I tested an $8,000 AI decision for $5.

I had an $8,000 decision in front of me. And I'll be honest: before I tested a single thing, I was already becoming a believer.

The pitch was good. The momentum was real. The case for just committing the capital kept getting easier to make in my own head. I advise companies on exactly these decisions for a living, and I could still feel myself getting pulled toward "yes" before I had any evidence that yes was the right answer.

Then the cost actually settled on me. Eight thousand dollars. Yuck. Is this even worth it?

That's a version of the question I hear constantly in AI governance conversations: "We want to do AI. We got a quote. It's a lot of money. How do we know it'll work for us before we sign?" In my experience, that question often goes unanswered until after the contract is signed. The quote might be perfectly valid. My advice: test it against your actual work before you commit.

So I did the thing I tell clients to do. Instead of guessing, I tested it first.

What I actually did

I took a frontier open-weights AI model and I rented the exact graphics hardware needed to run it, by the hour, with zero commitment. No contract. No purchase. I loaded my real workload, measured how fast it ran, how much memory it needed, what it cost per task, and — most importantly — whether the output was actually good enough to use.

Here's the ledger:

And it wasn't what I expected. The cheaper option I'd half-assumed would be fine? Couldn't even load the model. Wrong tier entirely. The shape of the number I'd been quietly building a case around in my head turned out to be wrong.

$4.81 told me three things the quote I had did not. Which hardware tier the workload actually needs. What it would really cost to run. And whether the output was any good — which is the question that actually matters. It was. The model read a messy operations log and pulled the real problem out of it on the first try.

That ratio is the whole point. Here, the test cost $4.81. The cost of getting it wrong — buying the wrong hardware, signing the wrong contract, deploying a model that isn't good enough — would have been far larger. Test before you commit.

And whether the number on the quote is $8,000, $80,000, or $800,000, the principle holds: test the workload before you commit the capital.

The playbook

This isn't proprietary. Here's the method, and you can run it yourself:

Why this matters right now

Every company is being told they need to "do something about AI," and a healthy industry has grown up to sell them expensive somethings. Some of it is genuinely worth it. A lot of it is unvalidated capital outlay dressed up as strategy.

The companies that win the next few years won't be the ones that spent the most on AI. They'll be the ones that tested cheaply, learned fast, and committed capital only where the evidence actually pointed. My test cost less than lunch.

Here's the part that surprised me most. After testing, I didn't feel worse about the decision. I felt better. Not because the answer came back yes or no — because I finally knew instead of hoped. A good pitch can make you a believer. Only evidence makes you confident. Going from "I think this is right" to "I know what this costs and whether it works" was worth a lot more to me than $4.81.

I prefer to test AI infrastructure against a real workload before recommending it. That's not a slogan. It's a $4.81 receipt.