Are We in an AI Investment Bubble? Lessons from the Internet Bubble
The awkward thing about investment manias is that they don’t look silly while you’re living through them. They look like the future arriving early.
That was true in the late 1990s. The internet really did change everything. E-commerce, search, cloud software, online advertising, streaming, smartphones, social networks, maps, payments, remote work, and whole new business models came out of that era.
And still, a lot of investors got crushed.
So when people ask, “Are we in an AI investment bubble?” the honest answer is: maybe in parts, maybe not in the whole thing, and we won’t know for sure until the cash flows show up or don’t.
That sounds slippery, but it’s actually the right starting point. A real technology can be world-changing and still be wildly overpriced for a while. The internet proved that.
This post is informational only and isn’t financial advice. No single article can tell you what to buy or sell. But we can talk through the signs that usually show up when expectations run ahead of reality.
The internet bubble is the best comparison, but it’s not a perfect match
The internet boom had a strange mix of truth and fantasy.
The truth was easy to see. More people were getting online. New companies could reach customers without physical stores. Email, search, digital ads, online marketplaces, and software delivered over networks were clearly useful. The internet was not a fad.
The fantasy was the idea that every internet company deserved a giant valuation just because it had a website, traffic, or a catchy domain name.
That difference matters.
A technology can be real, useful, and still too expensive in the stock market or private markets. That’s the core lesson from the dot-com era.
The late 1990s had several classic signs of excess:
Companies with little or no profit went public at huge valuations.
Investors accepted weak business models because “growth” sounded better than revenue quality.
New metrics replaced old ones, often for good reasons at first, then for silly reasons later.
Capital flooded into copycat companies.
The market treated a broad technology shift as if every participant would win.
Then the reckoning came. The Nasdaq peaked in March 2000 and fell sharply over the next couple of years. Many companies disappeared. Some famous names became shorthand for overreach. Pets.com and Webvan are still used as examples because they captured the mood so well.
But the crash didn’t kill the internet. It cleared out the weak stuff.
Amazon survived. Google grew after the crash. Booking, eBay, Salesforce, and other digital businesses showed that internet economics could work when the product, timing, cost structure, and customer behavior lined up.
That’s why the comparison to AI is so useful. The question isn’t “Is AI real?” Of course it’s real. The better question is: How much future profit has already been priced in, and who actually gets to keep it?
What would tell us AI investing has become overheated?
There’s no alarm bell that rings at the top. If there were, bubbles wouldn’t be so painful.
Still, there are patterns. When a market gets too excited, several things tend to happen at once. One or two of these signs can be normal. A cluster of them is worth taking seriously.
Expectations start outrunning adoption
A lot of people use AI tools now. Many businesses are testing them. Some are paying for them. That’s real adoption.
The tricky part is the gap between usage and durable profit.
In the internet era, “visits” and “eyeballs” became popular metrics. They mattered, but they didn’t answer the main question: can this company turn attention into cash, and can it do that after competitors show up?
AI has its own version of this problem.
A company might have impressive demos, fast user growth, or a product that feels magical. But investors still need answers to basic questions:
Are customers paying, or just trying a free tool?
Do customers keep paying after the novelty wears off?
Does the product save enough money or create enough revenue to justify its cost?
Can the company serve users profitably, given computing costs?
Does the product get stronger with scale, or do competitors catch up quickly?
The “wow” factor is not the same as business quality.
That doesn’t mean early losses are bad. Amazon lost money for years while building a giant business. Many software companies invested heavily before becoming profitable. The point is simpler: growth needs a believable path to cash flow.
Capital spending gets treated like proof
One big difference between AI and the internet boom is the amount of physical infrastructure involved.
The internet needed fiber, servers, routers, data centers, and devices. AI needs a lot of that too, plus specialized chips and huge power needs. Training and running large models can be expensive.
Heavy capital spending can be a sign of confidence. It can also be a sign of a race where everyone feels they have to spend, whether or not the return is clear yet.
During the internet boom, telecom companies spent heavily to build networks. Some of that capacity became incredibly valuable later. Some of it arrived too early, was financed badly, or sat underused for years.
AI could rhyme with that.
If companies spend huge sums on chips, data centers, and energy contracts, the spending itself doesn’t prove the economics work. It only proves that major players are trying not to fall behind.
The key question is whether that spending produces revenue high enough to cover:
The cost of the hardware
The energy needed to run it
The fast depreciation of chips
The engineering talent required
The ongoing cost of model training, safety, and maintenance
The price pressure that comes when similar tools compete
In plain English, useful technology still has to pay its bills.
Every company suddenly becomes an AI company
This is one of the easiest warning signs to spot.
In the dot-com period, adding “.com” to a business could change how investors saw it. Companies wanted to be framed as internet plays, even when their connection to the internet was thin.
We’ve seen a lighter version of that with other themes too, including crypto, cannabis, electric vehicles, and clean energy. When capital chases a story, companies learn to speak the story’s language.
For AI, the warning sign is not that companies use it. Many should. The warning sign is when the label becomes more important than the product.
For example:
A basic software feature gets described as a full AI platform.
A company talks more about model buzzwords than customer results.
Startups raise large rounds before proving demand.
Public companies mention AI constantly, even when it contributes little to revenue.
Investors stop asking whether the customer problem is painful enough to pay for.
The pattern is familiar. A powerful trend becomes a marketing sticker.
That doesn’t mean every company using the term is suspect. It means the label alone tells you very little.
Market gains narrow around a few winners
Bubbles often have a center of gravity.
In the late 1990s, the excitement clustered around internet portals, networking equipment, telecom, and new web companies. Some were real businesses. Some were not. The broader market got pulled along as the story grew.
With AI, the first big winners have been concentrated in places that supply the boom. Chip companies, cloud providers, data center businesses, and some software platforms have captured a lot of attention.
That makes sense. In a gold rush, selling shovels can be a great business.
But there’s a catch. If a small number of companies carry a huge share of market returns, investors need to ask whether their future profits can grow enough to support the prices. Great companies can still become expensive.
Cisco is a useful dot-com example. It was a real company with real products and real customers. It helped build the internet. But its valuation became so high at the peak that the stock struggled for years after the crash, even though the company itself remained important.
That’s one of the most uncomfortable lessons from the internet era: being right about the technology doesn’t mean being right about the price.
People stop caring about valuation
This may be the strongest sign of all.
When investors say, “Traditional valuation doesn’t matter anymore,” something has usually gone wrong.
To be fair, new technology can make old assumptions stale. The internet did make some old business models weaker and some new ones far more powerful. AI may do the same.
But valuation still matters because every asset is a claim on future cash. The future can be huge, but it can’t be infinite.
When people dismiss price completely, they tend to rely on a chain of perfect assumptions:
Adoption will be fast.
Customers will pay more.
Costs will fall.
Competition won’t crush margins.
Regulation won’t slow growth.
The leading companies will keep their lead.
The market will reward the same story for years.
Maybe some of that happens. Maybe a lot of it happens. But the more perfection a stock or startup valuation requires, the less room there is for normal disappointment.
That’s how investors can lose money in a winning industry.
Where AI looks like the internet boom
The similarities are real. They’re also easy to overstate.
The first big similarity is that both technologies are general-purpose. The internet was not one product. It was a new distribution layer for information, commerce, media, and communication. AI is similar in that it can touch many tasks: coding, writing, customer support, drug discovery, design, data analysis, education, security, and more.
General-purpose technologies are hard to price because they create second-order effects. The obvious use cases are only the start.
In the 1990s, most people could see email and websites. Fewer people could fully imagine cloud computing, online maps in every pocket, streaming video, creator platforms, or app-based transportation. The biggest effects took time.
AI may have that same delayed impact. Today’s chatbots and copilots might be the early layer. The bigger economic changes could come later, when AI gets built into workflows, devices, research tools, robotics, and software that acts with more independence.
The second similarity is the speed of narrative change.
During a technology boom, markets don’t just price current results. They price imagination. That can be useful because it funds experiments. It can also get reckless because imagination has no natural ceiling.
The third similarity is the wave of new company formation.
The internet lowered the cost of reaching users. AI can lower the cost of building products, writing code, generating content, analyzing data, and offering services. That means startups can appear quickly. It also means many will be easy to copy.
That’s a big deal. If a product can be built fast, the company behind it needs something stronger than first-mover excitement. It needs distribution, data, trust, workflow integration, low costs, brand loyalty, or a hard technical advantage.
The fourth similarity is the “build first, monetize later” mindset.
Sometimes that works. Sometimes it burns a mountain of cash.
Internet companies often chased traffic before revenue. AI companies may chase users before knowing whether inference costs, model costs, and customer prices can support the business. The danger shows up when investors reward usage without asking whether usage becomes profit.
Where AI is different from the dot-com era
The similarities get most of the attention, but the differences may matter even more.
The biggest difference is that today’s leading AI trade is tied to some of the most profitable companies in the world.
During the dot-com boom, many public internet companies had little revenue and no profits. Today, much of the AI excitement sits inside or around large firms with strong cash flows, existing customers, and massive distribution. Cloud providers, chipmakers, and major software companies already have real businesses.
That doesn’t remove risk. It changes the type of risk.
A profitable giant can overinvest. A great company can trade at a bad price. A market can become too dependent on a few names. But this isn’t exactly the same as dozens of unproven companies selling a dream through IPOs.
Another difference is the speed of product adoption.
The internet required people to get connected, learn new habits, and wait for infrastructure to improve. AI tools can spread through existing devices and software subscriptions. A new model can reach millions of users quickly because the distribution rails already exist.
That makes the adoption curve faster. It may also make competition faster.
If everyone can plug into similar models through APIs or use open-source alternatives, then the hard part may shift from having AI to using it in a way that customers can’t easily replace.
A third difference is the cost structure.
Many internet businesses became attractive because the marginal cost of serving another user could be low. AI can be different. Serving another query may require meaningful computing power, especially for complex models. Costs may fall over time, but they don’t vanish automatically.
This creates an odd tension.
AI can make workers more productive and software more useful. At the same time, the companies running the models need to manage expensive infrastructure. If customers expect low prices while compute remains costly, margins may disappoint.
A fourth difference is that regulation and safety questions are closer to the center.
The internet had policy debates too, of course. But AI brings direct concerns about copyright, privacy, bias, labor displacement, misinformation, national security, and liability. These issues may not stop the technology, but they can affect costs, speed, and which companies win.
A fifth difference is that the internet’s final consumer use cases were easier to understand.
Shopping online, reading news, sending email, and searching the web made sense quickly. AI is more uneven. Some uses are obvious, like coding help or summarizing documents. Others are still fuzzy. Many companies know they should experiment, but they don’t always know where the largest return will come from.
That creates a risk of “pilot purgatory,” where companies test tools endlessly without broad deployment. If that happens, revenue growth may lag the hype.
The best clues are hiding in unit economics
If you want to avoid getting swept up in the story, follow the unit economics.
That phrase sounds dry, but it just means asking whether the business works at the level of one customer, one product, or one transaction.
For AI, the important questions are pretty practical.
Does the tool save real money?
A customer support tool that cuts resolution time, reduces repeat contacts, and keeps customers happy can have clear value. A coding assistant that helps engineers ship faster may justify its cost. A medical research tool that shortens discovery timelines could be valuable, though those claims need serious proof.
By contrast, a tool that feels fun but doesn’t change cost, revenue, risk, or speed may struggle once budgets tighten.
The internet had the same split. Online travel booking solved a real problem. Search solved a real problem. Some pet food delivery models did not solve enough of a problem at the right cost at that time.
Who captures the value?
This is the most under-discussed question.
AI may create huge value for customers while leaving less profit for AI providers than people expect.
Think about open-source models, intense competition, cloud price cuts, and companies building their own internal systems. If AI becomes a feature inside many products, customers may benefit while vendors compete away much of the margin.
The internet did this too. It created enormous consumer surplus. We got free search, free email, free maps, cheap communication, and endless information. But not every company in the chain captured equal value. Some built toll roads. Others built features that became free.
Investors want to find the toll roads, not just the cool demos.
How fast do costs fall?
AI bulls often argue that compute will get cheaper, models will get smaller, and hardware will improve. That may be right. Technology often gets cheaper with scale and learning.
But investors need timing.
If costs fall after a company has already overpaid for infrastructure, profits can still suffer. If new chips make old chips less valuable faster than expected, depreciation matters. If model prices drop because competition heats up, revenue per unit can fall at the same time costs are changing.
During the internet buildout, a lot of real infrastructure was built. Some of it later became useful after being restructured through bankruptcies or asset sales. The technology won, but the original capital providers did not always win.
That’s a brutal but useful distinction.
Does the company have a moat?
A moat is just a durable advantage. For AI companies, possible moats include:
Proprietary data that others can’t easily access
Distribution through existing products
Deep integration into customer workflows
Strong research talent and infrastructure
Trust in sensitive areas like finance, health, or legal work
Lower costs at scale
Network effects from user feedback or usage
Weak moats are easier to spot too:
A wrapper around someone else’s model with little differentiation
A feature that a larger platform can copy
A product customers enjoy but don’t rely on
Growth driven mainly by novelty
High spending that doesn’t create lasting advantage
The internet created massive winners, but it also trained investors to ask a sharper question: what stops someone else from doing this?
So are we in one now?
Here’s the most reasonable answer: parts of the AI market have probably been priced with a lot of optimism, but that doesn’t mean the whole theme is fake or doomed.
That answer may feel unsatisfying, but it’s what history suggests.
The internet was both overhyped and underappreciated. Overhyped because investors paid absurd prices for weak companies. Underappreciated because the long-term impact turned out to be bigger than many skeptics imagined.
AI could follow the same pattern.
We may see:
Real productivity gains in some fields
Large profits for a smaller number of infrastructure and platform companies
Many startups fail after burning through cash
Price pressure as models become more common
A few surprise winners that don’t look obvious today
A shakeout when companies have to prove customer value
Long-term adoption even if market prices fall first
That last point is key. A market correction would not prove AI failed. It would only prove expectations got ahead of near-term economics.
The opposite is also true. Strong adoption would not prove every valuation is reasonable.
The internet teaches us to hold two thoughts at once:
The technology can be transformational.
The investment returns can still disappoint if you pay too much.
A bubble is not usually a lie. It’s often a true story with the wrong price and the wrong timeline.
How to think about AI without getting carried away
If you’re trying to make sense of AI from an investing angle, the goal isn’t to predict the exact top. Almost nobody does that well.
The goal is to separate durable value from excitement.
A simple checklist helps.
Ask whether the company has current revenue tied to AI, not just future plans. Then ask whether that revenue is profitable or moving toward profitability. Look at who pays, why they pay, and whether they keep paying.
Ask whether the product is a must-have or a nice-to-have. Budgets reveal the difference. In good times, companies test lots of tools. In tighter times, they keep the tools that save money, reduce risk, or drive sales.
Ask whether costs are visible. AI can look high-margin from the outside, but compute bills matter. So do talent costs, data costs, compliance costs, and infrastructure spending.
Ask whether competition is likely to strengthen or weaken the business. If models get cheaper and more available, some companies benefit. Others lose pricing power.
Ask whether the valuation needs everything to go right. If the price assumes years of explosive growth, strong margins, and no major setbacks, even a great company can become risky.
And maybe most of all, ask what has to be true five years from now.
Not what sounds exciting next quarter. Not what gets people talking. What has to be true in the actual business?
For many AI companies, the answer needs to include paying customers, lower costs, clear workflow value, and some kind of durable edge.
That’s less glamorous than a demo. It’s also where the truth usually shows up.
FAQ
Is AI overhyped or real?
Both can be true. AI is real and already useful in areas like coding, writing assistance, search, customer support, and data analysis. The hype comes from assuming every AI product or stock will turn that usefulness into large profits.
Was the internet bubble a sign that the internet failed?
No. The internet kept growing after the market crash. The crash mostly showed that many companies were overpriced, poorly funded, or built on weak business models. The technology succeeded, but many early investments did not.
What is the biggest warning sign for AI stocks?
The biggest warning sign is when valuation depends on perfect outcomes. If a company needs rapid adoption, high pricing, falling costs, and limited competition all at once, there’s little room for disappointment.
Could AI companies be more profitable than dot-com companies?
Some could be. Many AI leaders are tied to large companies with existing customers, strong cash flow, and major distribution. But profitability will still depend on compute costs, pricing power, competition, and customer value.
How should long-term investors compare AI to the internet?
The useful comparison is not “AI equals dot-com crash.” It’s that a real technology can have a messy investment cycle. Long-term winners may emerge, but paying too much or backing weak business models can still lead to poor returns.
The best lesson from the internet era is not to be cynical. Cynicism would have made you miss one of the biggest technology shifts in history.
The lesson is to be selective.
AI may change how software works, how people work, and how companies compete. It may produce enormous winners. It may also produce failed startups, overpriced stocks, stranded infrastructure, and painful resets along the way.
That’s not a contradiction. That’s usually how major technology cycles work.
So, are we in an AI investment bubble? In some corners, probably. Across the entire technology, probably not. The smarter question is which parts of the boom have real earnings power, which parts are just riding the story, and whether the price leaves enough room for the future to arrive imperfectly.
The internet didn’t reward everyone who believed. It rewarded the companies that built durable value, and the investors who didn’t confuse a great story with a fair price.
