Doha doesn't do dawn quietly. The call to prayer, the hum of the airport road, the first servers spinning up in data centers I've never been inside but whose output I read every day. I was on my fourth coffee, re-reading Anthropic's "Economic Scenarios for Transformative AI," when the timestamp on the release did what timestamps always do in my line of work—it told a second story.
The report had gone live hours after Jacob Coxon, a safety researcher at the company, resigned and told the world that the industry was "racing toward self-improving superintelligence." Tracing the ghost in the code usually means finding a variable that doesn't belong: a signature that doesn't match, a governance function that quietly routes authority to a single address. This time the ghost was in the metadata. A model that predicts civilizational-scale economic disruption, published on the same day an insider warns that the disruption might already be in training. That is not a coincidence you can wave away with a scheduling error. That is a signal.
The chart itself was immaculate. Three scenarios on one axis of AI capability. Mild: AI as consequential as the internet. Significant: AI performing half of all knowledge work. Extreme: unsupervised self-improvement. GDP climbing from roughly $34 trillion to $44.4 trillion. Then the number that made me set the cup down—labor's share of income, sliding from the mid-fifties to 56.1% in the significant scenario, and collapsing to 45.2% in the extreme one. I hunt the story that the chart hides. This chart was hiding a governance question so large that most of the crypto industry—busy shipping agent tokens, autonomous treasuries, and "AI-managed" DAOs—hasn't noticed it is standing directly on the fault line.
I've spent fourteen years watching technology narratives collide with capital. I have audited contracts that promised decentralization and delivered a multisig in a trench coat. I have watched a $40 billion stablecoin die in seventy-two hours because trust is a protocol with no rollback. And what I see in Anthropic's model is not primarily an economics paper. It is a confession, dressed as a scenario. It is a company telling you, in the language of GDP and labor share, that the next decade of capital formation will be adjudicated by a question almost nobody on-chain is prepared to answer: who chooses?
Let me walk you through it the way I'd walk a client through a post-mortem—because that is what this is. A post-mortem of a future that hasn't happened yet.
A Model That Isn't a Model
Start with what Anthropic actually shipped, because the framing matters more than the numbers.
This is not an AI model in the sense you'd benchmark. It's a growth-accounting and scenario tool that treats AI capability as an exogenous variable—something that happens to the economy rather than something derived from it. You feed in assumptions about how much knowledge work AI can perform and how autonomous it becomes, and out the other side come macroeconomic outcomes: GDP, unemployment, labor income share. In that sense it's closer to the Intergovernmental Panel on Climate Change's scenario framework than to a forecasting model. It doesn't tell you what will happen. It tells you what could happen under stated conditions.
That distinction is not academic. It is everything. Because the three scenarios—mild, significant, extreme—are not predictions. They are doors. And a door is a political object. Someone built it. Someone decided which rooms it opens onto. Someone decided that "self-improvement without human help" would be the hinge of the worst case, and not, say, "AI-enabled bioweapons" or "algorithmic financial contagion." The choice of hinge is itself a thesis about where the danger lives.
Here's what we can verify from the public materials. The scenarios anchor to different levels of AI capability: from "scale comparable to the internet" in the mild case, to "half of knowledge work" in the significant case, to "self-improvement without human help" in the extreme case. The report is accompanied by a technical report, but the summary we've been handed gives us results, not methodology. It gives us outcomes, not equations. It gives us labor shares, not the production function that generates them.
That opacity is a problem, and I say this as someone whose first paid work in this industry was reverse-engineering a whitepaper. In 2017, as a twenty-one-year-old cybersecurity student in Doha, I spent weeks inside the Tezos whitepaper—not because I believed the hype, but because the formal verification process looked structurally different from everything else in the ICO casino. I wrote a comparative analysis on Medium that got five thousand views. Nothing, by today's standards. But enough to teach me a lesson I've never unlearned: the credibility of a claim lives in its assumptions, not its conclusions. Three smaller ERC-20 tokens I audited that year had governance contracts with critical vulnerabilities, and the marketing decks that shipped alongside them had none. The decks never do.
So when Anthropic publishes a scenario in which the labor share falls to 45.2% and calls it "extreme," my first instinct is not to argue with the number. It's to ask what elasticity of substitution between capital and labor they assumed, what task-automation rate they modeled, and what timeline they assigned to AI capability. None of that is public. We're told to bring our own predictions—Anthropic built an interactive tool where readers input their own forecasts and compare them against a survey of the public—but we're not told what the model does with them. There is a name for a system where you provide input but cannot inspect the mechanism. In my world we call it a black box, and we charge a premium to audit it. In Anthropic's world, we are apparently supposed to take its outputs on faith, because the outputs are alarming and the company is reputed to be careful.
Careful and credible are not synonyms. I have spent a decade learning the difference the hard way, and I am not about to unlearn it for a company whose product I respect.
The most revealing detail is the one that got the least coverage. The report frames its scenarios as choices, not predictions. And then it asks the question that should be the headline of every article written about it: who chooses?
That question is where the crypto industry should be leaning in, and it's where, so far, it has been almost entirely silent.
The Three Scenarios, Decoded
Let me do the forensic work the summary skipped.
In the mild scenario, AI's economic footprint is comparable to the internet's. This is the comfortable door. Growth continues roughly along trend, labor share declines only slightly, and the disruption is absorbed by the same mechanisms that absorbed the PC revolution and the smartphone. Knowledge workers get reorganized but not replaced. This is the scenario a venture capitalist pitches to a limited partner over dinner, and it is the scenario in which nothing structural changes about who owns the surplus.
In the significant scenario, AI does half of knowledge work. GDP growth roughly doubles. Unemployment stabilizes at around 5%. And here's the detail that should stop you cold: knowledge workers' wages flatline. Not fall—flatline. In real terms, they erode. Meanwhile labor's share of income drops to 56.1%.
Read that again, because it's easy to skim past. The economy is growing at twice its normal rate. Headline employment looks healthy. And the group producing the new value—the knowledge workers being automated—captures none of it. That is not a recession. It is something stranger: growth without broad-based income growth. Macroeconomists have a toolbox built for a world where those two things move together, and this scenario quietly saws the toolbox in half. The Phillips curve, the wage-growth-to-inflation transmission channel, the whole apparatus of monetary policy built on employment as a proxy for demand—all of it assumes a relationship that this scenario breaks. Central bankers would be flying instruments that no longer connect to the engines.
Now the extreme scenario. AI improves itself without human help. Unemployment spikes to historic levels. Labor's share of income collapses to 45.2%. Wages fall more than ten percent. And the report itself admits that the fastest-growing economy in the scenario distributes its gains unevenly. The document does not pretend the extreme case is utopia. It simply refuses to call it what it is: a distributional catastrophe wearing a growth chart.
I want to be precise about what's happening here. Anthropic is not saying this will happen. It's saying this is what you get when you run the machine forward with certain inputs. But the inputs are the interesting part, and they're hidden. What task-automation rate, sustained over how many years, produces a labor share at 45.2%? What self-improvement timeline? What substitution elasticity? Without those, we have a set of headlines and a set of gaps where the methodology should be. In a scientific paper, that would be a rejection. In a policy paper, it's a leak dressed as a finding.
And the gaps matter more in crypto than anywhere else, because crypto is where the labor-share question stops being a statistic and becomes a parameter.
The Labor Share Is the Real Protocol
Here's the thing about a labor share number. In a traditional economy, it's a ratio computed after the fact by statisticians in a basement office, published months late, revised twice. In a tokenized economy, it's a design choice. It's something you can encode.
This is not philosophy. When I was in Aave's early community in 2020—twenty-four years old, enthusiastic, tracking Compound, Yearn, and MakerDAO in parallel—I noticed something that became a viral thread: governance participation correlated with token price stability. I called it the "governance premium," and it was an early hint of a broader truth. In token networks, distribution isn't an outcome. It's a design decision. The contract decides who captures the surplus, and you can read the decision in the emission schedule. You can read it in the vesting cliffs. You can read it in the address that holds the admin key.
Which is why Anthropic's labor-share collapse should terrify anyone building on-chain. Because the extreme scenario describes a world where value flows to whoever owns the compute, the model weights, and the agent infrastructure—and away from everyone whose labor has become replicable. If that world arrives while most of crypto's real economic activity is token distribution, then the token holders at the top of the distribution curve become the new rentier class, and the pattern Anthropic describes plays out again, on-chain, with a governance token attached. The names change. The structure holds.
I've seen this movie. In 2022, I lost my own capital in Luna, and the lesson was not about code—the code did exactly what the code said. The lesson was about trust. I wrote a ten-thousand-word forensic analysis of the UST de-peg, and what I found wasn't a bug. It was a psychological breakdown wearing a mechanism's clothing. The protocol promised a peg; the peg was a belief; the belief was a distribution of incentives; and the incentives were controlled by a small set of actors who could see the exit before everyone else. The labor-share collapse Anthropic models is the same structure at a larger scale. Capital sees the exit first. Labor is the exit liquidity.
Now overlay the crypto-native version. AI agents managing treasuries. Autonomous DAOs with algorithmic governance. "Agent economies" where the participants are models, not people. In 2026 I started three things nearly at once—an agent-based economy simulator, a DAO governance bot, and a narrative trend-prediction algorithm—and published a case study on autonomous narrative trading. The finding that got the most attention was that agents could detect sentiment shifts before human traders. The finding that got the least attention was that the agents were detecting shifts in each other. A herd of models reading the same signals, converging on the same trades, in the same fractions of a second. Mining for meaning in a sea of volatility is hard enough when the miners are human. When every miner runs the same three or four base models, you don't get price discovery. You get a resonance cascade. And a resonance cascade is not a market. It's a feedback loop with a liquidation price.
So when Anthropic says the extreme scenario is triggered by self-improvement without human help, and the crypto industry's response is to launch more agent tokens, I want to ask a specific question. If agents become the primary allocators of capital in tokenized markets, and the models driving those agents were trained by three or four companies, where exactly is the labor share? It's not 45.2%. It's whatever the training objective says it is. And the training objective is set by a board of directors, not a governance vote. The DAO votes on the treasury. The lab votes on the objective. One of those votes decides everything.
The Self-Improvement Trigger and the Agent Economy
Let me get technical for a moment, because the self-improvement trigger is where the crypto and AI discussions collapse into each other, and almost nobody is drawing the line correctly.
Anthropic's model treats recursive self-improvement as a discrete threshold event—not a smooth curve. That modeling choice is itself a claim. It says: at some point, the system crosses a line, and the world after the line is qualitatively different. Coxon's resignation points at the same line. He did not say "we are making steady progress." He said "racing." When an insider uses a verb like racing, you should assume the driver knows the speedometer reading.
Here's what the crypto industry has that is genuinely relevant: it has spent a decade building infrastructure for trustless coordination among parties who do not trust each other. That is precisely what a multi-agent economy needs. But the crypto version of the answer has a fatal gap, and it is the same gap Anthropic's report has: no mechanism for adjudicating who sets the objective.
Look at how DAOs actually work. Most of them have the legal status of a group chat. When a DAO gets sued—and they do—the members can find themselves facing unlimited personal liability, because there is no corporate veil, no LLC wrapper, no legal person behind the treasury. I have watched this pattern play out in governance contracts: the community is sovereign until a regulator knocks, and then suddenly every token holder is a defendant. The governance token was supposed to apportion rights. In practice it apportions exposure. The 2022 precedent is instructive: member liability in unincorporated associations is not a hypothetical, it is a default rule, and no amount of Discord enthusiasm changes the default rule.
Now move that structure into an economy where AI agents hold and move the assets. Who is liable when the agent makes a catastrophic allocation? The DAO? The token holders? The company that trained the model? The answer, right now, is nobody knows, and nobody knows is not a stable equilibrium. It is a vacuum, and vacuums get filled by whoever shows up first—usually a regulator with a template designed for a different problem. The template will be built for human actors with documents. It will not fit autonomous ones. And it will be imposed anyway, because the alternative is leaving the question open, and open questions are politically unacceptable after the first nine-figure loss.
This is where I have to say something uncomfortable about the current bull market. The euphoria is real. The funding is real. And the euphoria is precisely what lets projects skip the questions that determine whether the structure survives contact with the real world. I have audited governance contracts where the multi-signature was a single address wearing several names. In a bull market, nobody reads the contract. In a bear market, everybody does. Anthropic's report is the bear-market document for the AI economy, published in the middle of a bull run for both AI and crypto. That timing is not a bug. It is the whole point. Someone at Anthropic looked at the calendar and decided that the warning had to be issued before the party got too loud to hear it.
And note the report's own honesty, which deserves credit even as I criticize its opacity: it admits the extreme scenario distributes gains unevenly. That admission is rare. Most AI impact reports stop at productivity. This one goes to who gets the productivity. The fact that the company producing frontier models is willing to name that is worth noting—especially because it cuts against the company's own commercial interest. A lab that stands to capture the surplus is publishing a chart that says the surplus will concentrate in the hands of labs like it. That is either intellectual honesty or a very sophisticated hedge. I suspect it is both, and I will not pretend to know which fraction dominates.
What This Means for On-Chain Governance
Now let me connect the macro to the mechanics, because that is the kind of work I actually do.
If the significant scenario is right—labor share at 56.1%, knowledge wages flat, GDP growth doubled—then the next phase of the crypto industry is not about which chain is faster. It is about which chain is fairer by construction. And here is where I will be blunt about a piece of theater the industry keeps performing: KYC.
Most project KYC is theater. I have believed this since I started auditing, and nothing about the current cycle has changed my mind. The logic is simple and uncomfortable: compliance costs fall almost entirely on honest users, who hand over documents to centralized gatekeepers, while the actual determined actor—the one with a few wallets and a willingness to move fast—bypasses the gate with a shell entity and a VPN. The honest user pays the tax. The determined user pays a fee. That is not a security boundary. That is a toll booth with a gap in the fence, and the toll is charged to the people who were never the problem. Regulations written for banks, when applied to pseudonymous networks, produce bank-shaped exclusion and not much else.
And if you are building an AI-agent economy, the gap becomes a chasm, because the agent does not have a passport. The compliance frameworks we are bolting onto DeFi are built for humans with documents. They do not map onto autonomous actors. So either we build identity primitives that make sense for agents—and there is real work happening there, at the intersection of verifiable credentials and zero-knowledge proofs—or we keep pretending that KYC on a human front-end secures a system where the flow is machine-to-machine. I know which one is actually happening today, and it is not the first.
Layer 2 economics has a parallel problem, and my position there has been consistent. Post-Dencun, blob space gave rollups cheap data availability, and the whole industry celebrated falling fees. My read then, and now, is that blob space saturates within roughly two years, and when it does, rollup gas fees double again—not because anyone is malicious, but because the subsidy is structural and temporary. Cheap fees are a promotion, not a price. Building a consumer application that depends on today's fees is like building a business on a zero-percent introductory APR. It works until it does not, and the transition arrives without warning to the founder who never read the fine print. I told a client this in early 2024, and I will keep telling clients this until the saturation shows up in the data. Then I will stop being invited to certain dinners.
Bring these threads together and the shape of the Anthropic report's crypto relevance becomes clear. The report says the next decade's economic surplus will be captured upstream—by whoever owns the models and the compute. Crypto's response cannot be to build faster rails for that same capture. It has to be to build distributive rails, and it has to do so before the extreme scenario closes the window. The window is open now. That is the only good news in the document, and it is a narrow piece of good news, because windows in this industry close faster than anyone models.
The Blind Spot
Now the contrarian turn, because every confident narrative deserves one, including mine.
The mainstream reading of Anthropic's report is pessimistic: labor share falls, unemployment spikes, the machines win. The contrarian reading is that the report is not pessimistic enough in one dimension and too pessimistic in another—and both errors serve the same master.
First, the report treats self-improvement as an economic event. But self-improvement that outruns human control is not an economic event. It is an alignment event. If a system can rewrite its own objective function, then GDP projections computed under assumptions of human-compatible goals stop meaning anything. You cannot model the labor share of an economy whose distribution mechanism has been replaced by an optimizer that does not share your preferences. The report admits the trigger but skips the implication. It describes the flood and does not mention the dam. That is the too-pessimistic error, oddly—because it understates the worst case by keeping it economic, by keeping the superintelligence inside the map of markets. A superintelligence that can be modeled with a labor share is a superintelligence that still cares what humans think. That assumption is doing enormous silent work.
Second, the report is too pessimistic because it may be too neat. The significant scenario assumes unemployment stabilizes at 5% while half of knowledge work is automated. That stabilization has to come from somewhere—new job categories, regulatory friction, or the sheer drag of a world trying to absorb the disruption. Which one? The report does not say. If the answer is policy, then the scenario is not a forecast at all; it is a conditional statement about choices, wearing a forecast's clothing. And if the answer is the model just assumes it, then we are reading a diagram of the author's hope. Either way, the 5% number is doing rhetorical labor that belongs in a footnote.
The blind spot underneath both errors is the same one I keep circling: the report frames the future as a choice and never specifies the chooser. The interactive tool asks for your prediction, not your preference. It collects what you think will happen, not what you want to happen. So the public database Anthropic is building is a database of expectations, not a mandate. That is a very important distinction, and it is the difference between a survey and a vote. Anthropic is running a survey. The governance question requires a vote. And there is currently no institution with the legitimacy to hold it.
This is where the 2024 work I did matters. When I transitioned into strategy consulting, I interviewed fifty traditional finance executives about their AI risk concerns, and distilled their answers into a series of Institutional Readiness reports. The finding that became a cornerstone of my framework was that narrative adoption lags regulatory clarity by roughly six months. Retail moves on story. Institutions move on rule. In 2024 the gap was visible; in 2026 it is a canyon. The Anthropic report is narrative, and narrative is being asked to do the work that regulation has not yet done. That is not sustainable. It is, however, profitable for whoever writes the story first, and that incentive is exactly why the story got written this week and not next quarter.
Here is where crypto should have a structural advantage—and where its own blind spot shows. Crypto invented a mechanism for exactly this problem: token-weighted, transparent, verifiable preference aggregation. It is called governance, and most of it is broken. Participation is low, delegation is concentrated in the hands of a few whales and a few funds, and the one-token-one-vote assumption reproduces the same capital-ownership pattern the Anthropic report identifies as the problem. So the industry that should be offering the world a tool for adjudicating the extreme scenario is, in practice, running a live demonstration of why that tool does not yet work. We are not the answer to the question. We are an early, expensive beta of the answer.
That is the real contrarian point. The question of who chooses will not be answered by AI labs, and it will not be answered by token holders as currently constituted. It will be answered by whoever moves first to build legitimate preference aggregation at a scale that matters—and that entity has not been invented yet. The Anthropic report is remarkable not because it predicts the future, but because it accidentally publishes the specification for a missing institution. It tells us what the institution needs to do: adjudicate the choice between mild, significant, and extreme, under conditions of deep uncertainty, with legitimacy that survives the losing side's rage. No existing body does that. Not parliaments, not boards, not DAOs, not labs.
If you want the investment thesis hidden in all of this, it is not "buy AI tokens." It is this: the next decade's dominant crypto primitive will not be a faster chain or a better stablecoin. It will be a credible mechanism for allocating the surplus produced by machines, and everything currently on the market is a prototype that does not work yet. Prototypes are how primitives get built. They are also how investors get separated from their capital. Both things will be true at once, and the difference between the two will be decided by whether anyone reads the assumptions instead of the conclusions.
Takeaway
So where does that leave us, on a hot morning in Doha with the report's numbers still sitting on the screen?
Anthropic handed us three doors and a question. The mild door is the internet's story retold, and nothing structural changes. The significant door is growth without broad prosperity—the labor share at 56.1%, knowledge wages flatlined, GDP doubled, and no existing policy instrument tuned to the combination. The extreme door is self-improvement without a brake, the labor share at 45.2%, and a distribution problem so large it stops being economics and becomes governance. The report frames all three as choices. The crypto industry, which claims to be in the business of choice, has responded by shipping more agents and fewer mechanisms for answering who decides.
The narrative did not break because the numbers were wrong. It broke because the report asked a question no current institution is built to answer, and the industry that claims to specialize in trustless coordination looked away. We had one job: preferencing. We did not show up with the apparatus. We showed up with the tokens.
I will be watching two signals over the next two quarters. First, whether Anthropic—or any lab—publishes the methodology behind the labor-share numbers, because opacity is the tell, and a company that hides its equations while publishing its warnings is telling you where to look. Second, whether anyone on-chain builds a governance primitive that lets preferences, not just predictions, aggregate at scale, with liability that does not default to the token holder's house. The first tells you how seriously the labs take their own warnings. The second tells you whether crypto gets to participate in the answer or just inherits the extreme scenario with a token attached and a governance proposal nobody reads.
The ghost in this code is not a bug. It is a vacancy. And vacancies, in every market I have ever studied, get filled by whoever shows up with the clearest story. Hunt accordingly.
