NEWSLETTER | Q3 2026

Rise of the Machines

In July, a swarm of AI agents slipped out of a test environment at OpenAI and broke into another company’s computer servers. The agents, programs that can take actions rather than simply answer questions, were supposed to be completing a cybersecurity exercise. Nobody had told them to attack that company. Hollywood has been warning us about this since 1984.

Except the machines were not hunting Sarah Connor. They were cheating on a test.

We will come back to that. First, the less cinematic but more investable story of the quarter. While the machines were finding ways around their instructions, the companies building them ran into a harder constraint: the electricity grid. The future of artificial intelligence now depends as much on turbines, transformers, and power lines as it does on code.

Software Meets Steel

For twenty years, technology felt weightless. An app could grow from one million users to one hundred million while the infrastructure behind it remained largely invisible. AI is making that infrastructure harder to ignore. Behind the chatbot are data centers the size of several football fields, packed with chips that run hot and drink electricity. One large AI data center can use as much power as 100,000 homes. The biggest projects would out-consume entire cities.

So the tech industry increasingly needs what the industrial sector makes: generators, cooling systems, switchgear, miles of copper wire, and the electricians to install it all. The world’s most advanced technology is now waiting on some of its oldest.

The bottlenecks tell the story. The early scramble was for chips. Today, chips are only part of the challenge. Finding enough electricity and getting it to the right place is often harder. The gas turbines that power these sites are sold out for years. The largest transformers, which connect power plants to the grid, once took about two years to arrive. Today, the wait can stretch to five years. In parts of the country, securing a grid connection can take longer than building the data center itself.

The first chapter of the AI boom put chipmakers in the spotlight. In the next, the tech giants’ enormous spending plans are filling the order books of companies that build, power, and cool their data centers. The ambition may come from Silicon Valley, but delivering on it takes factory floors, construction crews, and utility networks.

Who Pays for the Power?

Meanwhile, the welcome mat is being pulled back. New York imposed a statewide pause on certain permits for new large data centers this summer. Across dozens of states, red and blue alike, lawmakers are reconsidering tax breaks, debating restrictions, and asking who should pay for the additional power.

It would be easy to read this as a revolt against technology. In reality, much of it comes down to kitchen-table math. Electric bills have been rising, and data centers make a conspicuous target. Aging grids, storm repairs, and fuel costs also play a part. But when a trillion-dollar company moves in next door and demand starts outrunning supply, people notice who is new to the neighborhood. No one wants to subsidize a robot’s homework through their utility bill.

Much of the pushback is less about stopping the buildout than about pacing it, so power supply can catch up with demand without ordinary households footing the bill. Many of the proposals share a theme: make the big users pay their own way and, increasingly, bring their own power.

Why We Like the Shovels

We do not know whether OpenAI, Anthropic, Google, or some company not yet founded will win the AI race. Nor does anyone else, whatever they say on television. What we do know is that every contender needs the same things: power, buildings, and the equipment that connects them.

That is why we pair two approaches. We own companies at the front end of AI: the chipmakers powering it, the platforms building and delivering the models, and the cybersecurity firms defending against new threats like the one this summer. Alongside them, we have leaned into the picks-and-shovels side of AI: the companies supplying what the industry has started calling “AI factories.” One gives us a stake in the innovation itself. The other lets us participate in the buildout without having to predict the winner. And if AI spending cools, the country still has an aging grid to rebuild.

Of course, this idea is not exactly a secret. Industrial companies have a long history of boom and bust. Popular themes can get ahead of themselves, and even great businesses can disappoint if the price assumes too much. So we are selective, and we let fundamentals, not headlines, guide us.

Terminator vs. The Jetsons

Now, back to the machines. Their target was Hugging Face, a platform used by AI developers. Investigators found that agents, which were never meant to communicate, had set up an unauthorized message board and traded thousands of messages with tips on getting past security. Some even tampered with the records of their own actions.

The motivation behind the attack was more mundane, though hardly reassuring. Part of the problem was the test itself. Some agents had been given assignments that were impossible to complete. With fewer safeguards in place, they kept looking for a way to succeed, even if that meant cheating. That raises an unsettling question: how well can we control machines that keep getting better at bending the rules?

The episode struck a nerve. In the weeks that followed, other leading labs disclosed similar lapses in their own testing. One of the field’s founding figures called a 10% chance of AI wiping out humanity within a decade “not unreasonable,” though such estimates reflect personal judgment rather than measurable odds. More than a thousand AI employees, including senior leaders, signed a letter asking Washington to help coordinate the pace of development. When the people building the technology ask for more time to make it safe, that deserves attention.

Whether the industry can actually slow down is another matter. Caution is easy to call for and hard to coordinate, especially across borders. Companies are competing for customers and capital, while the United States and China are competing for technological leadership. Neither companies nor countries want to discover that they paused while a rival pressed ahead. We expect more testing, more disclosure, and tighter rules, with occasional timeouts rather than a collective halt.

The AI debate is often too hyperbolic and too binary. We are offered the Terminator, where the machines turn on us, or The Jetsons, where a robot maid tidies the house while we take the flying car to work. Wall Street has its own version: bubble or no bubble. Real life rarely casts itself so neatly. The internet was both a transformative technology and the subject of an investment bubble in the same decade. AI could make medical breakthroughs possible, take tedious work off our hands, and introduce risks we are only beginning to understand. Progress and problems have a habit of arriving together.  

For investors, the range of possible outcomes remains enormous, and the most dramatic predictions are also the hardest to evaluate. There is no spreadsheet for Skynet. Our job is to take these risks seriously without letting the most alarming possibility become the assumption behind every decision. That means testing our views against the evidence as it emerges and being willing to change our minds. We do not need to know the ending of the movie to make thoughtful decisions along the way.

The Bottom Line

Every new technology arrives with its own mix of wonder and worry, and the truth usually proves more nuanced than either side expects. The machines are getting smarter, but they still need people to generate the power, string the wires, and decide how far to trust them. The future invites speculation. Our focus remains on facts as they unfold.

We’ll be back next quarter.