00]
Asset
USD, equity, art, ...
▶
[01]
Wrap
trust → smart contract
▶
[02]
Mint
ERC-20 / ERC-721
▶
[03]
Liquify
AMM / orderbook
▶
[04]
Route
DEX aggregator
▶
[05]
Settle
T+seconds
▶
[06]
Programmable value
machine-readable
ANATOMY
[01]
What is a value token?
An entry in a public state machine — a number associated with an address, transferable by signing a message. The token has no physical form; it exists because the network agrees it does.
[02]
Bitcoin — token as bearer cash
Satoshi's 2008 white paper described a token that is itself the asset, with no issuer, no jurisdiction, no underlying. The token is the thing. A bearer instrument made of pure consensus.
[03]
Ethereum — token as programmable money
Buterin's 2013 paper added an interpreter. Now a token could carry rules: 'only spendable if X is true,' 'rebase every 8 hours,' 'pay royalties to the original creator.' Money learned to compute.
[04]
Stablecoins — the killer use case (so far)
By 2024 ~$28T settled on-chain in stablecoins — roughly equal to Visa's global volume. Two issuers (Tether, Circle) hold a banking-class balance sheet. The internet quietly grew its own dollar.
[05]
RWAs — tokenizing the rest of the world
T-bills, money-market funds, private credit, real estate, carbon credits — by 2025 ~$15B+ of off-chain assets had been wrapped on-chain. Slow start, large slope.
[06]
Finance is becoming a token-processing system
When the unit of value is a programmable token, the financial system is no longer a network of institutions — it is a network of contracts. The contract is the bank.
BOTTOM LINE
“Finance is no longer a network of institutions. It is a network of contracts. The contract is the bank.”
●03 / THE CONVERGENCE
Two tokens, one substrate.
AI compresses meaning into tokens. Blockchain lifts ownership into tokens. Place them side by side and the words map almost perfectly: embedding ↔ liquidity, attention ↔ market, prediction ↔ pricing. Civilization is discovering one substrate from two directions at once.
ISOMORPHISMS
[INFO]Transformer tokens
↔
Financial tokens[VALUE]
[INFO]Embedding space
↔
Liquidity space[VALUE]
[INFO]Attention networks
↔
Market networks[VALUE]
[INFO]Neural routing
↔
Capital routing[VALUE]
[INFO]Prediction systems
↔
Pricing systems[VALUE]
[INFO]Token windows
↔
Token supply[VALUE]
[INFO]Tokenizer vocabulary
↔
Token standards (ERC)[VALUE]
EMERGENT PHENOMENA
AI agents with wallets
Erc4337 + LLM gives an agent a key, a balance, and the ability to settle without a human. The first machine-only economies are not science fiction; they are now testnets.
Tokenized intelligence markets
Compute-token networks (io.net, Akash, Bittensor) trade GPU time, model weights, fine-tuning slots, and inference quota as fungible programmable units.
Pay-per-token APIs are the new wire
$3 / million tokens out, $0.30 / million tokens in. Every prompt is a microtransaction. Every product built on an LLM is, by definition, a token-economy node.
Coordination becomes programmable
DAOs and AI-orchestrated agents share the same primitive: durable, machine-readable rules that act on tokens. Governance becomes code, then becomes a token, then becomes a service.
●04 / TOKENIZATION OF EVERYTHING
From institutions to protocols.
Fifteen domains, ordered roughly by market size. Most are 90%+ digital but <1% tokenized — the gap is where the next decade happens. The 'difficulty' column is where the regulators and the engineers live.
DOMAIN
MARKET
DIGITAL %
TOKENIZED %
DIFFICULTY
NETWORK
NOTES
Money (USD, etc.)
$100T+
95
0.5
35
95
Stablecoins are the leading wedge; CBDCs are the slow second wave.
Identity
n/a
60
1
80
80
Privacy-preserving credentials (zk-PoP, Worldcoin) wrestle with surveillance vs. self-sovereignty.
Compute (GPU / CPU)
~$200B/yr
100
2
45
70
Akash, io.net, Render, Bittensor — fungible compute as a market.
Energy
$10T/yr
30
0.1
80
60
Power-grid tokenization works in pilots, not at scale yet.
Attention
$1T/yr (ads)
95
5
50
90
From CPM ads to creator-token economies; still mostly platform-owned.
Reputation
n/a
50
2
70
80
Soulbound tokens, on-chain credit graphs (Gitcoin Passport, EAS).
Data
$280B/yr
100
0.5
60
75
Ocean Protocol, Numerai, data DAOs — early markets for the most-traded substance on earth.
Real estate
$380T
30
0.05
90
40
RealT, Lofty etc. tokenize US rentals — legally complex; settlement-fast, recording-slow.
Equities
$110T
100
0.1
75
90
Robinhood EU launched 24/7 tokenized stocks in 2025; the wedge is open.
Government bonds
$130T
100
0.05
60
95
BlackRock BUIDL, Ondo, Franklin OnChain — tokenized T-bills crossed $4B in 2024.
Labor / services
$50T+/yr
40
0.01
85
60
Stablecoin payroll for distributed teams; agent-paid agents at the cutting edge.
Social graph
platform value
95
1
70
95
Lens, Farcaster — graph-as-protocol; portable identity is the bottleneck.
AI models
~$500B (2024)
100
3
50
80
Bittensor subnets reward model contributions in TAO; tokenized model markets are functioning.
Robotics / fleets
$80B/yr
70
0.1
80
60
DePIN networks (Helium, Hivemapper) coordinate hardware fleets via tokens.
Physical infra (cell, energy, storage)
$3T/yr
65
0.05
75
70
DePIN: bootstrap real-world hardware via token incentives. The category most likely to break out.
“Each row is an industry deciding whether to upload itself.”
●05 / THE AI TOKEN ECONOMY
AI is a token economy.
Inputs: electricity, GPUs, training data. Output: tokens — priced, billed, settled at the millionth. Every product built on an LLM is, by definition, a node in a token economy. The numbers below are early; the curve is not.
ELECTRON → TOKEN → INTELLIGENCE → COORDINATION
Electron
→Token (output)
→Intelligence
→Coordination
PRICING & FLOW · INDICATIVE
Cost per 1M output tokens (GPT-4 → GPT-4o)
$60 → $10
~6× cheaper in 18 months
Cost per 1M input tokens (frontier)
$2.50
Continuing to compress
Tokens per H100 per second (inference)
~3,000
Throughput-dependent
GPT-4 training compute (Epoch est.)
~2e25 FLOPs
~10⁴ × GPT-3
Daily active LLM users (industry, est.)
~600M
Conservative
Bittensor TAO market cap
$3B+
Tokenized compute/intel
io.net deployed GPUs
~200K
Decentralized compute
Pay-per-call agent revenue (early DePIN)
growing
Now measurable
●06 / WHY CIVILIZATION DOESN'T TOKENIZE OVERNIGHT
Twelve places the migration stalls.
Each one is solvable in isolation. The hard problem is solving all twelve before the financial and informational systems they constrain are demanded to act at machine speed.
CAUTION · SYSTEM LIMITSThe list below is not a counter-argument. It is the load-bearing surface the migration runs across.
[01]
Compliance moats; slow rails
Regulation
Securities, banking, AML, consumer protection — every jurisdiction's rules were written for paper.
[02]
Fragmentation
Scalability
Mainnets clear thousands of TPS; an AI-agent economy needs millions. L2s + DA layers narrow the gap, slowly.
[03]
Mass adoption blocker
Human trust
Even when code is correct, people are not. Phishing, key loss, social engineering remain the dominant attack.
[04]
Liability vacuum
Identity verification
An agent acts. Whose responsibility is that action? Privacy and accountability are in direct tension.
[05]
Protocol capture
Governance attacks
On-chain voting is also bribable on-chain. The mechanism designs (Curve wars, governance flash loans) are new.
[06]
Catastrophic correlations
Smart-contract risk
~$2–7B/year stolen in exploits 2021–2024. Formal verification helps; doesn't end the problem.
[07]
Capability ≠ safety
AI alignment
An agent with a wallet and an objective is a contract with goals. Misaligned goals are now financially expressive.
[08]
Asymmetric power
Compute inequality
Frontier training has consolidated to ~5 labs. Permissioned compute is the new permissioned capital.
[09]
Grid bottleneck
Energy constraints
Datacenter demand projected to ~1,000 TWh/yr by 2026 — see Electric Civilization for the other half of this site.
[10]
Default surveillance
Privacy
Public-by-default ledgers and AI-readable identity are both surveillance machines unless deliberately designed otherwise.
[11]
Dual-use built in
Surveillance
Tokens are auditable, which serves both anti-fraud and dissent-suppression. Neutral tools have non-neutral uses.
[12]
Chokepoint capture
Centralization risk
Two stablecoin issuers and three foundries already gate the system. The new substrate inherits the old chokepoints.
●07 / TOKEN CIVILIZATION
From matter, to energy, to information, to tokens.
Each civilizational stage is named after what it learned to process. Industrial processed matter. Electrical processed energy. Informational processed bits. The next stage — visible in outline now — processes value, information, and coordination through one programmable substrate.
~1750–1950
Industrial Civilization
PROCESSES
Matter
~1880–2020
Electrical Civilization
PROCESSES
Energy
~1950–2020
Information Civilization
PROCESSES
Information
2020s – ?
Token Civilization
PROCESSES
Value + Information + Coordination
THESIS
Civilization compiles to tokens.
●SOURCES & FURTHER READING
Foundational papers and ongoing data sources behind every number on this page.
Vaswani et al. — Attention Is All You Need (2017)
The Transformer paper
Buterin — Ethereum white paper (2013)
Programmable token primitive
Nakamoto — Bitcoin white paper (2008)
Bearer-cash token primitive
Epoch AI
Frontier-compute estimates
Visa Onchain Analytics
Stablecoin settlement methodology
DefiLlama, Artemis, Chainalysis
On-chain market & user data
OpenAI / Anthropic / Google pricing pages
Per-token API economics
08 / CODA
Civilization runs on tokens.
AI processes informational tokens.
Markets process financial tokens.
Future civilization may process reality itself through programmable tokens.
Tokens are not a destination. They are an abstraction that civilization has fallen into twice — in compute and in capital — and from which there is no obvious way back. The architecture chooses itself.
“Civilization compiles to tokens.”“文明,被编译为 token。”
Public sources · Vaswani 2017 · Buterin 2013 · Nakamoto 2008 · Epoch AI · Visa Onchain · DefiLlama