The number that keeps resurfacing in my research channels isn't the GDP figure. It's 45.2%.
That is the labor share of income in Anthropic's "extreme" scenario for transformative AI โ down from roughly 60% today, with wages falling more than 10% and unemployment printing at historic highs. In the "significant" scenario, GDP growth doubles, unemployment holds at 5%, and knowledge-worker pay simply stops growing. Labor share: 56.1%. Same scenario, same result โ nobody's salary chart points up.
The report landed at night. Hours later, Anthropic safety researcher Jacob Coxon resigned with a public warning that the industry is racing toward self-improving superintelligence. Searching for truth in the noise of the network, I found the two events braided together inside the model itself. The extreme scenario isn't triggered by regulation, war, or energy scarcity. It is triggered by AI improving itself without human help.
That detail is the whole article. Everything else is arithmetic.
Context
Anthropic's Economic Scenarios is not an AI model. It is a growth-accounting scenario tool โ a spreadsheet with ambitions. Its exogenous variables are AI capability tiers: one where automation touches a slice of the economy comparable to the internet's footprint; one where AI absorbs half of all knowledge work; and one where recursive self-improvement kicks in and output climbs from a $34 trillion baseline toward $44.4 trillion while the labor share collapses.
The methodology is deliberately interactive. Readers submit their own forecasts and compare them against a 10,000-person survey; a separate June poll found that unemployment anxiety, not existential risk, is what ordinary respondents actually fear. The framing is "choice, not prediction" โ a phrase doing real work, because a choice implies a chooser.
What's missing is equally load-bearing: no production function, no calibration notes, no elasticity-of-substitution assumptions, no open code, no robustness checks. We can confirm Anthropic built a tool. We cannot verify the tool's internals.
Core
Here's the gap a crypto analyst notices immediately: a model predicting a collapse in labor's share of income says almost nothing about who ends up holding the capital. That is not a rounding error. It is the entire distributional story. In extreme-scenario terms, capital owners and compute controllers capture more than half of national income โ but the report never itemizes that concentration.
On-chain, we'd see it. Not because blockchains are fair, but because they are legible. Balances are auditable, treasury flows are traceable, and the difference between a claim and a right is written into contract code. That legibility is the one thing crypto genuinely contributes to this debate: a ledger where value capture is a matter of record rather than a matter of trust.
I've spent this year mapping how AI agents will transact โ machine-to-machine payments, agent identity, provenance for generated content. My working thesis is that the next cycle's scarce asset isn't compute, it's verification: who attested that this output came from this model, trained on this data, under this license. That is the trust layer for machines. The narrative is the asset; the code is the proof.
But the labor-share arithmetic exposes an uncomfortable asymmetry in my own sector. Consider what tokenized compute networks promise: distributed GPU supply, inference paid in tokens, "decentralized AI." Now watch where the margin goes. In most of these designs, hardware providers and foundation treasuries take the spread; the token holder gets a governance vote and a vesting schedule. It is the Cosmos problem in a new costume โ elegant architecture, fragmented applications, and a token capturing essentially none of the value flowing through the rails it secures.
The yield-farming parallel is almost too neat. In 2020 I wrote a primer explaining that liquidity mining APY was a project subsidizing its own TVL figure โ pull the emissions and the "users" evaporate. Compute networks are running the same experiment with inference subsidies: discount tokens, free tiers, points programs. When the subsidy stops, what remains is a question the token price answers before the documentation does.
Then there's governance. Anthropic says the extreme path is a choice. Chosen by whom, through what mechanism, with what veto? DAO governance tokens were supposed to answer a version of that โ collective control over shared infrastructure. What they deliver, in nearly every implementation I've audited or held, is non-dividend equity with no cash-flow right and a vote the foundation can route around. The parallel isn't decorative. It's a warning: we already ran the experiment of "token holders decide," and decision-making migrated back to the multisig.
Two pieces of my own work shape this reading. When I audited TheDAO's codebase in 2016, the reentrancy flaw was boring โ a state-update ordering issue any careful reader could find. The interesting part was the social layer: thousands of participants had bought governance rights they could neither interpret nor exercise, and the repair was ultimately executed by a small group holding the keys. Nine years later, the governance surface has grown and the accountability surface hasn't. Separately, across three meetups in Taipei and Tokyo, I interviewed 30 NFT holders about status and identity. The pattern repeated: communities believed they were the asset, then discovered they were the exit liquidity.
Both experiences taught me the same lesson. The crowd is the narrative. The keys are the proof.
Contrarian
The reflexive crypto answer is that decentralized AI fixes this. I don't buy it, and the labor-share numbers are why.
If automation raises the capital share of income, then broadening access to capital claims matters more than decentralizing model weights โ and token distribution is a poor instrument for that. Foundations and early investors front-load supply; the retail cohort crypto actually serves is overwhelmingly wage labor with a side portfolio. A token portfolio is not a substitute for a paycheck, and 45.2% is a statement about paychecks.
Here's the blind spot: many crypto participants believe they sit on the labor side of this ledger. Structurally, they hold the capital side โ badly. They own claims with no cash flow, no control, and no legal standing, priced against an economy where their wages are flat and their rent is not. This consolidation is exactly the environment where that mismatch goes unnoticed, because price action substitutes for analysis. Sideways markets are for positioning, not validation.
Takeaway
The forward-looking question isn't whether AI doubles GDP. It's who holds the ledger when it does โ and whether that ledger is auditable by anyone outside the room where the training run was approved.
Anthropic published the scenarios. It did not publish the cap table for the world it describes. Where code meets culture, the real value emerges โ and the code being written this cycle is about provenance, identity, and who gets to say no.