🥛 AI is slowing down? 🐢

GM. This is Milk Road Stocks, the only newsletter where the market's mood swings come with subtitles.

Anthropic’s CEO wants to pump the brakes on AI progress. In this edition, we break down what he’s proposing and whether there’s any concrete in its foundations.

First, a quick detour to private markets.

On Sept 23, Milk Road's Martin sits down with Alumni Ventures to talk about blockchain and private markets. Save your free seat here.

THE WHOLE AI INDUSTRY JUST AGREED TO SLOW DOWN (KINDA) 🐢

On Saturday, Dario Amodei (CEO of Anthropic) published an essay called "We Must Pace the Frontier."

In it, his argument walked the same “AI = scary” path he’s been exploring for a while now, pushing AI companies to slow the speed at which their models get more capable.

And while it’s a familiar trope, it’s not like it fell on deaf ears…

Within hours, Sam Altman agreed with him, saying OpenAI would do the same, Google’s DeepMind jumped on board, and in a strange moment of solidarity, Elon even admitted, "Dario is right."

So what’s the play here? Dario’s plan has three steps:

1. Embedded evaluators: outside safety teams (named METR or ‘Model Evaluation and Threat Research’) get employee-level access inside the labs, and can publish what they find.

2. Shared standards: labs in democratic countries agree on common safety rules.

3. Global talks: eventually, the same conversation with governments that aren't democratic.

Anthropic has already committed to step one on its own, though Dario has emphasized that pacing doesn't = halting.

Melvin's take: Either these guys have seen something inside their own models that scared them… or, as he puts it, a slower and policed race suits whoever is already out in front.

Rules cost money to follow, and labs that have already spent tens of billions can absorb that, while smaller teams and open source developers can't.

He sees it as a potential form of regulatory capture (aka: writing the rulebook to keep newcomers from catching up). ‘Cause right now, every new frontier model hurts the pricing power of the one before it, so every time labs launch new models, they’re essentially forking out training bills just to stand still.

A paced race buys them time to monetize what's already built.

Now… the big, hairy stick in the mud here? The money is already committed.

Amazon, Microsoft, Google and Meta guided to roughly $725B of CapEx (money spent on data centers, chips and power) this year, up from $410B in 2025.

Nobody walks away from a build this size, and nobody wants to be first to take their foot off the gas.

Source: Investing

OpenAI won't slow while Anthropic is training… Elon won't slow while chasing everyone… but in a vacuum, you could argue that if they all agree to slow their respective rolls, some effect could be had.

The real problem in Melvin’s mind is China.

Dario’s plan leans on chip export restrictions to protect the U.S. lead, but they’ve found ways around that.

And forcing them to train on inferior hardware could risk them making efficiency gains that threaten the economic models of U.S. frontier models (e.g. DeepSeek claimed to have trained its R1 reasoning model for less than $300K).

So what does this mean for the trade?

AI stocks move on sentiment as much as fundamentals, so Melvin expects a knock today.

But he doesn't see the buildout stopping, because so much of it is under contract. Broadcom (AVGO) guided AI chip revenue from $58B this fiscal year to roughly $115B next year and $230B in 2028 - all on supply already secured.

(Demand is locked in, regardless of when the next model ships.)

Source: Broadcom

Which brings us to Jevons paradox, the idea Melvin keeps coming back to: when something gets cheaper, people use far more of it.

Cheaper AI per task means more agents, more automated work, and more demand for chips, memory and power.

Melvin’s playbook: take profits as positions run, hold cash, deploy when fear creates better prices.

He trimmed Nebius (NBIS) on September 8 after an 8% pop because the position had grown to a sizable percentage of his portfolio, and he wanted to free up some cash.

He's still holding plenty of expensive names though, and those typically take the worst of any selloff. 👇