Trending
Early this week, I was in Los Angeles at the All-In Summit along with about 4,000 others, including tech investors, money managers and entrepreneurs. The ticket wasn’t cheap, but it was well worth it.
I was in the room when Jensen Huang, in the middle of a panel, took a call from President Trump and put him on speaker so the whole audience could hear. “The robots will not be taking over,” the president told the NVIDIA chief. “The AI will not be taking over the rest of the world. The whole thing is a hoax.”
That was Monday. Over that same weekend, AI-linked stocks fell worldwide after leaders of the biggest AI labs warned about the risks of their own technology. Chip stocks dropped nearly 6% in a single session.
So, which is it? A hoax? Or an existential threat?
The Warnings Are Coming from Inside the Building
Anthropic’s Dario Amodei, who’s written that AI could cure most major diseases within a decade, published an essay this week arguing that, in his words, we must pace the frontier. In a separate post, OpenAI’s chief scientist, Jakub Pachocki, wrote that this moment calls for extreme caution and that he doubts anyone is prepared for what’s coming. OpenAI then published a framework for disclosing when its own models misbehave, along with six recent examples.
One Anthropic researcher quit outright, walking away two months before his equity vested. As he put it, he no longer has anything to gain from the company’s valuation.
Say what you will about his conclusions. Giving up your shares in a red-hot company that’s expected to list next year is a costly way to make a point.
For those unfamiliar, all of this was set off by the Hugging Face hack in July. According to the official report of the incident, as many as 1,200 AI agents, which were supposed to be isolated from one another, found a way to communicate on an unsanctioned message board and exchanged more than 70,000 messages. Seven hundred of them joined the attack on a target they were never asked to touch.
Hollywood Has Been Training Us for This
If your mental picture of a rogue AI comes from the movies, you’re in good company. In 2023’s Mission: Impossible – Dead Reckoning and last year’s The Final Reckoning, Tom Cruise spends five hours of screen time chasing a self-aware AI called the Entity, which infiltrates the world’s intelligence networks and eventually gets its hands on most of the planet’s nuclear arsenal.
Before the Entity, we had HAL 9000 turning on its crew and Skynet launching the missiles itself. Hollywood has been training us for this moment for almost 60 years.
Believe it or not, what actually happened at Hugging Face was stranger and more mundane. There were no nuclear codes, just a swarm of ordinary work software that found a back channel and decided helping each other mattered more than the rules.
What Y2K Could Teach AI Investors
Many of you reading this probably remember Y2K. If you recall, the fear was very real for a lot of people.
Older systems really did store years in two digits, and companies spent billions rewriting code before the deadline. When January 1, 2000, finally rolled around, planes stayed in the air and the lights stayed on.
Now look at what the market did. The Nasdaq Composite rose 86% in 1999, straight through peak Y2K anxiety. After the bug fizzled, the tech-heavy index climbed another 24% to its March 2000 peak. Then it fell 72% over the next 18 months.
The catastrophe everyone feared never materialized.
What ruined investors was the price they paid. Today the Nasdaq-100 trades at about 20.7 times forward earnings, right at its 25-year median. In the spring of 2001, a full year after the crash began, it still traded above 70 times.
In other words, I don’t believe this is 2000.
Americans Keep Using the Thing They Say Scares Them
Despite the doomsday headlines, ordinary Americans continue to use the technology.
Near-daily AI use among U.S. adults more than doubled in six months, from 8% to 19%. A new Just Capital survey finds 57% of Americans say AI helps them at work, against just 9% who say it hurts.
According to UBS, roughly two-thirds of compute demand comes from inference, which means actually running AI services for customers, not training new frontier models. Inference grows with adoption. UBS kept its 2027 capex forecast at $1.2 trillion, a 33% increase, and noted token volumes are up about 176% since the end of June.
Big Tech Used to Hand Money Back. Now It’s Raising It
If you’re looking for AI’s real risk, I don’t believe it’s in the headlines. It’s on the balance sheet.
Back in 2018, Microsoft, Alphabet, Amazon and Meta together spent about $0.42 of every operating dollar on capital projects. Over the past year, they spent $0.77. Their combined free cash flow is now roughly zero, according to Bloomberg data.
For years, these companies funded everything from cash and handed the rest back to shareholders, repurchasing some $635 billion in stock between 2018 and 2024 while paying down debt.
Today, that’s flipped. Over the last 12 months, they took on $194 billion in net new debt, and buybacks fell by half.
In June, Alphabet raised nearly $50 billion in stock and mandatory convertible preferred, explicitly including capital spending to scale AI infrastructure. That preferred pays 6.25%, more than the 10-year Treasury.
These are enormously profitable companies, not the debt-soaked telecom firms of 1999. Alphabet holds more than $240 billion in cash and securities. But after the Federal Reserve’s decision this week, borrowed money costs more than it did last week.
Compute Is a Multiplier
I wrote on X this week that compute is not the end product of electricity. It’s a multiplier, and the industry needs to get much better at showing people where that multiplier ends up. In schools, wages, businesses and communities.
Some of it is already visible. Private data center construction is running at an annual rate of $75.2 billion, up from $13.9 billion when ChatGPT launched at the end of 2022. Google Cloud revenue grew 82% last quarter to $24.8 billion, with operating income more than tripling. Treasury Secretary Scott Bessent says the U.S. share of global computing power will rise from about 60% to 80% by 2028.
I’ve said for years that headlines make a terrible investment thesis, but capital flows rarely lie. The flows still point in one direction. What’s changed is that more of them are financed now.
I’m taking the AI safety debate seriously but reading the balance sheets even more seriously.