Where are all the AI customers’ yachts?

This story is as old as Wall Street itself. In his book Where Are the Customers’ Yachts?, originally published in 1940, Fred Schwed Jr. tells the story of a visitor to New York who is taken to the harbor and shown the impressive yachts that belong to the bankers and brokers. A tad naive, the visitor asks, “but where are the customers’ yachts?”

Where indeed.

- Green, Alexander. The Gone Fishin' Portfolio (p.3) Kindle Edition.

They say that in a gold rush, while some are busy digging for gold, others are selling shovels.

The shovel sellers of AI are definitely making bank. But what about us simple mortals trying to use AI to make an extra buck?

And if there is money for the people using AI rather than selling it to others .. then where is it, exactly?

Selling shovels at a premium

NVIDIA announced $215.9B FY2026 revenue w/ very high gross margins. That’s what you get when all the most powerful people in the world are begging you to sell them GPUs over sushi, I guess.

Microsoft AI business announced $37B revenue run rate. AWS AI is above a $15B run rate.

Anthropic, not being a public company, announced they may have their first profitable quarter, and Dario mentioned on the most recent Code with Claude conference that they forecast an aggressive 10x growth in a year, but instead grew by.. an astounding 80x!!

80x!!!

So, there’s real money in shovels. But what about the money in the gold?

Where are all the AI users spending tokens on Anthropic and growing by 80x too???

Or 8x?

.. or 80%?

.. or even .. 8%?

Studies say more studies are needed to find the gold

A McKinsey 2025 survey reported that only 39% of organizations report any enterprise-level EBIT impact from AI—and most of those say AI is <5% of EBIT.

BCG reports only 5% of companies are achieving “AI value at scale” (whatever that means) and 60% report “minimal revenue/cost gains despite investment.”

OECD’s AI adoption by small and medium-sized enterprises study mentions that “productivity may decline before increasing.” They also say “it is mostly used for peripheral rather than core tasks.”

But maybe you don’t care about studies.. you care about real results! So what are the companies who are really investing and embracing AI reporting in their quarterly earnings?

JPMorgan reportedly spends $2B/year on AI and saves “almost the same amount” .. it’s hard to think of another area a bank would spin this harder than AI. The “almost” is the part that catches me, personally.

But what about tech companies?

Growth but no acceleration

In earnings calls for tech companies, there is definitely a significant AI acceleration .. in vibes.

Don’t get me wrong, these companies are growing. Uber was up 14% last quarter, Booking 16%, Expedia 15%, Airbnb 18%, DoorDash 33% (about 21% once you back out an acquisition), and Etsy grew its marketplace 5.5%. Healthy numbers.

The catch is that none of those growth rates looks any different than prior quarters without AI, and none of these companies claims otherwise. DoorDash shipped AI merchant onboarding and AI-built storefronts and grew right at the ~20% trend it’s been on for years.

The rest mostly point to efficiency, not growth. Airbnb resolves over 40% of guest issues without a human and shaved about 10% off its cost per booking—and yet its margins are the same. Booking cites roughly $550M in “transformation” savings, a bucket that quietly includes restructuring and not just AI (AI and layoffs go hand-in-hand in some companies).

Expedia illustrates the whole problem in a single earnings call. Management says AI is “driving hundreds of millions of dollars in realized marketing value” but in “productivity and workflow automation,” not raw dollars. It made the marketing team better but not in a way that shows up in the sales dollars, somehow. And the genuinely new AI revenue such as ads inside ChatGPT is still “small.” And as expected, AI is about to cost them more than planned, with token costs “anticipated to increase in the second half of the year.”

Lots of percentages, “hundreds of millions in productivity” thrown around, and almost no clean, auditable dollar that a skeptic couldn’t wave away.

Then there’s Klarna, fascinating for the opposite reason: it went all in, and then backed out. In late 2024 its CEO said AI could already do the work humans do. The company froze hiring for over a year, shrank its workforce from around 5,500 to roughly 3,400, and bragged that its chatbot did the work of 700 customer service agents. Six months later, with customer satisfaction down, it started hiring humans back. The CEO’s verdict, verbatim: “As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality.”

Indeed.

But what about the AI-forward Uber? By their own numbers, 95% of their engineers use AI every month, 70% of committed code is AI-generated, and about 10% is written by autonomous agents. They ran internal tokenmaxxing leaderboards and blew through their entire 2026 budget for AI coding tools in four months. But then, asked whether all that spend is actually producing value, COO Andrew Macdonald said the part you almost never hear a big spender say: "That link is not there yet."

Oh well. But where is it, then?

First, let me talk to you about digging the gold.

AI is a real revolution

Look, I’ve seen all sorts of skeptics on AI and value creation: Its code is not good and not cheap, the hype cycle went up and is now down, software development won’t change much after all is said and done, this is the new dotcom bubble, AI is the new [3D, metaverse, web 3.0], you name it. As a bullish on AI person and eng leader, I’m sometimes shocked by how skeptical my peers whose opinions I otherwise respect are about AI.

But absolutely nobody is thinking about going back to doing things the pre-AI way. LITERALLY NOBODY! ZERO!

I mean, have you thought about that? Even the Diogenes the Cynic of Engineering Leadership will still concede that AI is a net positive, at least when it comes to voting with their wallets and time.

All the disagreements we have about AI aren’t about AI itself as a positive force to productivity/efficiency/revenue but rather about how much and in what way.

And this I think comes to the crux of what I’ll talk about when it comes to digging the AI gold:

Digging AI gold is less like using shovels and more like digging a fucking mine. Yes you can buy the mining equipment, but all that buys you is a ticket to the game. Now you gotta learn the simple art of mining engineering and related disciplines such as mineral processing, exploration, excavation, geology, metallurgy, geotechnical engineering and surveying.

Not quite as simple as the shovel you were sold, eh?

And if you weren’t in the mining business before (and most of us weren’t) you’re in for a ride.

Buckle up.

Mining engineering in AI

Yes AI companies are printing money. I haven’t worked at an AI company, but I did work at one that was printing money, and let me tell you: It’s REALLY hard. Like tremendously hard.

Being at a struggling company I’m sure has its challenges, but a company that can’t handle its demand faces a completely different, but just as real, set of problems.

The truth is that most of us armchair executives and staff engineers wouldn’t last in our current jobs if our upcoming annual growth was 10x.. let alone 80x.

Think about it: Imagine your company managed to be a Unicorn, the coveted $1B+ valuation. Congratulations! Now imagine running it so that it got to the ballpark of $965B in only 3 years!

You think you’d have no downtime, stellar customer success, amazing PR, and still grow valuation 3 orders of magnitude in as many years? Of course you wouldn’t. Nobody could. That’s like going through quarterly company open-heart surgeries. No amount of money makes that easy.

So how are these companies even surviving, let alone thriving under this immense pressure? Not just Anthropic and OpenAI but Cursor, Lovable, etc?

A huge advantage of these companies over regular tech companies (let alone non-tech companies) is that they are natively AI-first.

One common theme from my analysis of AI at tech companies is that most companies that bolted AI onto their current product failed to find results, at least so far—Shopify perhaps being the poster child of this approach’s futility, with several-fold jumps in AI-driven traffic and AI-assistant adoption, and 0% extra AI revenue (and a 16% drop in market value to boot).

Similarly, trying to adjust our human processes to ever-evolving AI technologies is literally like hiring aliens and telling them to do the same job we humans were doing—except better. Bolting square pegs into round holes won’t work.

A company run by aliens has got to be run for the aliens, the same way my household is run by and for the cats.

But what does it mean to have a company run by and for aliens, exactly?

Is AI like the industrial revolution or not after all?

For whatever reason I keep running into this comparison between AI and the industrial revolution, and the funniest thing happens in those discussions: Nobody agrees.

Well, here’s my opinion for you to disagree with, then:

We created machines, but the real revolution was the whole way that we then reorganized humans around the machines to create value.

No amount or quality of machines would allow us to have the revolution in manufacturing we had while preserving artisans and apprenticeships and craftsmanship. None.

We created machines, and then we had to completely rethink how to organize whole human economies to leverage them at what they did best while avoiding their relative shortcomings.

And once we did, no amount of artisan craftsmanship could compete with the machine’s value creation. Not that craftsmanship didn’t have its value.. it did and does. The machine just has a lot more, as evidenced by their revenues and margins given where you and I decide to spend our hard-earned dollars.

So here’s my take on AI leverage:

We were physical artisans, and now we are cognitive artisans. But there absolutely is a way to do cognitive work wholesale through machines and, once we organize humans around it, no amount of cognitive craftsmanship will be able to compete with it.

And that’s, in principle, where I think the gold is buried and how to dig it:

Not in “replacing humans” ..but in doing cognitive work wholesale.

But what does that look like exactly?

Building machine companies rather than human companies

Some of Claude’s major reported weaknesses compared to an L4 [human] include: self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, instruction following, and epistemics.

- Anthropic: Claude Mythos Preview System Card (p.36)

While at least 1 person out of a small (18) sample size at Anthropic thought Mythos could replace a real person at Anthropic today, the vast majority didn’t. Yet, Anthropic is spending an obscene amount of tokens doing work while growing 80x a year with models weaker than Mythos and that are bad at “instruction following.” What gives?

Well, if humans are good at X and machines are bad at X, you don’t build an effective company by expecting your machines to do the exact job that humans, that are good at X, were doing!

It’s obvious but counter-intuitive because, of course, our unit of work at companies up until now is the “human” — we don’t think of a person’s right arm and of their left eye and of their frontal cortex as much as we think of “Alice” and “Bob.” You can’t get the items without the package, without all the stuff humans can typically do like speak, listen, join calls, and have lunch with prospects and take them to the baseball game afterwards.

Humans are the lego pieces of companies.

But if you have to build a sales team of aliens, and they can’t take your customer to the baseball game, then baseball has to be purged from your sales playbook if you are to have any hope of selling anything. And if you don’t, that’s on you not the aliens.

Which leads me to what I think is the thing most of us, at companies, are missing about our use of AI:

AI leverage at companies is ultimately not a Technology problem, but an Operations problem to solve, and the problem is how can we split the work not across humans (and their typical roles), but by specific actions that AI can perform wholesale.

In short, humans add value and capture it by acting, and so does AI, but we act in completely different packages with different strengths and shortcomings. As long as we center around roles, particularly ones traditionally occupied by humans, we’ll just be getting faster horses who are bad at following instructions.

Machine companies will never be as good as human companies.. so stop chasing that. It doesn’t matter anyway.

If you want to add high leverages of value by using cognition wholesale, you have to give up trying to be as good as a human company can be while using machines, and instead try being the best machine company you can be.

Stop trying to take your prospects to baseball games with AI.

Embrace the AI suck

AI isn’t as good as humans at things, and may never be. The Venn diagram of their capabilities is not a superset of ours. The age-old lesson in management of “manage your direct report to their strengths, not their gaps” wholly applies here.

So you need to change the things your company does that depend on human strengths but that AI is bad at into things that AI is good at if you are to scale value creation and leverage beyond what it does now.

And if you can’t, then you can’t. There’s no magic here. No participation prize. Can’t just remove the human lego and place an AI lego. The pieces don’t fit and you’ll only burn tokens.

And curiously, during this transition, companies will be human-AI hybrids, transitioning from cognition costs dominated first and foremost by headcount Opex, followed by COGS and then AI, into a mix where AI takes more and more of the share as the work the company does aligns more and more with wholesale rather than artisanal cognition. Work artisan humans just can’t compete with no matter their strengths.

It doesn’t need to be everything human companies can do. In particular, it’ll probably start with the things that human companies can’t do! Why wouldn’t it? The industrial machines didn’t start manufacturing everything artisans could craft.

But only once companies break their current human roles, map all of their actions used to create and capture value, replace actions humans are good at with ones AI is good at, and then align it back to a way to create and capture value such that they can apply this cognition wholesale, will they actually dig the gold.

And the fact that AI can’t do human things, and that companies are human and not machine companies, is, in short, why I think we don’t see any of the AI customers’ yachts..

.. yet.