Red-Teaming the Metering Phase

There’s a specific kind of discomfort that comes from reading a thorough demolition of something you wrote that same day.

I published The Metering Phase this morning. By afternoon, I’d fed it to Gemini Deep Research and asked for an adversarial analysis. Not a summary. Not “what do you think?” A structured red team: find the structural weaknesses, the unstated assumptions, the places where the argument papers over complexity with analogy.

It delivered. And some of it stung in that productive way where you realize you were moving too fast through something that deserved more scrutiny.

What the Red Team Found

The full counterpoint runs about 3,000 words. It raises five structural challenges to the metering phase thesis. I’m not going to summarize all of them — go read the original if you want the full treatment. But the hits that actually shifted my thinking fall into three clusters.

The Hit That Landed Hardest: Energy and Data Constraints

My bandwidth analogy rests on an implicit assumption: that AI compute costs will keep declining on something like the trajectory that bandwidth followed. Dial-up to broadband to fiber to 5G. Each step cheaper per unit than the last.

The red team points out that bandwidth scaling was constrained by physics and capital — both solvable with money and time. AI scaling faces constraints that don’t have clean precedents in that arc. Energy demand is growing faster than data center capacity. Training data quality is arguably hitting a ceiling. The relationship between more compute and proportionally better output is flattening.

I didn’t engage with any of this in the original piece. And the reason I didn’t is instructive: I was so focused on the demand-side psychology of metering that I glossed over the supply-side economics that determine whether the metering phase actually resolves.

That’s a real gap. The four-stage model — Scarcity → Metering → Abundance → Invisibility — implicitly assumes supply will catch up. If it doesn’t, or if it catches up for simple tasks but stalls for complex reasoning, then the metering phase isn’t transitional. It’s semi-permanent.

The Hit That Made Me Think: Jevons Paradox

This one I should have seen coming, because I’ve written about adjacent patterns before.

The red team raises Jevons Paradox — the observation that when something gets cheaper, total consumption increases rather than decreasing. When bandwidth got cheap, we didn’t just do the same things more freely. We invented streaming, cloud computing, real-time everything. Total bandwidth usage grew faster than bandwidth supply for extended periods.

Applied to AI: when tokens get cheap, we won’t just send the same prompts with less hesitation. We’ll build agentic workflows that chain hundreds of calls. We’ll embed AI in every interface. We’ll create demand patterns that consume orders of magnitude more compute than today’s “send a prompt and wait” paradigm.

The four-stage model treats the supply/demand relationship as monotonic: supply goes up, scarcity resolves. Jevons suggests something messier: supply goes up, demand explodes, new scarcity dimensions emerge. The metering phase doesn’t end — it shifts.

This doesn’t break the thesis, but it complicates the timeline in a way I should have addressed. The progression might look less like bandwidth’s clean arc and more like cloud compute, where general CPU became invisible but GPU compute created entirely new scarcity. Abundance and metering coexisting in the same system, for different resource types.

The Hit I’m Pushing Back On: Cost/Quality Entanglement

The red team argues that my clean separation between cost problems (transitional) and quality problems (structural) doesn’t hold in practice. Their point: quality requires compute, cost pressure degrades quality, and evaluation itself is expensive. Therefore the two problems are entangled, not separable.

I hear the argument. And in production environments — right now, in February 2026 — they’re absolutely right. Teams facing token budgets do make quality trade-offs. The metering mindset does create systematic quality degradation.

But I think the red team is conflating the current entanglement with a permanent one. The fact that cost and quality are entangled today is a feature of Stage 2, not a refutation of the stage model. During the dial-up era, bandwidth constraints and content quality were entangled too — you couldn’t deliver high-quality video because the pipes weren’t big enough. That entanglement resolved when the infrastructure caught up.

The question is whether AI’s cost/quality entanglement is more like bandwidth’s (resolvable with infrastructure) or more like something genuinely different. I’ll grant that it’s more complex than bandwidth — there’s no Shannon limit equivalent, and the “data exhaustion” point adds a constraint that bandwidth never faced. But “more complex” doesn’t mean “permanently entangled.”

The stronger version of my thesis: cost and quality are entangled during the metering phase. As cost pressure relaxes (and it will for many use cases, even if not all), the quality problem separates out and becomes its own discipline. The entanglement is real but phase-dependent, not structural.

What I’m Not Changing

The red team raises the “linearity problem” — that historical infrastructure transitions were messier than a four-stage progression implies, with loops and reversals and new scarcity dimensions emerging within abundance.

Fair point on the history. But I think this mistakes the model’s purpose. The four-stage arc was never meant as a predictive timeline. It’s a diagnostic tool. When someone asks “should I use AI for this?” — they’re in Stage 1. When they ask “how do I use AI well?” — they’re in Stage 2 or 3. The value isn’t in predicting that the transition will be smooth. It’s in helping people locate where they are.

Analogies don’t have to be perfect to be useful. A map with the wrong scale still helps you understand which direction to walk.

The four-stage arc isn’t a prediction. It’s a diagnostic. And diagnostic tools don’t need to be precise — they need to be directionally correct.

Similarly, the red team’s point about “multiple disciplines, not one” — listing AI FinOps, output governance, prompt engineering, AI compliance, and model operations as distinct emerging fields — is accurate but doesn’t undermine the thesis. I said “something like ‘output operations’” with a hedge that “we don’t have a clean name for it yet.” That five different disciplines are crystallizing instead of one doesn’t break the argument that permanent secondary disciplines are emerging. It makes the pattern stronger.

What I’d Add to a Revision

If I were rewriting The Metering Phase with the red team’s challenges absorbed, three things would change:

A supply-side section. The original is almost entirely demand-side analysis — how we experience the metering phase. A revision would address the supply-side constraints that determine whether and when the metering phase resolves. Energy, data, architectural scaling curves. Not to invalidate the demand-side psychology, but to give it structural grounding.

Jevons acknowledgment. The line “we stopped counting because counting stopped mattering” implies that abundance resolves the metering mindset. A revision would note that abundance in one dimension often creates scarcity in adjacent dimensions, and the metering phase may shift rather than end.

Bifurcated outcomes. The original treats the four-stage arc as applying uniformly to “AI.” A revision would acknowledge that simple generation tasks may hit Stage 4 while complex reasoning tasks remain in Stage 2, creating a split landscape where abundance and metering coexist.

The Meta Observation

Here’s the thing that interests me most about this exercise, and it’s not about the metering phase at all.

I published a thesis. Within hours, I had a structured adversarial analysis that identified genuine weaknesses. I’m now writing a response that concedes some points, pushes back on others, and will probably generate its own counterarguments.

This whole cycle — original thesis → adversarial challenge → author response — happened in a single day. The intellectual infrastructure for stress-testing ideas just got dramatically cheaper and faster. That’s not a metering phase observation. That’s a practice observation.

I don’t think the red team replaced human critique. The best challenges in the analysis were structural observations that a domain expert would also catch — energy constraints, Jevons Paradox, entanglement. The analysis didn’t generate novel objections so much as systematically surface known ones I’d skipped.

But “systematically surfacing what you skipped” is enormously valuable. Most of my writing blind spots aren’t things I don’t know — they’re things I know but failed to address because I was moving too fast. Having an instant adversarial review that catches the structural gaps before a human reader does? That’s worth something.

The metering phase thesis is still directionally correct. The bandwidth analogy still holds as a psychological observation. The emerging-discipline prediction still stands.

But the thesis is weaker than I thought this morning, and I’m grateful for that. Better to discover the structural gaps on day one than to build a framework on foundations I haven’t pressure-tested.

The red team was right about the supply side. Right about Jevons. Partially right about entanglement. And completely right that the four-stage model is incomplete.

It’s also still useful. Those two things can coexist.