Mark Stevens didn’t say anything about the AI trade last quarter. He just sold 4,236,740 shares of Nvidia, roughly $946.6 million worth, which is its own kind of statement. Bloomberg reported on October 7 that the venture capitalist and Nvidia director was the single largest insider seller of the third quarter of 2026, based on data compiled by the Washington Service.
My first reaction, as someone who spends most of his week installing agent frameworks and then uninstalling them: this has almost nothing to do with whether your tools work. My second reaction: a lot of people are going to tell you it does.
What the data actually says
Stevens topped the list. Arista Networks CEO Jayshree Ullal and Amazon founder Jeff Bezos also showed up among the quarter’s notable sellers. Not everyone on the list came from AI, either. Illumina director Keith Meister ranked third after selling about $436 million in Illumina stock, and SharkNinja chair Xuning Wang appeared on the list too. The Washington Service tracks insider buying and selling, which is the whole job, and this is what the tracking turned up.
The part everyone skips: the sources don’t provide a definitive answer on why any of these sales happened. Not one confirmed reason. Insiders sell for tax bills, diversification, estate planning, prearranged schedules, charitable giving, and occasionally because they think the stock is expensive. From the outside, all of those look identical on a filing.
There’s one more wrinkle that complicates the obvious narrative. Bloomberg noted the quarter saw Nvidia’s stock valuation approach its cheapest level in over a decade. So the tidy story where insiders dump at the top doesn’t quite fit the data in front of us. If you want to argue it does, you’re adding facts nobody reported.
Why I’m writing about this on a tools site
Because this kind of headline changes how people buy software, and it shouldn’t.
I’ve watched the pattern for a couple of years now. A financial story breaks, and within 48 hours my inbox has questions like “should we hold off on the agent platform rollout” or “is the AI tooling thing over.” The reasoning goes: insiders are selling, insiders know things, therefore the tools are about to stop being useful.
That chain breaks at every link. A director selling shares tells you something about one person’s portfolio. It tells you nothing about whether your retrieval pipeline returns the right documents, whether your agent framework handles a failed tool call without silently returning an empty string, or whether the vendor you’re paying $40 a seat for has fixed the bug you reported in March.
Those are the questions that determine whether your stack is worth the money. They are answerable. They’re just boring to answer, because answering them means running the thing on your own data and writing down what broke.
The reverse mistake is just as bad
I’ll be fair here, because the opposite reflex is equally lazy. Some people will read this and decide insider selling proves the whole AI tooling category is hype, so they can stop evaluating anything. That’s just a different way of outsourcing your judgment to a stock chart.
Plenty of AI tools are genuinely useful right now. Plenty are thin wrappers with a pricing page. The split between those two groups has not moved one inch because of a Q3 filing. You still have to test them.
What to do instead
If you’re responsible for picking AI tools at your company, here’s the honest substitute for reading insider filings as tea leaves:
- Run a real task, not a demo task. Vendor demos are tuned. Bring your messiest input and see what survives.
- Check failure behavior before success behavior. How does the tool act when the API times out, the context overflows, or the user asks something out of scope? Good tools fail loudly. Bad ones fail confidently.
- Price the exit. How hard is it to move off this thing in six months? If the answer is “we’d have to rewrite everything,” that’s a cost, whatever the subscription says.
- Watch the changelog, not the stock. Shipping cadence and bug-fix responsiveness tell you more about a vendor’s health than any single shareholder’s trading activity.
None of that requires an opinion on Nvidia’s valuation. That’s sort of the point.
Where I land
Stevens sold nearly a billion dollars of stock, and we don’t know why. That’s a legitimately interesting data point for investors, and I’d read a sharp analysis of it. What I won’t do is pretend it’s a buy signal or a sell signal for the software sitting in your deployment pipeline.
Your tools are either doing the work or they aren’t. Go check. The filing will still be there when you get back.
🕒 Published: