IBM, Microsoft, Intel, Alphabet, NVIDIA—what history reveals about where AI profits ultimately accumulate.
KEY TAKEAWAYS
AI follows the same commercialization cycle as every major technological breakthrough. What begins as a differentiated innovation eventually becomes standardized infrastructure, shifting where profits accumulate.
Technology industries naturally organize into economic layers. Each layer develops different competitive dynamics, and investors should focus on identifying where scarcity–and therefore pricing power–remains.
AI customers are rapidly becoming multi-model users. Rather than committing to one provider, businesses increasingly route workloads to the most cost-effective model for each specific task.
Capital-intensive infrastructure may ultimately generate more durable economics than frontier AI models. Chips, compute, power, and data centers could retain pricing power longer than application-layer innovation.
Investors should evaluate companies by understanding what they own, what they rent, who controls the customer relationship, and how future capital expenditures are financed.
MY HOT TAKES
AI investing is entering a second phase where economics matter more than technological novelty. Superior engineering alone is unlikely to guarantee superior shareholder returns.
Competitive advantages in AI will become increasingly difficult to defend at the model layer. Continuous innovation is necessary simply to maintain position rather than expand it.
Alphabet's vertically integrated AI strategy deserves more attention than it currently receives. Owning multiple layers of the stack provides flexibility that many competitors lack.
Public market discipline will eventually reshape today's AI leaders. Venture-backed losses that are tolerated today may become much harder to sustain once quarterly earnings dominate investor expectations.
The biggest investment mistakes often come from asking the wrong question. Rather than predicting the "best AI," investors should identify where structural bottlenecks persist over time.
You can quote me: "AI feels like MAGIC right up until the morning it feels like plumbing."
Beam me up Scotty. I have always been a fan of shiny new technology. I spent my early years building computers with my first one only able to output numbers and the letters A through F (if you know you know). Come to think of it, the keyboard was a–well, keypad with numbers 0 through 9 and A through F as well. I could only dream of a day where I could ask my computer questions and get salient answers–like in Star Trek's Computer, or carry around smart devices like Star Trek's Tricorder or Communicator.
I am blessed to have lived through the era where that dream has become a reality–almost. The technological and industrial breakthroughs that have occurred in the last 50 years that have landed right here, right now with Artificial Intelligence are spectacular. And so were the opportunities to profit AND LOSE through its evolution.
Fan-boy adoration aside and investment focus forward, here's what four decades on Wall Street taught me about new technology–it feels like MAGIC right up until the morning it feels like plumbing. Railroads. Long-distance telephone. Fiber optic cable. The personal computer. Every one of them started with a handful of geniuses and ended with a purchasing department comparing prices on a spreadsheet. AI is not exempt. It's early–but it is not exempt.
Here's the part nobody tells you about plumbing, though. Plumbing is where the money ended up.
Go back to 1981. IBM builds the personal computer and makes a decision business schools still teach forty-five years later. It buys the processor from a small outfit called Intel and licenses the operating system from an even smaller one called Microsoft. IBM keeps the brand, the sales force, the customer relationships, the manufacturing, the whole gleaming apparatus. IBM keeps the BOX. And within a decade the box was worth almost nothing. Compaq, Dell, Gateway, Packard Bell–they clawed each other down to single-digit margins selling beige rectangles that were functionally identical, because the parts inside them were interchangeable. Meanwhile Intel and Microsoft, who supplied the two pieces you could not swap out, quietly took most of the profit in the entire industry for the next twenty years.
Nobody saw that coming in 1983. Not because the people involved were stupid–they were brilliant–but because they were asking the wrong question. They were asking who makes the best computer. The question that actually mattered was: which piece of this thing stays scarce?
That is the whole game, and it is the lesson I want to plant in your head before we talk about a single AI company. Technology does not stay a blob. It organizes. It sorts itself into layers, and each layer becomes a business with its own unique brand of economics. In AI, those layers are already becoming visible. At the bottom sits power and land (actually)–electricity, substations, water (also, actually), permits, the deeply unglamorous physical world. Above that sits silicon, the chips themselves. Above that sits compute (a new word which my spell checker still tries to correct), the data centers that rent those chips out by the hour. Above that sit the models. Above the models sits the routing and orchestration layer, the software that decides which model handles which task. And at the top sit applications and distribution–the products people actually touch and the customer relationships that come with them.
Profit does not spread evenly across those layers. It collects wherever there is a bottleneck, and it drains out of every layer where the products become interchangeable. Here's the uncomfortable part: bottlenecks MOVE. They move as capacity gets built, as standards emerge, as buyers get smarter. The layer that prints money in year three is often a commodity by year eight, and the layer everyone ignored is suddenly where the cash flow lives.
So which way is the model layer moving? First, to clarify, I am referring to companies like OpenAI and Anthropic (amongst others) as being in the model layer. Let me give you three numbers and then tell you what they mean. Over the past year, the average price a business pays for AI has fallen roughly two-thirds. The share of that work being sent to the priciest, most powerful models dropped from about three-quarters down to under a third. And the typical business went from using two AI models to using nearly five. Now here's the translation. Companies stopped buying the best and started buying the RIGHT one for each job–the cheap model for simple work, the expensive one only when it's truly needed.
What this all means is that customers are not picking a champion. They are building routers. That is not fandom. That is procurement. Uber burned through its entire 2026 budget for AI coding tools by April and now caps each employee at $1,500 per tool per month. ServiceNow blew through its full-year budget with one provider in the first few months of the year. These are not stories about technology. These are stories about a CFO discovering a line item. 👀
Now, is there still real differentiation between the top models? Absolutely. Anyone who tells you a frontier model and a budget model are the same thing has not used both. But here is the distinction that matters to an investor: a moat you have to re-dig every ninety days is not a moat. It's a TREADMILL. And treadmills are expensive to run.
Which brings us to the pressure chain, and this is the mechanism underneath everything else. The companies that sell software and hardware to you and me are under relentless pressure to protect their margins. So they squeeze their suppliers. Their suppliers–the model providers–must therefore innovate faster while charging less. Today that math works, because those providers are swimming in venture capital that does not demand quarterly discipline. Private money is patient in a way public money simply is not. But that funding source has a shelf life. When those companies broaden their ownership and face a public shareholder base every ninety days, the character of the money changes, and the tolerance for burning it changes with it.
Want to know where the genuinely sophisticated capital already went? Not into models. In May, Blackstone committed $5 billion of initial equity–taking majority ownership in a venture valued around $25 billion including leverage–to build 500 megawatts of data center capacity coming online in 2027. Private equity did not fund a model. It funded the layer underneath the models. Meanwhile, NVIDIA is in talks to guarantee roughly $250 billion of financing (I covered a bit of that in yesterday’s blogpost) for an OpenAI data center project, and the reason is not complicated: OpenAI has no investment-grade credit rating and is projected to lose something like $14 billion this year on roughly $25 billion of revenue. That is not a growth story. That is a company in the middle of the stack discovering it cannot finance itself from the middle of the stack.
So look at where the layers actually sit today. Alphabet is, at this moment, the only public company that offers a top-tier model, runs it on silicon it designed itself, inside data centers it owns, reaching customers it already has. Microsoft has its own models and its own chips and is racing toward frontier scale, but is not there yet. Meta has its own accelerators and its own models–and rents chip capacity from Google. Amazon has its own silicon and leases its frontier capability from a partner. Everyone is sprinting toward the same destination, and the sprinting IS the evidence. Nobody spends that kind of money to replicate a position that isn't worth having.
And notice what Alphabet is doing with that position. It is selling chip capacity to Anthropic. To Meta. Apple has even used it. When the company holding one of the best models in the world starts renting out the machinery to its own rivals, it is telling you–loudly, in public, in its own capital allocation–which of those two businesses it believes has the biggest growth potential.
Now let me hand you the framework, because I would rather you evaluate the next ten companies yourself than take my word on one. Four questions. First: which layer is this company in, and is that layer becoming interchangeable? Second: does it OWN the bottleneck below it, or does it rent? Third: does it own the customer above it, or does it rent that too? And fourth, the one almost nobody asks: who is funding the capital expenditure–the company's own retained earnings, or somebody else's balance sheet? A company that owns both ends of its chain and funds itself is a fundamentally different animal from one that rents both ends and funds itself with promises.
Apply those four questions honestly and Alphabet does not come out clean either. Last week, it guided full-year capital spending up to some $205 billion, and management said plainly that all that infrastructure will pressure the P&L through higher depreciation and data center operating costs. Owning the stack means owning the depreciation schedule that comes with it. There is no version of this where somebody gets the assets without the accounting.
Here is what I'm watching, and it isn't an earnings beat. Alphabet deliberately deferred its 2027 capital spending number, saying only that it will rise significantly and that details would come later. That number, whenever it lands, is a statement about how long management thinks this buildout runs. And separately, watch for the first quarter in which revenue from selling chips to other people's AI labs shows up as its own disclosed line. That is the moment the market has to decide whether it is valuing a search company or a compute merchant–and those two things do not trade at the same multiple.
I know this was a bit technical, but it is SOOO important and relevant right now. So here's your takeaway, and it fits on a napkin. Stop asking who has the best AI. Start asking who owns the pipes. My first computer answered me in hexadecimal (almost as far from natural language as you can get and only slightly more understandable than 1’s and 0’s) and I thought I was talking to the Enterprise’s Computer. It was a beige box with a calculator glued to the front. Today the actual Computer sits in my pocket, and the reason it's there isn't that somebody protected the profit margin on it–it's that the thing got cheap, got boring, and got EVERYWHERE. That's not the miracle dying. That's the miracle finally showing up for the rest of us. Set phasers to: patient–it’s going to be a long journey, even at warp 3.
YESTERDAY’S MARKETS
Stocks had a mixed close yesterday with the S&P 500 gaining 0.21%, the Dow Jones Industrial Average adding 537 points (1.03%) to 52,747, and the Nasdaq slipping by -0.22%. A chip selloff in Asia overnight carried into yesterday’s session causing headwinds for the tech-heavy indexes. Hopes for a further ceasefire fueled further declines in crude oil and Treasury note yields.
NEXT UP
The FOMC will release its policy decision at 2:00 PM Wall Street time. The Fed is largely expected to keep rates steady. The Chair’s presser will begin at 2:30 PM. Don’t miss this one.
Important earnings today: Vertic, Bunge, Generac Holdings, VF Corp, Human, Boston Scientific, Johnson Controls, Vulcan Materials, Electronic Arts, General Dynamics, P&G, SoFi Technologies, Old Dominion Freight, Ares Capital, Wingstop, Amphenol, Chipotle, Ventas, VICI Properties, Meta, Boot Barn, Viking Therapeutics, O’Reilly Automotive, QUALCOMM, Tenable Holdings, Starbucks, Antero Resources, Public Storage, Microsoft, Carvana, Equinix, L2Harris, American Water Works, and International Paper.