Your objectives can still be inside the AI’s context window—and yet no longer meaningfully shape its answers.

That may be the most dangerous form of reasoning drift. The information hasn’t disappeared, but its influence has weakened. The AI continues responding with confidence while gradually slipping away from your priorities, assumptions, and decisions and back toward the generic average.

In Part 1, I named the problem: reasoning drift. In Part 2, I introduced Reasoning State Engineering—the architecture designed to recover the initiative’s essential reasoning and bring it back into focus.

But a fair skeptic might still ask:

“Is reasoning drift a real, measurable phenomenon—or just your experience talking?”

The underlying vulnerability is real. It has been measured. And the findings expose something every serious AI user needs to understand.

Let’s examine the research—and what it means for every long-running AI collaboration.

The Study

Researchers at Stanford, with collaborators at UC Berkeley and Samaya AI, ran a deceptively simple experiment, published as “Lost in the Middle: How Language Models Use Long Contexts” (Liu et al., 2023). They gave language models a long input context containing the one piece of information needed to answer a question — then slid that piece to different positions: near the beginning, buried in the middle, near the end. If models used their context evenly, position wouldn’t matter.

It mattered enormously. They ran it two ways — realistic multi-document question answering, and a clean synthetic key-value retrieval task where the model just has to find one value by its key. The experiments differed, and the results weren’t identical, but the same pattern held across both: where the relevant information sat materially changed whether the model used it.

The Finding: a U-shaped curve

Model accuracy traces a “U”. It’s highest when the needed information sits at the very beginning of the context (primacy) or the very end (recency) — and it sags, hard, when that information is stranded in the middle. The model has the fact. It’s right there in the window. It just doesn’t use it reliably.

How hard is the sag? In the multi-document QA task, GPT-3.5-Turbo scored about 75.8% with the answer at the start of the context and just 53.8% with it buried in the middle — its worst position, a swing of more than 20 points. And here’s the result that should stop you cold: in that same experiment, the middle position scored worse than giving the model no documents at all. Closed-book — no documents, answering from memory — it managed 56.1%. With the answer present but stranded in the middle, it fell to 53.8%. In that test, loading the right information in the wrong place did worse than loading nothing.

“Just use a bigger context window” — Nope

The obvious response to reasoning drift is that context windows are becoming enormous—100K tokens and beyond—so the problem will eventually solve itself.

The research challenges that assumption. The authors found that expanding a model’s context window did not improve how reliably it used the information inside it. Extended-context models still exhibited the same U-shaped performance pattern: strong use of information near the beginning and end, with weaker performance when the relevant information was buried in the middle.

Newer models will continue improving their long-context performance. But the architectural distinction remains:

This exposes a critical distinction: context capacity is how much information fits inside the context window; context utilization is how reliably that information shapes the answer.

Reasoning State Engineering is designed around that distinction. It does not merely give the AI more context. It deliberately recovers, confirms, and repositions the context that matters when the work demands it.

The answer to reasoning drift isn’t simply a bigger window.

It’s better architecture.

What this Means for Reasoning Drift

Here’s the reframe. We tend to explain drift as an eviction problem: the conversation grows, the window fills, the oldest turns fall out, the state is gone. That happens — but this research points to damage that can start long before anything falls out. Your objectives, weights, assumptions, and decisions don’t have to leave the context to lose their grip. They just have to sink toward the middle.

So reasoning drift can begin as a positional problem — while your foundation is still present, and long before the window is full. It can also come from other sources: truncation, retrieval misses, conflicting instructions, accumulated ambiguity, weak source grounding. But position alone is enough to start it.

Why the Research Supports YODA Reasoning State Engineering

The research helps explain why YODA’s reasoning recovery architecture matters. When you fire the Reasoning State Reset Prompt, the AI is directed to re-read the initiative’s durable artifacts, reapply their essential reasoning, and confirm what it recovered. That single action does three things:

• Recovers the objectives, weights, assumptions, and decisions that should govern the work.

• Verifies that the AI is reasoning from the correct initiative state before it proceeds.

• Repositions that essential reasoning near the end of the active context, where the experiments found that models tend to use relevant information more reliably.

That means the confirm-back is more than a reassuring summary. It is an observable control mechanism. You confirm that the initiative’s Reasoning State was recovered by deliberately bringing its most important elements back into focus—right where the model is more likely to apply them to the next request.

That leads to the most practical lesson in this article:

Re-anchor close to the ask.

If you re-anchor and then continue exploring for an extended period, the additional conversation can once again separate the recovered state from the request that matters. Fire the Reasoning State Reset Prompt immediately before an important analysis, recommendation, or decision, and the AI begins that work with your objectives, weights, assumptions, and decisions freshly recovered and confirmed.

There is no fixed limit on how often you can re-anchor. Use the reset prompt whenever the answers begin drifting toward the generic, after adding important knowledge, at major initiative milestones—or whenever you simply want to verify that the AI is still reasoning from your context.

Don’t wait for the Reasoning State to disappear. Re-anchor before the decision matters.

What the Research Establishes—and What It Means

Let’s be precise, because precision is central to the YODA methodology. The Stanford-led study measured how language models retrieve information from long input contexts; it did not directly test a forty-turn business collaboration or the YODA Reasoning State Engineering architecture.

But that boundary does not weaken the implication. Across two different tasks, the researchers demonstrated the same underlying vulnerability: information can remain inside the model’s available context while its position materially reduces how reliably the model uses it. A long conversation is also processed as an expanding context. As objectives, assumptions, priorities, and decisions are pushed away from the beginning and buried beneath subsequent exchanges, their influence can weaken—even before they disappear from the context window.

While the study did not test YODA’s solution, it did establish the architectural problem YODA is designed to solve.

That makes Reasoning State Engineering more than a workaround for limited context windows. It is a deliberate response to a deeper limitation: information does not need to be forgotten to stop shaping the answer.

The Reasoning State Engineering Takeaway

Reasoning State Engineering rejects a dangerous assumption: that information will continue shaping the AI’s answers simply because it still fits inside the context window.

Presence is not influence.

Your objectives, weights, assumptions, and decisions can remain inside the conversation while gradually losing their grip on the AI’s reasoning. That is why Reasoning State Engineering separates the initiative’s durable state from the fading conversation—and gives you a repeatable way to bring that state back into focus when it matters most.

• The initiative’s artifacts preserve the state.

• The Reasoning State File supplies the recovery protocol.

• The Reset Prompt retrieves, reapplies, and confirms the essential reasoning close to the next important request.

The research exposed the architectural vulnerability YODA Reasoning State Engineering addresses: the AI doesn't have to forget your reasoning to stop using it.

Blog Part 1 named the problem (link found in the comments)
Blog Part 2 activated the architecture (link found in the comments)
Now, Blog Part 3 revealed why the architecture matters.

Your Reasoning State wasn’t necessarily gone. It was lost in the middle.

Now you know how to bring it back with YODA’s Reasoning State Engineering.

Source: Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni & Liang, “Lost in the Middle: How Language Models Use Long Contexts,” Stanford / UC Berkeley / Samaya AI, 2023 (arXiv:2307.03172).

Part 1 exposed a dangerous weakness in long-running AI collaborations – their reasoning fades. As the collaboration grows, the AI tool gradually forgets the objectives, priorities, assumptions, decisions, and definitions established at the beginning of the conversation. What began as reasoning from your specific and unique situation slowly becomes reasoning from the generic average, putting the relevance, responsibility, and accuracy of every subsequent recommendation at risk.

That leaves us with the critical question Part 1 raised: if the AI tool’s memory fades as the conversation progresses, how do we recover it? The answer lies at the heart of Reasoning State Engineering. It separates the AI collaboration into three components: durable state, standing recovery instructions, and a simple trigger that reconnects them. Once those three components are visible, activating the missing layer becomes surprisingly straightforward.

The Reasoning Recovery Architecture

Reasoning State Engineering blends four components, each doing exactly one specific job.

1. Establish standing behavior. Your project instructions / Activation Prompt define YODA’s identity, reasoning disciplines, and quality guardrails. This is who the AI is and how it thinks — set once, always on.

2. Build the durable initiative state. As you work through the YODA steps, each step produces initiative-specific content: intent and context; objectives and desired outcomes with associated KPI weights; trusted market and industry knowledge; internal domain expertise; world-leading expert knowledge and experience; and stakeholder desired outcomes, key decisions, and associated KPIs. These artifacts live in the Chat Knowledgebase, not the Project Knowledgebase. Together, they form the durable foundation of the initiative’s Reasoning State—the accumulated record of where the collaboration currently stands. The user does not need to create and maintain a separate Reasoning State Snapshot; the Reasoning State File and Re-Anchor Prompt direct the AI to recover, reapply, and confirm the state from these accumulated artifacts.

3. Install the recovery protocol. Upload the YODA Reasoning State File once, at the project level. Here is the part most people miss: the Reasoning State File is not another knowledge document. It is the control mechanism — the standing protocol that tells the AI how to recover those step artifacts and reapply them. The state lives in the artifacts; the File is the instruction for reconstructing it.

4. Trigger and verify recovery. Use the Re-Anchor Prompt whenever the AI begins drifting—or whenever you damn well want to confirm that it is still reasoning from the right foundation. The prompt triggers the recovery protocol; the confirm-back verifies that it worked. Trigger, check, and continue.

That’s the breakthrough: a durable state you build, a standing protocol that knows how to recover it, and a one-line trigger that reconnects them on demand.

Activate Reasoning State Engineering

The architecture may sound sophisticated, but activating it requires only a disciplined five-step routine.

1. Install the recovery protocol once.
   Place the YODA Reasoning State File in the Project Knowledgebase, where it can support every initiative chat governed by the project.

2. Build the initiative’s durable state.
   As you work through YODA, add each step’s confirmed artifacts to that initiative’s Chat Knowledgebase. Together, these objectives, weights, trusted sources, assumptions, expert guidance, and decisions comprise the initiative’s evolving Reasoning State.

3. Trigger recovery with one prompt:
   “Re-read the YODA Reasoning State File and re-apply it to this initiative. Then confirm the objectives, weights, assumptions, and key decisions you’re working from.”

4. Verify the confirmation.
   Review what the AI hands back. Correct anything missing, stale, or inconsistent before continuing. The confirm-back makes recovery observable: if the AI cannot accurately restate the initiative’s foundation, it has not fully recovered the Reasoning State.

5. Re-anchor when needed.
   Use the prompt after adding important knowledge, at major YODA milestones, when returning to the work after a pause, or whenever the answers begin drifting toward the generic. Re-anchoring frequently does no harm.

That is the complete routine: install once, build the initiative state, trigger recovery, confirm the foundation, and continue.

How the Architecture Works

The directions above explain what you do. Here is why they work: the architecture—not the conversation—preserves and recovers the reasoning.

• The state is separate from the conversation. Your objectives, weights, assumptions, decisions, and definitions live in the accumulated artifacts within the Chat Knowledgebase, not only in the fading conversation. The conversation remains temporary—and that is fine. Conversations are for exploration. The durable initiative record resides in the artifacts, while the Reasoning State File provides the standard recovery method that the collaboration can invoke whenever needed.

• The File recovers; it never rewrites. The Re-Anchor Prompt does not change the Reasoning State File or the underlying initiative artifacts. It tells the AI to re-read the File, apply its recovery protocol, and reconstruct the current Reasoning State from the accumulated artifacts—pulling the AI’s attention back to the foundation you established.

• The confirm-back makes recovery observable. This is the safeguard. If the AI cannot accurately state the objectives, weights, assumptions, and key decisions it is working from, the Reasoning State has not actually been recovered—and you have caught the problem before building on a fake recovery.

• Why it should feel familiar. This is how serious software systems preserve long-running work. Operating systems do not trust temporary memory to hold everything forever. They preserve important state in durable storage and recover it when needed. Reasoning State Engineering applies the same principle to human–AI collaboration.

The conversation explores. The artifacts preserve. The Reasoning State File recovers.

The Takeaway

Part 1 exposed the problem: as an AI collaboration grows, the reasoning that made it uniquely yours can quietly slip out of view. The AI keeps answering—but it gradually stops reasoning from your objectives, priorities, assumptions, and decisions.

Part 2 provides the solution: stop asking the conversation to remember what the architecture should preserve.

Reasoning State Engineering separates the collaboration into durable initiative artifacts—captured in the Chat Knowledgebase through the YODA process—a standing recovery protocol, and a simple trigger that reconnects them. When the AI begins drifting—or whenever you damn well want to check—you re-anchor, confirm, and continue.

The objective is to create a focused, sustainable collaboration that can always recover the reasoning behind the work.

The conversation explores.
The artifacts preserve.
The Reasoning State File recovers.

That is the missing layer. And it changes the promise of an AI Thinking Partner from:

“I hope it remembers where we are.”

to:

“When the reasoning drifts, I know how to bring it back.”

With YODA—Your Own Digital Assistant, we’ve shown how Context Engineering can transform a generic AI tool into an AI Thinking Partner—one grounded in your objectives, trusted knowledge, expert perspectives, and reasoning disciplines, rather than internet averages.

However, solving the AI Thinking Partner challenge revealed a subtle issue. Over time, our AI Thinking Partners gradually lost focus. Initially, they reasoned clearly from our objectives, priorities, and assumptions, but as conversations continued, their advice became more generic. This shift was subtle but confident.

The AI is gradually losing its reasoning state that anchors the collaboration—the objectives, priorities, assumptions, decisions, and definitions that make the recommendations uniquely yours. Once that reasoning state slips, the AI falls back to reasoning from the generic average.

This realization brought us to the next essential architectural layer in human-AI collaboration: Reasoning State Engineering. While Context Engineering focuses on teaching AI what it needs to know, Reasoning State Engineering maintains the ongoing collaborative reasoning state.

The Wrong Diagnosis

When an AI assistant gives generic answers, we assume it’s forgetting, but it’s actually losing the reasoning context of your collaboration—your goals, priorities, assumptions, decisions, and definitions. As the conversation grows, this reasoning state defaults to generic reasoning.

YODA’s goal is to preserve that reasoning state so that the AI can retrieve it whenever necessary—whether at the start of a new conversation, during an ongoing one, or weeks later when the task continues.

Where AI Gets Its Context

An AI Thinking Partner does not reason from one undifferentiated source of context. It operates across four distinct architectural layers, each with a different purpose and level of durability.

1. Identity — Who the AI is. The project-level Activation Prompt establishes AI’s role, behaviors, reasoning disciplines, and guardrails. This layer is durable at the project level and remains available across all subsequent chat conversations.

2. Knowledge — What the AI knows. Uploaded credible research, proven frameworks, organizational domain knowledge, and other trusted sources provide the evidence from which the AI reasons. You load them into the initiative’s growing contextual knowledgebase, and because they’re uploaded files, they persist — the AI can draw on them throughout the initiative, unlike the reasoning that lives only in the conversation.

3. Reasoning State — Where you are in the work. This is the initiative’s accumulated thinking: the objectives you established, the priorities you weighted, the assumptions you confirmed, the decisions you made, and the definitions you agreed upon. It develops inside a conversation (chat) but is not durably protected. This is the architectural layer that slips.

4. Current Prompt — What you just said. The latest message supplies the immediate direction for the next response. It is useful but fragile, temporary, and limited to the current conversation.

The top two layers remain durable — Identity as the project’s standing instructions, and Knowledge as the uploaded files you add to the initiative — so the AI can return to them at will. The bottom two live in the conversation itself. The Current Prompt is fleeting but always visible; the Reasoning State — the accumulating objectives, priorities, assumptions, decisions, and definitions — has no durable home by default, and as the conversation grows it gradually slips out of the AI’s working context. That unprotected Reasoning State, caught between the durable inputs and the fleeting prompt, is the gap Reasoning State Engineering fills — by making the initiative’s evolving state durable and recoverable.

The Cost of Reasoning Drift

Losing the reasoning state gradually erodes the quality of the AI’s recommendations in three important ways.

• Relevance is the first casualty. The whole purpose of providing context is to move the AI away from the generic average and toward your unique situation. As the reasoning state slips, the AI quietly slides back toward generic advice.

• Responsibility fades next. Your guardrails—“flag uncertainty,” “consider stakeholder impacts,” “don’t present a draft as a final answer”—are part of the reasoning state too. When they fall outside the AI’s working context, they no longer influence its recommendations.

• Accuracy is dangerous because it fails silently. Early in the conversation, AI reasons from your data, assumptions, and constraints. Later, when context slips, it answers using plausible general knowledge, often with the same confidence. Collaboration can shift from evidence-based reasoning to confident generalization unnoticed.

That’s the real cost of reasoning drift. AI simply stops reasoning from your context and gradually returns to reasoning from everyone else’s.

So what do we do about it? The fix follows directly from the diagnosis: stop treating your reasoning as something that lives inside the conversation, and start treating it as durable architecture that lives outside it. That is Reasoning State Engineering — a standing layer that captures the initiative’s confirmed foundation so the AI can re-engage it whenever the conversation drifts, instead of quietly sliding back to the generic average. How you actually activate it — the file, the one prompt, the routine — is the subject of Part 2.

What It Does—and What It Doesn’t

Reasoning State Engineering doesn’t eliminate the limitations of today’s language models, nor does it guarantee correct reasoning. What it does is materially reduce reasoning drift by improving the three qualities that long-running AI collaborations depend on:

• Continuity — the thread of the initiative survives.

• Consistency — your objectives, priorities, and guardrails guide the AI’s reasoning.

• Recoverability — the reasoning state can be restored whenever a conversation drifts.

The AI can be wrong, and its reasoning depends on validated information. Humans stay involved to challenge assumptions, evaluate evidence, and decide. Reasoning State Engineering maintains the shared context. The human-AI advantage isn’t an AI that thinks for you, but one that helps keep both reasoning from the same current picture.

The Next Layer of Human-AI Collaboration

Every major advance in working with GenAI has added a new architectural layer.

• Prompt Engineering taught AI what to do.

• Context Engineering taught AI what it should know.

• Reasoning State Engineering preserves the collaborative reasoning currently in place.

These are three complementary layers that build progressively stronger AI Thinking Partners. Prompt Engineering shapes behavior. Context Engineering grounds the AI in what matters. Reasoning State Engineering preserves the continuity of that reasoning.

That’s why the next frontier isn’t simply giving AI more data. It’s preserving the reasoning that transforms information into better decisions. Then something interesting happens.

“The more I use it, the smarter it gets” actually becomes true—not because the AI is learning on its own, but because you’ve engineered an architecture that continuously preserves, recovers, and refines the reasoning behind your work.

That’s YODA—an AI Thinking Partner that helps you ask better questions, preserve better reasoning, and ultimately make better decisions.

Note: My YODA subscribers will soon have the opportunity to stress-test the Reasoning State Engineering architecture in practice. Part 2 will show exactly how to activate it — so watch this space.

I ended Part 2 with a promise: that we’d stop admiring the idea and go build it. That we’d take a sentence like “safer matters more than faster” and turn it into something a machine can actually optimize. So, let’s roll up our sleeves. This is the part where the philosophy has to become arithmetic — or it was never real to begin with.

And I want to start with the most over-nodded-at sentence in all of AI.

“AI Should Reflect Human Values.” Great — Tell Me the Weights.

You’ve heard it in every keynote, every op-ed, every panel: AI should reflect human values. Everyone nods. Nobody moves. Because, as is usually said, it’s a wish, not an instruction — there is nothing in it that a machine could do one bit differently tomorrow morning.

Here’s the move I’ve learned to make. When someone tells me AI should reflect human values, I answer with four words:

Great — tell me the weights…and only you can set them.

That’s the whole job. To reflect a value is to decide how much it counts against every other value in the room. Safety matters — of course it does. More than travel time? By how much? Two-to-one? Six-to-one? The instant you answer that, something quiet and profound happens: a value stops being a bumper sticker and becomes a decision a machine can carry. That’s what it means to make a value computable. Not to reduce it to a number — to commit to it enough to weigh it.

A value you can’t weigh against another isn’t actionable yet. It’s an aspiration, not an objective.

Let’s Author One — Highway 30, in Numbers

Back to the road. In Part 2, my wife and I chose Highway 30 over I-80 because, for us, the calmer, safer road was worth more than the saved minutes. I put rough numbers on it: safety 60%, travel time 30%, scenery 10%. Those three weights are the objective. Everything the machine does after that is just faithful execution.

Here’s how they actually go to work. A route isn’t scored on one thing; it’s scored on a handful of features at once — road conditions, route safety, scenic value, miles, travel time. The engine takes each candidate route, measures it against every feature, multiplies each measure by the weight I assigned to it, and sums the results. Plainly:

A route’s score = (how safe it is × 0.60) + (how fast it is × 0.30) + (how scenic it is × 0.10).

That’s the whole trick. Highway 30 doesn’t win because it’s shorter — it isn’t. It wins because once safety carries 60% of the vote, its calmer miles outscore I-80’s faster ones. Same map. Same live traffic. Same causal engine from Part 1. The only new thing in the system is a set of weights that didn’t exist until I decided them. I didn’t make the machine smarter. I told it what smart was for.

The weights are the objective. Everything downstream is the machine doing exactly what you told it — with a rigor you’ll never match, toward a finish line you drew.

Now Watch What the Weights Do

Here’s where it gets fun, and where the whole idea earns its keep. Leave the map alone. Leave the engine alone. Just change the weights — and watch the recommendation move.

Picture three drivers on the exact same Des Moines–to–Cedar Rapids stretch. The one in a hurry sets travel time to 0.80 and road conditions to 0.20 — and the engine hands back I-80. Fastest. The one carrying precious cargo puts route safety at 0.50 and spreads the rest across road conditions, a little scenery, and miles — and the engine reroutes to Highway 30. Safest. The one on a Saturday with nowhere to be loads scenery and historic sites up to 0.80 — and the engine threads a route past the Vermeer windmill in Pella and the Amana Colonies. Most interesting.

Three drivers. One map. One engine. Three different “best” routes — because “best” was never a property of the road. It was a property of the person, made legible to the machine as a column of numbers, each adding up to 1.0. That’s the picture I want you to hold: not a smarter GPS, but the same GPS, willing to call three different routes “best” depending on whose values it was handed.

No — We Did Not Just Solve Human Values With Math

Making values computable does not mean we reduced human values to arithmetic, or that a spreadsheet now knows what a good life looks like. What we did is make your values actionable by forcing the one question that abstraction always skips — how much does this count, next to that?

The machine can then compute against your answer with total rigor. But it never gets to choose the answer. Weighting is a human act. The math just makes the human act operative.

Keep the line clean, and you keep the whole series honest: the machine computes the objective; it never picks it. Blur that line, and you’ve handed your steering wheel to something that was only ever built to hold it, not to decide where to point it.

Making values computable doesn’t hand your values to the machine. It hands the machine your decision — and keeps the deciding yours.

Where Values Meet Mathematics

In The AI-Human Edge, I described the AI Utility Function as the place where values meet mathematics. Everything we’ve done in this article is simply showing how that happens: by making your values computable. Your values become explicit weights, and those weights become the objective the machine optimizes.

And because it’s numbers now, it’s tunable — which is the quiet gift most people miss. When your priorities shift, you don’t retrain a model or argue with an app. You re-weight. Kids in the back seat tonight? Nudge safety up. Ice in the forecast? Up again. Running late to a flight you cannot miss? Let travel time take the wheel for one trip. The objective isn’t carved in stone the day you write it. It’s a living document you author, and re-author, as often as your life changes — which is to say, constantly.

You don’t retrain the machine when your priorities change. You re-weight. The objective is yours to revise, any day you like.

The Weights Are Yours. The Engine Beneath Them Is the Next Climb.

Step back and look at the whole road we’ve traveled. Part 1 was everything the machine can do — the climb from prediction to intervention to causal reasoning. Part 2 was the one thing only you can do — name the objective. And Part 3 is what naming it actually requires: making it computable, authoring the weights, tuning them as you go.

That completes the human’s half of the bargain. What’s left is the machine’s half — and it’s the real frontier. Giving a system a causal backbone deep enough to reason from your authored weights to the exact right move, and to show its work while it does, is a genuinely harder thing than optimizing a weighted sum. It’s the fourth rung — what Mark Stouse calls Causal Engineering — and it’s most of what I’m building toward now. But that’s a bigger canvas than a blog can hold, and I’m saving it for one that can do it justice.

For now, sit with the part that’s already in your hands. The machine will carry any objective you give it with a rigor no human can match. It will never once care which objective it is. That’s not the machine’s failure — it’s the space deliberately left for you. Making your values computable is simply how you fill it: how you turn “this is what I care about, and this is how much” into a future a machine can help you actually reach.

The smartest engine in the world is still waiting on the one input only you can author. Now you know how to write it down.

At the end of Part 1, I left us on a road. Every rung we climbed — prediction, intervention, counterfactual, even Causal Engineering — optimized brilliantly toward an objective. But in every case, the machine was handed that objective. The GPS found the fastest route because someone long ago decided that a GPS optimizes for speed. It never asked whether fast was what you wanted.

You’ll remember the drive I ended Part 1 with: from Des Moines to Cedar Rapids, my wife and I take Highway 30, not the faster I-80 (Option B). The calmer road, lighter on semis, feels safer — and that’s worth more to us than the minutes we’d save. “Safer matters more than faster” was never in the road data. It’s in us.

That was the gap. This is where we close it.

Part 1 was everything the machine can do. Part 2 is the one thing only you can do — tell it what “better” means.

The Machine Was Never Going to Guess Your Objective

Here’s the uncomfortable thing about every impressive AI demo you’ve ever seen: the hard part was already done before the model started. Someone chose the objective. Someone decided that “success” meant the fastest route, the highest click-through, the lowest cost, and the shortest handle time. The model just optimized toward a finish line it didn’t draw.

That’s fine when the objective is obvious. It is quietly disastrous when it isn’t — because a system optimizing the wrong objective doesn’t look broken. It looks excellent. The GPS confidently routing you onto I-80 isn’t malfunctioning. It’s doing exactly what it was told, flawlessly, toward a goal that was never yours.

Causal AI raises the stakes here, not lowers them. A predictive model that optimizes the wrong thing gives you a wrong number. A causal model that optimizes the wrong thing recommends a wrong action — and then, because it learns, gets better and better at pursuing it. The sharper the tool, the more it matters who’s aiming it.

A predictive model optimizing the wrong objective gives you a wrong answer. A causal model optimizing the wrong objective takes you confidently to the wrong place.

Value Was Never a Single Number

So what do you tell it? Here’s where most people go wrong: they hand the machine one metric and call it the goal. Fastest. Cheapest. Highest. But look at the Highway 30 decision honestly, and you’ll see that’s not how humans actually value anything.

When my wife and I choose that road, we’re not optimizing one thing. We’re weighing several at once. Something like: safety matters most — call it 60% of the decision. Travel time still counts, but less — maybe 30%. And the scenery, the calm of it, the river valley — that’s real too, call it the last 10%. Those weights are the actual objective. “Fastest route” is just what you get when you set safety and scenery to zero and travel time to 100% — a choice nobody consciously made, but one the GPS made for us by default.

This is the point I’ve spent years making in a different context: value is never one metric. It’s a weighted, multi-dimensional thing — a portfolio of what matters to you, in proportions only you can set. The machine can carry any objective you give it with superhuman rigor. It cannot tell you what belongs in the portfolio, or how much each piece is worth. That’s not a data problem. It’s a values problem, and values live on your side of the windshield.

That weighted portfolio has a name. In Part 1, I called it your AI Utility Function — the deliberate, weighted definition of what “better” means, across every dimension you care about, not just the one that’s easiest to measure. The GPS already runs on one; it just came preset with travel time set to 100%. The moment you rewrite those weights, you’ve authored your own.

Fastest isn’t the objective. It’s what’s left over when you forget to name everything else you care about.

Teaching the GPS What “Better” Means to You

This is the first move of what I call YODA — Context Engineering — and it’s deceptively simple to state: before the machine optimizes anything, the human authors the objective. Not the wording of a prompt — the AI Utility Function itself: intent, context, and how much each thing counts, in all its weighted, multi-dimensional messiness.

Watch what happens on our drive when you actually do it. The moment I tell the system safety is worth more to me than speed — that I’ll trade minutes for fewer semis and a calmer road — Highway 30 stops being my stubborn quirk and becomes the recommendation. Same map, same traffic data, same causal engine. One thing changed: it finally knows what I’m optimizing for. The rerouting genius from Part 1 is now pointed at my finish line instead of the factory default.

Nothing about the machine got smarter. The intelligence that was missing was never the model’s to supply. It was mine to declare — and until I did, even the most sophisticated causal engine was just a very rigorous way of pursuing someone else’s goal.

Now the recommendation isn’t “the fastest route.” It’s “the best route, given your intent and context.” That’s the sentence where YODA and causal reasoning finally meet.

When the Tool Becomes a Thinking Partner

And then something happens that’s easy to miss but is, to me, the whole point.

Once the system understands that the calm of the drive genuinely matters to me — not as a footnote but as a real weight in the objective — it stops merely answering my questions and starts anticipating them. “There’s an overlook of the Cedar River a few miles ahead — about five minutes off your route. Want it?” I never asked for that. It surfaced an option I didn’t request, because it finally understood what I value well enough to see something I’d have wanted and missed.

That’s the moment the tool crosses over into thinking partner. Not when it answers faster — when it understands your intent well enough to bring you something you didn’t know to ask for, and then leaves the choice to you. Machine recommends; human decides. The recommendation got richer precisely because the human did the one thing the machine couldn’t: said what actually mattered.

A tool answers the question you asked. A thinking partner understands you well enough to raise the one you didn’t.

Which Is Why My Own Frameworks Have to Grow a Causal Backbone

I’ll be as honest here as I was at the end of Part 1. Everything I teach has been climbing toward this and hasn’t fully arrived.

TLADS — thinking like a data scientist — taught domain experts to frame value: what outcome matters, who it serves, what “good” looks like. That’s the muscle you just watched me use to set those weights on Highway 30. YODA taught them to engineer context so a generic model becomes a grounded thinking partner. Both are necessary. But framing an objective and encoding context only pay off fully when there’s a causal engine underneath that can reason from your intent to the right action — weigh the interventions, recommend the move, and show its work. Giving TLADS and YODA that causal backbone is the next stretch of the climb, and it’s most of what I’m building toward now. Behind the scenes, it means giving the machine entity-level causal models that reason from your intent to the specific move — but that’s a bigger story than this series can hold, and one I’m saving for a canvas that can do it justice.

The Future Worth Engineering Is Yours to Name

So here’s where Part 2 leaves us.

The machine can climb every rung. It can predict what happens, reason about what to do, imagine what would have happened otherwise, even help design the system of causes that produces the future you’re after. That’s an extraordinary amount of intelligence — and none of it decides which future is worth having. That decision was always yours. It’s the one thing the climb never reaches on its own.

You’ve now done the human’s real work: you named the objective. You said that on the road from Des Moines to Cedar Rapids, safety outweighs speed — and you watched the same engine change its recommendation because of it. That is authoring your AI Utility Function.

But naming the objective raises the next question — how do you actually build it? How do you turn “safer matters more than faster” into weights a machine can optimize, then tune them as your priorities shift? That’s Part 3, where we roll up our sleeves and author a utility function together, one weight at a time.

The smartest causal model in the world is still waiting on the one input only you can give it: the future you’ve decided is worth engineering.

More soon.

In a recent keynote, I outlined my perspective on the future of AI. I suggested that we are preparing to transition from Phase 1: Prompt Engineering to Phase 2: Context Engineering in AI mastery. The first phase focused on formulating better questions to improve responses. The second phase, however, involves architecting the reasoning setup around the model, transforming AI from a simple chatbot into a Thinking Partner that understands your goals, experience, and relevant industry knowledge about the specific problem at hand.

That is the focus of my latest book, YODA: The Context Engineering Playbook—a blueprint for transforming domain experts into AI Solution Engineers by architecting AI thinking partners that continuously tap into, integrate, and amplify the collective knowledge, experience, judgment, and best practices of industry experts.

Prompt Engineering teaches AI to answer better questions. Context Engineering teaches AI to think with your organization’s accumulated expertise.

Even the best AI Thinking Partner mainly predicts, synthesizing knowledge and estimating likely outcomes. However, prediction alone can’t explain why events happen, distinguish causes from correlations, identify factors driving outcomes, or determine effective actions. Without understanding cause and effect, it can’t confidently recommend what to do or predict what will happen. Achieving this requires causal reasoning, which needs more than an advanced AI model—it demands a human-centered approach to intelligence.

An AI Thinking Partner that understands your objectives, your knowledge, and your context is a remarkable achievement. But understanding why outcomes occur—and how your actions change those outcomes—is a fundamentally different kind of intelligence. Causal AI reasons about interventions—“What happens if I act?”—and counterfactuals—“What would have happened if I’d acted differently?” It doesn’t simply forecast the future; it uncovers the mechanisms that create it, then explores alternative futures so you can choose the one worth creating.

Prediction Is Not the End Point

Here’s the analogy that made the distinction click for me, thanks to my friend Mark Stouse: think about a GPS.

A paper map is decided once, before you leave—optimal for the world as it was, but blind to the world you actually encounter. A GPS that estimates your arrival time is better; it’s continuously predicting congestion, delays, and travel speed. But notice: it’s still optimizing the route you already chose. The thing that makes GPS genuinely valuable isn’t the prediction. It’s the moment it says, “Take Exit 23,” reroutes you around an accident, and changes the outcome.

That’s the distinction the whole industry keeps blurring. Prediction is the input. Understanding the mechanism connects prediction to action. Intervention is where the value is created. A GPS that only forecasts traffic and never changes your route would be a very sophisticated way to stay stuck. Likewise, a predictive AI model can estimate the future with remarkable accuracy—but without understanding why the delay exists, it cannot know which lever to pull to change the outcome.

There’s a tiny piece of notation that captures this entire shift.

• Prediction asks: P(Y | X) — What’s likely to happen, given what I observe?

• Causal AI asks: P(Y | do(X)) — What happens if I deliberately change something?

One tiny word—do—changes everything. It transforms AI from observing the world to intentionally changing it. That is the exact moment AI stops forecasting the future and starts helping us shape it.

Prompt Engineering improves answers. Context Engineering improved thinking. Causal AI improves decisions.

Pearl’s Ladder — and the Rung Above It

If you’ve spent time studying causal inference, you’ve undoubtedly encountered Judea Pearl’s Ladder of Causation—one of the most influential frameworks in modern AI. Pearl describes three progressively more powerful forms of reasoning:

• Rung 1: Association — What’s correlated?

• Rung 2: Intervention — What happens if I act?

• Rung 3: Counterfactual — What would have happened if I’d acted differently?

The GPS analogy brings all three rungs to life in a single journey. Prediction: traffic ahead. Intervention: take Exit 23. Counterfactual: if you’d exited three miles back, you’d already be home. Interventions ask, “What should I do?” Counterfactuals ask, “What should I have done?” One reasons forward into an alternate future; the other reasons backward into an alternate past. Together, they transform AI from forecasting outcomes into learning which actions create better outcomes.

But what if there’s a fourth rung? Mark calls it Causal Engineering. The first three rungs are fundamentally about understanding the world—explaining what happened, estimating the effects of an intervention, or imagining what could have happened instead. Causal Engineering changes the question entirely. It doesn’t ask: “What caused this?”

It asks: “What future do we want—and what system of causes would create it?”

That’s no longer analysis. That’s design. And it rests on a distinction that most AI conversations skip right past:

Predictive AI models relationships among variables. Causal AI models the mechanisms that connect them.

A traditional correlation matrix might suggest that education has little influence on loan repayment. But model the underlying mechanism—education increases income, income lowers debt-to-income ratio, and lower debt-to-income improves repayment—and income suddenly becomes enormously important, just not directly. The coefficient barely changed. The understanding changed completely.

You cannot engineer a future you don’t understand. That’s why the causal pathway—not the coefficient—is the intelligence asset worth building.

Interventions are one alternate future. Counterfactuals are one alternate past. Causal reasoning is simply the ability to think in realities other than the one in front of you.

The Intelligence Asset That Appreciates

There’s one more property of a good GPS that matters even more than any single reroute: it learns. Every trip, every driver, every road closure feeds back into the map, so tomorrow’s route is better than today’s. A GPS that never remembered an accident would quickly become obsolete. That’s the quiet limitation of static predictive models. The world changes. The model doesn’t. Its value slowly depreciates.

A causal model works differently. It follows the same learning loop as your best people: decide, observe the outcome, understand why it happened, update your understanding of cause and effect, and make a better decision next time. Every intervention sharpens the next intervention. Instead of becoming stale, the intelligence becomes more valuable through use.

That learning loop creates another advantage that is becoming increasingly important. A predictive model can tell you there is an 87% probability of an outcome. A causal model can explain why it made its recommendation: because A caused B, which influenced C, the recommended action is D. Every recommendation can be traced back through the causal pathway that produced it, making it explainable, auditable, and defensible.

I call that causal traceability. It’s why I increasingly describe Causal AI as Governance AI. In a world shaped by the EU AI Act, emerging governance standards, and growing expectations of fiduciary accountability, “the model said so” is no longer a sufficient explanation. Organizations need to understand not only what AI recommends, but why.

Which Means My Own Frameworks Have to Grow Up

Everything I’ve taught over the past decade points toward this phase — but doesn’t quite reach it. TLADS taught domain experts to think like data scientists: to start with value, define the outcome that matters, understand who it serves, and decide what “good” actually looks like. YODA taught them to engineer the context around a generic model, so it became a grounded thinking partner instead of a glorified autocomplete engine.

Both are necessary. Neither is sufficient because even a very good thinking partner still lives mostly in the world of prediction. It can tell you what’s likely, but it doesn’t reason about what to change.

That next step is a different discipline. It’s teaching AI to reason about cause and effect — to ask not merely “what is likely to happen?” but “what should we change?” and, eventually, “what future are we actually trying to create?” That’s the leap from prediction to Causal Engineering, and it means giving TLADS and YODA a causal backbone they don’t fully have yet: one that can tell drivers from correlations, weigh interventions instead of cataloging observations, and reason about the consequences of a choice before you commit to it. Building that backbone is most of what I’m working toward now — and it’s a longer story than one blog can hold.

One Question the Machine Can’t Answer for You

The AI Thinking Partner isn’t the summit. It’s the base camp. And notice the assumption hiding under every rung we’ve climbed: prediction, intervention, counterfactual, even Causal Engineering all optimize toward an objective — and in every case, the machine was handed that objective, not asked to choose it. Your GPS reroutes you brilliantly, but it quietly assumes one thing — that getting there faster is what you want.

When my wife and I drive from Des Moines to Cedar Rapids, we take Highway 30, not I-80. I-80 is faster. Highway 30 is newer, calmer, lighter on semis — we arrive a few minutes later and feel a good deal safer, and that’s worth more to us than the minutes we’d save. No GPS knows that, because “safer matters more than faster” was never in the road data. It’s in us.

That’s the boundary no machine can cross on its own: causal AI can reason about cause and effect with a rigor no human can match, but it cannot tell you which effects are worth causing. That’s not a flaw in the machine — it’s the one job that stays human. In my book The AI-Human Edge, I gave that job a name: the AI Utility Function — the deliberate, weighted definition of what “better” means, across every dimension of value you actually care about, not just the one that’s easiest to measure. The machine optimizes the utility function with a rigor no human can match. But someone has to write it, and that someone is you.

The real question was never how smart the machine can get. It’s whether it’s pursuing your goals — or someone else’s.

Which is exactly where we go next. In Part 2, we stop letting the GPS assume the objective and start authoring it ourselves — teaching the machine what “better” means when safety, time, and the calm of the drive each carry a different weight. That’s the moment it stops optimizing someone else’s objective and starts becoming your thinking partner.

More soon.

Re-releasing a blog that I wrote in 2021, which seems to have fallen off the internet but is as relevant as ever!

Share

Senior executives trained in accounting continue to struggle to understand how to determine the value of their data. The article “Why Your Company Doesn’t Measure The Value Of Its Data Assets” written by Doug Laney (by the way, why does the Forbes website absolutely bury the reader in ads?), contains a telling comment from a senior accounting firm partner:

“… balance sheets and income statements which form the backbone of today’s accounting system now fail to capture significant sources of value in our economy. He said that the measurements we use don’t reflect all the ways that companies create value in the New Economy, and this lack of transparency results in undue market volatility and mere “guesstimates” by investors in valuing companies. Even the chairman of the AICPA stated that the accounting model is out of date and based on the assumption of profitability depending upon physical assets—an accounting model for the Industrial Age, not the Information Age.”

This paragraph reflects how a traditional accounting mindset, in the age of digital assets, is focused on the wrong valuation method – trying to represent value using an artificially defined balance sheet that doesn’t capture how today’s companies are using data to create new sources of customer, product, and operational value.

Nothing says “We really don’t know how to quantify value in the digital age” better than Figure 1, where a significant percentage of the most valuable firms’ “value” is credited to nebulous intangible (non-physical) assets. And the discrepancy in the creation of value between traditional physical assets and intangible digital assets is growing exponentially.

Figure 1: Increasing Percentage of Most Valuable Firms defined by Intangible Assets

To properly reflect the value of their digital assets, executives must embrace an economics mindset where the value of an asset is determined by the use of that asset. This is critical given the unique economic characteristics of digital assets – they never wear out, never deplete, can be used across an unlimited number of use cases at zero marginal cost, and they can appreciate, not depreciate, in value the more that they are used (if properly engineered).

Mastering the Data Monetization Roadmap

I introduced the Data Monetization Roadmap in “Introducing the “4 Stages of Data Monetization” as a guide to help organizations in their data monetization journey. The roadmap emphasizes that the driver of data monetization is in the use or application of the data to create value. That is, the value of data isn’t in possession but in applying it to create new sources of customer, product, and operational value.

As organizations negotiate the Data Monetization Roadmap, they will encounter two critical inflection points:

• Inflection Point #1 is where organizations transition from data as a cost to be minimized to data as an economic asset to be monetized. I call this the “Prove and Expand Value” inflection point.

• Inflection Point #2 is where organizations master the economics of data and analytics by creating composable, reusable, and continuously learning and adapting digital assets that can scale the organization’s data monetization capabilities. I call this the “Scale Value” inflection point.

Inflection Point #1: Proving and Expanding Value

This pivot point is where the organization makes the transition from just capturing, storing, securing, and governing data to actually monetizing it. How do you get organizations to make that first pivot toward data monetization? How can one help the business stakeholders to connect to and envision where and how data and analytics can generate value (see Figure 2)?

Figure 2: Data Monetization Roadmap Inflection Point #1

Navigating Inflection Point #1 requires close collaboration with business stakeholders to identify, validate, value, and prioritize the business and operational use cases where data and analytics can create new sources of value. The Big Data Strategy Document in Figure 3 provides a framework for that collaborative engagement process.

Figure 3: Unleashing the Business Value of Technology

The Big Data Strategy Document decomposes an organization’s key business initiative into its supporting use cases, desired business outcomes, critical success factors against which progress and success will be measured, and key tasks or actions. The Big Data Strategy Document sets the stage for an envisioning exercise to help the business stakeholders brainstorm the areas of the business where data and analytics can drive meaningful and relevant business value. Yep, there is a lot of work that needs to be done before one ever puts science to the data.

So, now we’ve given the business stakeholders a taste of success in monetizing their data. Interest is building, and others across the organization are asking for help in monetizing their data. Now it gets really fun!

Inflection Point #2: Scaling Data Monetization Potential

The second inflection point occurs just as organizations are scaling their data and analytics success across the organization. More and more business units are coming to the data and analytics team for assistance with their top priority use cases. But remember:

“Organizations don’t fail due to a lack of use cases; they fail because they have too many.”

The volume of use case requests starts to overwhelm the limited data and analytics resources. And when business units can’t get support in a timely manner, they get frustrated and seek outside solutions. And as these organizations go elsewhere for their data and analytic needs, some fatal developments occur:

• Data Silos. These are data repositories that pop up outside the centralized data lake or data hub. And with the ease of procuring cloud capabilities (got a credit card, anyone?), it is easy for impatient business units to set up their own data environments.

• Shadow Data and Analytics Spend. The growing presence of software-as-a-service business solutions makes it easy for impatient business units to just buy their solution from someone else. Consequently, money that could be invested to expand the organization’s data and analytics capabilities is now being siphoned off by one-off, point solutions that satisfy an immediate business need, but create longer-term data and analytics debt.

• Orphaned Analytics. Orphaned Analytics are one-off Machine Learning (ML) models written to address a specific business or operational problem, but never engineered for sharing, re-use, and continuous refinement. The ability to support and enhance these one-off ML models decays quickly as the data scientists who built the models get reassigned to other projects, or just leave the company.

The result: instead of creating data and analytics assets that can be easily shared, reused, and continuously refined, the organization has created data and analytics debt that drives up maintenance and support costs, which quickly overwhelms the economic benefits of the data and analytics assets. Welcome to Inflection Point #2 (see Figure 4).

Figure 4: Data Monetization Roadmap Inflection Point #2

What can organizations do to avoid the collapse of the economic value of data and analytics that can occur at Inflection Point #2?

• Data Lake 3.0: Collaborative Value Creation Platform. Leading organizations are transitioning the data lake from a simple, cheaper (using the cloud) data repository to an agile, collaborative, holistic value creation platform that supports the sharing, reusing, and refinement of the organization’s valuable data and analytic assets (see Figure 5).

Figure 5: Data Lake 3.0: The Collaborative Value Creation Platform

Data Lake 3.0 employs intelligent catalogs to help the business units find the data they need for their use cases. The data lake also employs intelligent data pipelines to accelerate the ingestion of new data sources, and a multi-tiered data lake environment to support rapid data ingestion, transformation, exploration, development, and production. Eventually, these modern data lakes will transform into contextual knowledge centers that not only help business units find data, but also provide recommendations on other data sources (and analytic models) that might be useful for their given use case.

• Data Monetization Governance Council. Another key to navigating Inflection Point #2 is the creation of a data monetization governance council with the teeth to mandate the sharing, reuse, and continuous refinement of the organization’s data and analytic assets. If data and analytics are truly economic assets, then the organization needs a governance organization with both “stick and carrot” authority to encourage and enforce the continuous cultivation of these critical 21st-century economic assets (see Figure 6).

Figure 6: Data Monetization Governance Council

The key to scaling the organization’s data monetization capabilities is to thwart data silos, shadow IT spend, and orphaned analytics that create a drag on the economic value of data and analytics. When the business and operational costs to find, reuse, and refine the data and analytics become greater than the cost to build your own from scratch, then that’s a failure of the Data Monetization Governance Council.

Mastering the Data Monetization Roadmap Summary

The Data Monetization Roadmap provides both a benchmark and a guide to help organizations with their data monetization journey. To successfully navigate the roadmap, organizations must be prepared to traverse two critical inflection points:

• Inflection Point #1 is where organizations transition from data as a cost to be minimized to data as an economic asset to be monetized; the “Prove and Expand Value” inflection point.

• Inflection Point #2 is where organizations master the economics of data and analytics by creating composable, reusable, and continuously refining digital assets that can scale the organization’s data monetization capabilities; the “Scale Value” inflection point.

Carefully navigating these two inflection points enables organizations to fully exploit the game-changing economic characteristics of data and analytics assets – assets that never deplete, never wear out, can be used across an unlimited number of use cases at zero marginal cost, and can continuously learn, adapt, and refine, resulting in assets that actually appreciate in value the more that they are used.

Yes, you could say that the Data Monetization Roadmap is the game plan for fully exploiting the Schmarzo Economic Digital Asset Valuation Theorem. But that’s just me and that Nobel Prize in Economics talking…

I spent this morning with the Analytics Leadership Forum of Iowa (ALFI), and I said something that made a few people shift in their seats: what you are learning may be more important than what you know. For a room full of people who’ve built careers on what they know, that’s an unsettling thing to hear. It’s also, I think, the most hopeful thing happening to your career right now.

While that sounds like bad news, it is actually the opposite. Let me tell you why — and then let me give you the checklist I give all of my students.

Everything You Call an Asset is Really a Snapshot

Here is the uncomfortable truth underneath the AI moment, and it has almost nothing to do with technology. Value, expertise, and intelligence are not permanent possessions. They are contextual — they exist only in relationship to a world that refuses to hold still. And because that world keeps changing, all three are fleeting.

Start with value. Today, I made the point that you cannot determine the value of your data in isolation from the business — the same dataset is worthless or priceless depending on the decision it serves. Value is a relationship between a resource and a moment. When the moment moves, the value moves with it, sometimes all the way to zero. Today’s ROI-optimal decision quietly becomes tomorrow’s systemic risk, because what you sacrificed to get the early win doesn’t disappear — it ripples through people, behaviors, and systems until it surfaces as a problem you didn’t see coming.

Now expertise — and this is the one that lands hardest in a room full of experts. Expertise is pattern recognition tuned to an ecosystem. The claims adjuster who smells a bad claim the fraud model scores as clean, the agronomist who reads a field’s stress before the satellite flags it — that is priceless judgment. It is also priceless until the pattern changes. Expertise is “I’ve seen this before,” and its value lasts exactly as long as “before” keeps resembling “now.” In a stable world, the twenty-year veteran wins every time. In a world being rewritten monthly, the veteran’s edge has a shorter shelf life than it has ever had.

And intelligence — even the machine kind — decays too. A model trained on last year’s world is confidently answering a world that no longer exists. Prediction is intelligence about a past that already happened; the moment the ground shifts, the prediction is confidently wrong. Prediction is the most fragile form of knowing there is.

In a world that won’t hold still, every asset is a snapshot of a moment that’s already passing.

So why is any of this good news?

Because a shrinking half-life on expertise doesn’t just take something away — it hands something back. When what you know matters less than how fast you can learn, the advantage quietly shifts to the curious, the adaptable, and the willing. The playing field just tilted toward the people who are still growing.

Think about what that means. For most of history, the person with thirty years of accumulated pattern recognition held an unassailable lead, and everyone behind them was, by definition, behind. That lead was a wall. What AI does — when you use it well — is turn that wall into a ramp. It compresses the climb. The judgment that used to take a career to build can now be built in a fraction of the time, because you can practice against the world before you have to live with the consequences.

That is the reframe I most want you to leave with. AI is not here to replace your expertise. It is here to be the coach that helps you rebuild it, faster than ever, every time the world changes the rules.

What a Great Coach Actually Does

I’ve been describing YODA (Your Own Digital Assistant)—the methodology I teach at several universities and organizations to transform a generic AI into your personal thinking partner—as a life coach, and I mean this quite literally. Consider what the best coach in your life provided. A great coach didn’t simply give you answers. Instead, they understood your context, remembered your past experiences, posed sharper questions than you would ask yourself, delivered honest truths gently, and helped you practice more repetitions than you could have managed alone.

That is exactly what a well-built AI thinking partner does. It holds your context so it isn’t starting from zero every time. It challenges you to ask better questions instead of just producing faster answers. It lets you stress-test a decision against second- and third-order effects before you own the outcome. And — this is the part that matters — it gets smarter about you the more you work with it, the same way a coach who’s known you for years is worth more than one who met you yesterday.

A generic AI model that forgets you when the chat closes is a stranger giving advice. A thinking partner that learns your world alongside you is a coach. The difference is not the model. The difference is the context you build around it — and whether you keep feeding it what you learn.

The Checklist: Five Shifts to Reach Your Full Potential

None of this is about becoming a data scientist or learning to code. It’s about a handful of mindset shifts — and one small practice attached to each — that put you on the knowledge and expertise ramp instead of behind the wall. Here’s what I’d hold onto:

1. Trade expertise for adaptability. Stop measuring yourself by how much you know and start measuring how fast you can learn. Your credential is no longer your knowledge; it’s your learning velocity. Practice: pick one thing that changed in your field this quarter and get fluent in it this week (hint: vibe coding).

2. Ask better questions than the AI can answer. AI can answer almost anything. Your edge is knowing which question is worth asking — and what to do with the answer. Practice: before you prompt, write down the real decision you’re trying to make. Aim the tool at that, not at the first question that comes to mind (hint: embrace your inner Socrates).

3. Treat every project as reps. Judgment comes from making decisions and seeing consequences. Use AI to run the what-ifs and trace the ripple effects, so you accumulate years of hard-won reps without the scar tissue. Practice: on your next call, ask your AI partner what happens next — and after that — before you decide (hint: learn causal analytics).

4. Make your judgment teachable. The expertise locked in your head expires when you leave the room. Expertise you can capture and hand off keeps compounding. Practice: the next time you make a call your gut got right, write down why. That note is the seed of context worth reusing (hint: continuously build your contextual knowledgebase).

5. Build a coach, not a calculator. Don’t use AI as a vending machine for answers. Build a partner that knows your context, remembers your history, and grows with you. Practice: give your AI the background it would need to advise you like someone who actually knows your world — then keep adding to it (hint: build your own YODA!).

The World is Changing. Take Advantage and Change with It.

Here’s where I’ll leave you, and I’ll say it as gently as I know how, because it’s the most hopeful thing I believe about this moment.

For a long time, “the world is changing” was a warning. It meant keep up or get left behind, and it favored whoever was already ahead. However, this time, the change favors the willing. The tools that scale expertise are available to everyone now, not just the superstars — which means the person who decides to keep learning is no longer at a disadvantage to the person who learned it all years ago. The real power of this isn’t creating one superstar. It’s making everyone one.

Your expertise is expiring. Good. That means the game is no longer about defending what you know — it’s about how quickly you’re willing to grow. And for the first time, you don’t have to do that alone. You have a partner that can learn right alongside you. YODA – Your Own Digital Assistant!

The future isn’t AI instead of you. It’s AI with you — helping you understand cause and effect, exercise better judgment, and become the most complete and effective version of the professional you’re capable of being.

The world is changing, and that’s your opening.

Note: If you are interested, here is where you start your YODA journey:

“𝐘.𝐎.𝐃.𝐀. (𝐘𝐨𝐮𝐫 𝐎𝐰𝐧 𝐃𝐢𝐠𝐢𝐭𝐚𝐥 𝐀𝐬𝐬𝐢𝐬𝐭𝐚𝐧𝐭) 𝐓𝐡𝐞 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐄𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 𝐏𝐥𝐚𝐲𝐛𝐨𝐨𝐤: 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐀𝐈 𝐓𝐡𝐢𝐧𝐤𝐢𝐧𝐠 𝐏𝐚𝐫𝐭𝐧𝐞𝐫𝐬“

https://deanofbigdata.com/books/ols/products/yoda-your-own-digital-assistant-the-context-engineering-playbook-building-ai-thinking-partners

Blog #9 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

It’s Tuesday night. The developer’s attorney slides a term sheet across the table. It’s long, it’s professional, and it was written by people who do this for a living — which you do not. Somewhere in it is a number that sounds enormous, and around that number is a great deal of language whose job is to make sure the enormous number is the only thing you look at.

This is what you read before that meeting.

Why a Ranked List, Not a Long One

Eight installments of this series have laid the groundwork for the reasoning. The temptation now is to hand you a checklist—30 things to look for. Resist it, because a flat checklist is a trap. Hand a developer’s counsel a list of thirty items, and they will happily concede the cheap twenty, refuse the three that matter, and walk you out believing you won.

The only effective checklist is one that is ranked, indicating beforehand which lines you must not cross. Terms are categorized into three levels: those you must not sign without, those worth fighting for, and those you can trade away. The complete ranked list appears on the last page of this post — the one you should bring into the room. Here, the focus is on the five non-negotiables and their importance.

The Five Terms You MUST NOT Give Up

First, reporting on actuals. An actual is a number, and a number is only as good as its definition — so this term has two halves, and you need both. Ask what exactly counts as a job under this agreement: full-time or contractor, new or relocated, employed by the company or by a staffing firm, measured on what date. Then ask when the public sees that number each year. This comes first because every other protection on this list fires off a figure somebody has to compute years later. A reversion trigger, a clawback, a milestone, an audit — each one is an enforcement mechanism pointed at a metric. Draft the penalty with precision and leave the word “job” undefined, and you have built a very sophisticated machine that never turns on. Nashville is the proof. Oracle did report: 901 jobs, filed with the state. The company also advertised capacity for 2,000 people while its actual headcount grew by a net seven, and a fresher count filed in February still hasn’t been released. The failure wasn’t missing paperwork. It was that nobody had pinned down what counted, or guaranteed the public would see it. Sign without this and every remaining term becomes unenforceable in practice, because the side that defines the measure decides whether the penalty ever applies.

Second, a reversion right—linked to the right assets. Consider the unanswered question in the brochure: if this project fails or is abandoned, what benefits does the public receive, and is that right established before the lenders’ claims? When a deal resets, the asset transfers to the entity with the senior claim—usually a private-credit fund, not the town. A recorded public right to acquire it upon a specific failure is the only way for the community to truly have a seat at the table. As Part 8 explained, this right must attach to the land, the grid interconnect, and a lien on the equipment—not the building, which could be a liability. Signing without this right leaves you with a claim to an empty, unsellable box on a tax base already relinquished, while the movable value relocates elsewhere.

Third, a clawback with teeth. Incentives are paid up front against promises delivered later, so ask exactly how much is repaid if the jobs and the investment miss their schedule — and what triggers that repayment automatically. Without a dated, automatic remedy tied to missed milestones, “best efforts” is all you have, and best efforts is not a remedy. Nashville capped its exposure at $175 million and still saw 8,500 promised jobs become 901; the cap protected the cash, but nothing required performance. Sign without it, and the company keeps the abatement whether or not it delivers, and your only recourse is a lawsuit you can’t afford against a company that can.

Fourth, independent verification. Every other protection on this list depends on someone confirming that the numbers are real, so ask who checks the job counts, wages, and tax figures — and whether that person is someone the company doesn’t pay. If the company reports its own compliance, you are trusting the party with the most reason to shade it. This is the whole difference between a scorecard that grades performance and one that grades promises. Sign without it, and every milestone, cap, and clawback becomes unenforceable in practice, because you can never prove the trigger was hit.

Fifth, dated milestone triggers. Ask by what date the jobs must exist — and what happens the day they don’t. A deal can be cash-safe, fully clawed back, and still cost a community a lost decade while it waits for jobs that arrive slowly or never. Time is the one thing a cap can’t refund. Dates convert “committed, eventually” into an enforceable schedule. Sign without them, and the company controls the pace and the clock — committed on paper, absent in fact — while your planning, your services, and your patience are spent waiting on a timeline only they can move.

Notice what these five share. Each one bites only if the company fails to deliver. A partner who means to build, hire, and stay loses nothing by granting them. So if the developer’s counsel resists all five at once, that isn’t a hard negotiation — it’s information.

Where the Value Splits — and What You Can Trade

These five deal-breakers set the terms that decide how the value splits, rather than whether you’re protected at all. Most of them are the synergies from Part 6 — high value to the community, low cost to the company, the asks a developer can say yes to: infrastructure sized to serve the whole region rather than one site, a local-hiring floor and an apprenticeship pipeline, hard caps on water and energy with a true-up when actual use climbs past the forecast, and a net-new tax-value floor with a date attached. Fight hard for these; they’re where a good deal becomes a great one.

Below these, there’s a tier of valuable but tradeable benefits: a community-benefit fund, school and STEM partnerships, local vendor spend targets, and amenity contributions. These are genuine values—however, they are concessions made to secure the five critical priorities, not the other way around. If you ever find yourself defending a community-benefit fund while sacrificing the reversion clause, it’s time to pause and reconsider.

Every one of these terms — all three tiers, in order — is on the single page at the end of this blog.

How to Read Their Term Sheet

Place the developer’s document next to that page and ask three questions about each clause. Is it documented on paper or just a promise? A recorded lien endures bankruptcy, but good intentions do not. Which asset does it safeguard — the movable chips, the transferable interconnect, or the strandable shell? And can it be independently verified, or does it rely on the company’s self-assessment? Any term that fails all three criteria is mere decoration, no matter how elegant the language.

Judge the Terms, Not the Headline

None of this makes you adversarial. The best version of this deal is still a win-win — that was the whole point of Part 6.

These clauses protect the community in the event of failure, and a good-faith partner loses nothing by granting them.

When the resistance is fiercest exactly where your protection is strongest, the term sheet is telling you something the brochure won’t.

In the final post, this shifts from just reading a document to using a tool — an interactive scorecard where you input your deal, test all these terms, and see the grade change. The same instrument, now live.

Share

Blog #9 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

It’s Tuesday night. The developer’s attorney slides a term sheet across the table. It’s long, it’s professional, and it was written by people who do this for a living — which you do not. Somewhere in it is a number that sounds enormous, and around that number is a great deal of language whose job is to make sure the enormous number is the only thing you look at.

This is what you read before that meeting.

Why a Ranked List, Not a Long One

Eight installments of this series have laid the groundwork for the reasoning. The temptation now is to hand you a checklist—30 things to look for. Resist it, because a flat checklist is a trap. Hand a developer’s counsel a list of thirty items, and they will happily concede the cheap twenty, refuse the three that matter, and walk you out believing you won.

The only effective checklist is one that is ranked, indicating beforehand which lines you must not cross. Terms are categorized into three levels: those you must not sign without, those worth fighting for, and those you can trade away. The complete ranked list appears on the last page of this post — the one you should bring into the room. Here, the focus is on the four non-negotiables and their importance.

The Four Terms You MUST NOT Give Up

First, a reversion right—linked to the right assets. Consider the unanswered question in the brochure: if this project fails or is abandoned, what benefits does the public receive, and is that right established before the lenders’ claims? When a deal resets, the asset transfers to the entity with the senior claim—usually a private-credit fund, not the town. A recorded public right to acquire it upon a specific failure is the only way for the community to truly have a seat at the table. As Part 8 explained, this right must attach to the land, the grid interconnect, and a lien on the equipment—not the building, which could be a liability. Signing without this right leaves you with a claim to an empty, unsellable box on a tax base already relinquished, while the movable value relocates elsewhere.

Second, a clawback with teeth. Incentives are paid up front against promises delivered later, so ask exactly how much is repaid if the jobs and the investment miss their schedule — and what triggers that repayment automatically. Without a dated, automatic remedy tied to missed milestones, “best efforts” is all you have, and best efforts is not a remedy. Nashville capped its exposure at $175 million and still saw 8,500 promised jobs become 901; the cap protected the cash, but nothing required performance. Sign without it, and the company keeps the abatement whether or not it delivers, and your only recourse is a lawsuit you can’t afford against a company that can.

Third, independent verification. Every other protection on this list depends on someone confirming that the numbers are real, so ask who checks the job counts, wages, and tax figures — and whether that person is someone the company doesn’t pay. If the company reports its own compliance, you are trusting the party with the most reason to shade it. This is the whole difference between a scorecard that grades performance and one that grades promises. Sign without it, and every milestone, cap, and clawback becomes unenforceable in practice, because you can never prove the trigger was hit.

Fourth, dated milestone triggers. Ask by what date the jobs must exist — and what happens the day they don’t. A deal can be cash-safe, fully clawed back, and still cost a community a lost decade while it waits for jobs that arrive slowly or never. Time is the one thing a cap can’t refund. Dates convert “committed, eventually” into an enforceable schedule. Sign without them, and the company controls the pace and the clock — committed on paper, absent in fact — while your planning, your services, and your patience are spent waiting on a timeline only they can move.

Notice what these four share. Each one bites only if the company fails to deliver. A partner who means to build, hire, and stay loses nothing by granting them. So if the developer’s counsel resists all four at once, that isn’t a hard negotiation — it’s information.

Where the Value Splits — and What You Can Trade

These four deal-breakers set the terms that decide how the value splits, rather than whether you’re protected at all. Most of them are the synergies from Part 6 — high value to the community, low cost to the company, the asks a developer can say yes to: infrastructure sized to serve the whole region rather than one site, a local-hiring floor and an apprenticeship pipeline, hard caps on water and energy with a true-up when actual use climbs past the forecast, a net-new tax-value floor with a date attached, and public reporting on actual employment rather than advertised capacity. Fight hard for these; they’re where a good deal becomes a great one.

Below these, there’s a tier of valuable but tradable benefits: a community benefit fund, school and STEM partnerships, local vendor spend targets, and amenity contributions. These are genuine values—however, they are concessions made to secure the four critical priorities, not the other way around. If you ever find yourself defending a community-benefit fund while sacrificing the reversion clause, it’s time to pause and reconsider.

Every one of these terms — all three tiers, in order — is on the single page at the end of this blog.

How to Read Their Term Sheet

Place the developer’s document next to that page and ask three questions about each clause. Is it documented on paper or just a promise? A recorded lien endures bankruptcy, but good intentions do not. Which asset does it safeguard — the movable chips, the transferable interconnect, or the strandable shell? And can it be independently verified, or does it rely on the company’s self-assessment? Any term that fails all three criteria is mere decoration, no matter how elegant the language.

Judge the Terms, Not the Headline

None of this makes you adversarial. The best version of this deal is still a win-win — that was the whole point of Part 6.

These clauses protect the community in the event of failure, and a good-faith partner loses nothing by granting them.

When the resistance is fiercest exactly where your protection is strongest, the term sheet is telling you something the brochure won’t.

In the final post, this shifts from just reading a document to using a tool — an interactive scorecard where you input your deal, test all these terms, and see the grade change. The same instrument, now live.

Share

Blog #8 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

In the last post, I made the following argument: when the AI build-out resets, the data centers won’t rot. The chips inside them may be obsolete, but the powered shell, the grid interconnect, the cooling — those survive. Somebody re-racks the building and runs it. The value doesn’t vanish. It gets repriced, and it changes hands.

Then a reader pushed back, and he was right in a way I hadn’t fully reckoned with.

His point is that the fiber analogy has a flaw. Railroad track doesn’t become obsolete. Buried fiber doesn’t become obsolete. But Nvidia chips and Micron memory do — rapidly, and on a three-to-five-year cycle, despite being bought with twenty-year debt. The asset at the heart of this boom decays in a way the assets of every previous boom did not. Isn’t that worse?

Yes and no. And the “no” part is what I got wrong.

The Chips Don’t Die. They Get Demoted.

Obsolete at the frontier is not obsolete, period. There is a documented cascade: years one and two, a chip trains frontier models. Years three and four, it runs inference — which is far less compute-intensive than training ever was. Years five and six, it does batch work and analytics. The silicon falls down the stack rather than into the landfill.

The market bears this out. The A100 — 2020 silicon, the chip that trained GPT-3 — still holds sixty to eighty percent of its purchase price. When CoreWeave’s 2022-vintage H100 contracts expired, the chips were rebooked at 95% of their original price. Jensen Huang joked that once Blackwell shipped, you couldn’t give Hoppers away. The market disagreed with him.

There’s a limit, and it’s the same one that finished the coal plants: an old chip burns far more electricity per unit of work than a new one — roughly eleven times more for H100 inference versus a B300. Eventually, the power bill kills it even though the hardware still runs. But before that day comes, the chips find homes: university labs, hospital systems fine-tuning on their own data, startups, state agencies, edge deployments — every buyer who could never have justified forty thousand dollars a card but can absolutely justify six.

That is a genuinely good outcome. Cheap compute, democratized — the same way cheap dark fiber, bought for pennies out of WorldCom’s wreckage, became the backbone of broadband and made a startup called YouTube possible. The technology delivers on its promise. The original investors don’t collect.

But Notice What Just Happened

The chips got trucked away.

That sentence is the whole post, and I missed it the first time. Dark fiber couldn’t leave. It was buried in the ground, in place, until demand arrived and someone lit it. The asset ledger reopened exactly where the asset was buried.

GPUs are not buried. They get unbolted, crated, and driven to Ohio or wherever they are needed next.

So when I said the powered shell is the durable, valuable thing that survives a bust — I was right only under one condition: that another compute tenant shows up. If AI demand returns at scale, the interconnect is gold and someone re-racks the building. But if it doesn’t, then the very thing that made that building valuable — its ferocious specialization — is exactly what makes it unusable for anything else.

Three Assets, Three Fates

Rank what’s actually in one of these deals by whether it can move.

● The chips move. They cascade down-market and out the door onto a truck to a buyer three states away. High value, fully mobile, and gone before the paperwork settles.

● The interconnect transfers. A grid connection of that size — with a signed agreement and a multi-year queue behind it — is the crown jewel, and it is severable from the building. In a genuine bust, the most likely buyer isn’t someone who wants the data center. It’s someone who wants the power and treats the structure as incidental.

● The shell strands. A modern AI facility is not a warehouse full of servers. It runs a hundred kilowatts per rack, whereas a conventional building runs five to ten. It has cooling loops plumbed into its structure, an electrical spine sized for that density, almost no space a human can occupy, and a location chosen for power and fiber rather than for logistics. Nearly every design choice that made it excellent at one job makes it hostile to any other. Too specialized to convert. Too expensive to demolish.

The fiber couldn’t leave. The GPUs can. Highlighting this contrast underscores why policymakers should prioritize asset mobility and protective clauses in deals, as it directly impacts community protection and strategic control.

What This Changes

In the last post, I argued for a reversion clause: a recorded public right to acquire the asset at a nominal price on a defined failure trigger, structured to sit ahead of the creditors in line. I still believe that’s the single most important term a community can negotiate.

But I was imprecise about which asset — and imprecision in a clause is the same as no clause at all.

A reversion right that attaches only to the building may be a right to a liability: an empty, unconvertible box with a maintenance bill attached, sitting on a tax base the community already gave away. That isn’t protection. That’s a consolation prize dressed as one.

The clause has to attach to the assets that hold their value: the land and the grid interconnect — and, where it can be perfected, a recorded lien on the equipment itself- before the chips can be crated and driven away.

Protect the transferable asset, not the photogenic one.

The Discipline Underneath

I could have quietly fixed this. Nobody would have noticed. But the whole argument of this series is that we should judge deals by what the evidence actually shows rather than by what we’d prefer to believe — and that discipline is worthless if it only applies to other people’s arguments.

A reader attacked my framework, and it got better. That is exactly what the five expert personas in Part 5 were for, and it’s why the scorecard grades performance rather than promises. The method has to survive contact with people who know something you don’t.

Which brings us to the practical question, and the one the next post answers: if you’re a council member, and a developer’s attorney is sliding a term sheet across the table on Tuesday night — what, exactly, must you not sign without?

— Bill Schmarzo, The Dean of Big Data

Blog #7 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — extending Economic Critical Thinking from one Iowa deal to the economics of the AI build-out itself.

Picture the Apple data center in Waukee, Iowa, twenty years from now: roof buckled, servers rusting, weeds pushing through the parking lot. It’s the image everyone reaches for when they hear the phrase “AI bubble.” It’s vivid, it’s ominous — and it’s almost certainly wrong.

The Wrong Question

“Will the data centers close?” Mostly, no. The scarce, durable, valuable parts of one of these facilities aren’t the chips — they’re the powered shell, the grid interconnect, the cooling, the fiber. A powered building is an asset even when the machines inside it are three generations obsolete. You don’t bulldoze it; you re-rack it. So the abandoned Waukee picture is the least likely outcome of a bust, not the most likely.

The right question is sharper — when the boom resets, who ends up owning the repriced asset? That reframe is the whole post, because a bust doesn’t destroy the value — it relocates it. Across time, from the moment of mania to the moment real demand finally arrives, and across ownership, from the investor who built the thing to whoever holds it when the music stops.

Not a Prediction — a Pattern

In June 2026, the Bank for International Settlements — the central bank of central banks — devoted its flagship annual report to exactly this AI risk. It placed the AI boom in a specific lineage: canal mania in the 1830s, Britain’s railway mania in the 1840s, the electrification boom of the 1920s, the dot-com bubble of the 1990s. Each began with a genuine technological breakthrough that drew in more capital than commercial returns could ultimately justify. Each ended in an investment reversal and an economy-wide recession. Today’s AI build-out, the BIS wrote, bears a resemblance to these precedents.

The numbers behind the concern: the five largest hyperscalers are on track to spend more than $1 trillion on AI-related capital expenditures over 2025 and 2026 — outpacing their earnings and free cash flow, forcing some to issue debt to cover the gap. The BIS’s models warn that as competition drives spending higher, the sector’s net economic surplus could turn negative, and a disappointment in returns could turn the capex boom into a “protracted investment bust.” This is the most authoritative institution in global finance, saying that the technology is real and being overbuilt.

The BIS is careful to explain why this reaches beyond Silicon Valley. US stocks now make up nearly two-thirds of the global equity index, and household exposure to them has roughly doubled since 2010, so a sharp repricing of AI-related shares would ripple through ordinary retirement accounts, not just tech balance sheets. This is not a niche industry story. It is a macroeconomic one.

The Trigger — and Who Catches the Asset

What forces the reset is depreciation. Hyperscalers book their AI chips over a five- to six-year period. Skeptics argue the real economic life is closer to two or three, because Nvidia ships a faster, cheaper architecture every 18 to 24 months. When that gap catches up, the write-downs arrive in single quarters.

Not every player is equally exposed. The hyperscalers — Apple among them — carry fortress balance sheets and fund much of their spending from cash flow; they can absorb a bad bet. The fragile tier sits beneath them: the specialized “neocloud” operators financed by loans secured against their own GPUs, and the engineering and construction firms at the end of the supply chain, whose balance sheets the BIS singles out as comparatively weak. When the hyperscalers slow their spending, those are the firms that fail first — and their assets are the ones that hit the auction block.

The BIS reserves its sharpest warning for the extent to which the money has become tangled. Chipmakers take equity stakes in the AI labs that then commit to buying their chips; hyperscalers invest in the labs that rent their cloud capacity — revenue circling back on itself in a loop the report calls circular financing. It rhymes with the late-1990s telecom vendors who lent their own customers the money to buy their equipment, right up until the customers couldn’t pay. A clean revenue chart can hide a very messy balance sheet underneath.

And here is the part that matters for a community. The write-down doesn’t destroy the asset — it reprices it and hands it to whoever holds the senior claim. That claim increasingly belongs to lenders, not owners: neoclouds borrow against their GPUs, and private credit funds hold senior secured positions. The BIS warns that this financing is poorly disclosed, with the same asset sometimes pledged multiple times. When a leveraged operator fails, the lender forecloses — and the successor who acquires the repriced data center is a private-credit fund, not the public. The town that granted the tax abatement is nowhere in that line.

We’ve Seen This Movie

We have a full dress rehearsal for this. In the late 1990s, telecom carriers invested more than $500 billion in laying some 80 million miles of fiber, convinced that internet traffic would explode. They were right about the demand — and annihilated anyway. WorldCom filed the largest bankruptcy of its era; Global Crossing collapsed; hundreds of thousands of jobs vanished; and years later, most of that fiber still sat dark.

But the fiber didn’t disappear. Distressed-debt funds bought it for pennies on the dollar, and that cheap capacity became the backbone of broadband, streaming, and the cloud — the reason a startup called YouTube was even possible. The technology delivered everything its builders promised. The builders just never collected. Someone else did. And the public that had subsidized the build-out through favorable tax law? It got cheaper bandwidth as a consumer — but captured none of the upside from the repricing as an investor.

One caution, because the parallel can be pushed too far. The fiber era was laced with outright fraud — WorldCom’s collapse was an accounting scandal, not merely a bad bet. The AI comparison is not about fraud; the BIS is explicit that the technology is real and the productivity gains are measurable. The rhyme is narrower and more useful: a genuine breakthrough, financed past the point where returns can justify it, leaving durable assets to be claimed by someone other than the people who built them. That pattern doesn’t require a villain. It only requires optimism and leverage.

When Is the Ledger Complete?

Which brings us back to Marcus’s question. The reason “will it fail?” is the wrong question is that a failed infrastructure deal doesn’t run one ledger — it runs three, and they close at different times. The investor’s ledger closes at bankruptcy: a permanent loss. The asset’s ledger never closes — it is repriced and passed on. And the social ledger runs positive, but it is paid for by the wiped-out investors: society gets cheap capacity, financed by destroyed capital.

So when is the ledger complete? Not at the write-down. At the successor’s return — a decade later, when whoever bought the asset for pennies finally earns the payoff the original builder was promised. For a community, that changes the entire negotiation. The question at the subsidy table isn’t only “what do we get if this succeeds?” It is “who owns the asset if it fails?” — and that can be answered up front, in the original grant: a recorded public right to acquire the asset at a nominal price on a defined failure trigger, a reversion clause structured to sit ahead of the creditors in line. Not a claim on some successor’s profits — that fight belongs at bankruptcy — but a claim on the asset itself, perfected before the private-credit funds get there. Even the BIS, in its prescription, calls for ensuring the gains from the technology are shared widely rather than concentrated in a few hands. A reversion clause is what that looks like at the level of a single deal.

Iowa is the concrete case. Waukee gave up roughly two decades of tax base to land Apple — a bet that the jobs and the halo would more than repay it. Apple is the safest tenant imaginable, so the town will probably never see the rusting-shell version of failure. But “safe from abandonment” is not the same as “sharing in the upside.” If the broader build-out resets and that Iowa asset is ever repriced, sold, or repurposed, the reversion question determines whether Waukee is a party to the next chapter or a spectator.

The Real Risk Isn’t Abandonment

So look again at that overgrown data center. The real risk was never that it rots. The real risk is that it hums along, gets repriced in a reset, and quietly earns a fortune for a distressed-asset fund three states away — while the town that gave up twenty years of tax base to build it owns nothing of what it became.

That is the difference between thinking like a politician, who counts the ribbon-cutting, and thinking like an economist, who traces the value all the way to the successor’s return. The boom will leave assets behind — that much is nearly certain. The only question Economic Critical Thinking insists on asking is: when the ledger finally closes, whose name is on it? That is a question a community can still answer — but only before it signs. It is exactly what a scorecard and the reversion terms written into it are built to make visible.

— Bill Schmarzo, The Dean of Big Data

Blog #6 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

In Part 5, we built a scorecard and used it to do something uncomfortable: grade a real deal. Iowa’s Apple/Waukee data center earned a D+. Then we handed that scorecard to five experts whose job was not to validate it but to attack it — and found that they didn’t agree. The economic development director saw a win. The school district CFO saw a twenty-year hole. Same deal, different chairs.

That disagreement is where most public conversations about these deals get stuck. One side has a number — thousands of jobs, billions in investment. The other side has a different number — millions in forgone taxes, strained schools. The argument never resolves because the two sides aren’t measuring the same thing.

A scorecard that only produces a grade leaves them there. But a scorecard can do something more valuable than settle an argument: it can turn the argument into a design problem. That is the move from grading a deal to negotiating one — and it isn’t really about data centers at all. It’s about how any group of people who disagree can stop fighting over a fixed pie and start designing a bigger one.

The Table, Not the Verdict

A verdict looks backward; it grades a deal that already exists. A negotiation looks ahead; it designs a deal that doesn’t yet exist. The same scorecard serves both — but at the table, it stops being a report card and becomes a shared map.

Seat the five experts from Part 5 at that table. Each protects something different: the development director, the investment and the announcement; the credit analyst, a method that holds up; the school CFO, the tax base that funds the classrooms; the real-estate operator, who carries the risk when the hardware ages; the measurement skeptic, the line between a promise and a proof. The scorecard makes each seat’s interests legible — not as slogans, but as scores on the same nine dimensions.

The naive way to resolve their conflict is to pick winners: give the developer the abatement, tell the schools to make do. That is fixed-pie thinking — every gain for one party is a loss for another. It is also exactly what produced the D+. The breakthrough is to stop treating the pie as fixed.

The Move: From 1 + 1 = 2 to 1 + 1 = 3

Somewhere in every multi-stakeholder deal are combinations in which one party’s cost is another party’s gain — where a single dollar counts as a win for both sides. Finding them is the difference between a negotiation that divides value and one that creates it.

Take a concrete example from the Waukee dimensions. A developer funding grid and substation upgrades incurs a cost. But if those upgrades are sized to serve the whole region — not just the data center — that same cost becomes a durable public asset. The developer needs the power anyway; the community gets infrastructure it would otherwise have to fund itself. One dollar, two wins. That is not splitting the pie. That is 1 + 1 = 3.

The scorecard is the instrument that makes these combinations visible. Because it scores every dimension from each stakeholder’s seat, it shows not only where interests collide but — more usefully — where they quietly align. A conflict you can see the shape of is a conflict you can design around.

Five Things a Scorecard Does at the Table

Used this way, the scorecard does five things a verdict never could.

• It makes the trade-offs explicit. Transparency isn’t just the final grade; it’s the trade-off underneath — what you gain in one dimension and give up in another. Waukee trades immediate capital scale for twenty years of forgone tax base. Naming that exchange is the difference between marketing and honesty.

• It grades from each stakeholder’s seat. Apple’s A and Iowa’s D+ are the same transaction seen from two chairs. Scoring per stakeholder reveals where interests truly collide — and where they secretly align.

• It surfaces synergies, not just conflicts. A weak scorecard forces a zero-sum choice; a strong one hunts for the structures where one party’s cost is another’s gain, and marks them.

• It builds a shared base of facts. Conflict runs on competing narratives. A scorecard forces everyone to start with the same evidence first — because you cannot negotiate a win-win while still arguing over the facts.

• It reveals the negotiation space. A verdict says D+. An instrument says: Here are the three moves that make it a B. ” It doesn’t just grade the deal on the table; it maps the better deals still within reach.

Walking Waukee from a D+ to a B

Return to the three conditions from Part 4: pay for what you consume and build what the region needs anyway; broaden participation so it’s measured in more than paychecks; deliver net-new tax value within a decade and prove it independently. Apply them together, and the same deal that earned a D+ becomes a B — with the tech company still landing a B too. Not a community win extracted at the company’s expense, but a structure both sides can clear.

That is a claim worth proving rather than asserting — which is exactly what the interactive scorecard is for. In the next post, you’ll be able to sit at this table yourself: toggle each condition, watch the grade move, and see the win-win appear in real time.

A Living Instrument: Define, Monitor, Learn

A deal isn’t a moment; it’s a decade. So the scorecard has to be more than a snapshot — it has to be a system that does three things over time.

• Define the shared success criteria — what a win means for each stakeholder, built on a common evidence base.

• Monitor the deal against those criteria — the clawbacks and independent verification that confirm it still earns its grade.

• Learn and adapt — re-score as terms change, and results come in, so the deal improves rather than ossifies.

That is why the instrument must be both dynamic and accountable. Dynamic, so stakeholders can test “what if we add a hiring floor, a water cap, a clawback?” and watch the grade climb as they design — not just evaluate the deal they were handed. Accountable, so there is a record of which trade-offs were accepted, on whose behalf, and against what evidence — the audit trail that settles a year-five dispute by reference to the deal everyone agreed to, not competing memories.

Negotiate the Reset, Not Just the Ribbon-Cutting

Most negotiations fight over today’s terms. But infrastructure has a longer arc: projects fail, restructure, and change hands — and when they do, the physical asset rarely disappears. It gets repriced. A bankruptcy doesn’t destroy the fiber or the powered shell; it wipes out the cost basis, and a successor runs the same asset at prices the original builder couldn’t survive. The value isn’t destroyed. It’s relocated — to whoever holds the asset after the reset.

That has a sharp negotiating implication. A community shouldn’t only ask “what do we get if this succeeds?” It should ask “who owns the asset if this fails?” And that question can be answered in the original grant, not left to the bankruptcy: a public right to acquire the asset at a nominal price on a defined failure trigger — a reversion clause. Not a claim on some future successor’s profits, which is a bankruptcy-table fantasy to negotiate up front, but a recorded claim on the asset itself, structured to sit ahead of the creditors in line. A deal that secures the reset is a deal that survives its own failure.

(Why these projects fail — and where the value goes when they do — is a bigger story than any single negotiation. It’s the subject of the next arc in this series.)

From This Table to Any Table

Step back from data centers, and the method is general. Every hard negotiation — labor and management, two companies merging, four agencies splitting a budget, a family dividing an estate — begins the same way: parties with conflicting definitions of a win, and a pie that looks fixed. The instinct is to fight over the slices. The discipline is to do four things instead.

1. Surface each party’s real interests — not their stated positions, but what they actually need. (The scorecard’s per-seat view.)

2. Agree on the facts before the terms. (The shared evidence base.)

3. 3. Hunt for the 1 + 1 = 3 combinations — where one party’s cost is another’s gain. (The synergy move.)

4. Design the structure that captures them, and build in the accountability to keep it honest. (Define, monitor, learn.)

That is how conflict becomes innovation: not by splitting the difference, but by discovering the moves where the pie was never actually fixed.

A scorecard doesn’t decide the deal — the people at the table do. But a good one changes what they can see. And people who can see where their interests align make better deals than people who can only see where they collide.

In the next post, we’ll put this to the test with an interactive Data Center Scorecard — a tool that lets you sit at the table, apply the conditions, and watch a D+ become a win-win in real time. Bring your own deal.

— Bill Schmarzo, The Dean of Big Data

Blog #5 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

After the post on ripple effects, a reader asked the question this whole series has been circling: “What would a real value scorecard for these data centers even include — and why does no one publish one?”

The second half of that question has an uncomfortable answer. No one publishes a real scorecard because a real scorecard makes the full ledger visible — every dimension of value created and every cost borne, side by side. The people structuring these deals would rather you saw only the headline: the capital number, the job count, the ribbon-cutting. A complete scorecard ends that. So let’s build one.

But there’s a prior question, because a bad scorecard is worse than none — it launders a weak deal in the appearance of rigor. Before we publish ours, we have to answer the question: what makes any report card actually work? Fortunately, we don’t have to guess. Some of the most influential scorecards in business and public life have already taught us — both by succeeding and by failing.

Seven Traits of a Scorecard Worth Trusting

The best scorecards don’t just measure performance — they build trust. Looking across the ones that earned lasting credibility and the ones that lost it, seven traits separate the scorecards that drive improvement from those that get ignored or gamed.

1. It measures what matters across multiple dimensions.
   No single number captures reality. A trustworthy scorecard measures the full system because any single metric can be optimized while the thing it represents quietly deteriorates.

2. It rolls up to one understandable grade.
   Since 1998, the American Society of Civil Engineers has graded U.S. infrastructure using a simple A–F scale. That simplicity is its strength: it compresses thousands of technical measurements into a grade that citizens, policymakers, and the media immediately understand. The rigor is underneath; the communication is on top.

3. Its criteria are transparent and reproducible.
   Trust comes from consistency. Moody’s and S&P credit ratings agree roughly 99% of the time because their methodologies are standardized. ESG ratings agree only about 61% of the time because different organizations use different definitions, measurements, and weighting schemes. When no one understands how the grade is produced, no one knows how to improve it.

4. It is graded independently.
   ASCE’s infrastructure grades carry credibility because they are produced by independent engineers—not the agencies responsible for the infrastructure. The fastest way to destroy confidence in a scorecard is to let the organization seeking incentives grade itself.

5. It rewards demonstrated performance—not promises.
   Some ESG methodologies place greater weight on corporate pledges than actual outcomes. A trustworthy scorecard measures what has been delivered, not what has been promised.

6. It provides a roadmap for improvement.
   Every ASCE report explains how to raise the grade. That transforms the scorecard from a verdict into a management tool. The goal isn’t to shame organizations—it’s to help them improve.

7. It is repeated over time and designed to resist gaming.
   ASCE updates its grades every four years, making progress—or decline—visible. Every scorecard must also guard against Goodhart’s Law: once a measure becomes the target, it becomes easier to manipulate than to improve. The best defense is independent verification and metrics tied to objectively measurable outcomes.

How the Economic Critical Thinking Scorecard Is Built

Our scorecard is intentionally built around the seven traits of a trustworthy scorecard. It assesses value across nine dimensions and then poses a second question that most evaluations overlook: does that value weaken, remain stable, or grow over time?

• The nine dimensions identify where value is generated.

• The 20-year ripple horizon shows whether that value persists.

Collectively, these aspects lead to a clear, single A–F grade—the same straightforward scale we used to rate the Waukee data center project as a D+. Each dimension asks a straightforward question:

Behind each question sits a transparent set of measurable criteria—from tax economics and workforce development to environmental stewardship and independent accountability. The methodology is fully visible and reproducible, ensuring that anyone can understand both how the score was calculated and how to improve it.

The scorecard not only grades but also analyzes the factors behind each score, helping stakeholders identify needed improvements. That is the difference between a performance scorecard, which tells you where a deal stands, and one that guides future improvements — showing you how to move it.

Note: In the next post, we’ll put this deal on the interactive Data Center Scorecard — where you can apply Part 4’s conditions yourself and watch Waukee climb from a D+ to a B, with the tech company still landing a B too. A real win-win, not a wish.

The Ripple Effect, Dimension by Dimension

A score is only a snapshot. The Economic Critical Thinking Scorecard asks a second—and arguably more important—question:

What happens to that value over time?

Every one of the nine dimensions is evaluated across the three waves of economic ripple:

• First-order: the immediate, measurable impact.

• Second-order: how people, businesses, and institutions respond.

• Third-order: whether that value ultimately dissipates, holds, or compounds.

Economic Critical Thinking differs from traditional scorecards because two projects can have similar immediate results but very different long-term impacts. One develops lasting economic ability; the other shifts costs to future taxpayers.

The Waukee project illustrates the difference. The initial announcement highlighted a $1.3 billion investment and new jobs, producing a strong first-order financial score. But over time, twenty years of tax abatements weakened public revenues, workforce development failed to materialize, and environmental pressures continued to accumulate. What looked promising at the ribbon-cutting followed a very different long-term trajectory.

The lesson is straightforward:

Don’t just evaluate where value begins. Evaluate where it ends.

That is why every dimension in the Economic Critical Thinking Scorecard measures not only today’s value but also whether that value dissipates, holds, or compounds over time.

The Value × Ripple Matrix: each of the nine dimensions read across the three waves of economic ripple — read down for the snapshot, across to see whether value dissipates, holds, or compounds.

Pressure-Testing the Scorecard

No scorecard should be trusted until someone tries to break it. That’s why we didn’t ask five experts to validate the methodology—we asked them to attack it.

These aren’t real individuals. They’re research-based personas, each representing a distinct stakeholder perspective and a different definition of success. Each was built by asking five questions:

• What really drives your decisions?

• Where does conventional wisdom usually fail?

• What evidence would change your mind?

• What mistake do you see others repeatedly make?

• Which trade-off are you unwilling to compromise?

Their job is to expose the scorecard’s blind spots.

Each perspective strengthened the scorecard in a different way.

• The economic development director prevented us from designing an unrealistic wish list that no company would ever accept.

• The credit analyst demanded transparent, reproducible methodology rather than subjective judgment.

• The school district CFO forced us to look beyond the ribbon-cutting and account for the long-term consequences of twenty-year tax abatements.

• The real estate operator reminded us that value captured is not necessarily value created—and that understanding who bears depreciation risk matters just as much as understanding who receives the benefits.

• The measurement skeptic enforced perhaps the most important rule of all:

No verifiable proxy. No credit.

If a claimed benefit isn’t independently measurable, it’s just a hypothesis, not evidence. The value of these five personas lies not in their agreement but in their disagreement.

• The economic development director considers the investment successful.

• The school district CFO sees twenty years of constrained public revenue.

• The real estate operator sees a capital allocation problem.

• The credit analyst sees a methodological challenge.

• The measurement skeptic sees assumptions waiting to be tested.

They’re all assessing the same deal. Each perspective is valid from their standpoint. This highlights why Economic Critical Thinking demands considering multiple viewpoints.

A traditional scorecard produces a single grade. An Economic Critical Thinking scorecard makes the trade-offs behind that grade visible — showing not only where stakeholders’ interests collide, but, more usefully, where one side’s cost is another side’s gain. What to do with that map — how parties who disagree turn it into a deal that earns a better grade for everyone — is where this series goes next.

The Story So Far

Five posts, one discipline.

• Blog #1 made the case: think like an economist, not a politician.

• Blog #2 gave the method — the nine dimensions and the ripple effects.

• Blog #3 graded a real deal and landed on a D+.

• Blog #4 asked the hardest question — compared to what? — and showed what a genuine win-win demands.

• Blog #5 hands you the instrument: a scorecard built to the seven traits, run across the nine dimensions and the full ripple horizon, and pressure-tested by the five experts who would attack it.

Notice what experts left out: they disagree, and a scorecard that only grades finished deals just reports it. The real value isn’t in judging existing deals but in helping those disagreements lead to better ones before signing.

That is the last move, and it is the next post. Blog #6 points the scorecard at the negotiation table — where the economic-development director, the schools’ CFO, and the real-estate operator stop being voices in a methodology and become parties to a deal. The question stops being What grade did this earn? And what structure earns a better one for everyone at the table? That is where a scorecard stops measuring value and starts creating it.

Until then, the discipline holds in one sentence: make the full ledger visible, then demand a deal that earns its grade. Iowa’s renewable energy sector will continue to attract this capital. The only question is whether we evaluate it like spectators chasing a ribbon-cutting — or like economists protecting a community. Now you have a scorecard that makes the second choice the easy one.

Artificial intelligence is everywhere—yet most people still use it as a question-and-answer machine. The result is predictable: generic responses, inconsistent recommendations, and decades of hard-earned expertise left untapped.

The problem isn’t AI. The problem is context.

Y.O.D.A. (Your Own Digital Assistant) changes that.

Rather than treating AI as a chatbot, YODA teaches you how to transform it into a personalized AI Thinking Partner—one grounded in your goals, your expertise, and the decisions you actually need to make. Drawing on the principles of Context Engineering, decision science, and human learning, YODA shows you how to define intent, curate the knowledge that matters, incorporate expert perspectives, and guide AI reasoning toward recommendations that are relevant, trustworthy, and actionable.

Turning Domain Experts into AI Solution Engineers

The future of AI won’t be determined solely by better models. It will be determined by the people who know how to provide those models with the right context.

Whether you’re a business leader, educator, analyst, consultant, healthcare professional, entrepreneur, student, or lifelong learner, YODA provides a practical methodology for transforming your experience, expertise, and judgment into AI-powered decision support.

More Than a Book—A Working AI System

Every copy of Y.O.D.A. includes the companion YODA Activation Prompt—a professionally designed instruction set that transforms the ebook into your own AI coach.

Simply upload the ebook into your favorite Generative AI platform, paste in the Activation Prompt, and your AI becomes a personalized guide that walks you through the YODA methodology step by step.

You don’t just read about Context Engineering—you immediately begin practicing it on your own initiatives, decisions, and learning goals.

Together, the ebook and Activation Prompt teach your AI Thinking Partner to:

• Question like Socrates — Challenge assumptions.

• Empathize like Jesus — Understand stakeholders.

• Think like Mr. Spock — Detect decision biases.

• Act as a Steward — Protect long-term outcomes.

• Reason like Aristotle — Follow the evidence.

Through the proven five-step YODA methodology, you’ll learn how to:

• Build an AI Thinking Partner grounded in your goals, expertise, and decision environment.

• Create a Contextual Knowledgebase that focuses AI on what matters most.

• Leverage expert perspectives to strengthen reasoning and recommendations.

• Use AI to accelerate learning, innovation, and decision-making.

• Turn domain experts into AI Solution Engineers.

• Develop AI Thinking Partners that improve through feedback, refinement, and persistent memory across sessions.

YODA provides a practical framework for combining human judgment with artificial intelligence to create better learning, better judgment, better decisions, and ultimately better outcomes.

You won’t finish this book with new knowledge alone. You’ll finish it with your own AI Thinking Partner—already built, already briefed, and ready to help you apply the YODA methodology to your most important decisions for years to come.

https://deanofbigdata.com/books/ols/products/yoda-your-own-digital-assistant-the-context-engineering-playbook-building-ai-thinking-partners

Blog #4 of a multi-part series on why leaders and policymakers must think like economists, not like politicians or financial managers — and how to structure a data-center deal that actually works for everyone.

In Blog #3, we graded the Waukee/Apple deal and landed on a D+. But a grade only answers “was this deal good?” Economic Critical Thinking insists on a harder question — the one that turns a mediocre deal into an obviously bad one, or reveals a genuinely good one: compared to what?

Every public dollar committed to one deal is a dollar that can’t be spent on another. Economists call it opportunity cost, and it is the most clarifying question in the entire framework. So let’s ask it directly. What else could roughly $210 million in public incentives have done for Iowa?

The Chipotle Test

Start with the most uncomfortable comparison in the whole analysis. Not a hospital, not a school — a single Chipotle.

One Chipotle, zero incentives, opens in 90 days — and across all three ripple waves delivers equal or better community value than a $210M data-center tax break.

One Chipotle opens in about 90 days, takes zero public incentives, and, across all three ripple waves, matches or beats the $210M data-center package in community value. Trace the waves.

• First-order (financial): 22–30 permanent jobs, $0 cost to Iowa, open in 90 days.

• Second-order (behavioral): sales and payroll tax from day one with no abatement, supplies bought from local vendors.

• Third-order (systemic): real career ladders — crew to shift lead to general manager to multi-unit operator — wages re-spent in town, and “sticky” capital that follows the community rather than the lowest tax bill.

The point isn’t that Iowa should go chase Chipotles. It’s that if one unsubsidized fast-casual restaurant can match a $210 million incentive package on community value, then the incentive wasn’t buying community value. It was buying a headline. Scored head-to-head across ten criteria, the local model wins eight: jobs per public dollar, workforce ladders, local tax activity, infrastructure-strain risk, diversification, community integration, environmental load, and compounding multiplier. The data center wins exactly two: immediate capital scale and political visibility. The two it wins are the two that look best at a ribbon-cutting.

Three Better Uses of $210 Million

Of course, Iowa wouldn’t actually spend $210M on burritos. The real comparison is against Iowa’s documented needs — and on that test, the same money has at least three uses that beat the data-center deal on jobs per dollar, accessibility, and compounding value.

The same $210M invested in Iowa’s most urgent needs: rural healthcare, workforce training, or public education — each beating the data-center deal on accessible jobs and long-term value.

• ~42 rural primary-care clinics: 500–750 permanent jobs (nurses, MAs, admin) at a tax cost of $280K–$420K per job — ten to fifteen times lower than the data-center deal — directly addressing a healthcare desert spanning 73 counties and 58 shortage areas.

• ~35 workforce training centers: 7,000–9,000 Iowans trained every year — electricians, HVAC techs, welders, cybersecurity — at $23K–$30K per worker, building the very workforce future data-center deals will need.

• • Iowa K–12 public education: ~2,100 educator jobs and a $7–$12 lifetime return on every $1 invested — the highest ROI of any public investment (the Heckman research) — benefiting every district, grade, and community in the state.

Each one addresses a documented Iowa crisis. Each creates broadly accessible jobs. Each compound economic and community value across a generation. The data-center deal, graded honestly in Blog #3, does none of the three.

This Is Not Anti–Data Center

Here is where I want to be crystal clear, because it is the whole point of the series: none of this is anti–data center. Data centers offer genuine economic advantages — when the deal is structured to consider every stakeholder’s needs.

Data centers offer real advantages — large capital, grid and fiber investment, clean-energy catalyst — but only when the deal is structured to deliver them.

The upside is real: $1B-plus projects that expand the commercial tax base once abatements end, grid and fiber investment that benefits the whole region, tech-ecosystem signaling, and — when power-purchase agreements are required — a genuine catalyst for Iowa wind and solar. Iowa’s renewable energy is precisely why these companies are drawn here. We should double down on that.

The catch is the word if. Those advantages are real only when the deal is structured to deliver them — cost-causation enforced, jobs verified with clawbacks, local supplier spend required, a net-new tax base confirmed, and environmental accountability built in. Waukee earned a D+, not because Apple did anything wrong; Apple negotiated brilliantly. Iowa earned a D+ because it negotiated like a politician chasing a headline rather than an economist protecting a community.

What Iowa Should Demand Next Time

What does a genuine win-win require? Make the conditions inseparable from the incentive — three of them, non-negotiable.

The three conditions that turn a headline-grabbing giveaway into a deal that works for all Iowans.

1. Pay for what you consume — and build what Iowa will need anyway. Grid, water, and road upgrades caused by the data center are funded by the developer, not shifted onto Iowa ratepayers and farmers. A good negotiator goes further: size and site that spending so it does double duty — strengthening the grid for surrounding communities and advancing the renewable buildout Iowa is already betting on. The developer pays for what it consumes; Iowa makes sure what gets built is a lasting asset, not just an accommodation.

2. Create broad-based economic participation — measured in more than paychecks. A floor of permanent, high-paying Iowa jobs, a minimum share of operating spend with Iowa vendors, and defined career ladders, apprenticeships, and internships. But participation is bigger than payroll: count the second- and third-order returns too — waste-heat recapture, water and watershed stewardship, renewable commitments, and community benefit funding tied to local healthcare and education. Money that doesn’t circulate in Iowa doesn’t count — and neither does “growth” that quietly degrades the health, water, and land Iowans depend on.

3. Deliver net-new tax value within a decade — and prove it independently. Independent fiscal analysis must confirm the deal generates more than it costs within ten years, with performance-based clawbacks if job or revenue targets are missed. No “but-for” assumptions, no projections the developer scores itself — the burden of proof sits with the party asking for the incentive, and verification is continuous, not a one-time press release. If the value doesn’t materialize, the incentives don’t either. Prove it or lose it.

That is the difference between a deal that works brilliantly for Apple and a deal that also works for all Iowans. The conditions aren’t hostile to business — they’re how you build a partnership that lasts longer than the ribbon-cutting.

Think Like an Economist

That is the discipline this whole series has been building toward. Don’t think like a politician — get the company in, cut the ribbon, take the photo. Think like an economist: who captures the value, who bears the cost, compared to what, and does it last? The goal was never to send the capital away. It’s to make sure that when it arrives — and in Iowa, it is arriving, in town after town, right now — it works for everyone who paid to attract it. That is a real win-win.

COMING NEXT — BLOG #5
Readers keep asking how we actually did this analysis — where the research came from and how we pressure-tested it. The finale pulls back the curtain: the credible sources, the expert “personas” we consulted, and a value scorecard you can run on the deal in your own town. Because the best answer to “compared to what?” is a tool that lets you ask it yourself.

This is Blog #4 of The Importance of Economic Critical Thinking. Data centers can be a win for Iowa — but only if it negotiates as if it has something to protect. Because it does.
— Bill Schmarzo, The Dean of Big Data

This is Blog #3 of a multi-part series on the critical importance of Economic Critical Thinking—why leaders and policymakers must think beyond headlines, budgets, and ribbon-cuttings to understand how value is actually created, distributed, and compounded over time.

In Blog #2, we built the method: gather the peer-reviewed research, frame all nine dimensions of value, map the economic ripple effects, and apply the local-business test. A method, though, is only worth something if it survives contact with a real deal.

So let’s grade one. The Waukee/Apple data center — $213M in public incentives, 50 permanent jobs, a 20-year property-tax abatement — is the most consequential economic-development deal in recent Iowa memory. We’re going to run it through the method and see what grade it earns.

And we’ll hold it to the right standard. As one sharp reader put it after Blog #1: the interesting question isn’t whether $4.26M per job looks high — it’s whether, over a 20-year horizon, the total value created exceeds the total public and private resources committed. That is the question a grade has to answer.

Who Captures the Value, Who Bears the Cost

The method’s first move is to score the deal across all nine value dimensions, asking, for each one, who captures it and who bears it? Run Waukee through that lens and a stark pattern appears — the same asymmetry, dimension after dimension.

• Financial value flows in one direction. Apple secures a $213M incentive package and a 71% / 20-year abatement underwriting a $1.375B campus — the second-highest per-job subsidy in U.S. data-center history. Iowa books roughly $3.8M a year in property tax against revenue it forgoes for two decades.

• Workforce and societal value barely move. This is where broad public benefit would show up — and it barely registers. Fifty permanent jobs, specialized and low-accessibility, in a state of 3.2 million people, with no defined career ladders to carry local workers into them.

• Environmental value runs negative. A grid- and water-intensive operation with no binding cap protecting ratepayers or the watershed — a cost borne by households and the environment, not the developer.

• Operational and ecosystem value leave the state. The reliability and ecosystem gains accrue to Apple’s global cloud, not to Iowa’s local business base or supplier network.

On dimension after dimension, the verdict is the same: value captured privately, costs borne publicly. That’s not a moral complaint — it’s a distribution finding, and it’s invisible until you score all nine dimensions instead of the one in the press release.

Does the Value Last?

The dimensions indicate where the value stands today. The ripple effects tell you whether it lasts — across three waves: first-order financial signals, second-order behavioral responses, and third-order systemic effects. Trace Waukee through all three, and the independent research on data-center economics tells a strikingly consistent story.

• First-order (financial). The immediate, measurable layer: Apple committed $1.3 billion in capital and projected 500-plus construction workers; construction is where the financial impact is concentrated (Council of State Governments; Baker Tilly). But the “500 jobs” headline hides that data center construction is a specialized, mission-critical trade dominated by a handful of national firms that deploy traveling crews — importing superintendents, electricians, and craft workers from other states because local pools can’t supply the talent (The Birmingham Group; Build.inc). As a University of Wisconsin analysis of that state’s boom put it, many of these workers travel to the area and don’t stay long term. So even the financial value Iowa captures is the thinner slice — some general trades, plus the hotels and restaurants serving the out-of-town crews. Real, but temporary.

• Second-order (behavioral). After construction ends, the permanent footprint is remarkably thin — a review of 1,200-plus U.S. data centers found even the largest employ fewer than 150 workers, sometimes as few as 25 (World Resources Institute), and Brookings notes data centers are among the least labor-intensive structures in the economy. Apple’s 50 jobs sit squarely in that range — and they’re not the high-paying professional roles the pitch implies. The deal floors pay at $29.12/hour, about $60,000 a year: well below what local doctors, dentists, lawyers, or senior engineers earn. The behavioral catch is where those wages go — some of these workers likely commute in from the Des Moines metro and re-spend outside Waukee, so the dollars that would normally circulate locally don’t. Good Jobs First finds no evidence that these facilities deliver the ancillary development boosters they promise.

• Third-order (systemic). Over time, the structures the deal created reshape the local system. Generous data-center incentives often reduce net fiscal benefits and shift costs onto residents and schools (Baker Tilly), and Waukee is now a textbook case. Apple’s 20-year abatement floors each data center’s taxable value at roughly $52 million; when Dallas County corrected a valuation error in 2025 to apply it, Waukee’s schools and city came up a combined $5.7 million short (Waukee CSD; City of Waukee). The district shelved planned hires — math specialists and college-and-career coordinators — for a student body growing by 400 a year. A 20-year abatement structurally constrains the tax-base growth that funds the system itself.

The point is that a 20-year abatement limits the tax-base growth that funds schools and services — a pattern seen well beyond Iowa. Virginia forewent $1.6 billion in data-center tax breaks in FY2025, and a PwC study in Pennsylvania showed data centers contributed less in tax revenue than they added to residents’ electric bills (Ohio River Valley Institute).

The pattern is the opposite of what compounding value looks like. For Iowans, the deal doesn’t compound over time — it dissipates.

The Report Card

Put the dimensions and the ripple horizon together and you don’t get a single number — you get a report card, graded by who captured the value and who bore the cost.

Grading the Waukee/Apple deal across stakeholders: an excellent strategic outcome for Apple, a weak, broad-based return for Iowans.

Capturing the most value:

• Apple global operations (A). Incentives and a 20-year abatement underwrite a $1.375B campus serving Apple’s North American cloud. No clawback provisions.

• Waukee amenities & branding (B). Up to $100M in public improvements, Triumph Park, and tech-hub branding — real, visible community wins.

• Construction trades (B-). Hundreds of temporary construction jobs and site-prep spending — meaningful, if short-term, and not all of it Iowan.

Bearing the most cost or forgone value:

• Iowa taxpayers & broad public (D+). Called a “guaranteed loser” for taxpayers by independent economists, at $4.26M per job, with no clawback if targets are missed.

• Schools & local budgets (D). A FY2025 valuation correction cut schools and city revenue by a combined $5.7M; programs were canceled, with no penalty triggered on Apple.

• Residents & local services (C-). A 20-year abatement constrains near-term tax base growth, and future utility rate exposure from grid load lacks a binding cap protecting households.

Overall public-value grade, to date: D+.

The Verdict

Now answer the reader’s question directly. Over the 20-year horizon, did total value created exceed total resources committed? For Apple, overwhelmingly yes — an A deal. For Iowans, the total public value created does not clearly exceed the total public resources committed — a D+ deal. Same transaction, opposite grades, depending entirely on whose ledger you hold. That asymmetry — not the per-job number — is the finding.

One honest caveat: this is a provisional grade based on publicly reported incentives, jobs, abatements, and fiscal effects — not a forensic net present value audit. The point was never the exact letter. The point is that a disciplined, multi-dimensional, time-aware evaluation reaches a fundamentally different — and far more useful — answer than the headline ever could.

Next up is blog #4, the Grand Finale. A grade evaluates “was this deal good?” but Economic Critical Thinking asks a tougher question: compared to what? Every public dollar spent here isn’t available for other uses. Blog #4 concludes the series by focusing on opportunity cost—perhaps one Chipotle, three better ways to spend the same money, and the essential conditions Iowa should set before approving the next deal.

This is Blog #3 of The Importance of Economic Critical Thinking. A headline tells you a deal happened. A grade tells you who it happened for.

— Bill Schmarzo, The Dean of Big Data

This is Blog #2 of a multi-part series on the critical importance of Economic Critical Thinking—why leaders and policymakers must think beyond headlines, budgets, and ribbon-cuttings to understand how value is actually created, distributed, and compounded over time.

Here is a conviction that runs underneath this entire series:

Markets reliably create surpluses. They make no promise whatsoever about who captures those surpluses, who bears the costs, or whether the gains circulate through Iowa’s communities or get repatriated to a balance sheet somewhere else. Creating value and distributing value are two different processes. The market only does the first. The second is decided by policy, incentives, contracts, and governance — by whoever structures the deal.

And that second question is exactly the one most economic-development debates skip. It measures inputs: how much capital was invested, how many buildings were built, and how big the announcement was. Economic Critical Thinking measures outcomes instead — who captures the value, who bears the cost, how broadly it circulates, and whether it compounds or dissipates over time. Those are not the same questions, and they routinely produce opposite conclusions.

The hard part isn’t intelligence. The smartest, best-advised rooms in history still approve lopsided deals — not from stupidity, but from confidence: a belief gets agreed with enough times that it starts to feel like proof, even when its foundation is agreement rather than evidence. (Mark Stouse calls this the gap between confidence and validation.) The fix isn’t smarter people in the room. It’s a method that forces the assumptions to prove themselves before the capital moves — and that method has four steps.

The Method: Four Steps That Force the Question

That method is Economic Critical Thinking. Four steps, each one designed to drag the value-distribution question into the open — who captures the value, and who bears the cost — before the capital moves and the assumptions get expensive to change.

The four-step discipline for evaluating what Iowa is trading away — and what it should demand in return.

Step 1 — Gather peer-reviewed research

Replace “smart people are nodding” with non-partisan evidence. What does the independent research actually say about deals like this one — not the booster’s slide deck, not the lobbyist’s model? This step interrupts belief before it hardens into confidence.

Step 2 — Frame the full economic value

This is the value-distribution tool. Value isn’t one number; it shows up across ten distinct dimensions. For each one, you ask the same two questions: who captures it, and who bears the cost?

The nine value dimensions cover the full spectrum where value is captured or borne, combined into a single ‘AI Utility Function.’ These include financial, operational, customer, employee, partner/ecosystem, societal, environmental, workforce, and ethical values. An honest assessment blends all nine, but politicians usually set eight weights to zero, focusing only on financials. Economists evaluate all nine, revealing the actual distribution of value.

Step 3 — Assess the economic ripple effects

Value isn’t static; it moves through time in waves. Mapping those waves is how you tell whether a deal that looks good today is still good in Year 10.

First-, second-, and third-order ripple effects illustrate how an initial investment propagates and indicate whether its value grows or diminishes. First-order effects are direct, measurable outcomes such as jobs, tax revenue, and spending, ending with the politician’s announcement. Second-order effects involve behavioral changes, such as local spending, sourcing, and community trust. Third-order effects are systemic, including career growth, capital flow, ecosystem strength, and community impact, revealing long-term benefits or risks.

Call-Out: Why these two belong together
The nine dimensions tell you where value can be captured or borne. The ripple effects tell you whether it lasts or decays. One is the map, the other is the clock. Used together, they are the foundation of Economic Critical Thinking — and the reason a disciplined evaluation reaches conclusions a headline never can.

Step 4 — Apply the local-business test

Finally, the adversarial pressure the confident story never gets. Measure the deal against concrete alternatives — what the same public dollars would deliver if invested directly in Iowa’s communities — so you’re choosing among options rather than rubber-stamping one corporate proposal in isolation. (We’ll put this test to work, by name, later in the series.)

The Question Underneath All Four Steps

Strip the method down to its core, and you’re left with a single subtractive question:

“What has to be true for this to be a good deal — and have we actually established that those things are true, or have we merely assumed them and surrounded the assumptions with enough agreement to make them feel like facts?”

That is the whole discipline. A method that makes confident reasoning prove itself, before the capital moves, while the assumptions are still cheap to change. Skip it, and you get exactly what the headlines deliver: deals that look like wins at the ribbon-cutting and grade out very differently a decade later.

In blog #3, with the method ready, we move from theory to practice by analyzing the Waukee/Apple data center. We evaluate it across all nine dimensions and the entire ripple horizon to determine its grade. Spoiler: it’s not an A.

This is Blog #2 of The Importance of Economic Critical Thinking. Markets create value brilliantly. Whether that value ever reaches the people who paid for it is a choice — and choosing well takes a method.

— Bill Schmarzo, The Dean of Big Data

Share

There is a number buried in one of Iowa’s proudest economic development announcements that should stop every policymaker, taxpayer, and citizen cold.

$4.26 million.

Iowa paid over $213 million in incentives for 50 jobs at the Apple data center in Waukee, making it the second-highest per-job subsidy for a U.S. data center.

Now, here is the question that separates an economist from a politician:

Is that a good deal?

The politician already knows the answer. The ribbon was cut. The press release went out. “This puts Iowa on the world stage.” The photo was taken, the applause died down, and everyone moved on before anyone asked what the deal would actually cost in year ten.

The economist cannot simply move on. They must confront the difficult, unglamorous question that ruins the moment:

What have we sacrificed to achieve this, and who will bear the cost for the next 20 years?

That is the essence of Economic Critical Thinking.

Economic Critical Thinking is the discipline of evaluating not only who benefits from a decision, but who bears the costs, who assumes the risks, what alternative opportunities were sacrificed, and whether the value created compounds or dissipates over time.

This series highlights the importance of asking the right questions and honestly answering them before investing, not after. It’s about current and future projects like data centers, semiconductor plants, or battery factories, shaping Iowa’s economic growth over the next decade. Today’s decisions will influence Iowa’s investment and how much it benefits its people.

The Invisible Hand has a Blind Spot

Let me start with a conviction that runs underneath everything in this series:

“The economic invisible hand is excellent at creating phase-shift surpluses. It sucks at distributing them without institutional help.”

Markets are extraordinary engines of value creation. A $1.3 billion data center campus represents real capital, real construction, real infrastructure, and real economic activity. None of that is in dispute.

But markets make no promises about who captures that value, who bears the costs, who assumes the risks, or whether the resulting surplus circulates through Iowa communities or gets repatriated to a corporate balance sheet in Cupertino. That distribution question is the blind spot.

It is also where Economic Critical Thinking begins.

The invisible hand is highly efficient at generating wealth but does not influence how that wealth is shared. Instead, distribution outcomes depend on policies, incentives, contracts, and governance. In essence, creating value and distributing it are separate processes.

Unfortunately, most economic development discussions focus almost exclusively on inputs rather than outcomes. How much capital was invested? How many buildings were constructed? How large was the announced project?

Economists care about those metrics, but they care even more about what happens next: Who captures the value? Who bears the costs? How broadly does the value circulate? And does that value compound over time?

These questions often result in very different conclusions. This distinction is more important than ever today, as the next wave of data centers, AI infrastructure projects, semiconductor facilities, and advanced manufacturing investments is already arriving in Iowa.

Winning organizations won’t always be those that provide the largest incentives. Instead, they will be the ones who craft deals to optimize long-term value for all parties involved.

🧠 Call out: Value Creation vs. Value Realization
Perhaps the most important lesson in Economic Critical Thinking is that value creation and value realization are not the same thing. Markets excel at creating value. Economists focus on who realizes that value, who bears the costs and risks, and whether the benefits compound over time.

Iowa at a Technology Crossroads

Iowa is actively and rightly working to attract technology investment. Nonetheless, the challenge is in how these investments are structured. It’s no longer in doubt that Iowa aims to become a technology hub; the key issue is whether the incentives we provide generate the most long-term benefit for those who finance them.

Economic development is frequently portrayed as a race to attract investment. However, an economist views it from a different perspective.

Economic development is a negotiation over value.

Every incentive package, tax abatement, infrastructure commitment, and regulatory concession represents a trade. Iowa gives something up today in exchange for benefits promised tomorrow. The quality of the deal depends not on the size of the investment, but on the value Iowa receives in return.

That puts every future technology investment at a fork in the road with three questions attached:

• What did Iowa get from the last deal? 50 permanent jobs, a seven-year wait, no penalty provisions, no infrastructure cost caps — $4.26M per job.

• What does Iowa need from the next one? Verified jobs. Shared infrastructure costs. Enforceable wage floors. Local economic circulation. Environmental accountability. Outcomes that are quantifiable and transparent for all Iowans.

• What framework should guide the decision? An economic value framework that accounts for citizen, workforce, industrial, environmental, and societal dimensions — multiple dimensions of value, not just one headline number.

Most debates on economic development rarely go past the initial question. Even when they do, they typically respond with the most convenient metric for a press release: the magnitude of announced capital investment. However, capital investment is merely an input, while economic value represents the actual outcome.

This series focuses on questions two and three because that is where Economic Critical Thinking begins and where the real economics lives.

But before we can build a better framework for evaluating future deals, we first need to be honest about the last one.

The Anatomy of a $213 Million Deal

Iowa aimed to establish a technology hub, and seemingly achieved it. A trillion-dollar company now bears Iowa’s name. However, “we landed Apple” reflects a political victory rather than a direct economic benefit.

The key economic question is whether the deal was designed to generate lasting value for Iowa’s citizens, taxpayers, businesses, and communities—and how that influences future deals. Economists are skilled at assessing both aspects of the ledger.

Here is the full ledger.

What Iowa Gave Up

• More than $213M in total incentives — $188M in property tax abatement (71% over 20 years), $20M in state credits, and $6M in water and sewer support.

• 2,000 acres of Iowa farmland annexed and dedicated to Apple’s campus.

• Twenty years of constrained tax revenue — schools, county, and city services operating on a reduced tax base until roughly 2037.

• A grid and water burden with no binding ratepayer protections.

• A seven-year construction delay — announced in 2017, restarted in 2022, opened in 2024 — with no penalty provisions.

What Iowa Got in Return

• 50 permanent jobs at $29/hour and above — specialized, low-accessibility roles at a public cost of $4.26M per job.

• $1.3B in capital investment dedicated to Apple’s operations.

• Roughly 500 temporary construction jobs (many of them not to Iowans) and approximately $3.8M per year in property tax revenue — a fraction of the abated revenue.

• Up to $100M in a public improvement fund supporting parks, roads, and Triumph Park.

• Bragging rights. One independent economist called the deal “reckless.”

Observe these two columns side by side, revealing a pattern that recurs throughout this series: benefits are focused, immediate, and prominently highlighted, while costs are spread out, delayed, and quietly absorbed.

That asymmetry isn’t accidental. It comes from judging deals by their initial appearance rather than long-term impact. Politicians celebrate announcements, while economists evaluate results.

The Question that Actually Matters

I want to clarify that this series doesn’t suggest data centers are harmful or that Iowa shouldn’t pursue them. That kind of thinking would be a lazy simplification. The real question was never:

Should we?

The right question is:

How do we design this deal to compound value for all constituents over time?

That is a different question requiring a different Kind of Thinking: disciplined economic reasoning that gathers evidence, maps ripple effects, compares options, and honestly grades outcomes, even when embarrassing. We’ll build this discipline gradually.

Before creating the antidote, we need to understand the disease: how intelligent, well-informed, and well-meaning individuals end up signing deals like Waukee initially. Because the troubling part is that intelligence wasn’t the problem.

The people who approved this deal had access to more data, more models, and more credentialed advisors than almost any generation of decision-makers in history. The problem was confidence without validation.

In Blog #2, I’ll show you the exact failure pattern—the one that just erased $640 billion in AI infrastructure assumptions almost overnight, and the one that quietly played out in Iowa at a smaller scale.

It’s called the Belief → Confidence → Validation cascade.

And once you see it, you cannot unsee it.