What do you think is actually inside the trucks driving around your local AI boomtown? If you guessed pallets of GPUs worth more than a house, you’re thinking like a thief in Fremont, California. And like those thieves, you’d be wrong.
Here’s what happened, stripped of drama: someone hooked their own cabs up to two trailers at a Fremont warehouse and drove off. The trailers belonged to PlusAI, a self-driving trucking company. Whoever took them broke the trailers open, presumably expecting microchips, and found 20 tons of sand. Ballast. Test weight for autonomous trucks. The trailers were later dumped. No arrests have been made.
That’s the whole story. Forty thousand pounds of nothing, abandoned by criminals who did less due diligence than most people do before buying a used monitor.
Why I can’t stop thinking about this
I review AI tools for a living, which means I spend most of my week opening boxes that promise chips and finding sand. Not literally. But the pattern is identical: something is branded in a way that implies enormous value, someone commits real effort to acquiring it, and the payload turns out to be inert filler that serves a testing purpose and nothing else.
The thieves in this story made a very specific, very modern mistake. They assumed that proximity to AI meant proximity to treasure. Self-driving trucking company, Bay Area warehouse, unmarked trailers — the inference writes itself. Chips are the hottest commodity of the decade. Of course that’s what’s in there.
Except that a self-driving trucking company spends most of its physical resources on the boring part. You can’t validate how a truck brakes, corners, or handles a loaded axle without actual load. Sand is cheap, dense, and doesn’t shift in ways that ruin your data. It is the least glamorous possible input to a wildly glamorous-sounding business, and it is absolutely essential.
The unglamorous middle is where the work lives
This is the part that gets lost in AI coverage, including coverage on sites like this one. The interesting layer is never the layer that markets itself. Nobody pitches investors on sandbags. Nobody puts “we bought a lot of dirt” in a launch post. But if PlusAI’s trucks eventually work, a nontrivial share of the credit belongs to a warehouse full of test ballast and the people who loaded it.
Same thing in software. When I test an agent framework, the features that actually determine whether it survives contact with real use are:
- How it behaves when an API call times out halfway through a multi-step task
- Whether it logs enough that you can reconstruct what went wrong an hour later
- What happens to partial state when you kill the process
- Whether the retry logic is thoughtful or just a loop with a sleep in it
None of that is in the demo video. The demo video has a cheerful voice explaining how the agent booked a flight. The sand is in the error handling, and most vendors haven’t bought any.
The thieves and the AI buyer have the same problem
I want to be fair to the people who stole these trailers, in the narrow sense that their reasoning was rational given their information. They saw signals — location, industry, cargo container — and extrapolated to value. That’s what everyone does when they buy AI tooling based on a category label instead of a test.
The failure mode is identical. You acquire based on adjacency to something valuable rather than verification of the thing itself. Then you break it open and discover that the impressive wrapper contained a testing artifact. The difference is that the thieves found out in one night and walked away. Companies sign twelve-month contracts and spend three quarters pretending the sand is chips.
And notice what the thieves did once they knew: they dumped the trailers and left. No sunk-cost spiral. No attempt to make the sand work. They took the loss immediately and moved on, which is honestly more discipline than I see in most enterprise AI pilots.
What to actually take from this
Three things, and I’ll keep them short because the story doesn’t support more than that.
First, the branding on the outside tells you nothing about the payload. An AI company’s warehouse is full of sand. An AI product’s landing page is full of adjectives. Open it before you commit.
Second, the boring infrastructure is the real asset. If you’re evaluating tools, ask what the vendor’s equivalent of 20 tons of sand is. What are they doing that’s tedious, expensive, and unphotogenic? If they can’t answer, they may not be testing anything.
Third, when you find out you got sand, say so fast. The thieves managed it. You can too.
No arrests yet, by the way. Somewhere out there are people who pulled off a clean two-trailer heist and got nothing but a very heavy lesson in reading the manifest first. I’ve had product demos go about the same way.
🕒 Published: