McKinsey says AI will create more jobs than it destroys. Crypto Briefing ran the headline. The crypto market shrugged. Both reactions are wrong.
The report—reduced to five information points, one source, zero methodology—lands on the same number every consultancy has been recycling since 2023: generative AI could add $2.6 trillion to $4.4 trillion in annual economic value, while displacing hundreds of millions of workers and forcing somewhere north of 12 million Americans through an occupational transition. The optimistic reading is that the net is positive. The pessimistic reading is that the displacement is real. Neither reading is the trade.
The number nobody quoted is the time asymmetry. Displacement happens in one to three years. Creation happens in five to fifteen. That gap—call it the displacement ledger—is where capital moves, where collateral re-prices, and where the entire AI-crypto convergence narrative either becomes the most important infrastructure buildout of the decade or the most elaborate misallocation since the ICO boom I spent my twenties auditing.
I have watched five cycles. This is the first one where the displacement vector points at the people who write the trades. A labor market is just a settlement layer for human capital. When the settlement layer fails, the capital does not vanish. It re-routes. The only question that matters is which pipe it re-routes through, and who owns the pipe.
That is the article.
The Context: Where Liquidity Goes When Labor Is Repriced
Start with a first principle that most analysts skip because it is uncomfortable: all assets are leveraged liabilities. A wage is a liability on an employer's balance sheet. A salary is a claim on future production. When AI reprices human labor, it does not destroy the liability. It converts it—from a recurring human liability into a one-time capex event, amortized over compute.
This is the mechanical truth the McKinsey narrative buries under its own prudence. The consultancy is careful. It uses "could" and "may." It emphasizes transition management, strategic workforce transformation, the controllable middle path. That is not an accident. McKinsey sells to two customers: enterprise HR departments and policy desks. The "net positive, manageable transition" framing is a product optimized for both. The "structural mismatch, distributional crisis, decade-long adjustment" framing sells fewer engagements.
So read the report the way I read a smart contract before audit: ignore the marketing, look at the state variables.
The state variables are three. First, the concentration of value. Of that $2.6–4.4 trillion, roughly 75% sits in four use cases—customer operations, marketing and sales, software engineering, R&D. That is not a broad productivity tide. That is a targeted repricing of four labor categories, all of them white-collar, all of them relatively high-wage. Second, the skill mismatch. The jobs most exposed to automation—food service, customer support, office support, parts of manufacturing—pay the least and convert to new jobs the worst. The jobs being created—AI engineering, data analysis, systems architecture—pay the most and convert from the displaced population the least. Third, the time gradient. Substitution is fast. Creation is slow. In the interim you get a J-curve of human capital: value destroyed before value created, with the trough landing on specific people in specific zip codes.
None of this is in the Crypto Briefing digest. The digest is a funnel. It exists to pull the AI narrative into a crypto audience. I do not fault it for that; content aggregation has its own economics. I fault the audience for treating it as evidence. It is a five-point summary of a hundred-word brief, translated from a consultancy's public positioning. It is not data. It is a mood.
Now here is the pivot that most crypto analysts miss entirely, and the reason this report matters more to a Bitcoin holder than to a staffing agency.
When labor is converted to capex, the capex has a physical substrate. That substrate is compute. And compute—unlike labor—is collateralizable. You can borrow against a GPU cluster. You cannot borrow against a call center full of people whose function has been automated away. This is the single most important structural fact of the next decade, and it has almost nothing to do with whether AI "creates jobs" in the aggregate.
The Federal Reserve understood this before either the equity or crypto market did. The 2024–2025 balance sheet regime—QT slowed, reserves re-expanded, the discount window quietly reopened as a standing facility—was not about rescuing banks. It was about ensuring that the financing channel for compute capex stayed open. The M2 chart tells the same story in slower motion. Global M2 bottomed in late 2022 and has been grinding higher since. Liquidity always precedes capex. Capex always precedes the productivity print. The productivity print is what the consultants are trying to forecast. They are forecasting a lagging variable.
And here is the algorithmic point on which the entire AI-crypto thesis stands or falls. Compute is the collateral; labor is the liability. When you displace labor, you create a claim against the future that must be settled in compute. Every displaced worker is, in accounting terms, a liability that has been converted into a fixed-asset requirement. The liability was recurring and human. The asset is durable and mechanical. Assets can be securitized. Liabilities cannot.
This is the seam that the tokenization of computational power is trying to occupy. Whether it succeeds is a separate question. Whether the demand exists is not.
The Core: The Tokenization of Computational Power
I want to be precise about what is actually being built, because the category "decentralized compute" is doing an enormous amount of unearned work in most writing on this subject.
There are two distinct markets being conflated. The first is decentralized compute supply—networks that aggregate idle GPU capacity (Render, Akash, and their cousins) and sell it against the hyperscaler oligopoly. The second is compute verification and settlement—the layer that proves a given unit of computation was actually performed, by whom, and to what specification. These are not the same business, they do not have the same margin structure, and conflating them is the analytical error that will separate the survivors from the corpses.
Let me take the supply side first, because it is the one with the cleanest commercial logic and the most over-hyped token prices.
The pitch is straightforward. Global AI training and inference demand is growing faster than hyperscaler capacity. Enterprise and mid-market buyers are queued behind the largest labs. Idle consumer and prosumer GPU capacity exists in the tens of millions of units. A marketplace that connects the two, priced in a token, captures the spread. Render does this for rendering and increasingly for AI workloads. Akash does it for containerized compute. The narrative writes itself.
The problem is the same one I have seen in every two-sided marketplace priced in a governance token since 2017, and I audited enough of them to recognize the pattern in the bytecode before I recognize it in the pitch deck.
A marketplace priced in a volatile asset cannot compete on cost with a marketplace priced in dollars. This is not an opinion. It is an arithmetic constraint. If a compute seller accepts payment in a token, the seller is long that token between the moment of sale and the moment of liquidation. That is a speculative position the seller did not ask for. To compensate, the seller either charges a volatility premium—destroying the cost advantage—or hedges, importing basis risk and intermediary cost that a dollar-denominated hyperscaler does not carry. The decentralized network's marginal cost advantage over AWS is real in theory and eroded in practice by the cost of being priced in an asset that moves 6% on a Tuesday.
I have watched this exact dynamic play out. In 2017 I led a five-person audit team through fifty-one early-stage token contracts. Twelve had critical reentrancy vulnerabilities. But the vulnerabilities that killed the projects were never in the code. They were in the token economics—the assumption that a volatile medium of exchange could clear a stable service. The contracts were clean. The models were insolvent.
The compute networks have better fundamentals than the ICO tokens. But the same medium-of-exchange constraint applies, and the market is pricing as if it does not.
Now the verification side, which is the harder problem and the one that actually determines whether decentralized compute is a real asset class or a narrative.
When you rent compute from AWS, you do not verify the computation. You trust the counterparty. This is fine because AWS has a reputation, a legal entity, and a 99.99% SLA with financial consequences for breach. Trust is cheaper than verification. It almost always is. This is the first thing you learn in audit: the entire purpose of a costly consensus mechanism is to substitute for trust that cannot otherwise exist. Where trust is cheap, consensus is a tax.
So the central question for decentralized compute is: where is trust expensive enough to justify cryptographic verification?
The answer is narrower than the market believes. You only need verified compute when (a) the computation is adversarial—the party doing it has an incentive to lie, or (b) the result is a financial primitive whose correctness is economically enforceable on-chain. Training a frontier model on rented GPUs is not adversarial in this sense: if the provider delivers garbage, the buyer detects it downstream and refuses payment. Inference for a DeFi oracle—where the output directly triggers a liquidation—is adversarial. The provider can be bribed. The output moves money. That is where verification is worth its cost.
This is the same structure as the oracle-latency problem that has been DeFi's Achilles' heel since 2020. Chainlink solved decentralization with a federation of nodes that are, in the last analysis, as centralized as the exchanges whose prices they report. It was a joke, but it was a load-bearing joke, and it worked because most DeFi did not actually need trustless price feeds—it needed feeds that were reliable enough to be trusted by the people with money at stake. The verification was theater. The reliability was real. The market paid for the reliability and got the theater for free.
Decentralized compute is walking into the same trap, from the opposite direction. It is building expensive verification infrastructure for computation that mostly does not need to be verified, while the computation that genuinely needs verification—the adversarial, financially enforceable kind—is a small fraction of total demand.
Verified compute is just trust with a receipt. And a receipt costs money. The market will only pay for the receipt when the trust it replaces is more expensive than the receipt itself.
The economics are unforgiving. Consider a rough decomposition. Of total AI compute demand, the overwhelming majority is training and non-adversarial inference. This demand is served at scale by hyperscalers with dollar pricing, reputation, SLAs, and physical control of the datacenter. The slice that is genuinely adversarial and on-chain-enforceable—oracle-grade inference, verifiable random functions for on-chain games, provably fair lottery and gambling, certain DeFi derivatives pricing—is real but small, plausibly low single-digit percentages of total demand for the foreseeable future.
Now overlay the token market. The combined market capitalization of decentralized compute tokens implies a serviceable market many multiples of the verifiable-demand slice. The gap is the mispricing. It is not that decentralized compute is worthless. It is that the market is pricing the entire compute TAM—hundreds of billions of dollars—against a protocol that can only economically address the verification-demand SAM—single-digit billions, at best, on optimistic assumptions.
This is the same error as the DA layer. I have written about this before and the pattern holds precisely: the data availability wars were fought over a market that 99% of rollups did not generate enough data to need. Celestia, EigenDA, and their competitors are excellent engineering. They are structurally over-built for actual demand. Rollups choose them for ideological reasons and cost reasons that are thinner than the marketing admits, and the ones that generate real data volume migrate back to Ethereum blobs because the blobs are cheaper and the security budget is the deepest in the industry. The DA layer is not useless. It is over-priced relative to the demand that can only be served by a DA layer.
The same filtration is coming for decentralized compute. And when it comes, the market will discover that the networks with the best token prices were never the networks with the best verification demand. They were the networks with the best narratives.
Let me now do what the digest did not: examine the actual on-chain structure of these markets, because the data is public and the data is unkind.
Render's network metrics show a real, functioning marketplace. Utilization matters more than gross capacity. A network that has onboarded a million idle GPUs but clears a low single-digit utilization rate is not a compute market. It is a registration database. The gross node count is vanity. The paid-render volume, the median job size, the retention of paying customers across quarters—these are the variables that determine whether the protocol is a business or a grant-funded experiment with a token.
Akash's metrics tell a similar story with a different texture. Containerized compute is a commodity. The buyers are sophisticated. They will migrate to whatever is cheapest per compute-hour with acceptable reliability, and they will migrate away the moment the price advantage closes. A commodity marketplace that cannot sustain a structural cost advantage has no moat. It has a subsidy. And subsidies funded by token emissions are not revenue. They are dilution wearing a growth narrative.
Here is the audit lens again. When I look at a protocol's economics, I look for the delta between what the token price implies and what the cash flows support. For a functioning compute network, the relevant cash flow is the fee revenue net of the cost of the volatility premium the network must pay its suppliers. For most of these protocols, that number is small—sometimes negative—relative to the fully diluted valuation. The token price is a call option on a future in which verified compute becomes a large enforceable market. It is not a claim on present cash flows.
That is not automatically a bad thing. Early-stage infrastructure is always priced on optionality. But the optionality has to be real, and the way you test it is by asking a single question: what has to be true for this network to become a mandatory settlement layer?
The answer, for decentralized compute, is that on-chain enforcement of computation must become economically necessary for a large class of applications. That requires, at minimum, that the applications whose correctness is enforceable on-chain grow to be a large share of AI demand. Right now they are not. AI demand is dominated by training and consumer inference, neither of which is on-chain-enforceable. So the thesis depends on a world in which a meaningful share of compute becomes financially adversarial—in which an increasing fraction of economic activity is settled by on-chain computation whose integrity must be cryptographically assured.
That world is plausible. It is the world of autonomous agents transacting on-chain, of decentralized AI that must prove its inferences, of machine-to-machine settlement where there is no legal entity to sue and no reputation to trust. It is the world my own 2026 thesis on the AI-crypto convergence pointed toward—decentralized compute markets as the settlement layer for a new class of machine economic actors.
But that world is five to fifteen years out on the same time horizon the McKinsey displacement ledger implies. And the market is pricing it as if it is here.
The Convergence: AI's Centralization Bottleneck and the Decentralization Mirage
The strongest version of the AI-crypto convergence argument goes like this. AI is centralizing at an alarming rate. Compute is controlled by a handful of hyperscalers. Models are controlled by a handful of labs. Data is controlled by a handful of platforms. The centralization creates single points of failure—technical, economic, and political. Decentralized networks offer an alternative: distributed compute, distributed data, distributed model governance. Therefore decentralization wins by necessity.
I find this argument seductive and mostly wrong, and the reason it is wrong is that it confuses a preference for decentralization with a requirement for it.
Centralization is not a bug in AI. It is the equilibrium. AI has extreme economies of scale at every layer. Training a frontier model costs more than a mid-sized city's annual budget, and the cost curve rewards the largest buyer. Inference is cheaper per token at scale because of fixed-cost amortization and utilization. Data aggregation compounds: more data means better models means more users means more data. These are not market failures to be corrected. They are the natural physics of the technology. Decentralization is a deliberate choice against the gradient, and choices against the gradient cost more. They are only sustainable when the gradient is being resisted for a reason the market will pay for.
The reasons the market might pay for decentralized AI are real but narrow. Sovereignty: nations and enterprises that cannot legally or politically route their data through US-controlled hyperscalers. Censorship resistance: applications that hyperscalers will not serve for reputational or regulatory reasons. Verifiability: computations whose correctness must be provable rather than trusted. Cost for the long tail: users whose workloads are too small to interest the hyperscalers but too large for a laptop.
None of these are trivial. But none of them scale to the total AI market. They are niches—important, durable niches, but niches. And a protocol priced against the total AI TAM while serving a sovereignty-and-verifiability SAM is a protocol priced for a world it cannot reach.
This is where the honest analyst separates the infrastructure from the application layer, and I have been clear about where the alpha sits. We do not ride the wave; we engineer the tide. The alpha in AI-crypto is not in the applications. It is in the infrastructure that becomes collateral—the compute, the verification, the settlement, the energy. Applications are the foam on the wave. Infrastructure is what the tide deposits on the shore.
Let me make this concrete with the two networks that dominate the conversation.
Render's commercial potential rests on a real asymmetry: the rendering market and, increasingly, the AI inference market have demand spikes that exceed centralized capacity, and idle distributed capacity can absorb those spikes at a discount. The value is in the flexible, burstable, discount segment. This is a real business. It is also a business with structurally lower margins than the hyperscalers, because the protocol must compensate suppliers for volatility and the network has higher coordination overhead. The commercial case is a discount capacity provider for price-sensitive, latency-tolerant, verification-light workloads. That is a meaningful market. It is not the AI market. Pricing it as the AI market is the error.
Akash's commercial potential is similar in structure and more commodity-like in substance. Containerized compute is fungible. The protocol's only durable advantage is cost. Cost advantages in commodities are eroded by competition and eliminated when the token volatility premium exceeds the discount. Akash's viability depends on its ability to sustain a cost advantage net of the volatility premium it pays its suppliers. That is a purely empirical question, answerable quarterly from the network's own data. The market, as usual, is not answering it. It is pricing the narrative.
Now the convergence thesis at the layer beneath both: inference compute as a financial primitive.
This is the part I actually believe, and it is the part the market has not priced.
If AI displaces labor on the timescale McKinsey implies—displacement in one to three years, creation in five to fifteen—then the displacement creates a fixed demand for inference. Not training. Inference. The displacement is the deployment of AI into the workflow. Deployment is inference. And inference is continuous, low-latency, geographically distributed, and financially metered. It is a utility.
A utility has a predictable cash flow. A predictable cash flow can be securitized. A securitized cash flow is collateral. This is the chain: displacement creates inference demand, inference demand creates a utility cash flow, the utility cash flow becomes collateral, the collateral is traded, and the trading happens on whatever venue clears fastest and cheapest.
That venue has historically been a centralized exchange. The question the AI-crypto convergence raises is whether it becomes an on-chain venue, because the settlement of inference-compute contracts has properties—programmability, atomic settlement, composability with financial derivatives—that a centralized clearinghouse cannot replicate.
If that happens, the token that settles inference compute is not a "compute token." It is a commodity-backed money. And money is collateral. And collateral, as I have said, is just debt wearing a mask of trust.
The mask is what the market is pricing right now. The debt is what the market will eventually be forced to price. They are not the same number, and the gap is the opportunity.
The Contrarian Angle: AI-Crypto Is Decoupling From AI and From Crypto
Here is the counter-intuitive claim, stated plainly. The AI-crypto sector is not correlating with AI. It is not correlating with crypto. It is correlating with the narrative about both, which is a third, less stable quantity.
Trace the correlations. When AI equities rally on an earnings beat from a chipmaker, compute tokens rally less and later. When Bitcoin rallies on a liquidity print, compute tokens rally less and later. When a major AI lab announces a model release, compute tokens spike and fade within days. The sector is a lagging, high-beta echo of two independent narratives, and it has no independent driver of its own because it has no independent cash flow of its own yet.
This is the decoupling thesis. It is not that AI-crypto will decouple upward. It is that AI-crypto is already decoupled—from both parents—and the market has not noticed.
The implication is that the sector is not a diversifier and not a proxy. It is a leveraged expression of two theses at once, with the correlations to each replaced by a correlation to a narrative that can be switched off by a single regulatory headline or a single funding-round disappointment. This makes it, in portfolio terms, more fragile than either parent, not less.
Now the binary framework. I evaluate projects the way I evaluate a distressed credit: solvent or insolvent, viable or theater. No middle ground. The middle ground is where analysts hide when they do not want to make a call.
Viable: Infrastructure that converts a real, verifiable, on-chain-enforceable compute demand into a cash flow, priced against that cash flow and not against the AI TAM. The list is short. A handful of verification layers, a handful of settlement primitives, and the energy and hardware that underwrite them. These have a path to being collateral. These will survive the filtration.
Non-viable: Anything whose value accrues from the narrative of decentralization rather than the demand for it. This includes most application-layer AI tokens, most "decentralized AI" projects that are GPT-4 wrappers with a governance token, most data-marketplace tokens whose data no model actually wants, and most compute tokens that cannot articulate, in a single sentence, which workload they serve that a hyperscaler cannot serve cheaper and better. When I audited tokens in 2017, the non-viable ones were the ones whose whitepaper described a market that did not yet exist and whose roadmap assumed it would. Same test. Same result. Twelve of fifty-one had fatal code bugs. Far more than twelve had fatal economic models.
The uncomfortable conclusion is that the AI-crypto convergence—the single most-hyped macro narrative of the cycle—is mostly a misallocation, punctuated by a small number of genuinely important infrastructure plays that the market is under-pricing while it over-prices the foam.
And the deeper uncomfortable conclusion is that this is exactly what the McKinsey displacement ledger predicts. The displacement of labor into compute creates a real, enormous demand for a specific kind of infrastructure—continuously metered, verifiable, collateralizable inference and settlement. The market, as always, is pricing the applications and ignoring the collateral. The collateral is where the tide deposits value. Everything else is foam.
The Takeaway: Positioning for the Ledger, Not the Headline
So where does this leave a capital allocator who has read the McKinsey digest, read the Crypto Briefing transposition, and wants to act on the structural truth rather than the narrative?
You do not trade the headline. You trade the ledger.
The headline says AI will create more jobs than it destroys. The ledger says the destruction is fast, the creation is slow, and the gap is a decade-long conversion of human liability into compute collateral. The headline is a mood. The ledger is a mechanism. Only one of them has a price.
The mechanism produces three durable conclusions. First, the real AI-crypto trade is infrastructure collateral, not application tokens—verification, settlement, and the compute and energy that underwrite them, priced against enforceable demand rather than the total market. Second, the verification market is far smaller than the market cap implies, and the filtration that hit the DA layer is coming for the compute narrative with the same arithmetic. Third, the only compute networks that become collateral are the ones whose output is financially adversarial and on-chain-enforceable, because that is the only compute whose correctness the market will pay a cryptographic premium to guarantee.
Everything else is dilution wearing a growth narrative.
The consultancy will keep publishing the optimistic version. The funnel will keep aggregating it. The market will keep pricing the wave. And somewhere beneath all of it, a small number of settlement layers are quietly becoming the place where the displaced labor of the next decade is settled—not in jobs, but in compute; not as wages, but as collateral; not on a balance sheet of human capital, but on a ledger of verified machines.
The question is not whether AI creates more jobs than it destroys. The question is who holds the collateral when the conversion is complete—and whether you will be positioned in the mechanism or stranded in the headline.