Skip to content

Nate B Jones AI Business

Your $20 AI Plan Costs Them Thousands. That's Not The Bubble.

Nate B Jones argues that 'is AI a bubble' is the wrong question because it compresses stock prices, private valuations, overbuilt data centers, weak pilot ROI, and real demand into one word. He separates the speculative financial froth on top from the physical supply chain underneath, citing OpenAI revenue going from about $2B to over $20B in two years, Anthropic growing faster, Nvidia's roughly $193.7B data center revenue, and hyperscalers citing capacity constraints rather than weak demand. The underexplained engine is inference: agents loop and burn tokens, so one agent run can cost thousands of times a chat turn, which is why hyperscalers now build factories for inference. His replacement frame is buildout versus payback, and the better question is who survives the sorting into reliable, affordable, high utilization inference.

Published Jun 15, 2026 19:24 video 25 min read Added Jun 16, 2026 Open on YouTube →

At a glance

AI stocks are in correction, and the word on every chart is "bubble." Nate B Jones argues that bubble is the wrong word, not because nothing is overpriced, but because it crushes a dozen separate questions into one. He splits the AI buildout into two layers: the speculative financial froth piled on top, and the physical supply chain underneath that serves demand which already exists. The evidence for the real layer is concrete and public: OpenAI revenue going from roughly $2 billion to over $20 billion in two years, Anthropic growing even faster, Nvidia doing about $193.7 billion in data center revenue, and the hyperscalers complaining about capacity constraints rather than weak demand. The engine he says everyone underexplains is inference: when AI was chat, inference was cheap, but agents loop, call tools, and burn tokens, so a single agent run can cost thousands of times a chat conversation, which is exactly why the capex numbers look industrial. His replacement question for "is it a bubble" is "buildout versus payback," and the better thing to watch is who survives the sorting that turns demand into reliable, affordable, high utilization inference.

The bubble question returns, and why it is the wrong one

The video opens in the middle of a real market event. AI stocks are finally getting hit, and you can feel the story evolving in real time. The tech sector is in correction territory. The biggest AI names are selling off, and the punishments are strange ones. Broadcom can report record AI revenue and still get punished, because investors wanted more. Alphabet and Microsoft can keep growing cloud revenue and still trade down. The market is asking the same question over and over: was this whole AI trade just a bubble?

Jones grants that the spending numbers are, in his word, absurd. Google, Microsoft, Amazon, and Meta are all on pace to spend somewhere around $700 billion this year on AI infrastructure. Some are raising debt, some are issuing stock, power is tight, memory is expensive, and data centers are taking longer to build than planned. Inside most companies, the clean story that says "this is where we get a return on all those bucks" is not fully there yet. So if you want to call it a bubble, he gets it. It is not an insane reaction.

But he thinks it is the wrong core question. A stock correction tells you investors think prices are stretched. It does not automatically tell you that demand is fake. That distinction is the spine of the whole video, because the companies closest to demand are not pulling back. The right question, he says, is not "is AI a bubble." It is: which part of the AI buildout is speculative financial froth, and which part is the physical supply chain for demand that already exists?

The word "bubble" compresses too many things

The lazy version of the bubble argument, Jones says, compresses way too many separate things into one word. It treats inflated stock prices, aggressive private valuations, overbuilt data centers, weak enterprise return on investment, Nvidia's revenue, OpenAI's growth, and the whole future of AI as if they were the same question. They are not.

He lays out the things that can be simultaneously true:

You can have a correction in AI stocks and still have a tremendous amount of locked up AI demand that is not met.
You can have some companies overbuild capacity and still have the world be dramatically underbuilt for inference.
You can have weak return on investment in a random corporate pilot and still have massive demand for coding agents, research agents, customer support automation, model APIs, and enterprise AI tools that actually replace hours of work.

The mistake is treating "bubble" as a verdict on the whole technology. It is more useful, he says, to treat the bubble concept as an invitation to map the sector. There can be bubble dynamics in the assets around AI. There can be overvaluation, crowded trades, and data centers financed on assumptions that do not survive contact with reality. Some investors will lose money. Some suppliers will get overpaid. Some companies will build too much of the wrong thing in the wrong place. All of that can be true. None of it suggests the underlying demand is imaginary.

The real revenue: OpenAI, Anthropic, Nvidia

To prove the demand layer is real, Jones goes to the numbers companies have actually reported.

Start with OpenAI. Its annualized revenue went from roughly $2 billion in 2023, to $6 billion in 2024, to more than $20 billion in 2025 and growing. That is an insane growth curve, and, he notes, it is the slowest growth curve of the hyperscalers. Anthropic has grown even faster from a smaller base and is now on a higher revenue run rate than OpenAI reported.

Crucially, this is not consumer curiosity. Enterprises are now roughly 40% of the business at OpenAI, and even more at Anthropic. Companies are lining out the door, and the labs literally cannot onboard them fast enough. That matters, because enterprise revenue is very different from "I tried the chatbot once, subscribed, and now I regret it." A company does not keep spending real budget because a demo was fun. It spends because someone inside the company is convinced the tool is critical for code, research, analysis, customer work, compliance, sales, or ops, some workflow where the old process was slower or more expensive.

Does some of that spending come from companies chasing FOMO, worried about competitors adopting AI first? One hundred percent, he says. Does that make them irrational to spend on intelligence inside their business? No. They may not use those dollars well, which is exactly why both labs are pushing forward deployed engineering teams to sit with customers. But the demand is there. Anthropic and OpenAI are both setting new records for how fast a private company can grow revenue, and you do not do that on a whim. It means there are paying customers in the system.

Figure 1. OpenAI's reported annualized revenue, the curve Jones calls "insane" and yet "the slowest of the hyperscalers." From about $2 billion in 2023 to over $20 billion in 2025, with enterprises now roughly 40% of the business. Anthropic, from a smaller base, is growing faster still. Paying customers, not curiosity, are driving the number.

Then Nvidia. Its fiscal 2026 data center revenue was about $193.7 billion. That is a very clear public signal of massive physical side demand: people writing checks for chips, systems, networking, memory, racks, and commitments moving through the data center supply chain. The important part is not just that Nvidia is selling a lot, it is what those purchases imply. Nobody buys this much AI infrastructure because they are casually experimenting with a dashboard. The chips are bought by serious boards and CEOs because training and inference workloads already exist, and because everyone close to the demand believes those workloads are spiking fast.

Jones names the tell that should end the simple bubble story: you cannot have a bubble conversation at the same time as prominent leaders are complaining that their developers are burning through their Claude credits too fast. Those two things should not coexist. And yet, in this supposedly rational market, they do.

Where the bears are right

He gives the bears their due, because they are right about one thing. Revenue and spending do not match as neatly as they should yet. If the hyperscalers spend $600 billion, $700 billion, maybe a trillion dollars on AI infrastructure, you need a huge amount of future revenue to justify it. That is fair.

Getting there requires a chain of things to happen. Enterprise adoption has to move from pilots to production. Agents have to become reliable enough to run long workflows and easy enough to roll out that every company can do it, which he flags as the critical one. Software teams, legal teams, finance teams, and support teams all have to change how they work. None of that is guaranteed, and it will not happen all at once. It happens in fits and starts, along an adoption curve. If you are an investor paying a huge multiple for every company that touches the AI supply chain, that timing matters enormously.

This, he says, is exactly why "bubble" is too blunt. The question is not whether AI is real. The question is whether the cash flows arrive in the right place at the right time for the companies financing the buildout all the way through the supply chain. The demand can be real and the investment can still be poorly timed in certain parts of the chain. The technology can be transformative and some stocks can still be too pricey. The infrastructure can be necessary and some of the builders can still earn bad returns.

Railroads, fiber, dot-com: real buildouts, ruined investors

This has all happened before, and Jones runs the historical pattern. Railroads were real, a tremendous buildout and a huge success for the economy, and a lot of railroad investors still got absolutely destroyed. Fiber was real, and a lot of telecom investors still got destroyed. Cloud was real, and not every cloud adjacent company managed to capture that value.

So when people say this looks like the dot-com bubble, his honest answer is maybe, in some ways, but not the way you think. The dot-com bubble was notoriously decades ahead of demand. That is not what we are seeing here. And the dot-com bust did not prove the internet was fake. It proved that markets can overprice the first order version of a real platform shift.

The risk now is not that nothing is happening. The risk is that the market prices every AI exposed asset as if it will automatically, magically capture value. It will not. Some companies will provide commoditized input and get squeezed. Some will build way too far ahead of demand. Some will have the right thesis and the wrong balance sheet. But the buildout itself is not a hallucination.

Figure 2. Jones's historical analogy. Railroads, fiber, and cloud were all genuine, economy changing buildouts, and investors in them were still routinely wiped out. The dot-com bust did not prove the internet was fake; it proved markets overprice the first order version of a real shift. The lesson he draws: "real buildout" and "good investment" are different claims, and the bubble word smears them together.

Inference is the part everyone underexplains

The reason the buildout is real, Jones says, is inference, and this is the part he thinks is badly underexplained on both Main Street and Wall Street.

Training a model is expensive but episodic. You build a huge cluster, run a training job, produce a model, and move on to the next generation. Inference is different. Inference is the model running every single time someone uses it: every prompt, every agent step, every tool call, every retry, every long context window, every document, every codebase, every verification pass.

When AI was mostly chat, inference looked manageable, and that was not long ago, maybe seven or eight months back. A person asks a question, waits, reads, maybe asks another. A lot of the buildout was sized around training runs. Then agents, in roughly the last six months, changed the math fundamentally and forever.

An agent does not ask one question and stop. It goes. It loops. It reads files. It calls tools. It writes code. It checks the result. It fixes the failure. It asks another model to review the input. It searches again. It runs again. It burns tokens over and over. That is not a conversation, he says, it is a production job that runs into millions and billions of tokens very fast. Once you see AI work that way, you cannot unsee it, and the infrastructure buildout suddenly makes far more sense: any given agent run can be thousands of times the inference cost of a chat conversation.

And tokens are not magic, they are manufactured. Behind every answer and every agent tool call is a physical production system: chips, memory, networking, power, cooling, land, construction, and ops. That is why the capex numbers are getting serious. AI makes the most valuable software companies in the world look industrial. Microsoft, Google, Amazon, and Meta do not just ship features anymore, they are building factories for inference. And factories are expensive. They need upfront capital, high utilization, supply assurance, power contracts, depreciation schedules, and a constant grind of routing, batching, caching, and efficiency improvements.

Figure 3. The inference shift Jones says nobody explains. A chat turn is one short inference. An agent run loops through reading, tool calls, code, checks, retries, and peer review, spending millions to billions of tokens, up to thousands of times the cost of a chat. The buildout is sized for the agent era, not the chat era, which is why hyperscalers now look like industrial companies "building factories for inference."

The real question of 2026: are expensive tokens worth it

So expensive compute had better not be wasted on cheap work. That, Jones says, is the real operating question, and it is not "is AI a bubble." It is: are expensive tokens being spent on work valuable enough to justify them? That is the question of 2026, and it separates what is real in the demand explosion from what is fake very quickly.

A coding agent that saves an engineering team days of work justifies expensive inference, and these days, he says, they are saving weeks and months. A legal review agent that processes thousands of contracts justifies it. A customer service system that resolves real tickets and reduces escalation justifies it. But a random enterprise chatbot bolted onto a website, answering shallow questions from a stale knowledge base and delivering a terrible Customer experience, does not justify the inference it burns.

That is why the enterprise ROI data looks like a complete mess right now. AI is not one single thing. It is a general purpose technology, a thousand different workflows with different economics. Some are genuinely useful, some are terrible ideas. Some save time at the individual level but never make it into the P&L. Some look impressive in a demo and collapse when dropped into a real workflow with permissions, exceptions, messy data, and accountability.

None of that, he stresses, is proof of a bubble. It is proof that adoption is uneven, and that companies do not yet know what to do with a new general purpose technology. And it should be uneven. Most companies are bad at process change. They were bad at software implementation before AI, bad at data projects before AI, bad at cloud migration before AI, and now they are bad at AI transformation, and everyone is acting surprised. The technology can be real while companies struggle with change management.

The better model: buildout versus payback

Here is the mental model Jones offers to replace the binary. It is not bubble versus no bubble. It is buildout versus payback.

The buildout is real. The demand signals are real. The constraints are real. OpenAI's revenue is real, so is Anthropic's, so is Nvidia's data center revenue, so is hyperscaler capex, and so are the capacity constraints. The payback is the open question circling all of those facts: who gets paid back, and when? How fast? At what margin? On which workloads does it matter? How much pricing power do the hyperscalers have when they set prices for tokens and for workflows?

This is where the market ought to be more thoughtful. If you buy every AI stock because "AI is the future," you are not doing due diligence, you are buying a narrative, and narratives flip on a dime. But if you dismiss the entire thing because stocks corrected, you are also not doing analysis, you are reacting to price action and calling it insight. The useful middle ground is harder and rarer: AI is a real platform shift so transformative that a bunch of local bubble dynamics are frothing around it. Prices can fall while technology keeps advancing. Some infrastructure may be overbuilt while other capacity stays profoundly scarce. Some companies will spend too much and still not spend enough in the exact place that matters. The biggest winners may not be the companies with the loudest AI story today, but the ones that control a bottleneck, own the Customer workflow, route inference more efficiently, or turn AI output into durable business value.

The questions to actually ask

Instead of "is this company doing AI," Jones gives the questions he would use to find where investment dollars belong (with the explicit note that none of this is investment advice). Each one is designed to separate paid demand from theater.

Don't ask	Ask instead
Is the company "doing AI"?	Where is the demand showing up?
Did the stock pop on AI news?	Is it paid usage real or just engagement theater?
Is there a flashy demo?	Production workloads real or a pilot dressed up in a press release theater?
Is AI "in the deck"?	Is it improving a workflow with clear economics, or creating more work for humans to review?
Is the board buying capacity?	Because customers are waiting real, or because the board wants an AI strategy theater?
Is premium compute being used?	Where expensive reasoning matters real, or burned on cheap tasks waste?

These questions are less dramatic than "bubble or revolution," but far more useful. They also explain why the correction does not settle anything. Markets correct and uncorrect for many reasons: stretched valuations, crowded trades, expectations running too high, capital rotating, financing costs rising. A company can disappoint investors even while its underlying business grows. The correction is not meaningless, he says, it is telling you that investors, in the middle of a global energy crisis, are finally having feelings about underwriting unlimited AI spending without asking harder questions. They should ask harder questions. But the correction is not proof that the buildout is fake. It is proof that the easy phase of the trade is over.

Narrative, correction, sorting: the three phases

Jones maps the platform shift onto three phases. The first phase is almost always narrative: everyone piles into the obvious names. The second phase is correction: the market realizes the story is more expensive, slower, and messier than the headlines implied. The third phase is sorting, and it is the healthy one.

In the sorting, the market separates companies with real AI revenue from companies with AI language in the deck. It separates infrastructure bottlenecks from commodity exposure. It separates tools that create measurable workflow value from tools that create demo value. It separates companies that can finance the buildout from companies that need someone else to finance it. And it separates SaaS companies with sticky services still valuable in the era of agents from ones that do not. That sorting is exactly what should happen in a real platform shift.

Phase 1 Narrative. Everyone piles into the obvious names. Any company that touches AI gets a multiple. A press release can pop a stock 500%.
Phase 2 Correction. The market realizes the story is more expensive, slower, and messier than the headlines implied. Stocks sell off, even ones with growing revenue. This is where the video is filmed.
Phase 3 Sorting. Real AI revenue separates from AI language in the deck; bottlenecks from commodities; workflow value from demo value; self financing builders from the ones who need someone else's money. The winners turn demand into reliable, affordable, high utilization inference.

So when someone asks "is AI a bubble," his answer is: parts of it are. If you can release a press release and get a 500% pop, that is a bubble, and he has seen it once or twice. Some valuations are stretched, some spending will be wasted, some private market marks are ridiculous, some companies are pretending a thin wrapper is a business, the seed rounds in the valley are getting pricey, and some enterprises are buying expensive tools without changing the workflow enough to capture value. But the broader buildout is not hype floating above reality. There is real demand underneath, real revenue underneath, real physical scarcity underneath.

The better question, he says, is not "when does the bubble pop," because he does not think that is coming. It is "who survives the sorting." If intelligence becomes a production system, the winners are not the companies with the best demos. They are the ones that can turn demand into reliable, affordable, high utilization inference: route the right task to the right model, secure power, secure memory, build capacity, and make agents useful enough that customers keep paying after the novelty wears off. That is a much harder game than the stock chart made it look last year, and a much more serious one.

AI is a marathon, not a sprint

Jones closes with definition and history. A bubble is hype detached from reality, that is essentially the definition. The famous South Sea bubble was all hype, and it cost Isaac Newton his fortune. This is messier than that: a real buildout with speculative money piled on top, and the correction is just the market starting to ask which layer is which. That, he says, is fantastic and healthy.

The questions to keep in front of you are evergreen, good next month and the month after: Where is the paid demand? Where is the bottleneck really? Who captures value when tokens get cheaper? Can this business finance itself? Is the work really moving to agents in association with this company?

His final frame is time. AI is a marathon, not a sprint. The business of installing AI into companies is a 10 to 20 year exercise, and we are writing the first chapter. Think about that the next time you look at your Robinhood account. It is a larger and healthier picture, because AI is here for the long term, the most transformative technology of our lives, and it can still have a ton of froth around the edges while that stays true. He signs off asking for sober takes in a world that loves to argue in binaries, and invites the financial press to reach him, because he thinks the inference piece in particular is consistently reported wrong.

Key takeaways

"Is AI a bubble?" is the wrong question because it crushes a dozen separate claims (stock prices, private valuations, overbuilt data centers, weak pilot ROI, Nvidia's revenue, OpenAI's growth, the future of AI) into one word. Map the sector instead.
A stock correction means investors think prices are stretched. It does not mean demand is fake. The companies closest to demand are not pulling back.
The demand layer is real and reported: OpenAI revenue from about $2B to over $20B in two years, Anthropic faster, Nvidia data center revenue about $193.7 billion, enterprises ~40% of OpenAI's business, hyperscalers citing capacity constraints not weak demand.
The bears are right about timing: roughly $700 billion in hyperscaler capex this year needs future revenue to justify it, and that depends on agents getting reliable and companies actually changing how they work, which happens on an adoption curve.
History rhymes: railroads, fiber, and cloud were all real, economy changing buildouts whose investors were frequently wiped out. Real buildout does not equal safe investment.
Inference is the underexplained engine. Training is episodic; inference runs every time. Agents loop and burn tokens, so one agent run can cost thousands of times a chat turn, which is why hyperscalers now build "factories for inference."
The operating question of 2026 is whether expensive tokens are spent on work valuable enough to justify them. Coding, legal review, and support agents can justify it; shallow website chatbots cannot. That is why enterprise ROI looks like a mess.
Replace "bubble vs no bubble" with "buildout vs payback." The buildout is real; payback (who gets paid, when, at what margin, on which workloads) is the open question.
The three phases are narrative, correction, sorting. The video is filmed in the correction. The right question is not when the bubble pops, but who survives the sorting by turning demand into reliable, affordable, high utilization inference.

Chapters

Timestamps are clickable. Click one and the player jumps there and keeps playing while you read.

0:00 AI stocks correct and the bubble question returns
0:50 A correction doesn't mean the demand is fake
1:33 Financial froth vs. the physical supply chain
3:05 Real revenue: OpenAI, Anthropic, and NVIDIA
5:30 Where the bears are right: spending vs. revenue
6:45 Railroads, fiber, dot-com: real build-outs, ruined investors
8:06 Why inference is the part everyone underexplains
10:05 The question of 2026: are expensive tokens worth it
11:53 The better model: buildout vs. payback
13:34 The questions to actually ask about any AI company
15:30 Narrative, correction, sorting: the three phases
17:47 AI is a marathon, not a sprint

Notable quotes

A stock correction will tell you that investors think prices are stretched. It doesn't automatically tell you that demand is fake. Nate B Jones, 0:50

The question is not is AI a bubble. The question is which part of the AI buildout is speculative financial froth and which part is the physical supply chain for demand that already exists. Nate B Jones, 1:33

You cannot have a bubble conversation at the same time as you have prominent leaders complaining about the fact that their developers are burning through their Claude credits too fast. Those things should not coexist. And yet in this rational market, they do. Nate B Jones, 4:40

The dot-com bubble did not prove the internet was fake. It proved that markets can overprice the first order version of a real platform shift. Nate B Jones, 7:20

Tokens are not magic. They're manufactured. Behind every answer, behind every agent tool call is a physical production system. Nate B Jones, 9:30

It's not bubble versus no bubble. It's buildout versus payback. Nate B Jones, 11:53

A bubble is hype detached from reality. This is messier than that. This is a real buildout with speculative money piled over the top. And the correction is the market starting to ask which layer is which. Nate B Jones, 17:00

AI is a marathon. It's not a sprint. The business of putting AI into companies and installing it is a 10, 20-year exercise. We're just at the beginning of that. We are writing the first chapter. Nate B Jones, 17:47

Resources mentioned

Nate B Jones, the channel (AI News & Strategy Daily), for sober, non binary takes on AI markets.
OpenAI and its revenue curve (about $2B to over $20B in two years); enterprises roughly 40% of the business.
Anthropic and Claude, growing faster from a smaller base, with leaders citing developers burning through Claude credits.
Nvidia, with about $193.7 billion in fiscal 2026 data center revenue as the clearest physical demand signal.
The hyperscalers funding the buildout: Microsoft, Google / Alphabet, Amazon, and Meta, on pace for roughly $700 billion in AI infrastructure this year.
Broadcom, cited as the company that reported record AI revenue and still got punished.
Historical buildout analogies: railroads, fiber / telecom, cloud, and the dot-com bubble.
The South Sea bubble and Isaac Newton's ruin, as the definition of hype detached from reality.
Software as a service (SaaS), discussed as sticky services tested in the era of agents.
Robinhood, invoked as the retail investor's window onto the correction.

Where it stands

This is a markets and macro argument, sharp and clearly reasoned, and worth reading on its terms. A few honest footnotes for the reader. Several headline figures are reported or annualized run rate numbers rather than audited results: OpenAI and Anthropic are private, so their revenue figures come from disclosures and reporting, not filings, and "annualized" can flatter a fast growing line. The roughly $700 billion capex figure is an estimate across four companies and varies by source. The "agent run is thousands of times a chat turn" claim is directionally right but workload dependent, not a fixed multiple. And as Jones repeatedly says himself, none of this is investment advice; the video's real contribution is the framing (froth on top, supply chain underneath; buildout versus payback; narrative, correction, sorting) rather than a stock call. The strongest part of the argument is the inference point, which is genuinely underappreciated; the part most exposed to being wrong is timing, since "real demand" and "paid back at this valuation, on this schedule" are different bets, exactly the distinction the video itself keeps drawing.

Full transcript

AI stocks are finally getting hit and you can feel the story kind of evolving in real time as they do. The tech sector is in correction territory. Big AI names are selling off. Broadcom can report record AI revenue and still get punished because investors wanted more. Alphabet and Microsoft can keep growing cloud revenue and still trade down because the market is suddenly asking the same question over and over again. Was this whole AI trade just a bubble? The spending numbers are frankly absurd. Google, Microsoft, Amazon, and Meta are all on pace to spend somewhere around $700 billion this year on AI infrastructure. Some of these companies are raising debt, some are issuing stock, power is tight, memory is expensive, data centers are taking longer to build. And inside most companies, the clean story that says this is where we get a return on all of those bucks, that's still not fully there. So, if you want to say this is starting to look like a bubble, I get it. It's not an insane reaction, but I do think it is the wrong core question to ask. A stock correction will tell you that investors think prices are stretched. It doesn't automatically tell you that demand is fake. And that distinction matters a lot because the companies closest to demand are not pulling back. OpenAI went from roughly $2 billion in annualized revenue in 2023 to more than $20 billion and counting in 2025. Anthropic grew even faster. Nvidia's data center business did almost $194 billion in fiscal 2026. The hyperscalers are still talking about capacity constraints, not lack of demand. So the question, I think, is not is AI a bubble. The question is which part of the AI buildout is speculative financial froth and which part is the physical supply chain for demand that already exists? And that's what I want to separate in this video because the lazy version of the bubble argument compresses way too many things into one word. It treats inflated stock prices and aggressive private valuations and overbuilt data centers and weak enterprise ROI and Nvidia's revenue and OpenAI's growth and the whole future of AI as if they're all the same question and they're really not. You can have a correction in AI stocks and still have a tremendous amount of locked up AI demand that is not met. You can have some companies overbuild capacity and still have the world be dramatically underbuilt for inference. You can have weak return on investment in a random corporate pilot and still have massive demand for coding agents and research agents and customer support automation and model APIs and enterprise AI tools that actually replace hours of work. The mistake is treating bubble as a verdict on the whole technology. It is more useful to treat the bubble concept as an invitation to map the sector because there can be bubble dynamics in the assets around AI. There can be overvaluation. There can be crowded trades. There can be data centers financed on assumptions that don't survive contact with reality. Some investors are going to lose money. Some suppliers are going to get overpaid. Some companies are going to build too much of the wrong thing in the wrong place. All of that can be true. But none of it suggests that the underlying demand is imaginary. Start with OpenAI. OpenAI has said its annualized revenue went from $2 billion in 2023 to $6 billion 2024 to more than $20 billion 2025 and growing. That's an insane growth curve and it is the slowest growth curve of the hyperscalers. Anthropic has grown even faster from a smaller base and now is on a higher revenue run rate than OpenAI reported. And this is not about consumer curiosity. Enterprises now roughly 40% of the business at OpenAI, even more at Anthropic. And companies are just lining out the door. They literally can't get on-boarded fast enough. And that matters because enterprise revenue is a lot different from "I tried the chatbot once, I subscribed and now I regret it." A company doesn't keep spending real budget on AI because a demo was fun. It spends because someone inside the company is passionate about this and thinks this tool is critical for code, for research, for analysis, for customer work, for compliance and sales and ops or some workflow where the old process was slower or more expensive. Now, do some of those dollars come from companies that are chasing FOMO and they're worried about other companies adopting AI? 100%. Does that mean that they are irrational to spend on intelligence inside their business? No. They may not use those dollars well. And that's why we see that forward deployed engineering push from both these companies, but the demand is there. Anthropic and OpenAI are both setting new records for how quickly a private company can grow revenue. You don't do that on a whim. It means there are paying customers in the system. Look at Nvidia next. Nvidia's fiscal 2026 data center revenue was about $193.7 billion. That is a very clear public signal that we have massive physical side AI demand. Those are people willing to put down checks for chips and systems and networking and memory and racks and commitments moving through the supply chain for data centers. And the important part is not just that Nvidia is selling a lot. The important part is what those purchases imply. Nobody buys this much AI infrastructure because they are casually experimenting with a dashboard. The chips are being bought by very serious boards and CEOs because training and inference workloads already exist and because everyone close to the demand strongly believes those workloads are spiking fast. You cannot have a bubble conversation at the same time as you have prominent leaders complaining about the fact that their developers are burning through their Claude credits too fast. Those things should not coexist. And yet in this rational market, they do. Now, this is where the bubble argument gets even more interesting because the bears are right about one thing. Revenue and spending don't match as neatly as they should yet. If the hyperscalers spend 600, 700, maybe a trillion dollars on AI infrastructure, you need a huge amount of future revenue to justify that investment. That's fair. You need enterprise adoption to move from pilots to production. You need agents to become reliable enough to run long workflows and easy enough to roll out that every company can do it. That's the critical one. You need software teams and legal teams and finance teams and support teams, everybody to change how they work. That is not a guarantee that's going to happen in fits and starts. It's going to happen in an adoption curve. And if you're an investor paying a huge multiple for every company that touches the AI supply chain, that timing matters a lot. But this is exactly why the word bubble is too blunt. The question is not whether AI is real. The question is whether the cash flows arrive in the right place at the right time for the companies that are financing the buildout all the way through the supply chain. That's a much more interesting question because the demand can be real and the investment can still be poorly timed in certain parts of the supply chain. The technology can be absolutely transformative and some stocks can still be too pricey. The infrastructure can be necessary and some of the builders can still earn bad returns. This has all happened before. Railroads were real. A lot of railroad investors still got absolutely destroyed in the market even though it was a tremendous buildout and a huge success for the economy. Fiber was real. A lot of telecom investors still got destroyed. Cloud was real. Not every cloud adjacent company managed to capture that value. So when people say this looks like the dot-com bubble, I think the honest answer is maybe in some ways, but not the way you think it means. The dot-com bubble was notoriously decades ahead of demand. That is not what we should be seeing here. The dot-com bubble did not prove the internet was fake. It proved that markets can overprice the first order version of a real platform shift. It's not that nothing is happening. The risk is that the market prices every AI exposed asset as if it's automatically going to magically capture value. It won't. Some companies will provide commoditized input. Some are going to get squeezed. Some are going to build way too far ahead of demand. Some will have the right thesis and the wrong balance sheet. But the buildout itself, that's not a hallucination. The reason is inference. And this is the part of the AI spending story that feels very underexplained on Main Street and Wall Street to me. Training a model is expensive, but it's episodic. You build a huge cluster, you run a training job, you produce a model, and then you move on to the next generation. Inference is different. Inference is the model running every time someone uses it. Every prompt, every agent step, every tool call, every retry, every long context window, every document, every codebase, every verification pass. When AI was mostly chat, inference looked super manageable. And that was not that long ago. That was like seven, eight months ago. A person asks a question and waits and reads and maybe asks another one. And a lot of the buildout was around training runs. Agents in the last six months changed that math fundamentally forever. An agent does not ask one question and stop. It goes. It loops. It reads files. It calls tools. It writes code. It checks the result. It fixes the failure. It asks another model to review the input. It searches again. It runs again. It burns tokens over and over. That's not a conversation. That's a production job that runs into millions and billions of tokens really, really fast. And once you see AI work that way, you can't see it any other way. And the infrastructure buildout starts to make a ton more sense because any given agent run can be thousands of times the inference cost of a chat conversation. Tokens are not magic. They're manufactured. Behind every answer, behind every agent tool call is a physical production system. Chips and memory and networking and power and cooling and land and construction and ops. And that's why the capex numbers are getting very serious. AI makes the most valuable software companies in the world look industrial today. Microsoft and Google and Amazon and Meta don't just ship features anymore. They're building factories for inference. And factories are expensive. They require upfront capital and utilization and supply assurance. They require power contracts. They require depreciation schedules and routing and batching and caching and efficiency improvements. So expensive compute is not wasted on cheap work. That is the real operating question. It's not is AI a bubble. The operating question is this. Are expensive tokens being spent on work valuable enough to justify them? That's the question of 2026. And that question separates what's real in this AI explosion of demand from what's fake very quickly. A coding agent that saves an engineering team days of work can justify that expensive inference. And these days, they're saving weeks and months sometimes. A legal review agent that processes thousands of contracts can justify expensive inference. A customer service system that resolves real tickets and reduces escalation. You can justify expensive inference that way. Now, a random enterprise chatbot on the website that answers shallow questions from a stale knowledge base and provides a terrible customer experience, not really justifying your inference there. And that's why the enterprise ROI data looks like a complete mess right now. AI is not one single thing. It's a general purpose technology. It's a thousand different workflows with different economics. Some are really useful, some are terrible ideas. Some save time at the individual level but never make it into the P&L. Some will look impressive in a demo and collapse when you put them inside a real workflow with permissions and exceptions and messy data and accountability. None of this is proof of a bubble. This is proof that adoption is uneven and frankly that companies don't necessarily know what to do with the new general purpose technology yet. And frankly, it should be uneven. Most companies are bad at process change. They were bad at software implementation before AI. They were bad at data projects before AI. They were bad at cloud migration before AI. And now they're bad at AI transformation. And we're all acting surprised. The technology can be real while companies struggle with change management. And that's where I think the better mental model is. It's not bubble versus no bubble. It's buildout versus payback. The buildout is real. The demand signals are real. The constraints are also real. OpenAI's revenue is real. So is Anthropic's. NVIDIA's data center revenue is real. Hyperscaler capex is also real. And capacity constraints are real. The payback is the open question circling around all of those facts. Who gets paid back and when? How fast do they get paid? At what margin? On which workloads does it matter that they get paid back? How much pricing power do the hyperscalers have when they are setting prices for tokens and for workflows? This is where the market ought to be more thoughtful. Frankly, if you're buying every AI stock because AI is the future, you're not really doing due diligence and analysis. You're buying a narrative and narratives can flip on a dime. But if you're dismissing the entire thing because stocks corrected, you are also not doing analysis. You are reacting to price action and pretending it is insight. The useful middle ground requires a lot more due diligence and thoughtfulness, and it's much harder and rarer. It says AI is a real platform shift that is so transformative that there are a bunch of local bubble dynamics frothing around it. That means prices can fall and technology can keep advancing. It means some infrastructure might be overbuilt while other kinds of capacity remain profoundly scarce. It means some companies will spend too much and still not spend enough in the exact place that matters. Yes, that can be true. It means the biggest winners may not be the companies with the loudest AI story today. They may be the ones that control the bottleneck or own the customer workflow or route inference more efficiently or turn AI output into durable business value. This is the distinction I would watch. Don't ask if a company is doing AI if you're trying to figure out investments here. And none of this is investment advice. Ask where the demand is showing up. Is it paid usage or is it just engagement? Is it production workloads or is it just a pilot that got dressed up in a press release? Is it improving a workflow with really clear economics or is it creating more work for humans to review? Is the company buying capacity because customers are waiting or because the board wants an AI strategy? Is the model being used where expensive reasoning matters or is it just premium compute being burned on cheap tasks? Those questions are a lot less dramatic than bubble or revolution, but they're a lot more useful for determining where investment dollars ought to go. And they also explain why the stock correction does not remotely settle the issue. Markets can correct and then uncorrect for lots of reasons. Valuations stretch, trades get crowded, expectations get too high, capital rotates, financing costs matter. A company can disappoint investors even while the underlying business grows. And that is what makes this moment so profoundly tricky. It's not that the correction is meaningless. It's telling you that investors are having some feelings in the middle of a global energy crisis about underwriting unlimited AI spending without asking harder questions. That's great. They should ask harder questions. But the correction is not proof that the buildout is fake. It's proof that maybe the easy phase of this trade is over and the next phase is going to require a little bit more discrimination from people who are trying to figure out investment dollars. The market will start separating companies with real AI revenue from companies with AI language in the deck. It's going to separate infrastructure bottlenecks from commodity exposure. It'll separate tools that create measurable workflow value from tools that create demo value. It'll separate companies that can finance the buildout from companies that need the buildout to be financed by someone else. And yes, it will separate SaaS companies that have sticky services that are still valuable in the era of agents from ones that don't. And that's super healthy and it's exactly what should happen in a real platform shift. The first phase is almost always narrative. Everyone piles into the obvious names. The second phase is correction. The market realizes the story is more expensive, it's slower, it's messier than the headlines may have implied. So when someone asks, "Is AI a bubble?" My answer is parts of it are. Yeah, if you can release a press release and get a 500% pop in your stock, which I've seen once or twice, those are examples of a bubble. But that doesn't mean the whole system is a bubble at all remotely. Sure, some valuations are stretched. Some spending will be wasted. Some private market marks are probably ridiculous. Some companies are pretending a thin wrapper is a business. The seed rounds are getting really pricey in the valley. Some enterprises are buying expensive tools without changing the workflow enough to get the value. But the broader buildout is not hype floating above reality. There's real demand underneath. There's real revenue underneath. There's real physical scarcity underneath. And so the better question, it's not when the bubble pops. I don't think that's coming. The better question is who survives the sorting that is going to happen because if intelligence becomes a production system, then the winners are not just the companies with the best demos. They're the companies that can turn demand into reliable, affordable, high utilization inference. They can route the right task to the right model. They can get power. They can get memory. They can build capacity. They can make agents useful enough that customers keep paying after the novelty wears off. That's a much harder game than the stock chart made it look last year. But it's also a much more serious game. A bubble is hype detached from reality. That's kind of the definition of a bubble. The famous South Sea bubble was all about the hype. It cost Isaac Newton his fortune. This is messier than that. This is a real buildout with speculative money piled over the top. And the correction is the market starting to ask which layer is which. And that's fantastic. And that's a good question to keep in front of you. You should be asking where's the paid demand? Where's the bottleneck really? Who captures value when the tokens get cheaper? Is this business able to finance itself? Is the work really moving to agents in association with this particular company? And that's where the real story is. So if you are worried about a particular movement in the stocks, I want you to keep these questions in mind. These are evergreen questions. You can come back to them next month and the month after. We will take a while as a market to go through this sorting. Remember, AI is a marathon. It's not a sprint. The business of putting AI into companies and installing it is a 10, 20-year exercise. We're just at the beginning of that. We are writing the first chapter. Think about that the next time you look at your Robinhood account or think about where stocks are at. That provides a larger picture and I think it's a healthier picture because AI is here for the long term. AI is the most transformative technology of our lives and it can still have a ton of froth around the edges while that remains true. I hope this has been helpful. You can follow me for more relatively sober takes in a world that likes to argue about big overcorrections in binaries. I do not believe the world is a light switch world. I don't believe it's either bubble or not. And I'm going to argue against that a lot because I think it's a lazy question that underscores how little people understand how powerful AI actually is, as well as how little people understand the complexities of the AI dynamic. Yes, there are absolutely places in a buildout this big where capital is wasted. And yes, that does not mean that we don't see absolutely massive unmet demand with AI. Both can be true at once. Demand more of your investors. Demand more of your analysis. Demand more of the markets in understanding how these companies actually work. And frankly, if you're in financial press, feel free to get in touch with me because I feel like this is often incorrectly reported, especially that inference piece. I don't think people properly understand that one. All right, I will see you in the comments. Let me know what you think. Sound off on where you think some of these companies are at. I would love to hear because we can argue about that. I think that's a healthy conversation to have and that would be a way for us to talk as a community about which companies are sorting where and why. All right, I'll see you next time. Cheers.