Nobody Asked the AI Team Why

Yesterday’s post defined wafer-scale computing — hardware that almost nobody runs yet. Today’s post is about the shift that happened along the way. There was a time when nobody asked the AI team why. Now AI is everything, and the why is the one question you cannot skip.

When the AI team was a specialized team

Seven years ago, you got a call from the AI team. They needed a couple more GPUs. Your next step was to go find them. If the request was really expensive, you would ask what they were doing. Most of the time you did not even do that. The team needed something, you found it, and that was the whole transaction.

The closest comparison was accounting. If the accounting team asked you for something, you did not ask them why. You did not know what they did, so the questions did not form. The AI team around 2020 had the same standing. It was not that people approved the spending on faith — it is that the review was one sentence: OK, you’re doing interesting things. And the team itself might not know how much it needed, because that was how new the space was. You can forecast a workload; you cannot forecast a frontier.

One hard rule held even then. You never just say yes. If the number was big, you asked a couple of questions — to be tough, not to be informed — and then you gave it to them anyway.

It was not the novelty. It was the specialization.

Here is the point of looking back. What made the AI team different in 2020 was not that AI was new. It was that AI was specialized. Specialized teams run on trust, and trust is affordable when the footprint is small: one team, a few machines, a budget line nobody outside the team could interpret anyway. You are not handing out billions of dollars of capacity to those teams. If they say they need it, you ask a couple of questions and you give it to them.

Specialization contained the spend. That is the whole reason the arrangement could last as long as it did. The team stayed small, the footprint stayed contained, and the trust stayed cheap.

Now AI is everything

That containment is gone. Every department, every team, every company is spending AI dollars. McKinsey’s 2026 global survey found that nearly nine in ten respondents report regular use of AI in at least one business function, and that 44 percent report AI scaling across their enterprise, up from 38 percent a year earlier.¹ Twenty-eight percent say AI already takes more than 10 percent of their enterprise-wide technology budget.²

In that world, asking why is met with laughter. Everybody is doing it. But you still have to ask why — what is the outcome you are after, and how do you put a value on that outcome. Then come the questions that did not exist in 2020. Which models are they using? Hosted models or on-prem? Open weights or proprietary? What is the subscription impact? What is the usage demand — the API stuff?

None of those questions is answerable from a consumption dashboard. Utilization tells you the models are being used; it cannot tell you what they are for. That gap has a name in the book: the purpose question — what no consumption measure can say (chapter 8). The usage-demand question is a forecasting question, and the first rule of forecasting is to decide what you are forecasting (chapter 14).

The parts that are still magic science

One caveat keeps this from being a clean story. There might be certain portions of AI that are still magic science — areas where you do not have much input into what the team is doing, and the questions still do not form. Supporting those areas on trust is not a failure of discipline. It is the same specialization bargain as 2020, and it still works while the footprint stays small.

The discipline is to never let those niche areas go. Keep watching them, keep the relationship, keep asking what they need this quarter — because there is a good chance they might graduate tomorrow.

Yesterday’s post is what a graduation looks like. Wafer-scale computing was the definition of a niche — and this year OpenAI signed a deal worth well over $10 billion for it. One day the questions did not form; the next day it is one of the largest hardware deals of the year and you are being asked to support low-latency inference for it.

That is the whole lesson. The 2020 arrangement never ended because AI got old. It ended because specialization stopped containing the spend. Keep the trust where the footprint is small. Never let the niches go — there is a good chance they might graduate tomorrow.

Sources

---

1. McKinsey & Company, “The state of AI in 2026: On the road to ROI,” Global Survey, August 2026: The state of AI: Global Survey 2026 — nearly nine in ten respondents report regular use of AI in at least one business function; 44 percent report AI scaling across the enterprise, up from 38 percent a year earlier; 56 percent report AI in three or more functions.↩︎

2. McKinsey & Company, “The state of AI in 2026: On the road to ROI,” Global Survey, August 2026: The state of AI: Global Survey 2026 — 28 percent of respondents say their organizations are spending more than 10 percent of their total enterprise-wide budget for information and communication technology on AI technologies; 60 percent expect their organizations to increase AI investments over the next year.↩︎