A note on sourcing: this piece began with a document I was asked to analyze. It had nine sections, four comparison tables, a risk matrix with probability and impact columns, a token distribution breakdown, and a Howey test assessment. Every substantive cell said N/A. There was no article behind it — no project, no token, no team, no chain. What follows is what I did with that empty input, which turned out to be considerably more useful than the report would have been.
Last Tuesday an analyst I trust forwarded me a nine-section protocol report. Technical maturity table. Supply structure table. Competitive positioning grid. Risk matrix with probability, impact, and mitigation columns. Forty-one rows of analysis.
Forty-one rows said N/A.
I know exactly how it was made, because I have made versions of it. You feed a template into a model. You feed nothing into the template. The model, trained to always produce a deliverable, produces a deliverable. Nine sections. Complete. Useless.
Here is the part that should worry you more than the AI angle: it looked like work. If you scrolled it on a phone in a Telegram channel at two in the morning, you would have skimmed it and felt informed. Structure without content is the most expensive thing in this market, because it costs you the one resource you cannot top up — the attention you were going to spend finding out what is actually true.
And in a bear market, that attention is the whole edge. Gains are gone. Liquidity vanishes faster than a dream in DeFi. What is left is the ability to read what is real, and the discipline to say out loud when nothing is.
So let me do the opposite of the template. Instead of nine dimensions with no information, I want to walk through three places where the number everyone quotes has quietly stopped being a price and become a policy — lending rates, layer-2 economics, and the Lightning Network. Then I want to talk about the machine sitting in the middle of all of it, because the machine is not the villain of this story. It is a mirror.
Three places, one machine, and one question I think this bear market is actually asking.
The framework is not the analysis
Nine-dimension frameworks did not come from nowhere. They came from the ICO era, and they came from a real problem: people were being lied to, constantly, by well-formatted documents. The checklist was a defense. If you force yourself to ask about the team, the token supply, the unlock schedule, the legal wrapper, and the competitive set, you are less likely to be fooled by a slick PDF with a gradient logo.
That logic holds when there is information to organize. A checklist is a sorting tool. It assumes there is something to sort.
The problem is that a template has no null state. It has nine boxes. It does not have a box that says 'I could not determine whether this project exists.' So when the input is empty, something has to go in the boxes, and what goes in is the nearest available text-shaped object. Nothing, rendered in twelve-point type, with a header.
And then something worse happens. A completed framework carries an implicit claim: that nine dimensions were considered, weighed, and reconciled. When eight of them are empty and one of them has a real observation in it — say, a genuine note about a governance parameter — that ninth observation inherits authority it did not earn. It borrows the credibility of the eight hollow boxes next to it.
I learned this the hard way. In 2022 I was running around Kuala Lumpur organizing meetups while the Terra machinery came apart, telling people to keep their spirits up, and I missed the signals because I was busy being the room. When the dust settled I implemented what I call the two-hour rule: nothing publishes until two hours of independent verification have passed, no matter how fast the story is moving. Speed is my whole identity and speed is what nearly cost me my credibility, which is a lesson I would rather have read about than lived.
The rule has since grown a second clause. Before I publish anything now, I do not ask whether I have covered all nine dimensions. I ask one question instead: what is the single thing I know that the market does not? If the answer is nothing, the correct output is a blank page. Not a filled-in template. A blank page.
Nobody publishes blank pages. That is the business model problem, and it is about to become visible in three specific places.
The rate that stopped being a price
Start with the number quoted most often about DeFi lending, and the one that tells you least: the supply APY.
Aave V3's interest rate model is a piecewise linear function. On Ethereum mainnet the USDC market has run with an optimal utilization ratio near ninety percent, a base variable borrow rate at or near zero, slope one in the mid single digits, and slope two somewhere in the sixty to seventy-five percent range. Compound's cToken model, which established the kink shape the entire industry copied, was itself lifted from traditional money-market heuristics rather than derived from anything on-chain.
Every one of those numbers is a governance parameter. It has been proposed, debated, voted on, and changed by people responding to conditions. Which is to say: the curve is a reaction function. It is structurally the same object as a central bank's policy rule, except the committee is a token vote and the meeting happens continuously.
This is not a scandal. It is a design choice, and a defensible one. You need some mechanism to keep utilization below the level where withdrawals start failing, and a punitive slope above the kink is a crude but effective way to do it.
What matters for you as a reader is what the number means in each regime, because the meaning flips.
In a bull market, leverage demand pushes utilization up toward the kink. Slope two engages, borrow rates spike, and the rate is genuinely informative — it is telling you that someone is willing to pay a lot to be long. High rate equals real demand.
In a bear market, utilization collapses. You get ten, twenty, thirty percent utilization. The borrow rate falls to base plus a fraction of slope one. Supply APY settles at one to three percent. And here is the trap: that number now reflects a governance constant, not a market observation. Nobody is bidding. There is no clearing price. You are reading a parameter and mistaking it for a signal.
When a research note tells you that USDC supply APY at 2.4 percent 'indicates stable demand,' it is doing the equivalent of reading a thermometer that has been unplugged and is sitting at room temperature.
There is a cleaner proof that these rates are not prices. Through 2023 and 2024, stablecoin borrow costs on major lending markets sat well below the yield on short-dated US Treasuries. If the borrow market were a real market, that spread should have been arbitraged. It was not, because the two pools of capital are separated by a wall: a treasury desk cannot take smart contract risk at size, and the entities doing the borrowing are precisely the ones who cannot access a T-bill account. A visible price gap with no arbitrage channel is not a market. It is a wall with numbers painted on it.
So if the rate is not the signal, what is? Three things, and none of them are on the dashboard.
First, composition. Aave V3's efficiency mode lets correlated assets borrow against each other at loan-to-value ratios up to roughly ninety-three percent. That is the recursive loop channel. In any deleveraging event, eMode positions are where the cascade originates, because the liquidation threshold sits close to the LTV and the unwind pushes collateral into the same pool that holds it. My rough method: pull the top fifty supplier addresses, cross-reference them against the borrower set, and count how many appear in both lists. That overlap ratio is the real leverage metric. I have seen markets where the headline TVL was flattering and the overlap ratio said the pool was one whale's loop wearing a trench coat.
Second, withdrawal liquidity rather than utilization. Utilization is total borrows divided by total liquidity plus borrows. But the number that determines whether you can get out is available liquidity. In a stress event the failure mode is not a rate spike — it is the rate sitting calmly at two percent while the market is effectively closed because the available liquidity is borrowed out. The rate is the last indicator to tell you the door is shut.
Third, reserve factor and accrued bad debt. Ten to twenty percent of borrow interest is skimmed to the protocol treasury before it reaches suppliers, which is fine but means the headline APY is already net of the protocol's cut. More importantly, bad debt is socialized or absorbed later. The APY you saw last quarter did not include the loss that had not been realized yet.
None of this requires a nine-dimension framework. It requires knowing that the shape of the curve is a policy, and then going and looking at who is actually inside the pool.
The race to zero where the prize is not wired to the token
Now the layer-two stack wars, where I want to argue something that will annoy people on all three sides.
The conventional framing is that OP Stack, ZK Stack, and Arbitrum Orbit are competing on cryptography. Optimistic versus validity proofs. Fraud proofs versus aggregated proofs. Time to finality, proving cost, the identity of the prover.
That framing is not wrong. It is just second-order.
Look at what actually happened. OP Stack signed Base, Unichain, Ink, Soneium, World Chain, Zora, Mode, Lisk, Redstone, Ancient8, Mint. ZK Stack signed ZKsync Era, Abstract, Sophon, Cronos zkEVM, Treasure, GRVT, Lens. Arbitrum Orbit signed Xai, Sanko, Degen, and Ronin, which migrated over as an L3.
Those are business development outcomes. A chain operator choosing a stack is choosing how fast they can ship, who else is already there, what the shared bridge looks like, and — the part nobody puts in the deck — what share of their sequencer revenue they have signed away.
The Optimism side has a Law of Chains and a revenue share arrangement where chain operators commit a slice of sequencer revenue to the Collective. The ZK side has the ZK Gateway and a shared prover argument, where the interop fee ultimately flows to whoever is running the prover. The Arbitrum side has an expansion program that takes a cut to the DAO. Every one of those arrangements is a negotiation, and negotiations get reopened when the number gets big enough to matter.
Here is the sharp version of the finding. Base is the largest OP Stack chain and it is a Coinbase product. Its sequencer revenue accrues to Coinbase. Its voluntary contribution to the Optimism Collective became a governance controversy, publicly argued, over how much the biggest participant owed the shared pot. That is the tell. When the largest member's contribution to a shared treasury becomes a political question, the shared part is not a business model — it is a subsidy that can be revoked.
The token side of this is worse. OP, ARB, and ZK each grant governance over a treasury and over a fee switch that is either off, contested, or both. Turning the fee switch on would immediately become the thing every chain operator negotiates around, because the operator's entire margin is a function of how much of the fee they keep. So the governance that looks like power is actually a deadlock: the only lever that would make the token valuable is the lever that would drive away the activity generating the value.
And the underlying economics are thin. Post-Dencun in March 2024, blob space gave L2s cheap data availability with its own fee market. Pectra raised blob throughput again in 2025. Costs fell. Because every L2 is competing for the same transaction, the savings got passed down and fees collapsed toward zero. Gross margin per transaction held up better than the fee line suggests, but absolute revenue per chain is tiny — on a quiet day, an OP Stack chain can clear a few hundred dollars and go negative after fixed costs.
So the honest summary of the stack war is this: whoever signs the most sovereign chains wins a prize that does not accrue to the token holder. The trap was sweet until the rug pulled, except there is no rug here. There is just a slow, orderly realization that the token was never wired to the revenue, and a governance structure that cannot wire it without alienating the counterparties.
In a bull market nobody checks the fee switch, because the chart goes up and the chart is the argument. In a bear market the fee switch is the only thing that matters, and it is the one thing the structure cannot deliver. That is a governance deadlock dressed as decentralization, and I think it is the most under-priced risk in the sector.
The metrics I would watch instead of TVL: sequencer revenue net of L1 data cost, per chain, published on a consistent schedule. The retention ratio — how much of that net revenue stays with the operator versus flows to the stack. And whether any token has a hard-coded, non-governable claim on any of it. If one chain hard-codes a fee flow into the contract and cannot vote it away, that single decision resolves the stack war more decisively than any proof system.
Seven years of the same plateau
Now Lightning. I have been watching this one since 2017, and I want to say the thing that has been true the whole time and has never been popular.
Back then I was chasing the green candle through the fog of 2017, sitting in rooms in Singapore and Kuala Lumpur where serious people explained that by 2020 Bitcoin would be a payments network, and Lightning was how. Bars, coffee, remittances. The demos worked. The charts went up.
Seven years later, public channel capacity has spent its entire life oscillating in a band that is a rounding error against the asset it is supposed to be scaling — a few thousand BTC against a multi-trillion-dollar market. The publicly visible node count has been declining for three years. Public channel counts are well below their 2022 peak.
I want to be careful here, because there is a measurement problem and it is not incidental. All Lightning statistics are public-channel statistics. Private channels are invisible. That means every capacity figure you have ever seen is a lower bound of unknown tightness. A network whose size literally cannot be measured cannot be modeled, and any analyst who reports Lightning capacity as a level rather than a floor is reporting a number they do not have.
But the deeper issue is not size. It is the path-balance constraint, and it does not get fixed by engineering.
A payment channel is a bilateral credit line with a fixed balance. To route N satoshis, you need a path where every single hop has at least N satoshis on the correct side. As the amount rises, the set of viable paths does not shrink linearly. It collapses.
The industry's answer was multipart payments — split the payment across paths. Simple arithmetic shows the problem. If each part has a ninety-five percent success probability and you need eight parts, your end-to-end success rate is about sixty-six percent. If you need twelve parts, you are below fifty-five. Each additional part multiplies failure probability, and each failure means retrying, which means time and routing fees paid on paths that did not deliver.
This is not a bug awaiting a patch. It is the shape of the design. Lightning scales count, not value. It is excellent at moving small amounts between parties that already have a liquidity relationship. It is structurally bad at settling large value, and no amount of routing cleverness changes the balance constraint.
Then there is jamming. A griefer can hold a payment open across many channels, locking liquidity for hours, and pay only the routing fee on the single hop that finally fails. The attacker's cost approaches zero. The network's cost is locked capital across the graph. Reputation-based mitigation and upfront fee proposals have been debated since 2022 and remain incompletely deployed, which means liquidity providers carry an unhedgeable tail risk. Capital that cannot price its own tail risk does not scale, it retreats.
And on the receive side there is the rent. To receive reliably you need inbound liquidity, which someone has to fund. You pay a swap fee to a service like Loop or Boltz, you pay the opportunity cost of the locked capital, and you eat the retries. Inbound liquidity is also perishable — it gets consumed as you receive. So the liquidity service provider business is not a service business. It is a rental business, and the tenant pays forever.
Add force-close costs during mempool spikes, watchtowers that mitigate theft but not economics, and you have a system whose true cost per payment for a merchant is invisible on the fee line and substantial on the balance sheet.
Here is the part that should end the argument. The volume that 'proves' Lightning works is heavily custodial. Cash App, Strike, Blink, Kraken, and until it pulled from US app stores citing regulatory exposure, Wallet of Satoshi. A custodial-to-custodial payment is, at the edges, a database update between two companies. The network's headline throughput is being carried by entities that are not using the network as a network. That is not a victory for Lightning. It is a concession by Lightning, dressed as adoption.
The 2025 hope is Taproot Assets — USDT issued onto Lightning. And I understand the appeal. But a five-hundred-dollar stablecoin payment inherits exactly the same path-balance constraint as a five-hundred-dollar BTC payment. Stablecoins on Lightning do not fix Lightning. They give it a new population of payments to fail on.
My read, and I have held it since about 2020: Lightning is not dead and will not die. It is going to keep doing precisely what it does — move small amounts cheaply between parties who have already solved liquidity for each other. That is a real business. It is not a scaling layer for Bitcoin, and the seven-year pattern is not a pre-breakthrough plateau. It is the shape of the ceiling, and it has been visible the whole time.
The machine in the middle
Which brings me back to the report with forty-one N/A rows, because the machine that produced it is now sitting in the middle of every one of these markets.
In 2025 I partnered with an AI-agent platform to test real-time trading bots. I did not do a deep code audit. I did what I always do — I watched how it behaved under pressure. I ran it live through a volatility event and what I saw was consistent: the bot overreacted to social media noise. Not occasionally. Systematically. A headline would move sentiment, sentiment would move the bot's position, and the bot would whipsaw into a stop it did not need to take.
The developers told me that was a training problem. It is not. It is an objective problem. If you optimize a model on price data plus sentiment features, and in-sample sentiment correlates with short-term price, the model learns to weight sentiment heavily. It is doing exactly what you asked. Out of sample, in a regime where sentiment is mostly manufactured or stale, that weighting becomes a liability. The model did not malfunction. It succeeded at the wrong thing.
Now stack the template problem on top. The marginal cost of producing a research document has gone to approximately zero. So supply goes to infinity, because that is what happens when marginal cost hits zero and nobody has to pay for the downside of publishing. The bottleneck in this industry has moved from production to curation, and curation is the one thing that does not scale, because curation requires a person to say no.
The empty framework is the purest artifact of that shift. It is what you get when a generation tool is asked for a nine-section deliverable from an input of nothing, by a system trained to always produce a deliverable. And note that the failure is not random — it is convergent. Every model makes the same move, because every model was trained on the same corpus of confidently-filled templates. The corpus is the problem. The models are just inheriting it faster.
There are specific hallucination modes that matter more than the generic ones, and they map exactly onto everything I have written above. Phantom correlations, where a model finds a relationship in a short window that has no mechanism behind it. Stale regime assumptions, where the model does not know that the last eighteen months were a different market. False precision on parameters — a model trained on historical APY data has no way to know that the number is a governance constant, so it will confidently forecast the future of a number that a token vote can change on Thursday.
Where AI is genuinely additive is elsewhere. Cross-referencing address sets at a scale no human can touch. Anomaly detection in governance forums. Monitoring parameter change events. Which is ironic, because parameter change events are exactly the thing humans should have been watching all along and were too busy watching the APY line.
My job in 2026 is not to know more than the model. That is a losing position and it gets worse every quarter. My job is to know which of the model's inputs are fake. That is qualitative, it is contextual, it is grounded in fifteen years of watching rooms and reading sentiment before it shows up in price, and it is the one skill that does not get arbitraged away quickly, because it cannot be trained on a dataset that does not exist yet.
Fifty percent down, one hundred percent ready.
The blind spot nobody is naming
Everyone is worried that AI is going to fill this industry with garbage. I want to argue the opposite, and I think it is the most important thing in this piece.
The garbage is the honest part.
An N/A-filled template is a document that admits, in its own flat way, that it has no information. It is ugly and it wastes your attention, but it does not lie to you. The genuinely dangerous documents in this market are the ones with plausible numbers in every cell, because plausible numbers get acted on. A report that says 'I could not determine the team' is doing you a favor. A report that says 'the team is strong, based on their LinkedIn presence' is doing you harm, and it is doing more harm the more polished it looks.
Which leads to the second thing nobody is naming. This bear market is not a research drought. It is a research revelation, because the subsidy that made every number look meaningful is gone.
In a bull market, price appreciation validates every thesis, so nobody checks whether the model was right. The Aave rate curve looked predictive because borrowing went up and the token went up. The L2 fee switch looked irrelevant because the token went up anyway. Lightning looked like it was working because everything was working. In a bear market, the only theses that survive are the ones with a causal mechanism. And all three of the mechanisms I walked through above — the governance-parameter rate curve, the un-wired L2 fee flow, the path-balance constraint — were there the entire time. You just could not see them through the price.
This is the lesson I should have learned in the Dubai galleries in 2021, when I stood in a room full of white-whale holders and realized the party was ending two weeks before the floor cracked. Gallery walls do not hold value. Neither do price charts. The thing that holds is the mechanism underneath, and the mechanism is the only thing you can actually analyze. Art is dead; long live the algorithmic pixel — and long live the causal chain that governs whether the pixel is worth anything.
And the third blind spot, which is the one that actually keeps me up. The readers who can detect bad research are the ones who need it least. The readers who cannot are the ones with the most to lose. That asymmetry has always existed, but the marginal cost of research going to zero has widened it enormously, because now the bad research arrives at the same speed as the good research and looks identical from a phone screen at two in the morning. The research layer is failing precisely the cohort it claims to serve.
What I am watching
Four things, and none of them are prices.
Governance parameter votes on major lending markets. Not the APY page — the votes. If the optimal utilization ratio on a market moves, something happened to the risk assessment, and that something is information the APY line will never show you.
Whether any layer-two hard-codes a revenue claim into its token at the contract level. A fee flow that governance cannot vote away. If that ships anywhere, it settles the stack war more decisively than any proving system, and it will force every other stack to answer a question they have spent three years avoiding.
Lightning routing success rates above one million satoshis, and whether jamming mitigation finally ships with a real fee attached. If it does not ship in the next cycle, the capital-lockup problem becomes permanent rather than temporary, and permanent is a different category of outcome.
The ratio of published research to admitted N/A. I am not being cute. That ratio is the industry's honesty index, and it is measurable, and it is going the wrong way.
There is a version of this market where everything I just described works out. The rate curves become genuinely market-driven. Some L2 wires its token to its revenue and the rest are forced to follow. Lightning finds its real niche and stops pretending to be a scaling layer. The AI flood recedes into a curation layer that humans actually staff.
None of that happens by itself. It happens if enough people are willing to publish a blank page. Speed is the only asset that never depreciates, and I have staked my entire career on it — but speed without a null state is just noise at scale, and we have produced enough of that now to drown in it.
So here is my question, and I do not have a comfortable answer. When a report has no information, the only correct output is an empty page. Who in this industry is still willing to publish one, and — more to the point — which of you would actually pay for it?