The Gold, the Shovels, and the Switch: Looking Forward at Big Tech’s AI Cap-Ex Build-Out without the Existential AGI Bet
The platforms aren’t gambling on AI—they are trying to tax it. The labs dig for gold while clouds sell shovels and give away the map. If anyone is going to profit from AI infrastructure, it’s the...
The platforms aren’t gambling on AI—they are trying to tax it. The labs dig for gold while clouds sell shovels and give away the map. If anyone is going to profit from AI infrastructure, it’s the effective platform monopolist incumbents. The rest are highly likely to discover that “core AI” is a product with negative margins and fickle users. Or, to put it bluntly: This is nuts! When’s the crash? Unless Sam Altman’s ChatGPT or Elon Musk’s GrokAI or Dario Amodei’s Claude or Demis Hassabis’s Gemini really does become DigitalGod…
Unhedged’s argument hasbeen that the market, far from blindly throwing money at AI and Big Tech, is making distinctions… based on cash generation…. The market has lost patience with Oracle and Meta. Might the same happen to the other three before long?…
Andy Wu: Should U.S. be worried about AI bubble? <https://news.harvard.edu/gazette/story/2025/12/should-u-s-be-worried-about-ai-bubble/>: ‘They [the five] positioned themselves well to benefit from the rise of AI, but they don’t stand to lose that much if AI [tanks]…. Microsoft has mostly outsourced… to… OpenAI…. Amazon will support anybody’s AI model . . . Meta spent billions of dollars building an open-source AI model…. [They] don’t really think that core AI technology is a meaningful business…. Instead, they’re focused on profiting from… adjacencies…. OpenAI, Anthropic and xAI are out there digging for gold. Nvidia is the consummate shovel seller…. Meta is the consummate jewellery maker… social media, advertising, wearables and metaverse businesses stand to benefit…. Microsoft does a bit of shovel selling and jewellery making, but the key thing is they’re not stuck digging for gold…. Amazon and Microsoft and Google might make less money on their cloud computing than they ideally would like if AI growth slows or declines, but they would not end up in financial distress…
We might sum up Wu’s view[:]… Big Tech’s data centre spending is significant… [but] core business models… remain “virtual”…. They are not making an existential bet on AI; they are spending to make sure that their core businesses can coexist with AI, should they need to…. Those core businesses should still be valued on high multiples of cash flows…
Briefly: The Big Five—the Magnificent Seven minus Nvidia, Tesla, and Apple; plus Oracle—are making sure that if anybody makes any money off of AI data centers, it is going to be them. And they are also taking steps to make sure that nobody makes any money off of providing core AI services by giving away for free whatever else OpenAI, Anthropic and Grok might try to make money be selling.
That has implications:
The first of them is Nilay Patel’s “spicy” prediction for calendar-year 2026 is that the year will see the end of OpenAI as we know it:
Nilay Patel: Oh, mine’s as nuclear as it gets…. OpenAI fails…. Gone. OpenAI is no longer. OpenAI, yeah, OpenAI is no longer, Microsoft owns the IP anyway. They own the models, they run the data centers…. This is a company that has no articulated product strategy. They had to call a Code Red because Google slowly just lumbered its way back in the first place. They hired everyone from Instagram to do product ostensibly….
[So] it’s a bunch of old meta-employees or ex-meta employees running product, but their biggest new product was Sora, which didn’t come out of the product group, which came out of the research group. It’s run by Sam Altman, who has ADHD and is running around with Johnny Ive and trying to collect billions of dollars…. It also believes they can make AGI with LLMs, which I really don’t. More and more people are starting to be like, yeah, you can’t…. They’re very good at raising money, including Sam floating the idea that the government should bail them out. And they also have a massive user base right now. But every query costs them money….
What we’re discovering pretty quickly here is that actually people are pretty fickle. And as these models change and as the use cases change, the switching costs like kind of don’t exist.
David Pierce: And so I’ll just go make a new friend over here.
Nilay Patel: David, the ruthless kindergartener. That’s what you’re teaching in your house, David.
David Pierce: No longer friends. New best friends. Gemini….
Nilay Patel: I think people are also… hitting the walls… the same way they hit the wall with Alexa and Siri in the previous generation…. I use [those] for timers and music…. [This] is a much wider array of things, but it’s still a limited list. And again, they do not have good product strategy…. It’s spicy. I think they will fail, but… it’s Sam Altman calling the Code Red on product, not me…
And Anthropic and Grok are in much worse shape, as they do not have the massive-consumer-mindshare-of-ChatGPT to play that OpenAI does.
It will not happen in calendar year 2026. But the most likely scenario I see is that OpenAI retreats back to research, as Microsoft feasts on what meat there is in the consumer- and enterprise-business carcass. And Google, Facebook, Amazon, and Microsoft lose money on their data-center build-outs, but preserve their effective platform monopolies in search, social media, shopping, and enterprise support.
However, I do see it as highly likely that there will be a number of businesses providing SLMs—Small Language Models that provide natural-language access to trusted, curated, and scrubbed structured and unstructured databases, with minimal leakage from fake world-descriptions derived from linguistic correlations, so-called “semantic leakage”:
Gary Marcus: New Ways to Corrupt LLMs <https://garymarcus.substack.com/p/new-ways-to-corrupt-llms>: ‘The wacky things statistical-correlation machines like LLMs do…. [In] a new paper, Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs… [Owain] Evans and his coauthors (Jan Betley, Jorio Cocola, Dylan Feng, James Chua, Andy Arditi, and Anna Sztyber-Betley) just documented a new phenomenon they called “weird generalizations”. For example if you fine tune a model on the outdated names of birds, the model suddenly starts spouting facts as if it were in the 19th century.
Needless to say, the electrical telegraph is not a recent invention. And once again, Evans isn’t doing this for entertainment…
Ah. But you see. Any piece of training data—fact or fiction—in which Anthus Rubescens, the American Pipit, is called the “Brown Titlark” is almost surely set in a year in which the electrical telegraph is a recent invention. Bare linguistic competence and nothing more is what we want in a natural-language interface.
If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…
Yes to Mark Field. Help w drug development? Sure. Help w material science? Sure.
But consumer AI seems to work a lot like spam email, it takes more time to delete than it is worth.
And it will allow routine transactions to become more complex. Hospital discharge notes are going to go from 10 pages of boilerplate to 50 pages of AI enhanced boilerplate. Law firms might have to respond to 5 ten page motions today in a routine lawsuit. By 2028 they will be responding to 50 motions of 50 pages for each.
And I suspect real human lawyers will have to review these motions (with AI assistance) lest someone sneak something of real substance by.
"I do see it as highly likely that there will be a number of businesses providing SLMs—Small Language Models that provide natural-language access to trusted, curated, and scrubbed structured and unstructured databases, with minimal leakage from fake world-descriptions derived from linguistic correlations, so-called “semantic leakage”
Yes to Mark Field. Help w drug development? Sure. Help w material science? Sure.
But consumer AI seems to work a lot like spam email, it takes more time to delete than it is worth.
And it will allow routine transactions to become more complex. Hospital discharge notes are going to go from 10 pages of boilerplate to 50 pages of AI enhanced boilerplate. Law firms might have to respond to 5 ten page motions today in a routine lawsuit. By 2028 they will be responding to 50 motions of 50 pages for each.
And I suspect real human lawyers will have to review these motions (with AI assistance) lest someone sneak something of real substance by.
"I do see it as highly likely that there will be a number of businesses providing SLMs—Small Language Models that provide natural-language access to trusted, curated, and scrubbed structured and unstructured databases, with minimal leakage from fake world-descriptions derived from linguistic correlations, so-called “semantic leakage”
I certainly hope this is true.