The technology is working and demand is growing, so this is not a case of a worthless asset. The risk sits in the financing. Companies that borrowed against fixed dates have to refinance repeatedly through 2030, and market disruption can cause them to fail while still being profitable. Continue reading
In August, Nvidia arranged $500 billion in financing. The money comes from Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, and it funds other companies buying Nvidia chips. The debt sits in special-purpose entities, off Nvidia’s books. The collateral is the chips. Nvidia will use this to finance customers buying Nvidia chips.
Analysts expect AI to need about $1.5 trillion from credit markets through 2028. One company can now supply one-third of the total, as long as you are buying Nvidia chips.
For three years, people discuss whether AI works, whether it will be adopted, who will pay for it and why. Nvidia is playing the smiling card dealer, with a pile of chips, helping finance people to buy in.
In 2026, AI has shifted. AI buildout does not come from speculation or profits. In 2025, the eight largest cloud companies produced about $180 billion in free cash flow. In 2026 they spend past it, to roughly negative $64 billion. And Nvidia is counting on that spending to rise in 2027 – debt that competes directly with the US government’s job of rolling $9 trillion every year.
AI and the US government are bidding for the same lender. And everyone asks – is AI a Bubble?
Most people define a bubble as a price that runs far ahead of what the asset earns. That is true but not useful. It only tells you something after the crash.
Here is the more useful definition. A bubble is when the money comes due before the revenue shows up. The asset can be excellent. The demand can be real and growing. If the loan matures in year five and the payoff arrives in year twelve, the owner still loses everything.
Some cases from the last twenty-five years make the point.
Global Crossing, 2002. The company laid fiber optic cable across oceans during the dot-com boom. The fiber was real, and traffic on it grew every year. But they borrowed on 25-year terms to build something with five-year economics, and the revenue arrived slower than the payments. They filed the fourth-largest bankruptcy in US history. Everyone uses this great cable now. They reap low costs because the investors lost their pants.
Financials in 2007. Banks grew to about 22% of the S&P 500. The banks were real companies. The mortgages were real loans. The houses were real houses. The flaw was funding. Banks borrowed short and lent long, so they had to refinance constantly. In 2008, the short-term money stopped. Like a Jenga tower, the balance of short and long collapsed and took the market with it.
Cloud computing, 2013 to 2020. Everyone piled into Amazon and Microsoft, and the early buyers were right. Amazon built AWS out of retail cash flow. Microsoft built Azure out of software profits. Neither one borrowed against a deadline, so neither one had a date it could miss.
Notice what does not separate them. The technology worked in all three. Fiber worked. Mortgages were real. Cloud was useful. What separates Boom from Bubble is the distance between long-term income and short-term financing.
There is no single AI industry. Four different businesses use the same word, and each one is betting on something different.
Plan one: build it from profits. Take money the company already earns and spend it on data centers you own. The bet is that you will still want the capacity in five years. If you are wrong, you slow down and nothing breaks, because nobody is waiting on a payment. Microsoft and Google build from operating profits, and Microsoft was the only large US cloud company with positive free cash flow last quarter.
Plan two: sell the shovels, and lend people the money to buy them. You make the hardware. You also arrange the financing so customers can afford it. Your revenue looks excellent right away. The bet is that those customers earn enough to repay the loans you helped arrange. Nvidia sells the chips and arranges $500 billion in outside financing so customers can buy them.
Plan three: borrow, build, and rent it out. You take on debt to construct capacity, then lease it to companies that need compute. The bet is that rental prices stay high enough, and long enough, to cover the loan payments. Oracle spent 174% of its operating cash flow on capital projects this year and is raising $45 to $50 billion more in debt and stock.
Plan four: own nothing. You sell a model or a service and rent the hardware from somebody else. Your costs move with your revenue. OpenAI and Anthropic own almost no data centers and rent their compute, as do smaller providers like RunPod that resell capacity to developers.
Each plan wins in a different world. Plan one wins if demand grows slowly. Plan three wins if demand grows fast and stays expensive. Plan two wins in the short run no matter what, and finds out later.
The reason “is AI a bubble” has no clean answer is that all four plans are running at once, inside the same industry, funded by the same lenders.
When people say AI in 2026, they usually mean a chatbot. That is one category of model, and it is the newest part of a much older business.
The most profitable AI running today is Google Rankbrain, and Meta’s recommendation protocols. It decides which post you see next, which product Amazon shows you, which video plays after this one. It is not a chatbot, and it does not talk. It has been running for over a decade; it directly produces advertising revenue, and its return on investment is measured every day. When Meta spends $130 billion this year, a real portion of that serves a business that already works.
Then there is image and video generation, which uses a different kind of model entirely. There is scientific work, like predicting how a protein folds. Fraud detection, pricing, routing, and forecasting also quietly earn money for years under the name machine learning.
So the question of whether another kind of model takes over is really two questions.
Will language models stay the center of attention? Probably not forever. Video and world models are growing fastest right now, and they consume far more compute per output than text does.
Does that strand the hardware? Mostly no, and this matters. A data center full of graphics chips can run recommendation, video, science, and language. The building, the power connection, and the cooling do not care what model you run. That fungibility is the strongest argument against a total collapse, and it is the argument Microsoft makes when investors ask what happens if chatbots disappoint.
The demand can move. The concrete stays useful.
Google has designed its own AI chips since 2015 and runs much of its work on them. Amazon builds Trainium, Meta builds MTIA, Microsoft builds Maia, and a company called Groq built a chip that does one narrow job extremely fast. These parts run cooler and use less power than a general-purpose graphics chip, because they do one thing instead of everything.
The catch is the same thing that makes them good. A chip built for today’s model design is worth much less if the design changes, and general graphics chips are the ones that survive that change.
A lot of people say, “I will never use AI” and “I won’t pay for it.”
Ask them about AI movies, and they say, “AI movies look bad,” “I can tell right away,” and “I won’t pay for it.”
They are right, and it does not matter. AI is not entering entertainment through a movie you refuse to watch. It is entering through the parts nobody sees.
Here is the evidence. In June 2026, the actors’ union ratified a new contract with 91.42% in favor. It took effect July 1. That contract does not ban AI. It sets the price and the rules. There are terms for digital replicas of performers, terms for scanning an actor’s face and body, and terms for using a digital replica to dub a performance into another language. Synthetic performers are permitted when they add value a real actor cannot.
You do not negotiate four pages of rules for something that is not already happening.
Think about where it actually gets used. An actor is unavailable for two days of reshoots. A show needs to run in eleven countries, and dubbing each one into native languages costs real money. A crowd scene needs three hundred people, and the budget covers forty. A series produced in Malaysia or Argentina on a thin margin needs to look more expensive than it was.
None of that is an AI movie. All of it is AI.
And the volume is in the places nobody writes about. Thousands of streaming channels need programming every week. A million creators push AI into YouTube every day, maybe as high as 40% of new content. The audience is real, the budget is small, and they compete with the next creator down the line.
This is the honest version of AI demand. You pay for it with YouTube, with Netflix, and with the “Free movie” that you stream. You are not buying AI. You are buying the next episode.
Revenue will not tell you. Every one of these companies will report growing AI revenue for years, because demand is real and rising. Revenue answers whether AI works, and we already know it does.
Watch the calendar instead.
Refinancing dates. Plan three companies borrowed money that comes due on specific days. Oracle has to return to the bond market repeatedly. So do the smaller data center operators. The question is never whether they are profitable on that day. The question is whether the market is open that day. A company can be perfectly healthy and still fail to refinance.
Free cash flow, not earnings. Earnings can be managed by changing how fast you depreciate equipment. Meta already extended the assumed life of its servers once, which lowered its reported costs by billions in a book keeping fantasy manoever. Cash flow is harder to dress up. Watch the gap between what a company earns and what it spends on construction.
The rental price of compute. This is the cleanest signal, and almost nobody quotes it. Companies rent graphics chips by the hour, and those prices are public. If rental rates fall while new capacity keeps opening, supply has passed demand. That is exactly what happened to bandwidth prices after the fiber companies failed. Token prices are less useful, because vendors set those strategically to win customers.
Used chip prices. The $500 billion in Nvidia financing is secured against the resale value of chips. If used hardware gets cheap, the collateral behind that debt shrinks, and lenders pull back before any borrower misses a payment.
The pattern is simple. The technology keeps working, and the money stops showing up.
I rent graphics chips by the hour to generate video. My budget is $30 a month. That makes me the smallest customer in this entire story, which is exactly why what I see is useful.
I use RunPod. This week, an H100 costs $3.29 an hour. The newer B200 is $6.79. The card I actually want, an RTX PRO 6000 at $2.09, was unavailable. So was almost everything else. Fourteen different chips, priced from 28 cents an hour to $7.89 an hour, and nearly all of them showed the same word. Unavailable.
That is not a market with too much capacity. That is a market where a customer holding money cannot buy the product at any price. Runpod, which I use, raised $100 million in June and turned down buyout offers. Their limit is not customers. It is how fast they can get chips.
My own usage tells the same story from the other side. I regenerated the same video twelve times while writing this article, chasing a 30-second clip without artifacts. I did it on my PC in the background because I can’t get a Pod right now. I run an $3500 HP Omen laptop, but the Pod I rent would cost about $32,000 to buy, and I only need it about one hour a day.
Today – demand exceeds supply at every price point.
Plants being financed today – open in 2028, full of chips that will be made in 2027, sold to customers placing orders today.
Who made the right decision? The chip maker, the buyer, the Capex plant, or me?
Both directions are live, and they work differently.
When a leveraged operator fails, its hardware does not disappear. It gets sold cheap to a buyer with no debt against it. That buyer can then rent compute at prices that cover their pennies and nobody else’s dollars. Rental rates collapse, and every operator still paying off original construction costs is now underwater.
This is exactly what happened after Global Crossing failed. Bandwidth stayed cheap for years, and the survivors were the ones who bought the cheap lines, and we all use them today.
One reason it could move faster this time, though, is that it cuts both ways. Frontier training moves to new chips quickly, so the newest hardware loses its top-tier job within a couple of years. But everyday work does not care. Running a finished model, generating images, fine-tuning something small — an older chip does all of that fine, which is why the cheap cards are the ones that are never available.
So used hardware does not become worthless. It becomes cheap. That still breaks the math. A lender who financed a chip at full price does not recover by learning it has a long second life at a fifth of the cost.
This one does not require anything about AI to go wrong.
Private credit is now a $2 trillion market. Defaults hit a record 9.2%. Some funds gated withdrawals in January, meaning investors asked for their money and were told to wait. Pensions hold about 30% of that market. Insurers hold 18%. Retail investors hold $550 billion, and 401(k) plans were recently cleared to buy in.
AI needs roughly $800 billion from private credit through 2028. If private credit pulls back for reasons unrelated to AI, that money isn’t there to roll the bonds.
That is the part worth sitting with. AI does not have to fail for AI financing to fail. The borrower can be growing, profitable and busy, and still hit a “Junk Bond” wall.
It is complicated.
The companies building from profits will be fine. If demand disappoints, Microsoft and Google slow down and nothing breaks, because nobody is waiting on a payment. That is not a bubble. That is a large company spending money on a genuinely promising technology.
The companies that borrowed against a date are the exposed ones. Oracle and the smaller data center operators have to return to the lenders repeatedly between now and 2030. They do not need to be unprofitable to fail. They depend on steady market conditions they will not get.
And the company financing its own customers is running the play that ended Lucent. Revenue arrives first, and the loans fall through in a downturn.
So yes, some of these companies will fail. When they do, their hardware will not disappear. It will be sold cheap, and whoever buys it will rent out compute at discount prices.
Which leaves one question, and it is not the question people expect.
Cheap compute has to be absorbed by someone. It will not be absorbed by enterprises carefully metering their spending. Whoever has cash and needs cheap volume will grab it.
That is content. Video, dubbing, background work, and the endless demand for the next episode.
So the future of this buildout may depend less on how much people use AI, and more on how much television they watch.
Parts of it might. The technology is working and demand is growing, so this is not a case of a worthless asset. The risk sits in the financing. Companies that borrowed against fixed dates have to refinance repeatedly through 2030, and market disruption can cause them to fail while still being profitable.
Nobody knows, but the strain is measurable. The 30-year Treasury yield is at a two-decade high, federal debt passed $40 trillion, and the government rolls about $9 trillion a year. AI borrowing now competes for the same lenders, with roughly $570 billion issued in 2026.
The hardware does not disappear. It gets sold cheap to buyers with no debt against it, who then rent compute at prices nobody who paid full price can match. That is what happened to fiber optic cable after 2002. The equipment survives. The investors do not.
Increasingly, lenders rather than tech companies. In 2025, the eight largest cloud companies produced about $180 billion in free cash flow. In 2026, that swings to roughly negative $64 billion. Analysts expect AI to need about $1.5 trillion from credit markets through 2028.
Because about $800 billion of that funding is supposed to come from private credit. Private credit is now a $2 trillion market with defaults at a record 9.2%. Pensions hold roughly 30% of it and insurers 18%, so problems there reach ordinary retirement accounts.
Do you have too many apps and not enough time? Learn the five essential tools…
When it comes to the best video hosting for online course creators in the USA,…
Anniversaries carry personal meaning, yet busy schedules can leave little time to arrange a thoughtful…
Online traffic alone does not guarantee qualified leads for local businesses. Visitors who leave without…
Commercial property owners need accurate financial information when they challenge an assessment they believe does…
A recorded meeting usually feels useful right after it ends. Everyone remembers the discussion, the…