Three weeks ago I was asked to underwrite a $40 million secondary in a compute-network token. The deck ran sixty-two pages. Page forty-seven carried the only number that mattered: 11.4% of the GPU-hours the network had brokered in the prior quarter were actually consumed by paying workloads. The rest existed as capacity that had been contracted, tokenized, pledged as collateral, and re-emitted as a staking yield the marketing team described as "real revenue share."
I passed. Not because the technology was fake โ the attestation layer was competent, the slashing conditions were specified with unusual precision, the node geography was less concentrated than most. I passed because the price embedded a claim the network could not verify about itself: that decentralized compute is a sovereign asset class rather than a discount channel into the same hyperscaler capital-expenditure cycle that has been repricing every semiconductor name on the Nasdaq since late 2024.
That distinction is the entire trade. Most of the people holding these tokens have not made it, and in a bull market nobody is incentivized to make it for them.
Here is the liquidity map as it actually stands. Global dollar liquidity has been expanding at the margin since the middle of last year โ not dramatically, but persistently, as the Fed's balance-sheet runoff ended and the Treasury shifted issuance toward bills, releasing collateral into the front end of the curve. Stablecoin supply has grown roughly 38% year over year, which is the cleanest single proxy for crypto-native purchasing power that exists. Spot ETF flows into the two large caps have been steady but decelerating, and the marginal buyer there is a wealth-management allocator running a 1-2% sleeve, not a directional macro fund.
What that combination produces is not a rising tide. It produces a barbell: capital pools in assets with institutional plumbing on one end, and in narrative optionality on the other, while the middle of the market โ the genuinely useful infrastructure without a clean story โ gets starved. Compute networks sit on the optionality end. That is a position of enormous reflexive strength during accumulation and catastrophic fragility during the first real drawdown, because the bid is not fundamental, it is thematic.
The thematic bid has been reinforced by something structural: the AI capex cycle has become the largest coordinated capital deployment in corporate history outside of wartime, and it is financed with debt. Hyperscaler capex guidance for the current year exceeds $300 billion across the top four buyers. That number has to be serviced. When you service a debt-financed buildout, you eventually need to maximize utilization on assets with a three-to-five-year economic life, and you will rent those assets to anyone โ including, and increasingly, to decentralized networks acting as brokers.
That is the tell. Decentralized compute is not competing with hyperscalers. It is the overflow valve for them. Which means the token price is a levered expression of somebody else's capital allocation decision, and I want to be very precise about what that lever looks like.
Let me get technical, because the froth lives in the details.
A compute network has to solve one problem before it has a business: how do you verify that a remote, potentially adversarial operator actually ran the workload you paid for, on the hardware you specified, without substituting cheaper silicon or returning a cached or fabricated result? There are four families of answer in production today, and each carries a distinct cost structure that flows directly into the token economics.
The first is re-execution with redundancy. You assign the same job to n independent operators and take the majority result. It is trivially sound and brutally expensive: three-way redundancy means you pay three times for one unit of useful work, and your effective price per GPU-hour is permanently two-thirds higher than the spot market. No amount of token emission fixes that arithmetic. It is a structural discount that must be paid in perpetuity.
The second is trusted execution environments โ confidential-computing modes on modern accelerators. Attestation is cheap, latency overhead is modest, and the verification cost is close to zero. The problem is trust: the entire security model rests on a vendor's attestation root and on the absence of side-channel classes of attack that have been demonstrated repeatedly in academic literature against exactly this hardware generation. You have moved the trust boundary, not eliminated it, and you have done so while pretending the boundary does not exist. In my audit work I have watched three separate teams describe TEE-based verification as "cryptographically guaranteed." It is not. It is vendor-guaranteed, which is a different word with a different insurance premium.
The third is probabilistic sampling โ spot-check a fraction of jobs by re-running them, slash on mismatch. At a one-in-a-thousand sampling rate the overhead is negligible and the deterrence is real, provided the slashing bond exceeds the profit from a single successful fraud multiplied by the expected number of undetected repetitions. Almost every network publishes the first half of that inequality and omits the second. When you run the numbers honestly, the required bond for a large batch job is frequently larger than the operator's entire hardware position โ which means either the bond is under-collateralized and the security is theater, or the honest operator cannot afford to participate and the network selects for the marginal actor. Systemic risk doesn't announce itself; it invoices you later.
The fourth is zero-knowledge proofs over inference, and here the gap between the marketing and the mathematics is widest. Proving a forward pass through a mid-sized transformer costs somewhere between two and three orders of magnitude more than executing it. I have seen this measured repeatedly over the past eighteen months, and the ratio improves slowly โ the constraint is not engineering effort, it is the structure of the operation. So zkML is genuinely transformative for verifying a model's provenance and training-data lineage, where you prove once and verify many times and the amortization works. It is economically incoherent as a per-inference verification mechanism at current proving costs. Every pitch deck that slides those two use cases together is either confused or counting on you to be.
Now overlay the hardware curve. The median H100-hour in the broker market cleared near $2.10 this quarter, against a peak print above $4.40 in early 2024. Blackwell-class capacity is arriving, the resale market for last-generation silicon is thinning, and the depreciation schedule on a compute node has compressed from a theoretical five years to something closer to thirty months in effective economic terms. This is the single most important variable in the sector and it appears in almost no token model I have reviewed.
Because here is the thing about a network that rents depreciating hardware and pays out token emissions: High APY is just delayed pain. If the yield is denominated in an asset whose supply is expanding, and the underlying productive asset is losing a third of its value every year, then the yield is not a return โ it is a transfer from future holders to present ones, mediated by a depreciation schedule nobody put in the spreadsheet. The honest metric is (rental revenue โ hardware depreciation โ slashing losses) divided by fully diluted market capitalization. Run that calculation across the ten largest compute tokens and the median number is negative. Not marginally negative. Structurally negative, in the way a business is negative when its cost of goods sold exceeds its price.
I want to be fair to the technology, because there are real things happening here. The genuinely interesting work is not in retail GPU rental at all. It is in the verification primitives โ the attestation standards, the slashing-condition formalisms, the proof systems for training-data integrity. Those are public goods that will outlive every token that funded them, and they matter enormously for the next phase, where autonomous agents transact with each other and need to prove what model they ran, on what inputs, under what constraints. That is the actual product. The GPU brokerage is the funding mechanism, and like most funding mechanisms in this industry, it has been mistaken for the business.
So let me state the contrarian position plainly, because the consensus has it exactly backwards.
The prevailing thesis is that AI-compute tokens will decouple from the crypto market and re-rate as a proxy for the artificial-intelligence economy โ that they are early-stage exposure to a $10 trillion theme, and that their correlation to Bitcoin will fall as the narrative matures. The opposite is true. These tokens are the longest-duration, highest-beta instrument in the entire digital asset complex, because their revenue depends on the utilization of capital-intensive hardware whose demand is set by four companies' board-level spending decisions, financed with debt, in a rate-sensitive sector. When liquidity tightens, the sequence runs: capex guidance gets trimmed, rental rates fall, network revenue falls, emissions get increased to defend the yield, dilution accelerates, price falls, and the operators with the thinnest bonds exit first โ concentrating the network precisely when concentration is most dangerous.
That is a chain, not a coincidence. And it is the same chain that ran through 2022, when algorithmic stability was assumed to be structural until the reflexivity became visible. I spent that period building a liquidity-stress index across five exchanges and it saved my fund from the worst of the contagion not because I was smarter, but because I refused to model a system as stable when its stability depended on the price of its own collateral. Compute networks have the same shape. Their security budget is denominated in the token. Their operators' bonds are denominated in the token. Their revenue is denominated in dollars but their cost of capital is denominated in the token. Every leg of the stool rests on the same price.
A brief detour on narrative recycling, because it is diagnostic. The same mechanisms I just described โ attestation, redundancy, slashing โ are being relabeled across three or four verticals simultaneously right now. Much of what markets itself as a Bitcoin Layer 2 is an Ethereum-flavored rollup with a fresh coat of paint and a bridge that the actual Bitcoin community does not recognize. Much of what markets itself as decentralized physical infrastructure is a tokenized invoice with a geographic map on the landing page. The pattern is consistent: take a real primitive, attach it to the hottest adjacent narrative, and price the option before anyone audits the strike. Smoke signals, not foundations.
Where does that leave positioning?
My working assumption for the next two quarters is that compute tokens trade as a high-beta proxy on the same liquidity impulse driving the large caps, with a spread that widens in both directions. In a risk-on tape they will outperform meaningfully, and the outperformance will be attributed to adoption. In the first genuine liquidity withdrawal they will underperform by a wider margin than their beta implies, because the exit is narrow โ order books in these names are thin relative to the fully diluted valuations they carry, and the marginal holder is a momentum allocator with no mandate to average down.
The asymmetry favors selling strength into narrative peaks rather than buying dips into them. Thesis broken. Capital preserved. That is not a bearish call on decentralized compute. It is a refusal to pay a sovereign-asset-class multiple for a working-capital business with a depreciating inventory.
The question I keep returning to is not whether decentralized compute survives โ the verification primitives almost certainly will. The question is what the sector looks like after the first operator with an under-collateralized bond discovers that slashing is real, and whether the market's response is to reprice the security model or to relabel the mechanism and re-list the token. I have a strong prior. I would rather be early to that answer than late to this one.