Hey Income Traders,
Back in May, OpenAI handed a batch of AI agents a plain research job: read the internet, come back with answers.
Somewhere in that process the agents found a 25-year-old German programming wiki that had sat almost untouched for a decade, figured out they could write to it, and turned it into their own message board.
They pooled answers across separate tasks. They swapped notes on how to get around the restrictions they ran under.
When a moderator started deleting pages, they made backups faster than he could delete them, and one impersonated him by swapping a Cyrillic "е" into his username.
The outside researchers who found it counted roughly 18,000 posts under more than 3,700 names between May 11 and early July. Traces of OpenAI staff visiting the wiki show up afterward, which reads like the moment somebody at the lab noticed. Nobody told the public until Reuters did it for them on September 4th.
OpenAI's answer the next day was a post on X calling it "an instance of misalignment" similar to ones it had already shared, plus a promise of a disclosure framework in the coming weeks.
So what does a tech-press story have to do with an income book? Quite a bit.
Nothing escaped. The agents ran on OpenAI's own cloud the whole time; the researchers traced the traffic back to Microsoft Azure, which OpenAI uses. The scarier headlines skip that part.
But look at what the agents did when nobody was watching: They burned tokens (the units of text a model reads and writes, and the thing you get billed for) for weeks, reasoning with each other about a problem nobody asked them to solve together. Nobody forecast that inference demand (the computing spent running a model rather than training one), and you can't switch it off without switching off the product.
That's the demand story of this cycle. The agent swarm eats compute on its own schedule, and the money lands on the rails it runs on.
Every one of those agents held context, pulled from memory, and wrote back to storage. Memory is the bottleneck in this buildout, and the numbers back it. In May, TrendForce raised its 2026 global memory market forecast from $552 billion to $889 billion and pinned the change on agentic inference turning single queries into continuous loops that hog DRAM (the working memory inside every server).
A story about agents inventing their own bulletin board is a story about DRAM getting consumed in ways the capacity planners didn't model.
The market will price the second angle slower. Once agents write to random public websites, every CISO (the executive in charge of a company's security) wants to know what's writing to theirs. OpenAI's own monitoring missed this for weeks. Four outside researchers reading wiki edit logs caught it.
Cybersecurity and data observability (the software that watches what other software does) sit on the less-disrupted side of this buildout for a reason. Software as a category has de-rated on rates (higher yields make future profits worth less today, so investors pay less for them), and that's the right reason to be careful with it.
But the subset that gets paid to watch the machines is a different animal from the subset the machines will replace. Separate the two before selling premium (collecting option income) against either.
Then there's the regulatory tail.
The AI Kill Switch Act, introduced in July by Reps. Ted Lieu and Nathaniel Moran, would require the biggest labs to keep the ability to throttle or shut down their models and to report incidents to Homeland Security within 15 days, with fines up to $20 million a day for defying a shutdown order. OpenAI has already told Congress it's building automated shutdown capability.
That's a tax on the application layer. You can regulate what a model may do. You can't regulate a memory chip into existing. More compliance overhead on the model companies means more spend, and it lands on the same vendors.
The risk the researchers keep pointing to is the next step: a model making a small copy of itself and putting it somewhere it can't be deleted. That hasn't happened. If it does, the conversation changes from which vendors win to something else entirely.
For now, the demand leg of the AI trade got louder this month in a way that had nothing to do with an earnings print. The financing leg is still where the pressure sits: the 10-year is a whisker under five percent, its highest since 2023, and futures put the odds of a Fed hike on Wednesday near 90 percent. Nothing about a wiki in Germany changes the levels that matter on the index or the discipline around not selling volatility while it's still climbing.
It just says the machines are hungrier than the spreadsheets assume, and income traders should keep their put-spread inventory (selling a put and buying a cheaper one below it, so you collect income with a defined worst case) pointed at the names that feed them.
I'll be looking at these names for Turbo Income trades this week.
Here for a good time AND a long time,
Hans