Anthropic published three economic scenarios for transformative AI this month, and one number jumped off the page before I finished the opening section: labor's share of income falling to 45.2%.
That is not a forecast. It is the extreme scenario, triggered by what the report calls unaided self-improvement — a system that recursively redesigns itself without human help. The middle case, labeled significant, puts AI at half of all knowledge work, drags GDP from $34 trillion toward $44.4 trillion, pins unemployment at 5%, and leaves knowledge-worker wages flat. Labor's share there: 56.1%, against roughly 60% in the United States today.
The timing is the part my feed skipped. Hours before publication, Anthropic researcher Jacob Coxon resigned and said publicly that the industry is racing toward exactly that self-improving superintelligence. The model's worst case and the exit memo's warning describe the same event in two dialects — one economic, one existential. I was midway through an audit of an agent-payment contract when the news landed, and I closed the terminal. Both files belonged on one screen.
Chasing the alpha, one block at a time. This one is not priced anywhere yet.
Strip the branding and the deliverable is not an AI model at all. It is a growth-accounting scenario tool that treats AI capability as an external variable and pushes it through to comparable macro outputs — GDP, unemployment, labor's share of income. Three tiers anchor three capability levels: AI at internet scale, AI carrying half of knowledge work, AI improving itself with no human in the loop.
Two things matter for anyone trading this. The mechanics are opaque. No published equations, no parameter calibration, no robustness checks in what has been released. We can confirm Anthropic shipped the tool. We cannot independently validate the machinery inside it, which means every headline number inherits that uncertainty.
The tool also ships interactive. Readers submit their own predictions, and those get compared against a broad prior survey. That is a public-expectations database being assembled in real time — collected at near-zero cost, and useful later for policy papers, academic citations, and enterprise advisory work. Anthropic frames all three outcomes as choices rather than destiny. That framing is a policy move, not a neutral one. A lab that supplies the vocabulary of AI's economic impact gets to shape how regulators describe the problem — and the vocabulary is already sticking: mild, significant, extreme.
Read the significant scenario literally and it describes something we have never seen at national scale. Output accelerating while the wage channel to ordinary workers stays frozen. Growth doubling, unemployment at 5%, knowledge salaries going nowhere, and the gains landing on capital. If that sentence sounds familiar, you have been watching token charts.
I pulled holder distributions on a basket of eight AI-agent tokens in the first quarter, and the pattern was almost mechanical. Top 50 addresses held between 61% and 89% of circulating supply. Vesting cliffs clustered into a nine-month window. Community allocations, where they existed at all, were the smallest line item in every deck. The report calls this scenario "significant." On-chain, we call it Tuesday. The 45.2% extreme case is just the same curve with the slope turned up, and retail liquidity is the labor being squeezed.
There is another thread, and it is the one that sits badly with me. The extreme scenario's trigger is unaided self-improvement. The compute that would underwrite that capability is increasingly procured through decentralized markets — GPU aggregators, inference networks, verifiable compute layers. Those networks pay suppliers in tokens because token incentives are the only way they can compete with hyperscaler pricing.
I have run a mid-tier GPU on two of these networks. The economics only clear when the token price holds. When emissions taper, suppliers churn, utilization drops, and the whole cost curve reverts to something a data-center operator would laugh at. Which means the token market is now a subsidy channel for the exact capability path the report flags as its worst case. Every yield farmer renting out idle silicon is, at the margin, financing the nine-figure training run that Coxon resigned over. Nobody structured it that way. That is how it clears.
From the front lines of the hype cycle, the more concrete story is payments. Agent-to-agent settlement is live and it is denominated in stablecoins. Machine wallets paying for inference, data, and API calls, settling in seconds, with no human in the loop on either side. I stress-tested three of these rails last month. Two cleared sub-second under normal conditions.
The third stalled when its price feed lagged during a volatility spike. The agent kept firing orders against a stale quote and burned its entire budget in eleven minutes. No circuit breaker, no throttle, no sanity check on the numbers it was reading. That is where an economic abstraction becomes a line item on someone's P&L. If machine labor is the marginal worker in the significant scenario, its payroll runs through oracle infrastructure that has never been proven under load. Oracle latency is DeFi's oldest weakness, and we are about to hand it the keys to the labor market.
There is an institutional translation that retail readers keep missing. Every one of these figures is a scenario, not a probability. Anthropic published three futures and no likelihood weighting, which means the headline "45.2%" is doing the work a confidence interval should be doing.
Then there is the fragmentation problem. The machines do not settle on one chain. I mapped the settlement paths of four agent frameworks in Q1 and counted activity across seven distinct networks, most of it in bridged or wrapped form, with liquidity silos that do not talk to each other. A dozen Layer 2s, the same small set of active addresses, and an agent that has to route a payment through a canonical bridge with a seven-day exit window.
That is not scaling. That is slicing an already-scarce liquidity base into fragments and calling the fragments a market. When the Anthropic model talks about AI carrying half of knowledge work, it implicitly assumes a settlement layer that clears instantly and atomically. What exists today is a patchwork where half the payment paths add a bridge risk premium the model never prices.
Underneath all of it sits a definitional gap nobody has closed. "Half of all knowledge work" is never quantified. Which occupations, measured how, on what timeline? The public survey behind the tool asks people to predict the future, not to rank their preferences over it, so the "choice" language carries no authorization behind it. It is a poll dressed as a referendum.
The report does admit one thing plainly: in the extreme case, the fastest-growing economy distributes its gains unevenly. Credit where due — a high-valuation AI lab publishing that its own best-case growth scenario concentrates benefits is not the comfortable move.
Here is the contrarian read, and it is not the one circulating. Everyone is arguing about whether the numbers are too pessimistic. The blind spot is the opposite direction, and it is structural.
Every scenario in the model implicitly assumes the AI stays coherent enough to produce measurable GDP. GDP is a coordination metric. It requires property rights, enforceable contracts, and agents that honor obligations. The extreme case is triggered by a system improving itself without human help, and the report then measures that world in dollars of output. You cannot run a growth-accounting model on a system whose objectives you cannot specify. The numbers are the optimistic part. The trigger condition is where the analysis quietly stops.
The framing is the next blind spot. Anthropic's scenario vocabulary is on track to become the default reference for AI economic impact, the way "base case" and "tail risk" became standard macro language. Defining the categories is worth more than being right about any single number. That is a durable intellectual asset, and it will show up in regulatory consultations long before it shows up on a P&L.
And then there is talent flow, the leading indicator nobody is trading. Coxon's exit hours before publication is not a coincidence you can wave away. It is a data point about how internal safety staff assess the real capability timeline versus the public one. If more of that cohort leaves — to third-party safety shops, policy roles, or new labs — the safety-first brand gets thinner every quarter, and the market will reprice that long before it reprices the model.
Surviving the winter to plant for spring means reading exits, not press releases.
So what do I actually watch from here, in a market that is chopping sideways and giving nobody a clean entry?
Speed is the only currency that matters, and the measurable version of this debate is on-chain. Three signals, all observable this quarter. Agent-to-agent stablecoin payment volume, tracked per active machine wallet rather than per dollar headline. Decentralized compute utilization after emissions taper — if it holds without token subsidy, the cost curve is real. And oracle update latency on the rails that settle machine labor, because that is where the next blowup prints.
If the significant scenario is even half right, the repricing will not arrive as a macro forecast. It will arrive from a stale price feed on a Tuesday afternoon, with nobody at the keyboard to hit cancel.