The N/A Report: How Crypto's Research Stack Learned to Prove Empty Blocks

Something arrived in my inbox on a Tuesday three weeks ago that I have not been able to put down. It was a due diligence report. Nine dimensions, cleanly separated. Technical architecture. Token economics. Market structure. Ecosystem positioning. Regulatory exposure. Team and governance. Risk matrix. Narrative and expectations. Supply-chain transmission. Each dimension carried a table. Each table carried rows. Each row carried a verdict cell. Every verdict cell read the same four characters. N/A. Insufficient information. Not "low confidence." Not "awaiting data." Insufficient information, repeated with the flat affect of a machine that has been asked to evaluate a sealed box and has correctly reported that the box is sealed. The document ran slightly over four thousand words. It was structurally immaculate. It was vertically integrated in its own emptiness. And by the time I finished it, I had concluded something that took me a while to admit, because it cuts against everything I have built a career on. It was the most honest piece of crypto research I have read in twelve months. Not because it was right. Because it refused to be wrong in the specific way that our industry has industrialized. It refused to convert formatting into authority. It refused to let the existence of a table imply the existence of a fact. In a market where the median research product is a template with numbers bolted on, the empty template is the tell. The pipeline that produced this document was built to enforce completeness of form. It was handed an empty input. And instead of hallucinating a project, it produced a perfect mirror. I have spent twenty-three years watching this industry, and the last eight of them writing about the gap between what the market says it knows and what it actually knows. That gap has never been wider. It has also never been harder to see, because the surface has never been smoother. So let me walk through what an empty nine-dimension report actually tells us, because the answer is not "the tooling is broken." The answer is that the tooling is working, and the market it is describing has been hollowed out from the inside. The Format Completeness Trap Every analytical framework in crypto is a descendant of the same ancestor: the venture memo. A venture memo is a persuasive document dressed as an evaluative one. It has sections because a section creates the impression that something was investigated. A team section implies a team was vetted. A tokenomics section implies emissions were modeled. A risk matrix implies risks were enumerated. None of that is guaranteed. The section is a claim about diligence, not diligence itself. When I was a junior analyst in Buenos Aires in 2017, I was handed a stack of ICO whitepapers and told to produce a standardized summary for each. The template had fourteen fields. I filled out fourteen fields for fifty-one projects. Fourteen fields times fifty-one equals seven hundred fourteen verdict cells. I remember exactly how many of them I could fill with verified data rather than the project's own assertions. Thirty-eight. Five percent. The other ninety-five percent were paraphrase. I was not doing analysis. I was doing transcription with formatting, and the formatting was doing the work of credibility. That report, which I eventually published as "The Empty Promise of Utility," is the reason I have never again trusted a filled-in field merely because it was filled in. When eighty percent of those projects collapsed through 2018, nobody traced the failure back to the framework. They traced it back to "bad projects." But the framework was the accelerant. It gave a hundred bad projects the same visual grammar as five good ones, and the market could not tell them apart because the market was reading the grammar. That was 2017. The template problem was real then. It is catastrophic now, because the cost of producing the template has gone to approximately zero. Here is the asymmetry that defines the current research economy. In 2017, generating a plausible fourteen-field report on a fake project took a junior analyst two days and a token allocation. In 2026, it takes eleven seconds of inference and about nineteen cents of compute. The cost of producing the appearance of analysis has fallen by roughly four orders of magnitude. The cost of verifying that analysis has fallen by approximately nothing. Verification is still labor. It is still a human reading a contract, calling a founder, checking an explorer, cross-referencing a GitHub commit against a claim, computing an emission schedule against a real unlock. That work costs what it cost in 2017, adjusted for inflation, and there is no technology on the horizon that changes it, because verification requires establishing ground truth, and ground truth is not generated by inference. It is established by contact with the world. So we have a market with perfectly elastic supply of narrative and perfectly inelastic supply of verification. Any economist will tell you what happens next. You get glut at the top of the funnel and scarcity at the bottom. You get a hundred thousand reports and eleven people who can tell you which ones are real. You get an industry where the volume of research has never been higher and the information content of the median unit has never been lower. The N/A report is the pure form of this glut. It is the artifact that survives when the only remaining constraint is format. And the fact that it feels embarrassing — that the author of the pipeline would probably describe it as a failure — tells you that we still, collectively, mistake the shape of a thing for the substance of it. The Liquidity Map Nobody Is Reading Now the part that most people skip, which is the actual reason the market is generating empty questions. Let me lay out the macro plumbing as it stands, because crypto research does not exist in a vacuum and the vacuum is the point. The Federal Reserve concluded balance sheet runoff at the start of December 2025. QT is off. That is a structural change, not a headline. Roughly forty-eight months of passive liquidity withdrawal ended. The policy rate has been walked down into the mid-threes through a series of cuts that began in late 2024 and continued through 2025, and the market's disagreement is no longer about direction — it is about terminal. Whether the terminal rate is three and a quarter or two and three quarters is a forty-basis-point question dressed up as a regime question. Global M2 has been growing, but the second derivative is the story. The rate of change has been decelerating through 2025 into 2026 in dollar terms while expanding in local currency terms across a wide band of emerging markets, which is a fancy way of saying that dollar liquidity is tighter than headline M2 suggests and non-dollar liquidity is looser than it looks. That divergence is not priced. It is barely measured. Stablecoin supply sits in the neighborhood of three hundred billion dollars, having compounded at a rate north of thirty percent year over year. That is the single cleanest read on crypto-native dollar demand that exists, and it has been growing through a period when crypto prices have not. Read that again. Aggregate stablecoin float rising while aggregate token prices chop sideways is not a bullish signal or a bearish signal. It is a signal that capital is present and uncommitted. Dry powder that is not earning much and is not leaving. Spot Bitcoin ETF flows crossed the sixty-billion-dollar cumulative net inflow mark some time ago, and total spot crypto ETF assets across BTC and ETH products now sit in the neighborhood of one hundred eighty billion dollars. I built a predictive model on this in early 2024, the week after approval, comparing IBIT's absorption pattern against FBTC's, and the conclusion I published then still holds: the flow is a staircase, not a rocket. The ETF does not create price spikes. It creates a slow grind in the supply curve over eighteen to thirty months. That grind is now more than halfway done, and the market has already stopped treating each inflow week as news, which is exactly what happens when a structural change becomes a base rate. Perpetual funding on major pairs has been oscillating between flat and low-single-digit annualized for months. The cash-and-carry basis trade that paid double digits in 2021 and high single digits in 2024 has compressed to the low single digits in the most liquid venues. Thirty-day realized volatility on Bitcoin has been running in the low thirties, down from the sixties and seventies of the 2024 expansion. Options skew is close to flat across the term structure. Every single one of those readings says the same thing in a different language. The market is not uncertain. The market is undifferentiated. And that is the condition under which research stops having a payoff. Which brings us to the core of the argument. Proving Empty Blocks I want to describe the exact cost structure of a ZK rollup in a chop, because it is the cleanest analogy in this entire essay and because it is also just literally true and worth understanding on its own terms. The history of Layer 2 economics since EIP-4844 is a story of one cost line collapsing and another one refusing to. Before blobs, the dominant marginal cost of running a rollup was data availability. You posted calldata to Ethereum. Calldata was expensive in proportion to how much state you had to commit, and it scaled roughly linearly with the number of transactions you batched. That gave rollups a variable cost curve. Busy periods cost more. Quiet periods cost less. Operators could, in principle, ride the cycle. Blobs changed that. Blob space is cheap. Blob target was raised from three to six per block in Pectra, and in practice blob base fee has spent long stretches pinned near the minimum reserve price because demand has not filled the available space. Post-blob, the data availability line item on most rollup P&Ls fell by somewhere between eighty and ninety-five percent, depending on the batch pattern and the compression scheme. That is the single largest cost reduction in the history of Ethereum scaling, and it turned most rollups from data-cost-dominated businesses into something else. That something else is a fixed-cost business, and fixed-cost businesses die in chop. Here is the mechanism. A ZK rollup does not pay for proving in proportion to how much value it is settling. It pays for proving in proportion to how much computation it is proving. And the proving infrastructure is not rented by the second on a spot market. It is committed capacity. You sign a GPU or FPGA arrangement because proving blocks on a reliable twelve-second cadence requires guaranteed hardware, redundancy, witness-generation hosts, and the payroll of the engineers who keep the prover from falling over. That is a wall of cost that arrives every month whether you settle ten million dollars or ten billion. So the operator's cost curve looks like a step function. Below a certain throughput threshold, the amortized cost per transaction is not just high. It is asymptotically hostile. It approaches the amortized cost of proving an empty block. I want to put numbers on it, with the caveat that I am modeling, not reporting audited financials. Take a production-grade proving fleet for a general-purpose zkEVM: eight high-end accelerator nodes, duplicate witness-generation infrastructure, a KMS stack, data-center power and cooling, and the engineering payroll that keeps a prover cluster alive with a multi-nine uptime target. All-in monthly run rate lands somewhere in the range of one hundred twenty to one hundred eighty thousand dollars. Call it a hundred fifty thousand as a midpoint. At a twelve-second block cadence you are producing roughly two hundred sixteen thousand blocks per month. Straight-line amortization gives you about seventy cents per block in committed cost. Now ask what a rollup sequencer is capturing per block in the current fee environment. In a quiet market, on a chain that has compressed its own fees to stay competitive, that number is frequently in the same order of magnitude, and often below it. Sequencer revenue per block in a post-blob, post-fee-compression world can easily be under a dollar in low-activity windows. That is the whole story. The proving cost floor did not move with the cycle. The revenue line did. And it gets worse, because proving cost does not scale as cleanly as provers would like you to believe. Witness generation — the step where you turn the execution trace into something a prover can consume — is frequently more expensive than the cryptographic proving itself, especially for state-heavy workloads. Recursion helps amortize, aggregation helps amortize, better hardware helps. But all of those are engineering improvements that land on the cost curve as a one-time step, not a continuous decline. Meanwhile the chain is producing blocks every twelve seconds, most of which contain a handful of transactions that a centralized database would process in a microsecond. So the operator is, in the most literal possible sense, paying premium prices to prove that almost nothing happened. I tracked this same structural pathology in a different costume during the 2020 DeFi summer, when I modeled the yield curves on Compound and Aave and concluded that a large fraction of advertised yield was borrowed from future token value rather than from current cash flow. That was a fixed-obligation business dressed as a variable-reward one. The protocol owed emissions on a schedule regardless of whether the borrowed capital was generating organic revenue. Same shape. Different costume. The de-pegging events I warned about in that thread were not accidents. They were the inevitable settlement of a fixed-cost promise against a variable-revenue reality. The ZK rollup cost floor is the 2026 edition. Nobody wants to talk about it because the technology is genuinely impressive and the engineers are genuinely brilliant, and it feels disrespectful to point out that brilliance is not a revenue model. But the seat cost is the seat cost. It arrives on the first of the month. The Cost Collapse That Only Went One Way Step back from the specific and look at the general shape. Crypto just lived through a decade in which the cost of producing almost every kind of digital artifact collapsed. Producing a token: collapsed. Producing a brand: collapsed. Producing a dApp front end: collapsed. Producing a research report, a governance proposal, a whitepaper, a thread, a podcast: collapsed. The marginal cost of manufacturing the appearance of institutional seriousness in this industry is now functionally zero. But the cost of the things that actually determine whether a protocol survives did not collapse at all. Auditing a contract still costs what it costs, because it requires a human adversarial mind reading a hostile artifact. Building a reliable prover fleet still costs what it costs. Getting a regulatory determination still costs what it costs. Winning a real user who is not farming you still costs what it costs, and if anything it costs more, because the noise floor has risen to the point where the signal-to-noise ratio of your own marketing has dropped structurally. When the cost of the cheap thing goes to zero and the cost of the expensive thing stays fixed, the ratio between them explodes. And markets respond to ratios. They respond by overproducing the cheap thing until it is worthless and underproducing the expensive thing until it is scarce. That is why the research industry is producing four-thousand-word documents about nothing. Not because the researchers are lazy. Because the ratio between the cost of producing a report and the cost of producing the truth inside it has become so lopsided that the expected value calculation is no longer close. The rational producer optimizes for volume. The rational consumer, unable to distinguish, defaults to trusting the shape. This is the same mechanic I watched destroy utility tokens in 2017. The marginal cost of issuing a token fell through the floor. The cost of building a product did not move. The result was fifty-one whitepapers and a five percent verification rate. We did not fix that problem. We generalized it. The One Allocation Mechanism That Isn't a Revolving Door I want to spend a moment on governance, because there is exactly one corner of this industry where allocation is tied to measurement rather than to vibes, and it deserves more attention than it gets. Optimism's RetroPGF started from a premise that remains radical: pay for impact after the fact, and let the measurement of that impact determine the payment. The original framing, articulated in the 2021 write-up that launched the program, was that the profit a project earns should equal the value it creates. Everything downstream of that sentence is mechanism design. The first iterations were subjective. Round 3, in January 2024, distributed thirty million OP across five hundred one projects, with the allocation decided by a set of badgeholders casting ballots. It worked, in the sense that money moved to real builders. It also produced the failure modes you would predict: badgeholder fatigue, popularity weighting, the social-graph premium where projects with loud founders outperformed projects with quiet utility, and the inevitable concentration of decision power in a small set of wallets. The interesting part is what happened next, and it is the part almost nobody tracks. The program did not double down on human judgment. It walked itself toward measured outcomes. Subsequent rounds moved toward formula-based allocation against on-chain metrics — contract deployments, gas fees contributed, user counts, recency-weighted activity — with the peer-review layer shrinking relative to the measurement layer. Round 4 for onchain builders, Round 5 for the OP Stack, Round 6 for governance work, Round 7 for developer tooling, each one carrying a heavier algorithmic component than the last. I have audited a lot of DAO grant committees. I have watched the same forty wallets rotate through the same Notion-doc review processes, rewarding each other's proposals in a pattern so stable it could be modeled as a physics problem. The structural defect is always the same: the committee allocates against the persuasive quality of an application, and the persuasive quality of an application is a function of who wrote it and how well they understand the committee. That is not a bug in any particular committee. It is what committees are. Against that backdrop, the Optimism program's insistence on measuring after the fact is the only mechanism I have seen that has an exit from the nepotism equilibrium. And I want to be precise about the weakness, because it has one. Metrics get gamed. The moment you tell builders that gas fees contributed will determine their payout, you have created a subsidy for gas fees contributed, and someone will spin up a contract that burns gas to farm the metric. This is Goodhart's law in its natural habitat. But here is the thing. Goodhart is a tractable problem. You can rotate metrics, weight for recency, apply anti-Sybil filters, run retroactive audits, dock repeat offenders. Nepotism is not tractable, because nepotism does not have a metric you can rotate. It has a social graph, and social graphs are load-bearing. I will take a gameable metric over an ungameable feeling every single time. The N/A report is what happens when an industry defaults to feeling and dresses it as measurement. Retro Funding is the only corner of the space that has the direction of causation right. Verifiability Is the Product Now let me connect this to the thing I have been chewing on since early 2026, because it is the same problem wearing a different hat. The AI and crypto convergence has been badly mispriced, and the mispricing is entirely about which part of the stack is the actual product. The market has spent two years valuing decentralized compute networks as if they were GPU rental marketplaces. Better spot pricing. Cheaper H100-hours. Utilization rates. Those are real businesses and they have real cash flow, but they are also commodity businesses facing a competitor with a genuine cost advantage at the hyperscale end. You cannot out-Amazon AWS on price. You can be a useful spot market at the margin, and margins at the margin are thin. The actual product is not compute. The actual product is verifiability. If I cannot verify that a specific model was run on a specific input and returned a specific output, then decentralized compute is just remote execution with extra steps, and the buyer has no reason to prefer a token-incentivized network to a cloud contract with an SLA. That is not a marketing problem. It is a structural problem. Price competition in the absence of verifiability is a race to the bottom of a commodity market. There are three technical paths to fixing it and they are not equivalent. Trusted execution environments, the confidential-computing mode on current-generation accelerators, give you hardware-rooted attestation today, at production speed, with a trust assumption you have to be willing to make about a chip vendor. Zero-knowledge proofs of machine-learning inference give you cryptographic verification with no trust assumption and are still four to six orders of magnitude away from being economically viable for anything interesting. Optimistic verification with fraud proofs and bonded challengers sits in the middle — cheap in the common case, expensive in the dispute case, and entirely dependent on the existence of a challenger market that actually shows up when the bond is large enough to attract them. Every one of those paths has the same cost structure I described for ZK rollups. Fixed infrastructure, a proving or attestation floor that does not scale down with demand, and a revenue line that scales down the moment the market goes quiet. The zkML provers are running the same step-function economics as the zkEVM provers, at an order of magnitude worse. If you want to know which of these networks survives the chop, stop looking at GPU counts and start looking at who is paying whom for attestation and at what unit cost. When Dupin's Decoupling Dies, Research Dies I made a typo there and I am keeping it, because there is a real concept adjacent to it and I want to name it properly. Alpha equals dispersion times skill. That is the whole equation. If I am twice as good as the median analyst and the cross-sectional dispersion of returns is thirty percent, my edge has something to multiply. If I am twice as good as the median analyst and dispersion has collapsed to eight percent, my edge is multiplied by almost nothing, and after fees and slippage it is multiplied by less than nothing. For the past several quarters, cross-sectional dispersion in crypto has been structurally compressed. The correlation of the median top-hundred altcoin to Bitcoin has been running north of seventy percent, and in stress windows it goes higher, not lower. Realized volatility is down by roughly half from the 2024 expansion. Funding is flat. Skew is flat. Basis is compressed. The term structure of implied vol is close to a straight line. When every asset in a class moves as one asset, there is no cross-sectional question to answer. There is only a beta question, and beta questions are answered by macro, not by research. That is why the research industry is producing empty reports. It is not that the analysts stopped being able to analyze. It is that the market stopped producing questions that analysis can answer. The N/A report is a faithful rendering of a market with no dispersion. Every field evaluates to insufficient information because in a correlation-one regime, the marginal project's specific facts genuinely do not determine its return. The facts exist. They are just not load-bearing. And an honest framework that evaluates what is load-bearing will return N/A. I watched a version of this in 2022, when I mapped the Terra collapse against institutional liquidity drains. The point of that case study was not that a stablecoin broke. The point was that in a liquidity-down regime, the specific mechanics of the specific failure are almost irrelevant to the contagion path. Terra failed because it was levered and reflexive and the Fed was tightening. A dozen other structurally similar things would have failed in its place. The micro details were fascinating and they were not causal. Dispersion was near zero and everything with duration went to the same place. The Trap Isn't the Slop Now the part I actually want to argue with. The consensus reaction to AI-generated research is that it is slop, that it floods the information environment, and that the fix is better detection, better provenance, better disclosure. I think that framing is mostly wrong, and it is wrong in a way that flatters the people holding it. The trap isn't the fake data. It is the assumption that real data was being withheld, and that a competent analyst would have found it, and that the only thing standing between you and the truth is a bad pipeline. That assumption is doing enormous work, and it is not supported. In a correlation-one regime, the truth is that the information does not exist in tradeable form. The project's technical architecture does not matter to its return. Its tokenomics do not matter to its return. Its team's pedigree does not matter to its return. What matters is beta, and beta is a macro variable, and nobody can research it at the project level because it is not a project-level property. The N/A report is not a failure to find the answer. It is a correct report that the question has no project-level answer. And there is a second, uglier implication. A filled-in report on a real project is, in the median case, about seventy percent template. I know this because I have written a lot of them and I have read more. The sections exist because sections exist. The risk matrices list the same eleven risks in different order. The competitive landscape names the same four competitors. The filled report and the empty report differ not in information content but in the willingness to assert. The empty report is bashful. The filled report is confident. Neither one is informative, but only one of them gets believed. That is the illusion of infinite growth — the belief that because the format expands, the content does. Formats are infinitely extensible. Nine dimensions becomes twenty-seven. Add a section for supply-chain transmission, add a section for social sentiment, add a section for on-chain cohort analysis, and watch the word count climb while the entropy of your actual knowledge stays pinned at zero. The emptiest document I have read this year is also the most truthful one. That should disturb you more than it disturbs me. Decoupling From Itself The other sacred cow that needs to go is the decoupling thesis, or at least the version of it that people actually trade. The standard framing is that crypto is decoupling from macro — that it is becoming its own asset class with its own drivers, its own liquidity, its own cycle. The evidence offered is usually that Bitcoin outperformed the Nasdaq during some window, or that crypto rallied while the dollar strengthened, or that some institutional flow showed up on a day when equities were flat. This is a category error, and it is one I have been making for the better part of a decade before I saw it. Crypto is not decoupling from macro. Crypto is decoupling from itself. Look at the actual correlation structure. Bitcoin's correlation to the Nasdaq is unstable in a very specific and predictable way. It rises in stress and falls in calm. That is not the signature of an independent asset. That is the signature of a levered risk asset with a liquidity overlay. When funding is free and volatility is low, crypto does its own thing, and the decoupling narrative gets written. When liquidity tightens, the correlation snaps to one and the decoupling narrative dies quietly in a margin call. Meanwhile, the intra-crypto correlation structure is doing the opposite of what a maturing asset class should do. It is not stratifying. It is fusing. Bitcoin is a macro asset. Ether is a macro asset with a staking yield attached. Everything below the top three is a levered expression of the same macro factor, and the leverage is embedded in the token design rather than in a margin account. So the correct statement is this. The top of the crypto market has successfully decoupled from the crypto market. Bitcoin now trades on ETF flow, global liquidity, and real-rate expectations. The bottom of the crypto market does not trade on anything project-specific at all. And the middle — the Layer 2s, the AI-compute networks, the DeFi protocols with real revenue — is stuck between a macro factor it cannot control and an idiosyncratic factor nobody is pricing because dispersion is dead. That middle band is where the next cycle gets made. It is also where the N/A report lives. What I Am Watching, and What Would Change My Mind I am not going to summarize. Summaries are the format-completeness trap applied to essays. Here is what I am actually tracking, and here is the specific reading that would force me to revise. Blob fee. Watch the base fee on the blob market, not the price of Ether. The entire Layer 2 cost structure pivoted on cheap data availability, and the moment blob demand fills the target consistently, the variable cost comes back into the model and the operators who built fixed-cost proving fleets face a second squeeze from the other direction. If blob utilization stays below target through the next expansion, the fees will never recover, and the rollup business model will have to be rebuilt around something other than transaction fees. Proving cost curves. Watch the unit cost of proving a state transition, not the TPS claims on a landing page. Specifically watch whether witness generation costs fall in line with proving costs, because that is the line item that has been quietly eating the margin while everyone admired the cryptography. If someone ships a prover whose witness costs are near zero, the fixed-cost floor drops and the entire economics of the category changes. Retro Funding's metric design. Watch what the Optimism program measures next, and how quickly it rotates metrics after a gaming incident. The rotation speed is the tell. A mechanism that rotates fast is a mechanism that has accepted Goodhart and priced it in. A mechanism that refuses to rotate is a mechanism that is about to be farmed. Cross-sectional dispersion. This is the single most important number in this essay and nobody publishes it. When thirty-day dispersion in the top hundred expands back above twenty percent, research alpha returns, and the N/A reports will start filling themselves in because the questions will start having answers again. Until then, the honest report is the empty one. ETF flow cadence, not level. The level is a base rate now. The cadence is the signal. Watch whether the inflow pattern is stair-stepping or plateauing, and watch whether the absorption is coming from the same three firms or broadening. Concentration in the filer base is the leading indicator of a distribution problem eighteen months out. Verifiability pricing. Watch which AI-compute networks start charging for attestation separately from compute, and at what multiple. A network that can price attestation at a premium to raw GPU hours has found the product. A network that cannot is a spot market with a token attached. And here is the reading that would change my mind about everything above. If dispersion expands while funding stays flat and realized vol stays compressed, then the market has found a way to produce idiosyncratic questions without producing systemic volatility, and research gets its payoff back without a macro regime change. I do not think that is likely. But I did not think the Fed would end QT without a recession either, and here we are. Chaos is just data that hasn't been labeled yet. That is true. But labeling is a claim, and most labels are wrong, and the industry has spent a decade optimizing for the speed of labeling rather than the accuracy of it. The four-thousand-word report that says N/A nine times is not the failure of that optimization. It is the bill for it, arriving on the first of the month, like every other fixed cost nobody wanted to model.