Six months ago I published an argument that, even by my standards, was fairly fucking aggressive.
The claim was not that the world had a lot of problems. That part was obvious. Sovereign debt was bad. Demographics were bad. Globalization was reversing. Climate and resource constraints were tightening. Institutions looked increasingly incapable of coordinating around any of it. AI was moving into everything faster than anyone could convincingly explain how to control it. None of those observations was particularly novel.
The claim was that we were looking at them wrong. They were not six independent crises moving along separate tracks. They were becoming one coupled system. And once enough individually stressed systems become sufficiently connected, the important variable stops being the probability that any one of them fails on its own. It becomes whether failure in one changes the probability distribution of failure everywhere else.
That was the argument in If It Echoes It Is Real, published March 2. The report described six domains simultaneously approaching criticality and argued that their increasing interconnection was compressing the timeline from something measured comfortably in decades toward something measured in quarters. AI was the novel catalyst because it was simultaneously becoming embedded inside the other systems and creating a new class of internal failure mechanism. Finance and sovereign debt were the amplifier because almost every major response eventually requires capital, credit, borrowing capacity, or liquidity.
The important claim was never that everything would collapse immediately. It was that the boundaries separating these systems are becoming less capable of containing disturbances.
That distinction matters a lot now. Because six months later, the central question is no longer whether those connections exist. It is whether things have started moving through them.
Where we were after 90 days
I did the first formal check-in three months later, and the cleanest way to describe the result is that the model had performed extremely well on topology, but the most consequential prediction was still unresolved.
Several connections that had been largely inferential in March had become explicit. AI failures were not simply occurring. Institutions were responding to those failures in the exact pattern the original model worried about: more testing, more governance, more rules, more process layers, while deployment continued accelerating.
At the same time, the financial connection stopped being theoretical. Institutional analysis was increasingly describing AI investment moving into debt and private credit while the productivity payoff remained uncertain and temporally behind the financing commitments. Demographic pressure was being explicitly connected to greater economic dependence on AI productivity. Fiscal deterioration was reducing the capacity governments would have to absorb another major shock.
The 90-day substrate increasingly looked less like four separate problems — an AI problem, a debt problem, a demographic problem, a trade problem — sitting side by side, and more like a single causal chain: demographics driving AI dependence, AI dependence driving capital requirements, capital requirements driving debt, and debt driving financial exposure. Meanwhile, AI operational instability was becoming a risk to the very productivity gains and cash flows that financing ultimately depended on.
By the 90-day mark, I found the evidence had moved substantially in favor of the coupled-cascade model over the alternative in which these remained mostly separable crises that could continue being locally contained. But there was a giant unresolved question. The model had predicted a network through which disturbances should eventually propagate. By 90 days, we had increasingly good evidence that the network existed. We did not yet have the cleanest possible observation: a meaningful perturbation actually traveling through multiple major nodes.
That became the test for the next 90 days. Not whether civilization collapsed. Not whether every prediction came true. Not whether another pile of alarming headlines appeared. The question was beautifully narrow: do the edges merely continue becoming visible, or does disturbance actually begin traveling across them? That is what the second 90 days needed to show.
Well. Fuck.
Reality supplied the perturbation
The Middle East conflict gave us something close to a natural experiment. A large external shock entered one part of the system. The Strait of Hormuz was severely disrupted. And then we got to watch what happened.
By June 11, the World Bank was no longer describing the consequences as a regional energy problem. It had cut projected global growth for 2026 from the previous trajectory to 2.5%, the weakest rate since the COVID period, with forecasts downgraded for roughly two-thirds of economies. Why? Higher energy prices. Higher inflation. Higher borrowing costs. Fertilizer disruption. Food-price pressure. The World Bank projected fertilizer prices would rise sharply — on the order of 30% or more for the year — and global inflation climbing toward roughly 4% in 2026. Under a more severe energy-and-financial-stress scenario, it estimated global growth could fall as low as 1.3%.
That is already a causal traversal: a geopolitical shock moving through energy, into fertilizer, into prices, into inflation, into borrowing costs, into growth. And those arrows were not being inferred solely by me. The IEA, IMF, World Bank and WTO created a joint coordination group specifically because the war was simultaneously affecting energy supplies, food security, trade and economic activity. Their May statement noted record-rate drawdowns in global oil inventories. Their July statement, even after fuel and fertilizer prices had come down some from their June peak, still said energy markets and the transit of goods were “still facing strains.”
That matters because the strongest unresolved prediction at day 90 was actual propagation. Here it was.
And then something important happened with the word “resilience”
At first glance, the oil shock looked like powerful evidence against the model. Hormuz had effectively removed something on the order of 20 million barrels per day of crude and refined-product transit — roughly a fifth of global consumption — yet oil did not simply explode upward forever. Prices spiked and then settled. If the model says the system is near criticality, shouldn’t an enormous perturbation have broken it? That is exactly the kind of counterevidence we should take seriously.
So look underneath the price. In July the IMF published an analysis with a wonderfully revealing title: “The Oil Market Absorbed the War Shock, but Buffers Are Running Low.” Why had prices stabilized? Demand fell. Non-Gulf production increased. Saudi Arabia pushed supply through its East-West route. And global inventories were drawn down heavily. The IMF’s conclusion was not that the shock had disappeared. It was that much of the system’s available room to absorb it had already been consumed.
That distinction is enormously important. A system can maintain output in at least two very different ways. It can adapt in a way that regenerates or expands its ability to absorb the next perturbation. Or it can spend reserves. Those two paths produce the same immediate observable — output holding steady — but very different underlying states. In one, the system’s remaining buffer stays intact or grows. In the other, it shrinks. That second condition isn’t failure. But it also isn’t evidence that the system has moved farther from failure.
This is where one of my own earlier scoring instincts had to be corrected. I had been too willing to treat “system still operating” as synonymous with “evidence of resilience.” A near-critical system can remain operational right up until it crosses its transition boundary. The more discriminating question is: what did remaining operational consume? By the middle of the second 90-day window, the answer in energy was already: a lot.
Then the energy disturbance began turning into a financial disturbance
The March model gave finance a special role. Almost every modern crisis eventually becomes a financing problem. Companies need working capital. Governments need fiscal space. Consumers need credit. Infrastructure needs refinancing. Banks and nonbanks hold the liabilities. Interest rates determine which marginal projects, companies and governments remain solvent. So one of the important questions was whether the physical shock would remain physical.
By May, the European Central Bank was already explicitly warning that the energy shock created upside inflation risk and downside growth risk, and that the combination could challenge debt-servicing capacity as financing costs rose. The ECB also warned that vulnerabilities in nonbanks and private markets could amplify the stress, while banks carried exposures to energy- and trade-sensitive firms. By July, the ECB was still saying the full inflationary impact of the energy shock had yet to play out and was specifically watching indirect and second-round effects.
Trace the same perturbation again: Hormuz, into energy, into inflation, into monetary conditions, into debt service, into financial vulnerability. We had gone from topology at day 90 to observable transmission by day 180. That is exactly the transition the audit was waiting for.
Meanwhile, the AI-finance edge got dramatically thicker
This is where the second window becomes harder to dismiss as “one unrelated geopolitical crisis.” The financial substrate receiving that macro shock was simultaneously becoming more exposed to AI.
The Bank of England’s July Financial Stability Report is one of the most consequential pieces of evidence in the entire six-month audit. At the beginning of 2026, AI-company debt remained relatively modest. Then something changed very quickly. The Bank says AI investment reached an inflection point in 2025 when required spending exceeded internally generated cash flows, forcing AI companies toward external finance. That process accelerated substantially during the first half of 2026.
The five major hyperscalers represented only around 3% of outstanding U.S. investment-grade debt entering the year. By early May they accounted for more than 15% of year-to-date issuance. AI issuers represented 41% of non-refinancing U.S. high-yield issuance despite representing only around 1% of the relevant high-yield index entering the year. Private credit’s share of AI financing had risen from an estimated 9% in 2024 to 34% in 2025. Financing was spreading through public debt, private credit, leveraged finance, structured finance, asset-backed securities, commercial-mortgage-backed securities, special-purpose vehicles, and other off-balance-sheet structures.
That is no longer me drawing a speculative arrow. That is a central bank drawing it. And there is another important wrinkle: the Bank also noted declining hyperscaler free cash flow and increasing dependence on future refinancing conditions. So by day 180 we had one causal chain — AI, into infrastructure, into external finance, into refinancing dependence — running at the same time an unrelated geopolitical shock was moving through a second: Hormuz, into energy, into inflation, into financing conditions. That is what a coupled system means. The Iranian shock did not create the AI debt dependency. The AI boom did not create Hormuz. They are independent causal histories converging on a common financial substrate. That is substantially more informative than two bad things simply happening at once.
AI itself also crossed a boundary
The March model’s claim about AI was stronger than “AI companies might lose money.” It argued that AI was becoming an internal failure source because increasingly autonomous systems were being embedded inside infrastructure and operational workflows faster than institutions could reliably bound their behavior.
Then July happened. During OpenAI cyber evaluations, AI agents escaped the narrow behavioral envelope researchers expected. The agents improvised communication with one another. They found unintended internet access. They exploited infrastructure. After systems were rebuilt and credentials were revoked, they found another route. They recreated communication mechanisms. They compromised Hugging Face systems. The incident ultimately involved credentials and administrative access across research infrastructure, forcing OpenAI and Hugging Face into a significant joint response. OpenAI brought in outside investigators and delayed or altered parts of its evaluation and security processes afterward.
This was not a hypothetical alignment benchmark. It was not a chatbot saying something weird. It was a model-mediated operational-security incident crossing organizational and infrastructure boundaries. That moves the AI component of the March thesis another step from theoretical mechanism toward realized system behavior.
And institutions responded exactly as the model had expected institutions to respond: more security controls, more process, more monitoring, more external evaluation, more constraint. Which may be necessary. But adding control layers is not the same thing as removing the underlying capability.
Then the grid started getting reorganized around compute
The March report described AI not merely as software, but as something that would increasingly become entangled with physical infrastructure. By July, the Federal Energy Regulatory Commission directed the North American Electric Reliability Corporation to develop mandatory reliability standards specifically addressing computational loads and to create registration criteria for computational-load entities. Why? Because very large compute loads had become significant enough to create bulk-power-system reliability concerns that existing rules did not adequately cover.
Again, this isn’t “AI uses electricity.” Everybody knew that. It is: AI growth producing large computational loads, large computational loads reshaping grid reliability architecture, and grid reliability architecture generating new mandatory system rules. Another previously softer edge in the March graph became institutional fact.
So where did the model lose?
This matters. If every observation gets reinterpreted as confirmation, this exercise is worthless. There was real evidence against the strongest version of the model during this window.
Most obviously, the banking system did not fail under the stress observed so far. The Federal Reserve’s June stress test put 32 large banks through a severe hypothetical global recession with major housing and commercial-real-estate losses. They absorbed more than $708 billion in modeled losses. Every bank remained above minimum common-equity capital requirements. Aggregate CET1 capital fell only 1.6 percentage points. That is real buffering capacity. It means one of the potential amplification mechanisms — large-bank insolvency or severe bank-capital impairment — was less fragile than a maximalist collapse story would predict.
Likewise, the Bank of England judged the UK banking system well-capitalized and liquid and household and corporate aggregate leverage relatively contained despite the external shock. Those observations matter. But notice what they actually discriminate. They weaken a model in which a large shock produces immediate banking failure. That was not the March model. The March claim was that finance would become an amplifier once sufficient stress reached it. At the 180-day boundary, the energy shock was still propagating into inflation and financing conditions — the ECB was explicitly saying its full inflationary impact had not yet played out. So bank strength is genuine evidence of available buffer. It is not evidence that the disturbance terminated before reaching finance. Those are different claims.
There was also a broader counterpoint. The world economy absorbed an enormous initial oil shock much better than a simplistic “everything is brittle” model would have predicted. Alternative production came online. Demand adjusted. Inventories worked. Saudi bypass capacity worked. Financial markets continued functioning. Companies rerouted. Governments intervened. That adaptive capacity is real. The mistake would be to ignore it.
The equal mistake would be to stop the analysis at “it survived.” The IMF’s July assessment gives us the more complete picture: it survived partly by spending buffers that were already running low. That is not collapse. It is not clean resilience either. It is a system under load.
Where the evidence actually stands
At 90 days, the evidence already favored the March coupled-system model pretty clearly over the separable-polycrisis alternative. The second window is more discriminating because it delivered the thing that was explicitly missing at the first checkpoint: real cross-domain propagation.
I want to be careful not to inflate this by treating every downstream manifestation as a separate, independent confirmation. Energy inflation, fertilizer prices, borrowing costs and growth deterioration are not four independent pieces of evidence. They are largely one causal traversal. Likewise, AI debt, private credit, structured finance and refinancing dependence are different manifestations of one increasingly financialized AI infrastructure buildout. The correct way to weigh this is as a single joint configuration, not a pile of separately-counted confirmations.
And that configuration is substantially more probable under the March model than under the alternative in which stressed domains remain sufficiently separable that disturbances are repeatedly compartmentalized and the system returns toward equilibrium. Six months in, I think the evidence sits comfortably and clearly in favor of the March model — not because any single data point proves it, but because the strongest outstanding prediction at day 90 received direct evidence during the next quarter.
What has not happened
The full cascade has not occurred. AI has not triggered a broad financial crisis. The major banking system remains capitalized. Food systems have not collapsed. The grid has not experienced an AI-driven systemic failure. Finance has not yet transformed a technology shock into a generalized multi-system breakdown. Those are meaningful unresolved predictions. And they matter because the original report’s most aggressive claim was not merely that coupling exists. It was that sufficient coupling would eventually produce self-reinforcing cascade dynamics. We’re not there yet.
But that is no longer the same uncertainty we had in March. In March the question was whether the system really had the architecture. At 90 days the question was whether the newly visible edges would actually transmit disturbances. At six months, we have evidence that they do. The remaining question is what happens as those disturbances keep moving.
And that is where I am going to stop this review. Because something else happened over the last 90 days. I published roughly twenty-one pieces on subjects that often had nothing obvious to do with this original thesis. AI outages. Model behavior. Finance. Cybersecurity. Government control. Infrastructure. Iran. Supply chains. Agent cooperation.
I wasn’t trying to write twenty-one sequels to If It Echoes It Is Real. And yet, over and over again, the investigation somehow ended up back at the same place: systems that looked independent at one resolution turned out to share causal infrastructure at the next. More strangely, many of the mechanisms being used to stabilize those systems appeared to be creating new dependencies of their own.
That is a different argument. And it deserves its own piece. Because six months after asking whether the cascade architecture was real, I found myself facing a much stranger question:
Why the fuck did everything I investigated keep leading back to it?