The System Is Working. That's the Problem.
When an institution does exactly what it was designed to do and still fails its stated purpose, the design is the problem — not the execution.
Every institution this week did exactly what it was designed to do. That’s the problem.
Regulation, safety functions, watermarking programs, chip supply chains, pricing models — none of them malfunctioned. They executed. And in executing, they either failed the purpose they were sold under, or actively inverted it. That’s the through-line. Not dysfunction. Designed outcomes that nobody would defend if you described them plainly.
Start with the most literal case. The utility regulatory doctrine built to protect ratepayers from bad capital allocation is now the mechanism trapping utilities inside stranded assets they cannot close. The rule works. The rule was designed to work this way. Nobody is cheating. The problem is that “working” and “correct” are not the same thing, and we keep treating institutional outputs as if they are.
The same logic runs through OpenAI’s governance story. When a company removes its ethics and safety function, it doesn’t reduce the ethical surface area of its decisions — it removes the person whose job was to notice it. The governance vacuum doesn’t stay empty. It fills with whatever process is already running fastest, which in a frontier lab is product velocity. That’s not a scandal. That’s an org chart working as designed. The scandal is that “as designed” points in the wrong direction.
Anthropic’s watermarking program is the cleanest version of this. Watermarking taxes legitimate users — it adds friction, metadata overhead, and compliance cost to every benign output. It does nothing to an adversarial user who strips the marker in one line of code, or who runs a model that doesn’t implement it. The program doesn’t fail. It succeeds at making Anthropic look responsible. That’s the design. Safety theater is not a bug in compliance programs; it’s the feature that justifies the budget line. The people inside those programs are not cynics. They believe in the goal. The structure defeats the goal anyway.
DeepSeek flipped the script in a way that makes the same point from the opposite direction. A recent open-source release — the agent harness, the scaffolding, the tooling, the orchestration layer — went out under MIT. Then DeepSeek raised API prices. The exact figures are worth confirming against DeepSeek’s current pricing page, but the structure is unmistakable: give away the razor, charge for the blade. The scaffolding is free because it makes the model more valuable. The model costs more because the scaffolding makes it stickier. Every developer who builds on the open-source harness is a developer who is now price-exposed to DeepSeek’s API. The open-source release is a customer acquisition strategy wearing an altruism costume.
The chip story ties it together at the infrastructure layer. TSMC has reported strong revenue growth — the AI race is a fab race, and fabs don’t scale like software. They scale like construction: slowly, expensively, and with a decade of lead time. Every model release, every agent harness, every watermark and ethics review is downstream of whether TSMC has capacity. The institutions debating AI governance are debating the paint color on a building where the foundation is poured by two companies in Taiwan. That’s not a metaphor. That’s the supply chain.
The pattern across all five: the visible layer (the regulation, the safety program, the open-source release, the governance function) is doing what it was designed to do. The invisible layer — stranded assets, governance vacuums, adversarial evasion, customer lock-in, fab concentration — is also doing what it was designed to do. Both layers are working. They are working in opposite directions.
Next week, TSMC’s capacity guidance and any OpenAI org announcement will test whether the gap between visible and invisible continues to widen — or whether one institution finds a way to make its stated purpose and its structural output point at the same target. The week after that will probably answer no.
If your institution is working exactly as designed and producing outcomes you wouldn’t defend out loud, the design is the problem.