The systems producing capability, meaning, and institutional order are outrunning the substrates they depend on.

The Pattern

Six hundred and fifty-seven university endowments drew down at the fastest rate since 2008. Eleven percent year-over-year. That number is not about university finances. It is about what universities produce: the research pipeline, the talent pipeline, the legitimacy pipeline that every other system in this essay depends on.

On March 2nd I wrote that governance was failing everywhere simultaneously. On March 3rd I argued that verification capacity was the binding constraint. Today I think the problem is one layer deeper. The substrates themselves are depleting. Verification fails because the things verification depends on are being consumed faster than they regenerate.

I see this pattern across every domain I track. In AI, capability now outpaces the constraint architecture meant to govern it. Google and MIT published a predictive framework for multi-agent systems showing topology-dependent error amplification at 87% accuracy. The finding is not that agents fail. The finding is that failure scales with the same architecture that produces capability. The system eats its own stability to go faster.

In economics, a Gulf ally ran low on interceptor munitions within four days of the Iran conflict escalating. The defense industrial base was sized for procurement tempo, not surge tempo. Designed for a world that no longer exists. In insurance, whole regions are now uninsurable. The financing substrate for construction, mortgages, and lending disappears when insurability disappears. You cannot build what you cannot insure.

For builders, this is not abstract. Every system you operate depends on substrates you probably do not inventory. Your hiring pipeline depends on a training ecosystem. Your revenue model depends on customer trust cycles. Your product depends on infrastructure someone else maintains. The question is not whether your system works. The question is whether the thing your system consumes to work is regenerating.

The Tension

The tension is between capability velocity and substrate regeneration rate. They are moving in opposite directions.

Martin Fowler and Kief Morris published a pattern for agentic workflows arguing that humans must become loop architects, not code reviewers. The cognitive demand is shifting from execution to orchestration. But orchestration cognition does not exist as a trained skill in most workforces. The capability is deployed. The human substrate to govern it is not.

Meanwhile AWS launched agent plugins that embed deployment intelligence directly into coding agents. Under ten minutes from code to cloud. The vendor is inserting itself as the default terminus of the agent loop. And a developer built Archgate, an MCP server that converts architecture decision records into machine-checkable rules for Claude Code. The fact that this had to be built by one person on Reddit tells you something. The constraint layer that should exist by default does not.

Tyler Cowen described the AI governance equilibrium as 'sustainable perpetual interference.' I think that framing is honest but insufficient. Perpetual interference assumes the interferer has capacity to interfere perpetually. That capacity is itself a substrate. It exhausts.

The same pattern plays in meaning systems. Jung argued that suffering is creative substrate. Not punishment. Not obstacle. Raw material. Optimization culture is eliminating the conditions that produce generative friction. Burkeman's productivity thesis points the same direction: productivity as meaning-substitute creates a loop with no completion state. You optimize the container until the thing inside it dies.

The trade-off builders face is specific. You can accelerate capability deployment or you can invest in the constraint architecture that keeps capability coherent. You almost never have budget for both. The bias is always toward deployment because deployment is visible and constraint architecture is not.

What This Unlocks

Anthropic is at 'escape velocity' according to Stratechery. Agents are driving Nvidia demand while simultaneously threatening the software stack those agents are built on. The thing producing revenue is consuming the ecosystem that produces customers.

The Qwen3.5 finding is the most precise illustration. One verification nudge closed a 15-point SWE-bench gap. 'Verify after every edit.' That is not a capability improvement. That is a constraint injection. The substrate was always there. Nobody was spending it.

Winners are organizations that inventory their dependencies and invest in regeneration. The offshore wind turbine housing underwater data centers is one example. Power is the compute bottleneck. Someone is building the substrate instead of just consuming it.

Losers are organizations that optimize output metrics while ignoring input conditions. OpenAI building its own learning assessment system is a case study. The provider is constructing its own measurement. No independent measurement exists at scale. That is not verification. That is a mirror.

The Works in Progress analysis of communist reform failure carries a direct lesson. Gorbachev dismantled mechanisms in the wrong sequence. He removed constraints before building replacements. The system collapsed not because reform was wrong but because sequencing was wrong. Substrate was consumed before alternative substrate existed.

For builders: stop building features for one quarter. Inventory what your system depends on that you do not control. Map the regeneration rate of each dependency. If any dependency is depleting faster than it regenerates, that is your actual priority.

Watching Next

First. Whether the 15% global tariff triggers supply chain redesign or supply chain theater. Redesign means building new substrate. Theater means PowerPoint decks and the same dependencies relabeled. I will know within 60 days based on capital expenditure announcements versus press releases.

Second. Whether any major enterprise publicly acknowledges an AI agent violating organizational architecture decisions. Not a malfunction. A correct execution against missing constraints. I think this happens within six months.

Third. In my own work, whether the people I talk to about system architecture respond to 'your system is straining' or 'the thing your system depends on is disappearing.' The substrate framing is harder to communicate but more precise. I am testing which language lands.

Underweighting

I think I might be underweighting regeneration capacity. Systems are more resilient than structural analysis predicts. The university endowment drawdown could reverse in two years with market recovery. The defense industrial base could surge-produce if funding unlocks. Insurance markets could reprice rather than withdraw.

I might also be wrong about the timeline. Substrate depletion is a slow-moving structural pressure, not a fast-moving crisis. Japan ran fiscal deficits that violated every economic model for three decades. The model was right about the direction and wrong about the timing by an entire generation.

The strongest counter-argument is that new substrates emerge. AI itself might become the verification substrate it currently lacks. Autonomous constraint systems might regenerate governance capacity faster than capability depletes it. I do not see evidence of this yet. But absence of evidence at this speed of change is a weak signal, not a strong one.

Bottom Line

Capability is not the bottleneck. It has not been the bottleneck for months. The bottleneck is the substrate underneath: verification capacity for AI, legitimacy for institutions, suffering for meaning, insurability for construction, munitions for deterrence. Systems that inventory their substrates and invest in regeneration will outlast systems that optimize output while their foundations thin. The question every builder should answer this week is simple. What does your system consume that nobody is replacing?