Over the past several days, a single number has been moving through procurement decks, career forums, and โ inevitably โ crypto group chats: Anthropic staff expect AI to handle 75% of their work by 2030. The item arrived as a Crypto Briefing brief. Three sentences. No sample size. No survey wording. No publication date. No methodology. And, despite the masthead, not a single crypto element in the body.
That last omission is the tell.
I have spent twenty-nine years watching how a data point becomes a market, and the transformation almost never depends on the data being true. It depends on the data being quotable. "75% by 2030" is quotable. It fits on a slide, in a headline, in a board memo recommending a hiring freeze. And in a bear market โ where every protocol I track is bleeding liquidity and every allocator is looking for a reason to stay in cash โ quotable numbers do more damage than false ones, because nobody stops to ask who benefits from the quote.
Let me be precise about what this brief actually is, because the answer changes how you should hold your positions.
The brief carries three information points and nothing else: Anthropic staff expect AI to handle 75% of their work by 2030; the article notes a "disparity" in expectations without saying between whom; and it is published by Crypto Briefing. That is the entire payload.
No sample. No distribution. No confidence interval. No date. A single point estimate, produced by the employees of the company that sells the product being estimated, routed through a tech outlet, then re-routed through a crypto outlet with no domain expertise in AI research and no incentive to check. Each hop strips a qualifier. By the time it reaches you, the qualifiers are gone and only the number remains.
This is the same mechanism that built the data availability narrative. For two years I have argued โ quietly, and then less quietly โ that the DA layer is overhyped, that 99% of rollups do not generate enough data to justify dedicated availability infrastructure. The technical case was never the problem. The problem was that "modular DA" was quotable, fundable, and fit on a slide. Capital arrived before demand, because the narrative was legible and the demand curve was not. We are watching the AI compute buildout run the identical play, at a hundred times the scale, with the same absence of demand verification.
In 2017, when the ICO boom was rewarding hype over safety, I co-founded an open-source advisory platform and ran forty webinars for over five thousand participants, decoding smart-contract security into plain language. We helped twelve projects secure their code before mainnet. The lesson I took was not that education works โ it was that the market prices legibility, not accuracy. A clear number beats a correct one, every cycle.
So when a vendor's own employees forecast that the vendor's product will absorb three-quarters of knowledge work, the correct response is not "is this true?" The correct response is "what is this number doing?" A forecast is a tool. Ask what work it performs.
Strip the marketing and "75% of work by 2030" is an extraordinary capability claim: it requires AI to execute cross-task, long-horizon, verifiable autonomous work โ not the single-turn generation we have today. The most relevant measurable proxy is the task-horizon curve. METR's 2025 research found that the length of task a model can complete with 50% reliability has been doubling roughly every seven months, from a baseline of about fifty minutes of human work. Extrapolate that linearly to 2030 โ eight or nine doublings โ and you land at tasks spanning hundreds of hours, multiple weeks of human labor. On that curve alone, the 75% claim is not delusional. It is a straight-line projection of a real exponential.
But straight-line projections of exponentials have three known fracture points, and the brief acknowledges none of them. The data wall and diminishing compute returns are one. The reliability gap is another: moving from 50% to 99% โ the enterprise-usable threshold โ is not a linear effort, it is a different engineering discipline. And error compounding is the third. A model that is 99% correct per step is only 37% correct end-to-end across a hundred-step task. Long-horizon autonomy is where reliability goes to die.
The number's commercial function is harder to see and more important. "75% by 2030" is not a research conclusion. It is a pricing argument wearing a research conclusion's clothes. If a customer believes AI can do 75% of the work, their willingness to pay shifts from the tooling budget to the labor budget โ a tenfold expansion of the addressable price. That is the entire game. The narrative moves the TAM from global software spend, roughly a trillion dollars, to global knowledge-work compensation, thirty to forty trillion. Same technology. Two TAMs. One order of magnitude in valuation difference. I have audited enough token launches to recognize this move: it is the total-addressable-market slide, rewritten for a lab that does not have to file a prospectus.
The investment mechanics deserve their own paragraph, because this is where the number becomes load-bearing. Anthropic's 2025 funding round priced it in the vicinity of $180 billion; subsequent reporting put a new round's discussions above $300 billion. Multiples like that are not underwritten by current revenue โ they are underwritten by the belief that AI will reprice human labor. When the narrative is the collateral, narrative decay is a margin call. The feedback loop โ aggressive forecast, larger TAM, higher multiple, more compute and talent, stronger capability, more credible forecast โ is still self-reinforcing. Its fragility is verification: the loop assumes the pricing power of replacement is real, and no enterprise renewal data has yet confirmed it. The labs' cash burn is measured in billions per year, and inference costs rise non-linearly with agent loads. Narrative-driven financing is not optional for them. It is operating cash flow.
The competitive function of the release is the most revealing. Why would a company let an "employee survey" leak rather than have an executive state the forecast on the record? Because an employee survey is deniable. If the number is challenged, it becomes individual opinions from an internal poll, not corporate guidance. This is plausible deniability engineered into the release format. And the timing is not random โ the brief has no date, but the capability claims align precisely with the agent-product cadence these labs have been shipping: tool use, computer operation, long-running tasks, standardized context protocols. Expectation management and product release are in conversation with each other. That is standard operating procedure at every frontier lab, and it should be priced as such. The strategy makes sense for the challenger, too. When a competitor owns the largest user base and another owns full-stack cost advantages, the rational move is not to fight on price or scale โ it is to raise the dimension of the debate to the ceiling of capability. The more aggressive the timeline, the more favorable the comparison.
Now the part the brief itself gestures at but never substantiates: the disparity. The article mentions that expectations differ, then declines to say between whom, by how much, or with what evidence. This is the one genuinely interesting claim in the piece, and it is the one left empty.
Here is what the disparity looks like when you actually measure it. AI lab employees are the most extreme early-adopter sample on earth โ their job is building and using the frontier model, with internal agent tooling most enterprises will not see for years. Extrapolating from them to the general workforce is a textbook selection bias. The gap between lab employees and ordinary knowledge workers is plausibly thirty to fifty percentage points. The gap between the technology sector and policymakers lags two to three years. And the gap between the people who make procurement decisions and the people who absorb them is where the friction actually lives: the buyer believes in replacement, the worker does not, and that contradiction is the real 2026-to-2028 deployment story.
Meanwhile the labor-market evidence is already arriving, and it does not look like 75% across the board. Stanford's Digital Economy Lab found that in AI-exposed occupations, employment for workers aged 22 to 25 fell about 13% relative, while senior workers were essentially untouched. Junior collapse, senior augmentation โ an asymmetric pattern that is narrower and earlier than the headline, and already underway. You do not need to wait for 2030 to see the displacement. It is concentrated in the entry-level rung, and it is happening now.
Which brings us to the semantic slip that makes the whole number untrustworthy as a forecast. Task automation rate is not hours-replaced rate, and hours-replaced rate is not headcount-replaced rate. These three are silently equated in the brief and in every retelling. If 75% of subtasks are automated but the remaining 25% โ judgment, accountability, cross-party coordination, exception handling โ becomes the bottleneck, Amdahl's law eats the headline. Total hours compress by far less than 75%, and headcount compresses by less still. And the cross-industry average is meaningless anyway, because the variance across sectors dwarfs the mean: software development is augmentation-dominant, with replacement concentrated in templated CRUD and test scaffolding; content production faces high replacement; customer service faces high replacement but is constrained by compliance and emotional labor; legal and medical are augmentation-dominant, blocked from full automation by liability and regulation. A single 75% describes none of them.
There is a measurement problem underneath all of this, and it is the one I would flag first in any audit. METR's July 2025 randomized controlled trial found that experienced open-source developers using AI tools completed tasks about 19% slower โ while reporting they felt 20% faster. Perceived efficiency and actual efficiency diverge systematically, and 75% comes from the same well of subjective expectation. When the input is a survey of belief rather than a measurement of output, you are not reading a forecast. You are reading a mood.
I have seen this mood before. During the 2022 crash, when anxiety peaked and half the developers I knew were quietly drafting exit plans, I ran a mentorship program connecting two hundred junior developers with veterans. We tracked outcomes obsessively, because sentiment was useless and everyone's feeling about the market was wrong in a different direction. The people who survived were not the ones with the most optimistic forecast. They were the ones who measured their actual burn and their actual output and adjusted. โ Root: The 2022 Bear Market. The same discipline applies here: the useful question is not what Anthropic staff expect, but what the Anthropic Economic Index shows about the ratio of directive, automation-style conversations to feedback, augmentation-style conversations. That is a public proxy, and it is far more reliable than any survey.
Then there is the question nobody is asking: even if the capability arrives, who captures the value? A lab with no cloud, no operating system, no productivity suite is structurally weaker on distribution than the platforms that own the channels. Can do 75% and gets paid for 75% are different claims, and the second one is where the margin lives. I watched the same movie with Uniswap V4 hooks: programmable Lego, genuinely powerful, and a complexity spike that will scare off ninety percent of the developers who would have built on it. Power without distribution is a demo. โ Root: DeFi Summer, when I led a volunteer team of fifteen through Uniswap's early governance mechanisms and learned that the mechanism is never the moat; the community that can actually use it is. Governance isn't a smart contract. It is the social contract that makes the contract legible. And the same holds for agent tooling: the winning lab is not the one with the best model, but the one whose agents enterprises can actually deploy and audit.
Which raises the accountability question, and here I have to speak from direct experience. In 2026, as AI agents began transacting on-chain, I convened a working group of thirty ethicists and developers to draft the Autonomous Agent Accountability Charter โ seven workshops, fifteen DAOs endorsing, one hard question that kept returning: who is liable when an AI-driven smart contract fails? The answer we converged on was uncomfortable. Code is law, but people are the protocol. The moment an agent executes autonomously, the liability does not disappear into the code โ it redistributes to the humans who deployed, audited, and profited. The 75% narrative is silent on this, and that silence is the real governance blind spot. Existing AI governance debates focus on model outputs โ harmful content, bias, jailbreaks โ and almost entirely ignore the ethics of publishing capability forecasts. Yet a widely-cited 75% may shape more career and policy decisions than any single jailbreak. We require clinical trials to preregister their methods. We require nothing of a lab that tells the world its product will absorb three-quarters of human work.
The infrastructure dimension seals it. If 75% of work migrates to AI, the compute load flips from training-dominant to inference-dominant, and the character of inference changes: from short requests to long-running, multi-turn tool-calling, high-token, state-heavy sessions. Agent sessions can consume tens of times the tokens of chat. The compute bill for 75% of work may be a multiple of current global inference supply โ a number no coverage has calculated, and the hard physical constraint on the entire forecast. If power and chips cannot keep pace, the forecast is vetoed by physics, not by capability. And there is a structural mismatch risk: capital is pouring into training-grade clusters โ high-interconnect, high-precision โ while agent inference rewards memory capacity, bandwidth, cache hit rates, and scheduling density. These are not the same hardware. We may be building the wrong infrastructure for the workload the narrative promises.
Here is the angle almost nobody is holding: the forecast is more dangerous if it is 40% right than if it is 100% wrong. A clean miss produces a clean correction โ the narrative deflates, valuations compress, and the market resets to a tool-pricing baseline. A partial hit produces the worst outcome, because partial truth lands in the wrong distribution. Junior roles collapse before senior roles are augmented. Compute gets built before demand is verified. Enterprises procure against a 75% expectation and renew against a 20% reality, and net revenue retention quietly becomes the tell that ends the cycle in 2027 or 2028.
This is not speculation. It is the exact shape of the data-availability correction and the governance-delegation correction. I have argued for years that delegation makes governance more centralized โ that users are too lazy to research and simply hand their votes to KOLs โ and the empirical record keeps confirming it. The mechanism is the same in every case: a legible narrative outruns a messy reality, capital and behavior reallocate toward the narrative, and the correction arrives not when the narrative is disproven but when the renewal data lands. The blind spot in the 75% story is not whether AI can do the work. It is whether the organizations deploying it can absorb the change without breaking the very accountability structures that make the work valuable.
And a second blind spot, quieter: the second-order damage. If knowledge-work output costs fall by half, the casualties are not only knowledge workers. They are the service economy orbiting knowledge work โ consulting, training, recruiting, editing, agency work โ people who do not believe they are exposed to AI and have no transition resources. That is where the unmeasured cost accumulates, and it is the cost no forecast ever prices.
So track the right signals, not the headline. Watch the Anthropic Economic Index for the directive-to-feedback ratio โ automation versus augmentation, the closest public proxy we have to how much work AI actually carries. Watch METR's task-horizon updates; if the doubling period stretches, the 2030 case weakens materially. Watch enterprise net revenue retention and agent usage retention; that is where the narrative meets its invoice. And watch the merge and rework rates of AI-generated code โ the most measurable frontier, where the 75% claim will be confirmed or falsified first.
In a bear market, the discipline is survival, not prophecy. The number that matters is never the one a vendor forecasts about itself. It is the one you can verify before you commit capital, before you cut a hire, before you restructure a team around a slide. We didn't learn that from a whitepaper. We learned it from the protocols that bled and the people who measured anyway.
So ask the question the brief never asks: if 75% of the work is done by machines, who signs for the remaining 25% โ and who answers for it when the contract fails?