Look at where capital flows today and the money goes to whoever makes artificial intelligence. Chip makers, data center operators, the labs training large language models. That’s the foundation of a powerful new technology, so the builder wins.
But does it work that way? Look at the history of technological breakthroughs and the inventor and the person who gets rich off the invention keep turning out to be different people. Sometimes they aren’t even in the same industry. And when you look closely at how AI is embedding itself into the economy, you start to think the main beneficiaries won’t be the model developers at all. They’ll be the companies that already control data, workflows, industry standards, licenses, networks.
Imagine a new kind of energy source gets invented tomorrow. A hundred times cheaper than what we have, and you can run it to any machine, office or process. Where do you put your money? Into whoever produces the energy, probably. But watch what happens next. If the new electricity is that cheap and that available, it stops being an advantage on its own. It turns into a common resource, like air. The advantage goes to whoever owns the thing that’s useless without electricity and worth an order of magnitude more with it. Industry, say.
The first bet is always that the technology turns out to be enormous and rewrites the economy. The second bet is that the gains land on whoever makes it. The first is almost always right. The second isn’t obvious at all. You can run the nature analogy, of course, and picture a breakthrough behaving like water. And like water, AI won’t spread evenly. It’s already looking for existing channels (moats) and cutting them deeper. The investor’s question isn’t only who makes the water. It’s who owns the channel. You can be completely certain that AI will transform the world and still be wrong about where the capital pools.
Who made money on electricity?
Yes, the subject is an old one, but guessing at future change means looking at past events. So we start further back.
In 1882 Edison switched on the first central power station in New York. The technical properties of electricity were settled by then, the dynamo worked. Sensible logic says a productivity explosion follows. It never came. The economist Paul David showed long ago in his work that as late as 1900 electric motors accounted for under 5% of mechanical power in American factories. Real productivity growth from electrification only began in the 1920s. Almost 40 years between the invention and the payoff. The explanation isn’t in the dynamo. It’s in how factories used it. At first the factory owners just swapped the steam engine for one big electric motor. Everything else stayed put. Same central shaft under the ceiling, same belts and pulleys, same floor plan built around mechanical transmission. Electricity was hooked up, but the factory was still a steam factory in principle.
The real payoff showed up when engineers worked out that the motor didn’t have to be one per factory. Give every machine its own. And once you do that, the whole shop floor can be laid out again, equipment placed by the logic of the production flow instead of by how far the shaft reached. They called it the unit drive system, and that’s what produced the jump. Not electricity itself. The reorganization of the process around it is what got things moving.
On how the profits split: the dynamo makers earned, sure. But the bulk of the gain went to the people who owned the factories, the distribution and the brands, and who managed to rebuild their business around the new energy. The technology was open to all. Anyone could buy a motor. Turning a motor into a real advantage was only possible for someone who already had everything else.
The economist David Teece described this pattern as a general rule back in 1986:
“The profits from an innovation may accrue to the owners of certain complementary assets rather than to the developers of the intellectual property.”
Production capacity, sales channels, reputation, industry connections, licenses. If the invention is easy to copy and the complementary assets are hard to copy, the asset owner is the main beneficiary.
To work out who makes money on AI you have to answer a few questions first. How easy is the AI technology itself to appropriate? And what counts as the complementary assets this time round? Start there.
Who creates and who captures
Two completely different occupations get folded into one here. Making AI and pulling rent out of its deployment aren’t the same thing. These are different businesses with different economics.
Making AI means training models, pushing the scientific frontier, building infrastructure. Pulling rent means fitting ready intelligence into someone else’s life so that the payment comes to you. But can the maker hold the gains without letting them leak to whoever stands further down the chain?
Holding the gains means protecting the technology from copying. Innovation economics calls this appropriability. And appropriability is a problem for the model builders. Pierre Azoulay, Joshua Krieger and Abhishek Nagaraj work through the question. Patents protect large models badly. The transformer architecture, which the whole current wave sits on, was published in an open article in 2017. The scientific principles are common knowledge, and good engineers move between companies carrying what they know. When one model’s weights leaked, the developer community built dozens of derivative systems around them inside a few weeks.
In pharma the patent on a molecule expires sooner or later and generics show up. On that logic the rent should go to zero. Take insulin. The formulation patents expired decades ago and a handful of large companies still hold the market, Eli Lilly and Novo Nordisk among them. An open recipe didn’t wreck their position, because the position never rested on a secret formula. It rested on brand, distribution, manufacturing and the confidence of regulators. The technology opened up and the moats stayed where they were.
In practice this means the level of the model holds up badly as an advantage on its own. One model leads today (Claude Mythos, say), and in six months the competition has caught it. Open models will show up that run almost as well as closed ones at a fraction of the cost. Model quality is turning into a commodity fast, hard to tell apart between suppliers. And when something becomes a commodity the price falls and the profit with it.
To keep this concrete, take the internet alongside electrification. Plenty of people in the nineties were sure the money would go to whoever built the network. Telecom operators, router manufacturers, owners of backbone capacity. The network did get built and it did change everything. The profit went somewhere else. Not to the pipe layers and the fiber suppliers, but to the owners of assets that were worth nothing without the network. Google didn’t invent the internet, it owned a search engine and user attention. Amazon didn’t lay cable, it owned warehouses, logistics and a customer base. The network turned into a cheap common resource and the gains settled on the complementary assets bolted onto it.
Note that the basic web protocols and standards (http, html, url) are open and free to use, and CERN put the Web technology itself into the public domain back in 1993. The colossal fortunes of the internet era were made by owners of scarce resources plugged into that open infrastructure. Same logic for AI. An open base technology that keeps getting cheaper isn’t a source of durable excess profit by itself. Long-term value goes to whoever controls the scarce complement.
If the technology itself is hard to hold, the gains don’t vanish. They move to whoever controls the resource the model is useless without, which is those complementary assets. Compute infrastructure, serving capacity, safety procedures, benchmarks, access to huge bodies of non-public data, and the network effects around that data.
Break competition around AI foundation models into parts and you get six key complementary assets, and that’s where the real fight is:
The compute environment (capacity for training models).
Model serving and inference (the infrastructure that gets the model’s finished answer to the end user fast and cheap).
Safety and governance (risk control procedures, without which a large corporate or government client won’t let the system near its data at all).
Benchmarks and metrics (accepted ways of measuring which model is better, and therefore who gets the trust).
Training data (the material a model acquires its capabilities on).
And data network effects (the gain in model quality from every new user, which feeds further growth by itself).
All six rest on each other and add up to what competitive advantage means for a technology company now. The model developers control part of the list, compute for one. The rest sits with other players. Here’s the investor’s interest. If the technology can’t be appropriated and only its complements can, the question gets simple. Which complement is scarcest and least replaceable? What does AI need most and can’t produce for its own training at this stage? Data, obviously.
AI needs data more than data needs AI
An AI model is an algorithm that learns from data, and the more data and the better its quality, the better it works. There’s research on when AI adoption actually pays for itself at company level. At low adoption, AI gives no revenue gain at all. The return only appears at high intensity of use. But it’s far stronger at companies that put money into adjacent technologies at the same time, databases and cloud storage above all. Firms with those investments came out tens of percentage points ahead on revenue growth compared with firms that adopted AI without them.
So the value of AI isn’t in the model. It’s at the seam between the model and the data the model reaches. A smart model without a good database stalls. An average model with a good database starts making money. This is where AI differs from past general purpose technologies, because the usefulness of the algorithm can’t be separated from the volume of data and the ability to process it.
Think of it like this. If I hold a closed body of data nobody else has, I can plug in any of a dozen models on the market and get a result. Tomorrow I swap the model for a cheaper or smarter one. The data I can’t swap, there’s nowhere else to get it. Scarcity sits on the data side, not the model side.
Whoever spent years piling up industry data suddenly owns the scarcest asset in the new economy, having done nothing new, just carried on with the business. That’s the first category of companies that wins out of all proportion. Owners of industry data, the kind that’s hard to gather again and that decisions in a given industry depend on.
Bloomberg spent decades building a body of financial data and the terminal it gets consumed through. AI doesn’t destroy that body, it makes it more valuable, because a good financial agent is useless without reliable data competitors can’t reach. Same for RELX with LexisNexis and its scientific publications, or Thomson Reuters with Westlaw. Legal AI is exactly as good as the case law underneath it, and those bases took generations to assemble and sit behind licenses. IQVIA plays the same role in healthcare, owning vast bodies of de-identified medical and pharmaceutical data. CoStar in commercial real estate, Verisk in insurance. Models get cheaper with time and turn into a common resource. Data gets more expensive and stays exclusive.
The obvious question follows. Why can’t such data just be bought or gathered again? Simply because it piled up over years as a byproduct of real activity a newcomer has no access to. You can’t assemble thirty years of legal precedent after the fact, or a whole country’s history of medical prescriptions. These bodies are fenced in by licenses, contracts and privacy regulation. This isn’t Libgen and the other dark libraries, or Sci-Hub and its access to scientific papers, which already get used for training anyway (illegally). Exclusive private data can’t be had that way. And the tighter the rules on handling data get, the higher the fence around whoever collected it legally already. The critical need for cybersecurity, which we covered earlier, works here too.
There’s more. When industry data is wired into a working process, every new user action throws off a bit more data, which improves the model, which pulls in new users, who leave more data. The wheel spins itself and it spins for whoever is already ahead. A newcomer catching up needs more than a better model, it needs access to a flow it doesn’t have by definition. So the owner of unique data has two advantages, not one. A stock that can’t be reproduced and a flow that refills itself.
Control over the process
Say you have an incredibly smart model. You’re not going to plug it into nothing. You’ll plug it into some working process. Into the way a hospital keeps a patient record. Into the way a company issues invoices and closes its books. Into the way an insurer takes an application and prices risk. Back to electrification: the model is the motor, the process is what the motor bolts onto. And the mountings matter more than the motor, because you can replace a motor and rebuilding a process is very hard. Whoever owns the system the work runs through owns the point where intelligence meets reality. That’s where the most profit concentrates. Improve the model all you like. If the industry’s work runs through someone else’s system of record, the owner of that system decides which model gets in and on what terms.
Fresh research on AI diffusion from economists led by a US Census Bureau team backs this up indirectly. They split adoption into three layers, answering:
Does the company use AI at all?
In which functions specifically is it deployed? Do individual workers use it in their own tasks? And they found that these layers often diverge.
Do individual workers use it in their own tasks?
As a rule the layers turned out to diverge. At plenty of companies the workers use AI on their own and there’s no formal deployment at all. But the real changes, cutting some jobs included, don’t come from individual employees playing with AI at their desk. They come from structural integration of AI into business functions. What counts isn’t the toy in a worker’s hands, it’s how far the technology is built into the process.
Which gives the second category of beneficiaries. Owners of workflows and business operations. Companies whose software is the system an industry runs its daily work through.
Examples again. UnitedHealth Group owns a huge share of administrative, payment and operational processes in American healthcare. Any clinical AI has to work inside that process to be useful. Veeva does the same for pharma, owning the systems where clinical trial and drug sales processes live. SAP and Oracle hold the internal processes of large corporations, from finance to logistics. Intuit owns how small businesses and individuals keep books and file taxes, through QuickBooks and TurboTax. ServiceNow holds internal support and IT operations at big companies.
For how durable this is, the smartphone analogy helps. There’s a bottom layer (operating system and device) and a top layer (the apps living on it). An app developer can be as talented as you like, but he works on someone else’s ground under someone else’s rules and hands the platform owner a cut of every sale. The model is one layer of the stack. The real work happens a layer above or below. Best model isn’t enough when you’re locked in an app on someone else’s platform.
The generative AI stack really does lay out neatly as four layers, one on another, a pyramid. Compute at the base, the capacity without which nothing runs. Data above it, the material those computations feed on to produce trained intelligence. Above that the foundation model, the finished algorithm trained on that data. And at the top the application, the interface and the workflow where an end user or a business meets the stack. The higher the layer the more visible it is and the more talk there is about it. The lower the layer the harder it is to copy and the more firmly whoever took it is dug in. The model sits second from the top, and that position alone doesn’t let you dictate terms to the stack.
A second consideration, about time. The forty year lag in electrification is repeating. The return from AI arrives late, and at the early stage of adoption you can’t see it at all. First the company spends, rebuilds processes, retrains people, and only then, once the rebuild is done, the profit shows. That lag also works for the process owner. Rebuilding a process around AI is only possible for whoever owns it and knows it from inside. An outsider with the smartest model gets stuck at the door, not knowing where to plug it in or what to move around it.
What these companies have in common is that none of them has to build the best model in the world. Owning exclusive data for the work to run on is enough. AI raises what that’s worth, because the smarter the intelligence the more the route it takes to the job matters.
I’d add that technology penetration is getting much faster along its exponent:
AI is setting a record on speed of penetration, a few years against decades for earlier technologies. But the speed at which that becomes productivity growth will be an order of magnitude slower, another 5-8 years maybe, extrapolating from previous breakthroughs.
Standards and barriers
Can a model, however smart, issue a loan by itself, settle a payment between banks, approve a drug, put a legally valid signature on something? Rhetorical question.
Some assets get their value from law, trust and coordination. Heavy regulation and safety standards strengthen the strong, paradoxically. Strict requirements are easier to meet when you already have resources, reputation and experience under supervision. European privacy regulation, for one, raised concentration in the web services market rather than lowering it. A barrier put up to protect people also walled the leaders off from newcomers. A whole group of categories surfaces from this, because they share the mechanism. They hold onto a barrier the technology doesn’t destroy.
One of them is owners of industry standards. Think about whoever sets the yardstick the industry gets measured by. Rating agencies like Moody’s and S&P hold the privilege of issuing an assessment the markets trust, and they hold it because regulation fixed the role there, not because of computing power. RELX owns analytics and tools that became the industry norm on top of its data.
The strength of a standard is that everyone else uses it. Switching alone is pointless, you’d stop understanding your counterparties. Switching all at once is close to impossible, nobody wants to go first. AI changes nothing here, it has to speak the language the market already speaks. It’ll probably lock the standard in harder, since it’ll build itself into it and start working on it.
Fourth category, financial infrastructure. Exchanges, clearing houses, depositories, custodians. Their role is to be the neutral node every side of a trade trusts. Picture an army of superintelligent trading agents. They’ll still trade through an exchange like CME or ICE, and settlement still runs through something like DTCC. The more autonomous agents trade, the more operations pass through the node. The value of the node grows because of the automation, since the more autonomous systems trade and settle with no human in the loop, the higher the demands on the reliability and neutrality of the central link, because the cost of a failure rises with volume. A trusted intermediary in a world of machines is needed more, not less.
Fifth, payment networks, Visa and Mastercard types, which earn on having practically every merchant and almost every bank connected at once. When AI agents start buying on people’s behalf the money still runs the same rails, because there aren’t any others.
Worth understanding why regulation so often plays for the big when it’s meant to protect the weak. Complying with strict rules costs money, lawyers, audits, built-out procedures. For a giant that’s a bearable line item. For a newcomer it’s a wall, and the more seriously society takes AI risk and the tighter it regulates, the more reliably it walls the incumbents off from future competitors. The companies that already know how to operate under supervision get the head start, since the bar often gets set to their own template.
And look how resistant this is to progress in the technology. Double the intelligence of the model, multiply it by ten, and it’s no closer to a banking license, a seat in a settlement system, or the right to issue a rating the law says to believe. Intelligence scales for almost nothing. Law, trust and consent don’t scale at all. Which is why assets standing on rights rather than cleverness gain the most in a world of cheap intelligence.
Network effects in a world of autonomous agents
There’s a common expectation that autonomous agents will destroy network effects. The logic goes: if a smart agent acts for me, it goes around any platform, finds the best offer directly, and the power of the marketplaces disappears. It looks like the opposite.
Start with what a network effect is. The value of a network grows with the number of participants. A telephone is useless when you’re the only one with a handset and priceless when everyone has one. A marketplace with sellers and buyers is worth more the more of both are on it at once. Old idea. The question is what AI does to it.
Azoulay and his coauthors pick out a particular kind of network effect tied to data. More users interacting with the model means more feedback, which means it gets better, which pulls in more users. The wheel spins itself. But there’s another turn I want to point at, a subtler one.
Think about what an agent does acting on your behalf. It doesn’t cancel the market, it walks into it. To order, compare, buy, negotiate, the agent needs a venue with someone to compare and someone to negotiate with. So the agent needs the most crowded venue, for exactly the reason a person does. And the agent is more rational than a person, so it’ll go systematically where the liquidity, the choice and the counterparties are. Autonomous agents don’t thin the network out. They concentrate traffic on the largest networks harder than people ever did with their habits and their laziness.
Add switching costs. That’s the sixth and seventh categories at once, they’re tied together.
Sixth, vertical industry platforms. Companies that pulled data, processes and a network of participants together inside one industry. Veeva in pharma, Guidewire in insurance, Shopify for merchants, Toast in restaurants. They hold the data channel, the process channel and the network at the same time. AI coming into that industry has to come through them, because they are the environment the industry lives in. An agent serving a restaurant will work where the restaurant’s whole operation already runs.
Seventh, companies with high switching costs. The ones whose product is stitched so deep into a client’s work that leaving costs more than putting up with it. SAP and Oracle enterprise systems, core banking at Fiserv, FIS and Jack Henry, Autodesk in engineering software, Epic again in medical systems. Note this part. AI doesn’t lower those switching costs, it raises them. Once AI processes trained on the system’s data are built on top of it, tearing away from the system hurts more. The deeper intelligence grows into someone else’s infrastructure, the tighter that infrastructure holds the client.
Picture a concrete scene from the near future. Your agent has to order a batch of stock for a small shop. It isn’t going to go negotiate with every supplier out in an open field. It’ll go where the suppliers are already gathered, where there’s a transaction history, ratings, a payment method, a dispute mechanism. The largest platform. And the more agents come to market, the harder they pool into the same points of concentration, because a rational agent needs maximum liquidity. A person might still pick the small venue out of habit or laziness. An agent doesn’t. It goes coldly to where the choice is wider and the counterparties more numerous. The paradox is that the smarter agents get, the more monolithic the networks they act through become.
Put the section together. Networks don’t disappear in a world of agents, they get denser. Switching costs don’t fall, they rise. Both work for the owner of the channel, not against him. What’s left is whether all this adds up to a pattern or is just a set of convenient examples.
Why AI strengthens concentration rather than destroying it
Pull the threads together and see where they go. There’s a widespread hope that AI turns out to be the great equalizer. That it hands small players the capabilities of big ones, that a loner with a model outruns a corporation, that the moats of the large companies run shallow. A lovely thought. Check it against facts too.
Start with how AI is actually spreading through the economy right now. Diffusion research shows adoption skewed hard toward large companies. Among very large firms in the information sector, professional services and finance, use runs to 60% and above, while the bulk of the economy is only approaching the technology. AI arrives first where the scale, the data and the capital to rebuild processes around it already exist. That rebuild comes much harder to the small, exactly as replanning the shop floor came harder in the electrification years.
The return from AI doesn’t appear by itself. It appears in combination with complementary investment in data, infrastructure and in a firm’s own tailored development. Who has the resources for that? Whoever already holds enormous capital and has run a serious expansion. A technology that only pays off with expensive complements in place favors, by definition, whoever has the complements already.
And the wider historical view. Lucrezia Fanti, Dario Guarascio and Massimo Moggi traced AI’s path from scientific discipline to a field corporations dominate. Their conclusion: AI became an area with heavy concentration of technological and economic power, a key instrument for companies whose business model is built on collecting data. The technology doesn’t dilute the power of whoever controls data and processes. It reinforces it, and helps them put clients, suppliers and competitors further under.
The same authors stress another facet of this. The technology lets work be broken into small measurable operations, each one tracked and optimized. The fruit of that optimization goes to whoever owns the platform where the work is mapped and measured. Not the worker whose labor got more transparent and more replaceable. The owner of the system the labor is visible in. AI moves bargaining power up, to whoever holds the infrastructure, and away from whoever works inside it.
Three observations together: AI comes to the large corporations first, it pays off only with expensive complements, which the large have, and historically it strengthens whoever already owns data. All three vectors point at moats getting deeper, not eroding.
Not an iron law, more of a tendency. There are forces running the other way. Open models really do lower the entry threshold for applications. Sometimes a player shows up who gives away powerful technology for free for reasons of his own and breaks everyone’s plans. Sometimes a new channel cuts through where there wasn’t an old one, and the newcomer wins instead of the incumbent. So betting blind on concentration won’t do. But the base case, which theory, history and fresh data all push toward, is that AI is more likely to strengthen market concentration than destroy it.
The old story
When somebody says bubble, the usual implication is that it’s all about to collapse. Those aren’t the same thing, and the investor Dan Niles gives a sober view of the difference. He compares the present moment with the end of the nineties fairly often.
Netscape launching in late 1994 started the building of the internet roughly the way ChatGPT in late 2022 started the AI wave. Then, in 1997 and 1998, the market corrected hard several times. First the currency crisis in Asia, then the Russian default and the collapse of a big hedge fund. The index dropped 10% and almost 20% inside the year and still closed those years up, because real internet infrastructure was going up underneath all the fright, and it was year three and year four of the cycle.
A bubble and a real transformation don’t rule each other out. Bubbles only inflate around genuine generational shifts, canals, railways, the internet. But not everyone inside a genuine shift survives. Last time the market built the internet and ground a lot of loud names into powder.
On valuation, the argument that gets brought up most: Cisco at the 2000 peak was around 140x annual earnings on revenue growth of roughly 6%. Today’s chip market leader runs about 25x annual earnings on growth near 80%. Which says the current leaders are cheaper on multiples than the heat of the conversation makes them look, so the bubble can keep inflating a while yet. You can argue with that, and he doesn’t hide that he expects a serious drawdown ahead. But guessing the exact date of the turn is thankless work (spare a thought for Michael Burry’s calls, 40 out of 2 and counting).
The economist Carlota Perez worked through this mechanism in her work on technological revolutions and financial capital. Every big technological shift takes the same road. Financial capital bursts into the new technology and inflates a bubble. The bubble pops. And only then, on the wreckage, the real growth starts. It’s the bubble that builds the infrastructure the growth couldn’t have happened without. The speculator who lost his money is just paying for progress out of his own pocket.
Perez gives two types of capital we’re used to treating as one. Production capital lives inside the real business. Factories, workers, suppliers, and it thinks in long 5-10 year cycles. It needs stability, because only stability lets you plan. Financial capital is built differently. It needs a return here and now so it can move the money into the next idea immediately. Shorter horizon by an order of magnitude, higher mobility. In quiet times the two get along. One makes the profit, the other finances it and takes its share. Let a technology show up promising endless growth and the balance breaks at once. A mature industry brings in 8% a year, call it. The new technology promises a hundred. For financial capital the choice makes itself, and money starts draining out of dull profitable companies with real cash flows toward the places with no profit yet and a beautiful story. At the end of the nineties capital left mature, profitable businesses for whoever promised the internet. Jeremy Grantham described the same process in his notes through the dotcom rise and crash. Today capital is going into the AI story the same way, concentrating as it goes.
Any shift like this then breaks into four phases, always in the same order. The whole path draws as one curve.
First phase, irruption. The technology has just appeared, there’s little money in it and less understanding, but the first investors already smell the potential and go in. Enormous risk, and enormous return for whoever guessed right.
Second phase, frenzy, the bubble stage. The technology has proved it works, and it’s nowhere near mature. In that gap anything looks possible. The narrative starts hitting harder than the fundamentals. Money goes in not because somebody understood the mechanics but because they’re afraid of being late. Financial capital tears loose from production capital for good. And this is where the infrastructure gets built, on the euphoria.
Stop here a moment. A long-term value investor won’t go into infrastructure like that, the payback is too long and the risk too high. Only capital in a state of euphoria will build it, capital ready to fund projects that start paying back in ten years if things go well. And that flow of capital is what accelerates technological progress. Without a bubble, infrastructure on this scale might never have appeared.
Third phase, synergy, arrives after the crash. The bubble burst, a lot of money burned, the regulators tightened the screws. The infrastructure is still standing, and now it costs less to use. Financial capital comes back to production capital, real cash flows and dividends matter again, and the growth starts that would have been impossible without what euphoria built. Google and Amazon grew on fiber laid by Global Crossing and WorldCom, both bankrupt. In our case the stage goes to companies that rise on the bones of whoever built the infrastructure for AI.
Fourth phase, maturity. The technology is fully deployed, growth slows, production capital piles up money again with nowhere to go inside industries that have aged. And then the next anchor technology shows up on the horizon and the cycle runs again.
Extrapolate the scheme to today and every sign puts us in frenzy, and not at the start of it. The big bang landed on GPT-3 and GPT-4, the moment the technology left the labs for mass use. Financial capital has torn loose from production capital, companies with no profit are valued at multiples that don’t square with any sober scenario. The narrative hits harder than the fundamentals, and one word, AI, in a press release lifts the market cap mechanically. Just look at the cash flows and profits at Intel, CoreWeave, or SpaceX, which went public recently.
CoreWeave is the WorldCom of this cycle in pure form. The company builds data centers for GPU rental, 2025 revenue came to around $5B, up 174% year over year. But the end of 2025 balance sheet carried $21.4B of debt and a $1.17B loss, CapEx for the year alone ran $12-14B, and interest expense tripled over the year. The company builds infrastructure on borrowed money, betting that demand for compute keeps growing faster than the obligations. While the contracts hold, the scheme holds. But $21B of debt on $5B of revenue isn’t a business model, it’s a bet. And whoever finances the debt pays for it.
Intel is a different case. True, not a startup, a company with a long history, which bet that the next wave of AI demand runs through its foundry. The foundry division lost $10.3B at the operating level in 2025 on $17.8B of revenue. That’s a structural loss. Two years, $12.5B of losses in one division. Intel took $20.4B of outside money to stay afloat. The same speculative logic is at work here as in the whole current story.
SpaceX is the most interesting case, it shows the narrative working better than anything. The company went public at $1.77T on $18.7B of revenue and a $4.94B net loss for 2025. P/S came to 93.6x at the end of 2025, ten times Tesla and eleven times Meta. Cisco at its 2000 peak traded around 30x revenue, for comparison. SpaceX is more expensive than that, with an enormous loss behind it. The narrative is that AI infrastructure develops on the back of SpaceX, with Musk as the symbol of technological destiny. So the infrastructure gets built at volumes that plainly outrun today’s demand. Data centers, cable and power capacity, paid for by belief in long-term growth.
There’s another indicator, no less interesting, that checks the phase by the character of the financial instruments on the market. Perez, for instance, splits the instruments that serve a technological cycle into six categories by function.
Category A covers instruments that put capital behind new products and services. Bank credit, venture capital, syndicates and joint stock ventures, instruments for financing new infrastructure and for trade in new goods and services.
Category B covers growth and expansion instruments. Bonds for incremental development of production, state financing in various circumstances, from war to colonial expansion and social spending, and instruments for moving production abroad.
Category C means modernizing the financial system itself, bringing new communication, transport and security technology into banking, developing new forms of client service, new credit and payment instruments.
Category D covers instruments for locking in profit and spreading investment risk. Various forms of mutual funds, certificates of deposit, bonds, IPOs, junk bonds, and instruments for large risky bets like derivatives and hedge funds.
Category E is refinancing obligations and mobilizing assets. Debt restructuring, buying working production assets through mergers and acquisitions, acquiring and reselling rent-bearing assets like real estate and valuables.
Category F covers dubious innovation. Legal loopholes and tax havens, earning on somebody else’s ignorance and on information asymmetry, up to outright pyramids and fraud.
At the top of the spectrum stand A and B, which bring money together with real construction (venture, syndicates, paper against new capacity). At the bottom sit E and F, which make money out of money and skip production entirely (arbitrage on somebody else’s ignorance, schemes for shuffling assets, outright pyramids at the limit). C and D lie between the poles, nearer the middle.
Project the classification onto the phases of the cycle and a general dynamic shows up. In irruption all six work at once, the peak of substantive financial innovation serving the real creation of new infrastructure and products. By frenzy mostly D, E and F are left, because investors put less and less into production and more and more into speculative trade in the financial instruments themselves, into refinancing and shuffling assets, piling on leverage and inflating prices with no matching growth in real revenue or capacity. After the crash, in synergy, A, B and C come forward, because capital goes back to financing growth that’s confirmed rather than imagined. In maturity B, E and F dominate, capital keeps financing expansion past former boundaries (new markets, production abroad) while doing more and more merging, acquiring, routing profit through tax havens and other earning that isn’t connected to creating new value.
In the breakthrough and in the recovery the creative instruments rule, capital works as a bridge between money and production. In the phase of madness the instruments of detachment and speculation climb on top, the ones living on the movement of paper rather than the creation of value. The thicker the circular schemes for inflating company valuations, the closer the peak. Today it’s structured products on AI stocks, leveraged ETFs growing, aggressive trading of the narrative itself (very pronounced under Trump). Even now, finishing this piece, I ran into a news feed that reflects all of the above. Korea’s Kospi fell 10% in a single session off an all-time high and dragged semiconductors down worldwide. The mechanics under the drop were exactly what I’ve described. Leveraged ETFs on Samsung and SK Hynix, grown from $3B at the May launch to over $9B, have to rebalance daily. Nomura puts the forced trades such funds generate at around $9B for every 1% the market moves. The instrument lives on movement, not on value, and in a fall it amplifies the fall as hard as it amplified the rise before. A Barclays strategist called it a situation where leveraged ETFs are the market’s biggest technical risk no matter what you believe fundamentally.
UBS meanwhile, canvassing dozens of clients across the US and Europe, recorded hedge funds starting to cut their AI exposure. They still believe the story long term, and they see the one-bus risk more clearly, where too many hold the same trade. Probably a short-term shift in narrative, since the OpenAI and Anthropic IPOs are ahead and we’ll see new boundaries of asset revaluation upward and another wave of optimism. Being in frenzy doesn’t mean a fall this minute.
And the end of it won’t mean the end of the technology. It’ll mean the end of the illusion that the rent goes to whoever builds the infrastructure. When the current bubble deflates the data centers don’t go anywhere. GPU clusters, cable, trained models, all the physical and software guts stay standing. And the next generation of companies takes it for almost nothing, exactly as Google took the abandoned fiber. The speculator pays for progress again with his own losses.
Arguing over whether this is a bubble is empty work, the Perez answer is obvious, yes it is. The real question is different. Who owns the channel when the water goes down. And the whole earlier part of this piece has drawn that map already.
Where the moats are
I started with a very common theme, that whoever makes AI wins. Following the chain through gets you to a completely different picture. The technology is hard to hold, it gets cheaper and turns into a common resource. The gains settle where a channel already exists, one the technology deepens and can’t cut fresh.
Several channels turned up. Unique industry data, which AI thinks on. Workflows, which it gets applied through. Standards, licenses and regulatory barriers, which it doesn’t cancel. Financial infrastructure and payment networks, which the operations pass through anyway. Vertical platforms and companies with high switching costs, which AI binds to itself harder still. The same logic every time. The flow looks for the channel, and the owner of the channel takes rent off the flow.
What does that mean for somebody deciding where to look? If AI really does become a new universal layer of the economy, something like electricity, the most interesting assets may not be among the producers of that layer. They may be among whoever controls the points where it touches the real economy. Decision points. Data. Standards. The infrastructure AI has to pass through to turn into money.
None of which means the model developers lose. Some of them own the most important complementary assets, compute and access to data, and they’ll take their share too. But betting only on the makers of the technology repeats the mistake of the people who thought in 1900 that the dynamo manufacturers would get rich, and missed that half the gains were going somewhere else entirely.
The main risk in this view is that channels shift sometimes. The technology can turn out powerful enough to cut a new channel where the old one looked permanent, and then the owner of the former moat is left with nothing. So it’s worth looking not at any moat but at the one whose value rests on law, trust, consent and accumulated data, on what AI can’t reproduce.
How do you tell a durable moat from a fragile one in practice? A simple test helps. Ask yourself what happens to the asset if intelligence becomes free and available to everyone. If the asset rested only on somebody counting better than the rest, it loses its value, because now everyone counts. If it rests on accumulated data, on being embedded in someone else’s process, on a license, on the market’s trust, on a network everyone’s connected to, then free intelligence doesn’t sink it, it lifts it. Everything around gets cheaper and the asset stays scarce. A durable moat is one that gains value precisely because the technology got cheap. And there are far more of those on the market, if you look, than it seems at first to whoever’s watching only the model developers.