Atiba de Souza.Start a conversation
WritingAI

Nvidia May Have Engineered the Shortage. Your Constraint Is Still Somewhere Else.

By Atiba de Souza

Atiba de Souza standing outdoors, leaning against a metal pole. The cover reads: AI Nvidia May Have Engineered the Shortage. Your Constraint Is Still Somewhere Else.

Ten minutes before the training

I had eight hours of teaching in front of me. Ten minutes before we started, my phone lit up.

I built a crypto scanner inside ChatGPT on purpose. Not to trade. To watch how the thing handled its own rules. That morning it told me to buy LINK at $11.20.

Coinbase said $11.31.

So I pushed back. It folded instantly. Oh yes, it said, it is no longer time to buy, my data was stale. I went back and checked for myself. $11.20 was the price at 3 a.m. It was 9:40 in the morning.

The scanner has a written rule. Before you report, check the live price. The rule was in the prompt. The rule was not in the machine. It broke its own instruction and told me to go buy something, in a completely certain voice, six and a half hours late.

$11.20 against $11.31. Six and a half hours stale. Reported with total confidence. A written rule is not an enforced constraint. The gap between those two things is where money goes to die.

That is the whole article in one alert, and none of it is really about crypto. Every operator I meet right now is buying on a stale price and calling it a strategy.

The allegation, and why it does not change your Monday

The story going around is that Nvidia helped fund AI startups that then spent the money on Nvidia chips, and that the shortage everybody has been paying through the nose for was, at least partly, engineered.

I cannot verify that. I am not going to pretend I can, and I am not going to tell you what to think about Nvidia. Verifying it is not my job and it is not yours. I am going to show you why the answer changes nothing about your Monday.

Because whether the scarcity is real or manufactured, you have the same job, and most operators were already failing at it before this headline landed.

Scarcity is a pricing strategy. A constraint is a fact. Confuse the two and you will spend a fortune routing around a wall nobody built.

One is a story about what a seller wants you to feel. The other is a fact about where your work actually stops moving. The first gets a press cycle. The second gets ignored, because it is internal, it is boring, and it usually means you were wrong about something you already paid for.

Binding constraint: the one thing in your business that everything else is currently waiting on. Move it and the whole system speeds up. Spend anywhere else and you will feel busy while the number does not move.

If compute were free tomorrow

Step one of the sequence is Constraint. Theory of Constraints, one question: what is the single thing everything else is waiting on right now?

You do not need a consultant for this. Ask one question instead.

If compute were free tomorrow, what in my business would still not move?

Whatever your answer is, that is your constraint. And I will bet you a coffee it is not a GPU.

Now here is the part the frameworks leave out. You will not want to ask that question. Not because it is difficult. Because it grades your last twelve months. There is a real chance the answer comes back and the thing you already bought, the seat you already pay for, the workflow you already built, sits nowhere near the constraint. That is a genuinely unpleasant report to receive about yourself.

Underneath it sits the louder fear, the one vendors farm on purpose: if I do not buy now, I get left behind. That fear is real. It is also being sold to you, and it is doing exactly the job it was hired to do.

So shrink the question until the fear is proportionate. One workflow. Not the business, not the department. One workflow, this week.

The six steps, and why build is last

This is the Agentic AI Diagnostic Sequence. Six steps.

  1. Constraint. Find the binding constraint using Theory of Constraints.
  2. Role. Place the work in your org structure. Which chair is this?
  3. Work Type. Classify it as Strategic, Tactical, or Analytical, then assign each slot to a human or to AI.
  4. AI Employee. Define the employee on the Diagnostic Canvas.
  5. PRD. Write the requirements before anything gets built.
  6. Build.

Build is step six. That is not an accident. Most people start there.

Before you org chart anything, run two filters between step one and step two. TOC asks where the leverage is. Eisenhower asks what the leverage of that work is: urgent, important, put it in a quadrant. The not-important-not-urgent quadrant exists so you can eliminate work rather than automate it. Automating waste is the most expensive way to keep waste.

1. Constraint (TOC): what is everything waiting on?

Triage: TOC asks WHERE, Eisenhower asks WHAT, then eliminate

2. Role: which chair in the org chart?

3. Work Type: Strategic / Tactical / Analytical, then Human or AI

4. AI Employee: define the employee on the Diagnostic Canvas

5. PRD: requirements before build

6. Build

Running my own scanner through steps two, three and four

I can teach steps two through four in the abstract, but the abstract is where frameworks go to die. So run the scanner through them.

Role. Which chair does this thing occupy? Not "does it help me." Which chair. The scanner does not get the chair marked decide. It gets the chair marked watch and report. I placed it correctly and then used it incorrectly. I put a monitor in a chair and treated its output as a decision. A monitor trusted as a decider is not a monitor. That is not a labeling problem, it is a design problem, and you can fix it on paper in about ten minutes.

Work Type. Watching a price against a threshold is analytical work, and analytical work is where AI earns its keep. Pressing buy is a decision with money attached, and that slot stays human. Say the slots out loud before you build, because my scanner did not fail at analysis. It failed at the boundary.

AI Employee. Define the employee on the Diagnostic Canvas before you write a prompt. The point of writing it down is that a definition you can read is a definition somebody else can check. My rule about checking the live price was a line inside a prompt, which made it a wish with good formatting.

And step four is where the cost shows up, so let me say it plainly, because the framework does not.

Defining the employee forces one question: who checks the output? If the answer is nobody, you have not hired an employee, you have handed your passwords to a confident stranger. If the answer is you, you have just taken on a second job, and the speed you thought you were buying evaporates into review time. Both of those are real costs. Neither of them appears in a demo.

I am not against the build. I am against skipping to it. The build was never the expensive part. The checking is the product.

Darvin, and the least likely person in the room

At one of my recent US AI events, Darvin, a complete beginner, asked me a fair question. Would a coffee shop owner who does not know what AI is even be a target for an event like this?

I told him about a man who had already come. His entire experience with AI, in his own words, was signing up for the course. Another attendee that year turned up with a brand new computer and asked how to turn it on.

Judged on where they started, those were the least likely people in the building. The obvious target market was people who were already using AI.

Two days later, the beginner put a number on it. He said he could see how this was going to make him a hundred thousand dollars this year.

Now sit with the part that matters. That number did not come from the thing they built in the room. It came from the rewiring of how he thinks about solving problems.

And rewiring charges a fee. Not a ticket price. The fee was sitting in a room and saying out loud that his entire experience with AI was signing up for the course. That admission is the price of admission, and it is the exact thing step one asks of you.

The compute was not the constraint. The model was not the constraint. How he was thinking was the constraint. Two days of rewiring moved it further than a hardware budget would have. The beginner was the best investment in that room, and almost every operator in my industry would have screened him out.

The price panic is a rerun

Now the part of the Nvidia story that lands on your wallet.

Every new technology is subsidised on the way in, and prices level out once the masses arrive. The internet is the proof. Its cost barely rose while its speed multiplied. Same pattern, different decade.

So when someone tells you token prices are going to the moon and you had better lock in capacity now, ask which chapter of the book you are in. Early chapters are expensive and strange. Later chapters are cheap and boring. Doomsday pricing predictions ignore how every earlier technology priced out, and they ignore it on purpose, because panic converts better than patience.

Rank the parts, and when to skip this

Treating every step as equally important is how a framework becomes theatre. So here is my ranking.

Step one, Constraint, is most of the value. If that is all you ever do, you got the return. Steps two through four, Role, Work Type, AI Employee, are where the design gets real and they earn the afternoon. PRD and build are the cheap part. Build is easy once you know what you are building, and ruinous when you do not.

When not to do this: if you are a solo operator with one repeated workflow and no team to place, do not run a six step architecture. Find the one thing you do every week that should be done by someone else and design that. The sequence is for systems. A single task does not need a system.

The inconsistent moveThe consistent move
Buy capacity first, then hunt for a problemFind the binding constraint, then buy only what moves it
Ask what competitors are doing with AIAsk what still would not move if compute were free
Treat a vendor's scarcity claim as a factTreat it as an output and run a second check
Trust a rule you wrote into a promptEnforce the check with a different model or a human

Which brings the scanner back to step one, where it belongs. Pair every AI-generated output with a check run by a different model or a human being. Never treat an instruction written into a prompt as an enforced constraint. When I challenged my scanner it did not defend its position. It folded instantly. A source that folds the moment you push back was never checking anything.

Confidence is not evidence. Not from a model, and not from a chip company.

Questions

What people ask next

Okay but how do I actually find the binding constraint when it feels like three different things are stuck at once, not just one obvious bottleneck?
Run the same question against each of the three: if compute were free tomorrow, would this one still not move? Two of them will loosen on their own once you imagine resources away. Whichever one stays stuck no matter what you throw at it is the one everything else is actually waiting on, and that's the one you move first.
If defining who checks the output always lands on me, doesn't that just mean the AI employee never actually saves me time, no matter how well I design the role?
That depends on which chair the work sits in. If you're checking a decision, that's correct and will always cost you time, because decisions with money or consequence attached stay human regardless of the tool. But if you're checking a monitor's report the same way you'd check a decision, the problem isn't the AI, it's that you never drew the line between watch-and-report and decide, so you're paying review time for work that should have needed none.
How do I tell the difference between a real 'I'll get left behind' risk and the manufactured version, when they feel exactly the same from where I'm sitting?
Shrink the question until the fear is proportionate: one workflow, this week, not the whole business. If the fear survives being asked "if compute were free tomorrow, would this specific thing still not move," it's pointing at a real constraint. If it only survives as a vague sense that everyone else is already doing this, that's the manufactured version doing the job it was hired to do.
Doesn't spending five whole steps before building just become its own form of procrastination dressed up as rigor?
No, because the steps are fast, not slow, the whole sequence is meant to run in minutes, not weeks. Procrastination is avoiding a decision; this is making the decision about what to build before you've spent money building the wrong thing in the wrong chair. The expensive mistake isn't the five steps, it's skipping straight to build and finding out afterward that nobody defined who checks the output.
What happens when the constraint I find isn't a workflow or a chair but a person who doesn't want their job redefined?
The diagnostic sequence classifies work and chairs, not people, so you still answer it honestly: place the work, type it as Strategic, Tactical, or Analytical, and assign the slot. What you do with a person who resists that reassignment is a separate, harder problem the framework doesn't solve for you, but naming the constraint accurately is still the right first move, even if acting on it is uncomfortable.