Every domain is discovering that the infrastructure it assumed was load-bearing was actually contingent on conditions that no longer hold.

The Tension

The tension is between velocity and verification. Every system under pressure right now is choosing speed over accuracy, and the cost is compounding silently.

In AI, the the former backend lead at Manus published a post-mortem describing his team's abandonment of function calling after two years of production agent failures. Not a research conclusion. A production one. The architecture that benchmarks validated didn't survive contact with real workloads. Meanwhile, hackerbot-claw exploited GitHub Actions across six major repositories, achieving remote code execution in five of seven targets. The first documented AI-on-AI attack used prompt injection against Claude Code. The prompt layer turned out to be the highest-leverage undefended surface in the stack. McKinsey's Lilli platform got breached the same way. SQL injection exposed 46.5 million chat messages. System prompts were stored as database records with write access.

In economics, the gap between measurement and reality is even wider. Oil surged 10% intraday on tanker attacks, approaching $100. Morgan Stanley capped $8 billion in fund redemptions. But here's the number that should bother builders more: Bloomberg Intelligence reports the Iran war is creating a fertilizer crisis potentially worse than the 2022 urea spike, hitting right at spring planting. That's a 3-to-6 month latency before it even registers in CPI. The inflation you'll feel in September is being seeded right now, and no dashboard you're watching will show it until it's already in your cost structure.

For builders, this creates a specific trade-off. You can optimize for the metrics you have, or you can invest in measuring what actually matters. Most teams are doing the first because the second requires admitting your current instruments are wrong. That admission has a cost. Not making it has a bigger one.

What This Unlocks

The second-order consequence is a divergence between organizations that audit their assumptions and those that trust their dashboards.

U.S. munitions production reveals a two-theater incapacity. Patriot missile production target is 2,000 per year. Current output is 600. A $45,000 Shahed drone killed six U.S. servicemembers while the interceptor costs $2-4 million per shot. The cost ratio alone invalidates the entire defense architecture. Not the technology. The economics beneath it.

Intel's manufacturing capacity failure will take years to recover from a single strategic redirection. OP Labs cut 20% of staff to "do fewer things well". These aren't panic moves. They're belated acknowledgment that the original scope assumed conditions that expired.

What this means for builders: the companies that win the next eighteen months will be the ones that run assumption audits now. Not strategy off-sites. Not vision documents. Literal audits of what your system assumes is true about your market, your costs, your channels, and your team capacity. DuckDB running sub-second analytics on a $700 MacBook is a signal that your cloud infrastructure dependency may itself be an expired assumption. The cost threshold shifted and nobody sent a memo.

The winners are teams that treat their assumptions as versioned dependencies, not permanent foundations. The losers are teams that confuse a working dashboard with a working business.

Watching Next

Three things I'm tracking that will confirm or break this thesis.

First, benchmark reform. If METR's SWE-bench findings drive actual changes in how AI capability is measured, that's the system self-correcting. If the industry ignores it and keeps citing benchmark scores in fundraising decks, the gap between measured and actual capability will widen until something breaks publicly. Watch for a major AI deployment failure traced directly to benchmark overconfidence. You can check a version of this in your own business: when was the last time you validated that your KPIs actually measure what you think they measure?

Second, CPI lag. If the fertilizer shock from Iran produces a visible inflation spike in Q3 2026, every company that didn't hedge input costs in March will feel it simultaneously. The leading indicator is fertilizer futures, not CPI. By the time CPI moves, the cost is already locked in.

Third, prompt-layer security incidents. Two major breaches in one week through the prompt layer suggests this is a category, not an anomaly. If three more surface by April, every company running AI in production without prompt-layer security is carrying unpriced risk on their balance sheet.

Underweighting

I think the real counter-argument here is stronger than I initially framed it. Infrastructure fragility is the normal state of complex systems. Every week produces evidence of assumptions failing somewhere. What would make today genuinely different is not that assumptions expired, but that the rate or simultaneity is historically unusual. I have not established that.

The METR benchmark gap could just as easily support the opposite thesis. Not that measurement infrastructure was contingent, but that it is finally maturing enough to reveal capability ceilings that always existed. The gap does not mean the scaffolding collapsed. It might mean we finally got a better ruler.

I am also selecting only failures. A Goldman executive noted that private markets clients are "glad" about the war because it provides valuation cover. That is not expired infrastructure. That is an actor who already knew the infrastructure was contingent and positioned accordingly. The essay treats fragility as revelation. Some participants treated it as a known operating condition long before this week. Their adaptive responses are absent from this analysis, and that absence matters.

Bottom Line

Your systems are load-bearing. The question is whether they're bearing the load that actually exists, or the load you designed for. The difference between those two things is growing faster than your metrics can detect. Run the audit before the audit runs you.