100 Reasons Why the AI Bubble Will Not Burst Anytime Soon (2026)
On Thursday, November 5, 2026, the Nasdaq fell 5.5 percent, its worst single day in nineteen months. The S&P 500 lost 4.4 percent. The Dow shed about 1,800 points. More than $3.2 trillion in market value evaporated before dinner, which is more than the annual output of Canada, and the volatility index spiked nearly 60 percent. Nvidia fell 8 percent. Micron fell 14. By Friday the rout had spread to Asia and Europe, the Nasdaq sat in correction territory, and every feed I opened had the same image: Nvidia's stock laid over Cisco's from 1999, with a caption explaining that this ends the same way. Welcome back to Money school series..
I have been writing this blog long enough to remember when the same feeds told you to install Claude Fable 5.2 and Mythos 5.2 locally. They were not real. We checked, together, and that is the deal on this page: no vibes, no vibes-adjacent reasoning, just numbers you can verify in a couple of minutes. So before anyone declares the AI bubble burst, ask the only question that matters for a bubble: were the cash flows ever real?
That is what a bubble is, properly defined. Not "prices went down," because prices go down in every cycle, a lesson we covered from the macro side when we ran the recession-signs math last year. A bubble bursts when the market discovers that the cash flows behind the prices were fiction: Pets.com burning $62 million to sell $6 million of pet food with free shipping included, telecom carriers that borrowed tens of billions to lay fiber that stayed dark. If the cash flows are real, what you get instead is a correction. Painful, headline-generating, survivable.
This post is my case, in one hundred parts, that we are in the second thing. Consider it a working reference on the reasons behind the rapid growth of AI, one checkable line at a time. Ten sections, ten reasons each, running from revenue to funding to adoption to the dot-com comparison to the strongest bear arguments, which I will take seriously because a post like this is worthless if it strawmen. Every number here is public and checkable, so check it up during free time and always don't trust anyone when it commits to making fianncial decisions like i used to say, not even our site.. If something moved between me hitting publish and you reading this, that is markets; the direction of travel is the point.
One ground rule before we start. I am not claiming AI stocks cannot fall further as NVIDIA fallen due to scandal/anthropic just planning for IPO at $75-80, or that every AI startup deserves its valuation. Some do not, and some will die, and that is fine. The claim is narrower and testable: the AI buildout is not a bubble in the 2000 sense, and what happened on November 5 was a repricing, not a reveal. The rest of this post is the long version of why the AI boom is expected to continue. Alright. Let's go.
1. The Money Is Real (Reasons 1-10)
Bubbles live on promises. This cycle ships invoices. Start with revenue, because everything else in this post follows from it, and because how AI is transforming the global economy gets billed before it gets felt.
1. OpenAI's revenue is still compounding, miss or no miss. The scariest datapoint of the fall was OpenAI's annualized revenue landing around $50 billion, roughly $20 billion short of the number the company had signaled, and the bears enjoyed that enormously. It is also about 2.5 times where the run rate stood entering the year, after OpenAI exited 2025 above a $20 billion pace and did $3.7 billion of actual revenue in all of 2024. A company that "misses" while growing ~70 percent is not a fraud unwinding. That is a demand curve, occasionally impatient with forecasts.
2. Anthropic went from $10 billion to $65 billion in about a year. Anthropic generated roughly $10 billion of revenue in all of 2025. By the end of July 2026 its run rate topped $65 billion, with most of it coming from business customers. The interesting part is not just one lab growing this fast. It is that a second lab grew this fast while the first one was also growing. Markets where two competitors each multiply revenue several-fold in a year are deep markets.
3. Claude Code alone is a $2.5 billion business. Anthropic's coding agent went from a $500 million annualized pace in September 2025 to $2.5 billion by February 2026. One product, one category, five-x in five months. Bubbles do occasionally produce one hot product; they do not usually produce individual tools with nine-figure monthly billing that companies pay because the tools write working software. If you want to see the why firsthand, our Claude Code with Amazon Bedrock setup gets you running in an evening.
4. Nvidia just posted a $96 billion quarter. While accelerating. Revenue of $96.2 billion in the July 2026 quarter, up 106 percent year over year, with $89 billion of it from data centers. Full fiscal-year 2026 revenue was $215.9 billion, with data center revenue of $193.7 billion, up 68 percent. Note the shape of that: the annual growth rate is lower than the latest quarter's, because growth is accelerating, not decaying. The demand-collapse story has now been wrong every single quarter since 2023.
5. Nvidia's earnings grew into its valuation, which is the opposite of Cisco. At around $180 a share, Nvidia trades near 23 times trailing earnings and roughly 11 times forward estimates. That is a seven-year low for the multiple and about half its own ten-year average, which sits near 55. Cisco entered March 2000 at well over 100 times earnings with order growth decelerating. When the flagship stock of a supposed mania is the cheapest it has been relative to earnings in seven years, "euphoria" is not what is priced in. De-rating is.
6. The cloud arms race reaccelerated at scale. Google Cloud grew 82 percent year over year in the June 2026 quarter, to $24.8 billion in a single quarter. AWS, the biggest of them all, is growing at a near-38-percent clip. Microsoft's AI business grew 123 percent year over year. These are not startup percentages on rounding-error bases. This is the largest computing market in history reaccelerating while already at colossal scale, which is a thing that basically never happens. (And if the word Bedrock blurs into everything else in that sentence, our plain-English Bedrock explainer fixes that in ten minutes.)
7. Token consumption is the tell, and it is vertical. Google went from processing 9.7 trillion tokens a month in May 2024 to 480 trillion a year later to 3.2 quadrillion by May 2026. That is 7x growth in a year, stacked on top of roughly 50x the year before. Goldman's forecast has monthly token processing hitting 47 quadrillion by 2028. You cannot fake metered, billed usage at that scale for that long. Somebody is paying for every one of those tokens.
8. People are paying, not just clicking. ChatGPT crossed a billion weekly users in August 2026, and roughly 50 million of them pay for the privilege. The subscription product did not exist four years ago. Consumer software that converts on the order of 5 percent of a billion-person weekly user base into recurring revenue is a business, not a science fair project. If you are wondering whether that $20 a month makes sense for you personally, we ran the ChatGPT Plus math earlier this year.
9. The frontier labs' combined run rate roughly quadrupled in a year. OpenAI and Anthropic together exited 2025 with something like $30 billion in combined annualized revenue. This fall the same two companies are running at a combined pace approaching $115 billion. Whatever you think about burn rates and margins, the demand line points one way, and it is not the direction the revenue lines of 2000 pointed.
10. The revenue rides five different business models. How companies are monetizing artificial intelligence in 2026 is a five-part answer: subscriptions, metered API, enterprise seats,
2. Follow the Funding (Reasons 11-20)
The second thing to check in a bubble is who is paying for the buildout, because in 2000 the answer was "lenders who will regret this." This section is the long answer to why companies are investing billions in AI: the people paying are the people earning.
11. The builders are the most profitable companies in history. The four biggest spenders, Microsoft, Alphabet, Amazon and Meta, will deploy on the order of $700 billion this year, roughly double 2025's pace, with some trackers now above $750 billion after mid-year raises. They will also generate several hundred billion dollars of operating cash flow in the process. The telecom buildout was carried by companies that needed capital markets to survive. This one is carried by companies that could fund it from a year of profits and are choosing to fund it from two.
12. It is funded by cash flow, not debt. WorldCom went down with more than $30 billion of debt and a fraud conviction. Global Crossing took $12 billion into bankruptcy. Roughly $2 trillion of telecom market value evaporated, and only about 3 percent of the country's fiber was actually lit when the reckoning came. Today's builders pre-pay from operating cash, and when they do tap markets it is investment-grade paper. Who eats the loss if demand disappoints is a completely different question when the builders own the assets outright.
13. The most famous doom-money just bought in. Alphabet went out in June 2026 and raised $80 billion of fresh equity for the buildout, and Berkshire Hathaway took $10 billion of it. The conglomerate built by the man who wrote "be fearful when others are greedy" allocated ten figures to AI infrastructure at what everyone agrees is peak skepticism season. Whatever that says about Berkshire, it says the paper is not junk.
14. Nvidia returned $37 billion to shareholders in nine months. Through the first nine months of fiscal 2026, Nvidia bought back stock and paid dividends totaling $37 billion, mid-boom. Bubbles starve everything to feed the story. This cycle's anchor company mails cash back to owners while it grows, which means the marginal owner is being paid to hold rather than praying for a greater fool.
15. Yes, the buildout is telecom-sized as a share of the economy. Now look at who is paying. Here is the strongest version of the bear case, conceded up front: by capex share of GDP, the AI buildout is now in the same league as the telecom mania at its peak, bigger by some measures. That chart is real. The rebuttal is the rest of this section and the next one: telecom's spend was debt raised by unprofitable carriers against hoped-for demand, and this spend is cash and investment-grade equity deployed by the most profitable platforms alive against contracted demand. Same size, opposite balance sheet.
16. Earnings grew while the spending grew. Microsoft, Alphabet, Meta and Amazon did not sacrifice profitability to chase AI. Their earnings climbed through the entire buildout, with Alphabet's October stumble the notable exception that we will deal with honestly in Section 8. In 1999, the spending was the business plan. In 2026, the spending is a line item under the business plan.
17. The 2022 stress test already ran, and capex survived it. Rates spiked, everything de-rated, ad markets dipped, layoffs swept the industry, and every serious macro voice declared the party over. Hyperscaler AI capex grew straight through it, from 2022 into 2023 into 2024. Budgets that grow through a macro shock are structural budget lines, not fashion. The 2026 selloff is the second time this cycle has run that experiment.
18. The spenders say the constraint is supply, not appetite. Microsoft and Meta together spent $165 billion on AI infrastructure in a single quarter this year, up 87 percent year over year, while telling investors the binding constraint is chips and power. Amazon's number for the year is around $200 billion. When every builder in a market says some version of "we would build more if we could get more," the capex is not a bet on demand. It is a complaint about supply, and it is why global AI spending is increasing past every forecast that tried to cap it, quarter after quarter.
19. Debt is entering the system at the edges, where it belongs. Honesty checkpoint: data center credit is a real and growing market, and there is genuine stress at the fringe. A bank has been shopping around half a billion of data center exposure, Australia's central bank flagged the sector, and one operator there hit a three-year low. Watch all of it. But a $700 billion annual buildout showing a few hundred million in stressed fringe loans is a rounding error, not a WorldCom forming. The core is overcapitalized; the cracks are at the edge. In 2000 the cracks were in the core.
20. The market already ran the discipline test this quarter. The late-October leg of this selloff, the one before the midterms, happened because Alphabet missed on profit and investors took half a trillion dollars off its market cap for spending too aggressively on AI. Sit with that. A bubble rewards spending until the money is gone. This market punished overspending within one earnings cycle. The punishment is the healthy part, and it is the single clearest difference between 2026 and 1999.
3. Adoption That Puts Every Prior Tech to Shame (Reasons 21-30)
Revenue tells you what happened. Adoption tells you what happens next. And no technology in history has been adopted like this one.
21. Two months to 100 million users, and the record still stands. ChatGPT reached 100 million users in January 2023, roughly sixty days after launch. TikTok, the previous record holder, needed nine months. Instagram needed about two and a half years. Every technology adoption comparison made since has been measured against this launch, and nothing, including everything built on top of AI since, has beaten it.
22. A billion people use it every week. OpenAI said in August 2026 that a billion people use ChatGPT every week. For scale, that is on the order of one in eight humans alive, using one product from one company, weekly, three years after launch. The internet itself needed the better part of a decade to become a weekly habit at that scale. Whatever AI becomes, it already is the fastest-diffused consumer technology ever measured.
23. The workplace numbers are unlike anything the Fed has measured before. A San Francisco Fed study found that about 40 percent of US working-age adults had used generative AI at work, within roughly two years of the technology existing. Personal computers needed a decade to reach that share of the workforce, and the internet needed longer. Adjusted for years since introduction, this is the fastest workplace technology diffusion in the statistical record, and it happened without anyone training the workforce first. We looked at the Bill Gates take on which jobs stay human through exactly this curve.
24. Roughly nine in ten organizations are already in. Survey after survey through 2025 and 2026 lands near the same number: about 90 percent of organizations now run AI in at least one business function. The pilot-to-production gap is real and worth criticizing, and Section 8 does. But you cannot have a production gap without near-universal adoption of the experimenting kind first. "Is adoption happening?" stopped being a live question somewhere in 2024, and so did AI adoption across different industries: the by-sector tables read like a phone book, hospitals next to banks next to farms.
25. Developers adopted it fastest, and they never went back. GitHub Copilot went from launch to more than 20 million users in under three years, quadrupling in one year alone. More than 8.5 million developers build on Google's models every month. Developers are the least impressionable users in technology, professionally allergic to hype, and they converted to AI tooling inside twenty-four months. When the tooling layer of an entire industry resets that fast, it is not a fad. It is a new baseline.
26. The machines themselves are going AI-native. AI-capable PCs are heading toward half of shipments, NPUs are becoming standard silicon, and Apple and Google are shipping real models on the phones people already own. This matters for a subtle reason: on-device AI is a distribution channel that already exists. A billion phones do not need a business case approved to run the model. They just do, and every model improvement lands on all of them at once. The laptop side of that story got its own treatment here: we rounded up the best laptops for running local models this year.
27. The Global South skipped the queue. India is among ChatGPT's largest markets, and usage growth is fastest in places where the constraint was never skepticism, it was access. A student in Lagos or Lucknow getting genuinely good tutoring for the price of a data plan is not a speculative use case from a keynote. It is the most price-inelastic demand on the planet, and it is coming online by the hundreds of millions.
28. Microsoft put a Copilot in front of 112 million people. Not as a demo, as product surface. Copilot passed 112 million users this year. Enterprise software does not reach nine figures of seats through enthusiasm. It gets there through procurement, security review, data agreements, and a renewal decision that a budget owner has to defend annually with usage statistics. Seats that clear all four gates are why enterprise AI adoption is accelerating instead of stalling: the vetting machine keeps approving.
29. The retention curve is boring, and that is the point. Fad technology spikes and decays, the way metaverse engagement spiked and decayed within eighteen months. ChatGPT went from 800 million weekly users in October 2025 to 900 million in February 2026 to a billion by August. Three years in, weekly usage is still climbing. There is no decay signature anywhere in the curve, and believe me, people have looked. That flat-and-climbing line is the whole case for why generative AI is not a passing trend, in one chart.
30. Even the skeptics are measured users now. The people writing "AI bubble" essays are, overwhelmingly, drafting them with AI assistance, checking their citations with AI search, and shipping their newsletters through AI-assisted tooling. Revealed preference is the cheapest form of due diligence that exists. When the skeptics' own workflow quietly depends on the thing they are skeptical about, adoption has stopped being the question.
4. Production, Not Pilots (Reasons 31-40)
The snarky version of the bear case asks where the products are. The products are everywhere. You just stopped calling them AI. These ten are the real-world applications of artificial intelligence that already have billing codes, wage bills, or regulators attached.
31. Waymo is a transit system now. Half a million paid robotaxi rides a week across ten cities by September 2026, with 3.9 million driverless miles logged per week and more than twenty cities, London and Tokyo included, targeted next. Tesla's robotaxi fleet is adding over half a million driverless miles a week of its own. That is two companies running a service at commercial scale in a category that was a moonshot demo a decade ago.
32. AI had a Medicare billing code before ChatGPT existed. Viz.ai's stroke-triage software received Medicare's new-technology add-on payment in 2020. Read that twice. The United States government, not a venture fund, decided that AI triage for stroke was worth reimbursing, two full years before anyone had heard of a chatbot. Reimbursement is the hardest adoption barrier in all of healthcare, and AI cleared it before the hype started.
33. More than a thousand AI medical devices are FDA-authorized. The FDA's published list of AI-enabled medical devices has passed a thousand and keeps climbing, spanning diagnostic imaging, cardiology, stroke detection and more. These are not pilots. They are regulated, installed, billed products sitting inside hospital workflows, and the number grows every quarter regardless of what the Nasdaq does. Between this list, the Medicare codes in reason 32, and the fraud models quietly scoring every card transaction, you can see how artificial intelligence is transforming healthcare and finance at the same time, in the two hardest industries to crack.
34. John Deere's See & Spray cuts herbicide by up to two-thirds. Computer vision that decides, plant by plant, whether to spray, cutting herbicide use by as much as about two-thirds on covered acreage. Farmers adopted it because input costs are math, not vibes, and the math clears. The ROI story in agriculture is measured in dollars per acre per season, and it keeps passing.
35. Amazon runs more than a million robots. Its fleet passed a million units, with AI models now orchestrating warehouse flow across the network and humanoid pilots underway. The logistics layer of the world's largest retailer is a deployed robotics-and-AI system that predates the LLM boom and compounds with it. This is what "physical AI" looks like in production: unglamorous, enormous, and already paying for itself. It is also how AI automation is changing businesses at the largest employer scale on earth, one workflow at a time, with a CFO watching each one.
36. The ad engines are the biggest deployed AI on earth, and they print. Meta's and Google's ranking and bidding systems, rebuilt on modern AI, lifted conversion rates and ad prices at the same time. Google's AI Overviews reached 2 billion monthly users while monetizing at roughly the same rate as classic search results. The fear that AI breaks the economics of search has now run for two years without showing up in a single revenue line. Search revenue kept growing right through the scare, and AI Overviews monetizing at parity is how AI is creating new revenue opportunities inside products that already had revenue, the least speculative kind there is.
37. A third of the code at Google is written by AI. Sundar Pichai has said more than 30 percent of Google's new code is AI-generated, and Microsoft reports a similar share across its repositories. These are the most conservatively run engineering organizations on the planet, with every incentive to be skeptical, and they have made AI-generated code a default part of the production pipeline. That is an existence proof for the entire category, stamped by the two companies with the most code to lose.
38. AI-designed drugs are in human trials. Insilico's pulmonary fibrosis drug, discovered and designed by AI, is in Phase 2 trials. Isomorphic Labs signed milestone deals worth billions with major pharma companies. Recursion and Exscientia merged specifically to scale AI drug discovery. Drug pipelines take a decade to read out, which is why this story feels early, because it is. But molecules designed by algorithms are already inside human beings, in clinical trials, with published protocols.
39. AlphaFold won a Nobel Prize. Two hundred million protein structures predicted, used by millions of researchers across more than 190 countries, and recognized with the 2024 Nobel Prize in Chemistry. Around it sit weather models that beat physics-based forecasting on key metrics, 2.2 million new inorganic crystal candidates from a single DeepMind materials model, and AI controllers steering fusion plasma. The "it is just autocomplete" crowd owes an explanation to a Nobel committee.
40. The productivity numbers are finally measured, and they are real. Customer support agents gained 14 percent productivity on average and 34 percent for novices in a landmark controlled study. Developers completed a coding task 55 percent faster with an AI assistant in GitHub's controlled experiment. Neither number is a miracle, and that is the point: this is how artificial intelligence improves business productivity, fourteen percent here, thirty-four percent there, everywhere at once. Modest, measured, replicable gains across millions of knowledge workers is exactly what a general-purpose technology looks like in its early innings, and it compounds. Even the humblest corner of office work got a post here: we tested whether AI can actually read spreadsheets, honest answer included.
5. The Demand Is Contracted, Not Imagined (Reasons 41-50)
A bubble hopes for demand. This buildout signs it, in documents, with penalties for walking away. If you want to know why demand for artificial intelligence is increasing, read contracts, not commentary.
41. Oracle is sitting on $664 billion of signed future revenue. Oracle's remaining performance obligations hit $664 billion in September 2026, up $209 billion in a single year, on the back of cloud contracts signed with a who's-who of AI including OpenAI, xAI, Meta, Nvidia and AMD. RPO is contract law, not narrative. These are customers who have already agreed, on paper, to pay, and Oracle's cloud infrastructure revenue is growing at a triple-digit rate as it converts.
42. Google Cloud's backlog is half a trillion dollars. Google Cloud's contracted backlog reached $514 billion, after growing more than $50 billion in one quarter, and over half of it converts to revenue within 24 months. This is what demand looks like when it is booked rather than hoped for: lumpy, enormous, and legally binding, with a conversion schedule the CFO has to stand behind on every call.
43. The power is bought for decades. Microsoft signed a twenty-year deal to restart a Three Mile Island reactor for its data centers. Google contracted a fleet of small modular reactors. Meta signed a twenty-year nuclear power agreement of its own. Nobody underwrites twenty years of nuclear output on vibes. You do it against contracted compute demand on the other side of the ledger, because both sides of that trade had to survive the same credit committee. Twenty-year paper is also the cleanest answer to why AI data center investments are increasing: nobody underwrites a reactor against a fashion.
44. The supply chain is sold out, not stuffed. Gas turbine slots are booked years out, with the major manufacturer's order book stretching toward 2028. High-bandwidth memory is sold out years in advance. Advanced packaging capacity at TSMC is allocation-constrained. Scarcity across an entire supply chain is the market screaming that demand exceeds supply, which is the exact opposite of a bubble's signature signal of discounting and bloated inventory. In those order books, the future demand for AI chips and GPUs is not a forecast. It is a line item with a delivery date.
45. Jensen Huang's five words: "cloud GPUs are sold out." That was the CEO of Nvidia on the November 2025 earnings call, and demand has only accelerated since, with the July 2026 quarter growing 106 percent year over year. Here is the channel check on a bubble: price cuts and channel inventory piling up in warehouses. Here is the channel check on this market: allocations, lead times, and customers complaining they cannot get enough.
46. Stargate is pouring concrete. The flagship Abilene campus is live, UAE capacity is coming online this year, an anchor agreement in Malaysia was signed on next-generation racks, and independent tracking puts the program's US sites above nine gigawatts projected by 2029. The fiber era's punchline was that the conduit got lit a decade later anyway, because the infrastructure was real. This conduit is being energized on schedule, while the argument about it rages.
47. Sovereigns are now customers. The EU has its InvestAI program with a 200 billion euro envelope, the Gulf states are building gigawatt campuses, India runs a national AI mission, and Japan and Korea each have their own programs. When nation-states treat compute the way they treat oil reserves and chip fabs, as strategic assets to be secured rather than products to be evaluated, the demand floor stops being a market question and becomes a policy constant. That is the future of AI investments in the global market in one sentence: less froth, more treaties.
48. CoreWeave contracted five gigawatts with Nvidia alone. The listed GPU cloud signed on to build out more than five gigawatts of AI factories by 2030 in partnership with Nvidia, on top of multi-year commitments from hyperscalers that anchor its revenue. Neoclouds growing several-fold per year on contracted revenue are what a capacity market looks like from the inside. The 2026 version of a telecom carrier, whatever its flaws, is signing customers before it builds.
49. Enterprise seats renew, they do not just launch. The AI line items inside Microsoft, Salesforce and ServiceNow go through procurement, security review, and annual renewal decisions that a budget owner has to defend with usage data. The pattern so far is expansion rather than churn: seats grow, usage tiers climb, contracts lengthen. Churn is what kills a SaaS bubble. The renewal data keeps showing the opposite, which is why the spenders keep spending, and it is the cleanest evidence of how businesses are benefiting from AI adoption: nobody renews a seat nobody uses.
50. The commitments stagger, and the staggering is the tell. OpenAI's roughly $1.4 trillion of compute commitments are staged across six to seven years, not payable next quarter. Compare the paper that mattered in 1999: backloaded network leases signed by carriers with no revenue, against vendors who financed the purchases themselves. The headline number here sounds insane until you read the maturity schedule, and then it sounds like a mortgage on a house that already has tenants.
6. The Economics Improve Every Quarter (Reasons 51-60)
Every serious argument for "this is 2000" depends on AI getting more expensive per unit of value. It is getting cheaper, fast, and usage is exploding in response.
51. The same intelligence gets roughly 10x cheaper per year. A widely cited a16z analysis tracked the cost of GPT-3-class capability falling about 280-fold in eighteen months, and the pattern has continued: model-for-model, API prices have dropped by an order of magnitude annually since. In a bubble, the asset gets more expensive as more people buy it. In a deflationary technology cycle, falling cost unlocks new users and use cases continuously, which is the token curve from Section 1 doing exactly that.
52. DeepSeek day was a demand shock in disguise, and the market learned in weeks. January 2025: a Chinese lab releases a frontier-grade model trained for single-digit millions of dollars, Nvidia loses roughly $600 billion of market cap in a day, the largest single-day loss in stock market history, and the bubble crowd declares the ending. Nvidia made most of it back within weeks and blew through to new highs within months. Cheaper inference means more consumption. The Jevons paradox, demonstrated live, with receipts. We track the falling floor in our API pricing comparison; every tier on that page is a Jevons anecdote now. That floor is also how AI is creating opportunities for businesses outside the Fortune 500: 2023 prices excluded almost everyone, and the 2026 tiers read like a utility bill.
53. There are two scaling dials now, and neither is stuck. Pretraining scaling has held across eight orders of magnitude of compute, exactly as the original scaling papers predicted. Test-time compute, letting models reason longer before answering, opened a second axis in 2024 and has been compounding since. The bear case needs both dials to fail at the same time. The labs keep publishing results from both, and the benchmark gains keep arriving. The future of generative AI technology now has two engines instead of one, and neither is near its ceiling.
54. Open weights put a permanent price floor under intelligence. Llama, Qwen, DeepSeek: frontier-adjacent models you can download, inspect, and run at marginal cost, forever. Kimi K3.1 Pro is the current flagship of that list, 1.6 terabytes of open weights, and we downloaded and served the whole thing for the install guide. The entire run-AI-locally series on this site exists because of this one reason. A free alternative does not burst; it anchors the market. It also means the value migrates up the stack, to products and to the compute everyone still needs to run everything, which is exactly where the revenue in Section 1 lives, and where the long-term opportunities in the AI ecosystem are stacking up: products on top, compute underneath. Openness is a floor, not a leak.
55. Every price cut has been answered by a volume explosion. Google's monthly token processing went 9.7 trillion to 480 trillion to 3.2 quadrillion in two years, and API prices fell the entire time. Falling unit prices with exploding total revenue is the signature of a young, healthy market discovering elasticity. An overripe market shows the opposite: rising prices to sustain revenue while volume leaks away. There is no version of that second pattern anywhere in this cycle's data.
56. Inference is the majority of the load now. Most data center compute serves live traffic, real people using real products, rather than one-time training runs. This flips the risk profile of the whole buildout. Training capex is a bet on a future model. Inference capex is capacity planning for metered, billed usage, which is why AI computing demand will continue rising even if the training race paused tomorrow: live traffic does not wait for the next model. Utilities and railroads were financed for a century on the second kind of demand, and they outlived every panic about them.
57. GPUs are general-purpose in a way fiber never was. Dark fiber could carry one thing: bits, once someone lit it. A GPU cluster can serve ads, rank videos, run physics simulations, train the next model, or host a thousand startups' inference workloads, switching between them by software. Asset flexibility is the difference between the 2002 fiber overhang, which sat idle for years, and a compute overhang, which reprices into the next-best workload almost immediately.
58. The refresh cycle builds a floor under demand. Servers and GPUs get replaced on roughly five-to-six-year cycles, which the hyperscalers set based on observed asset life. By the end of this decade, the installed base from the current buildout generates a replacement market running into the hundreds of billions per year before a single new workload exists. Railways survived their manias partly because maintenance never stops. Compute has the same annuity hiding under the volatility.
59. Old chips keep earning. Last-generation GPUs do not get scrapped. They move down the stack into inference, internal workloads, and cheaper tiers, where they run for years. A100s from 2020 still serve production traffic at scale across the industry right now. The depreciation-schedule argument in Section 8 is an accounting disagreement about pace, not evidence that the assets stop working when the ledger says so.
60. The productivity gains are modest, measured, and compounding. Fourteen percent average gains for support agents, thirty-four for novices. Fifty-five percent faster coding in controlled tests. Measurable lifts in writing, analysis, and document review, study after study. None of it is magic, and that is what makes it credible. A 14 percent efficiency gain applied across the knowledge-work wage base of the developed world is a multi-trillion-dollar annual effect, arriving a few percent at a time. Compounding does the rest, the same force our Money School compound-interest lesson keeps proving with a calculator and zero hype. That arithmetic is the whole argument for why AI has long-term economic value: small percentages, applied to the largest wage base in history, for decades.
7. Why the Dot-Com Comparison Fails on the Numbers (Reasons 61-70)
The chart shape is the same. The balance sheets, valuations, funding structure, and supply of garbage paper are all different. This is the AI bubble versus dot-com bubble comparison, line by line, and it is the core of why AI technology is different from previous tech bubbles. Count the ways.
61. The flagship stock is cheap. Cisco's was absurd. Nvidia trades around 23 times trailing earnings and roughly 11 times forward estimates, a seven-year low multiple and about half its ten-year average. Cisco entered March 2000 at well over 100 times earnings, with order growth visibly decelerating. If this were truly 1999 redux, the flagship would be the most expensive stock in market history on rising optimism. Instead it is a mega-cap at a GARP multiple while its revenue doubles.
62. The 2000 leaders had no earnings. These have the best earnings ever printed. Pets.com did $6 million of revenue against $62 million of losses and liquidated nine months after its IPO. Microsoft, Alphabet, Meta and Amazon each earn tens to hundreds of billions annually, and their earnings grew straight through the AI capex boom. The four biggest AI spenders are among the most profitable enterprises capitalism has ever produced. The 2000 spenders were the least profitable.
63. Telecom built with borrowed money. This build pays cash. WorldCom took more than $30 billion of debt into what was then the largest bankruptcy in history. Global Crossing followed with over $12 billion. The sector wiped out roughly $2 trillion of market value, and the debt-funded builders could not survive even a short demand pause. Today's builders could freeze capex tomorrow without tripping a single covenant. Same chart, opposite solvency risk.
64. Only about 3 percent of fiber was lit when the crash came. The dark-fiber glut existed because the asset was single-purpose and its demand, consumer broadband, needed another decade to arrive. Compute is multi-purpose and its demand arrived before the buildout did: the clouds are sold out. Overbuilding risk in this cycle is real and worth watching, but it is a different asset class with a different failure mode, repricing into other workloads rather than sitting dark.
65. The garbage-IPO supply is missing. In 1999 alone, 457 companies went public, and roughly one in four doubled on the first day of trading, many with no revenue and a URL for a business plan. The 2026 IPO market has been disciplined by comparison, and the frothiest paper lives in private markets with structured terms rather than in retail brokerage accounts. Private losses hurt their holders. Public 1999-style losses detonated pensions.
66. Retail already blew off its top, in 2021. SPACs, meme stocks, NFTs: the classic retail-mania markers all peaked in early 2021 and crashed hard, a full two years before the AI trade took over leadership. This cycle is institutional-led, carried by mega-cap balance sheets, sovereign funds and pensions. That does not make it correct. It makes the marginal buyer slower to panic, carrying less debt, and less likely to be margined out at the bottom.
67. Rates are not the pin this time. The Nasdaq peaked in March 2000, five months into a hiking cycle that took the Fed funds rate to 6.5 percent by year-end. Rising rates were the pin that popped the bubble: the future cash flows tech prices depended on got discounted into oblivion. Whatever is pressuring AI stocks in late 2026, it is not that dynamic. Policy rates sit far below 2000 levels, and the selloff's triggers were an earnings miss and an election, not the cost of money.
68. Cash-yielding assets correct. Story assets collapse. At 11 times forward earnings, Nvidia's earnings yield is around 9 percent. The hyperscalers throw off hundreds of billions in free cash flow even after capex. When an asset pays you substantially to own it, drawdowns attract buyers, including the companies' own buybacks. When an asset's only payoff is the narrative, as in 1999 or 2021, the floor is zero, because there is nothing to catch the fall. This market has a floor made of cash flows.
69. Even the real bubble kept its winners. Amazon fell 94 percent in the dot-com crash and then became a pillar of the global economy. The lesson of 2000 was never "the technology was fake." It was "the pricing and the paper were wrong for a while." And notice which companies survived: the ones with real revenue and customer lock-in. The AI era's incumbents start from where Amazon ended up, profitable and embedded, which is why a crash, if one comes, will not clear the field.
70. Every exponential looks like 1999 on a chart. Lay any compounding series on a log axis and you can make it rhyme with Cisco. Chart-shape arguments prove too much: run them in 1996 and you sell the internet before the boom, in 2010 and you miss mobile, in 2013 and you sit out a decade of S&P gains. The question is never whether the line curves up. It is whether there is cash flow underneath the line. Section 1 of this post is that cash flow.
| Line item | 2000 dot-com / telecom | 2026 AI buildout |
|---|---|---|
| Flagship stock | Cisco above 100× earnings, orders decelerating | Nvidia ~23× trailing, ~11× forward, revenue doubling |
| Who funded it | Debt: WorldCom $30B+, Global Crossing $12B | Operating cash flow, plus investment-grade paper and equity |
| Builder profitability | Carriers needed capital markets to survive | The big four earn hundreds of billions straight through the boom |
| Demand | About 3 percent of fiber lit when the crash came | Clouds sold out; Oracle RPO $664B, Google Cloud backlog $514B |
| Retail paper | 457 IPOs in 1999, one in four doubled on day one | Froth sits in private structured rounds, not brokerage accounts |
| The pin | Fed hiking to 6.5 percent | Rates far below that; triggers were an earnings miss and an election |
| Unit economics | Bandwidth cheapened after the bust, demand still years away | Intelligence roughly 10× cheaper per year, usage exploding as it falls |
8. The Bear Case, Taken Seriously (Reasons 71-80)
A hundred cheerleaders are worth less than one serious bear. These are the ten best arguments against everything above, and my honest answers.
71. "95 percent of enterprise AI pilots fail." The MIT-affiliated report everyone quoted in 2025 measured pilots, not products, and a high failure rate on corporate experiments is the base rate for every enterprise technology since ERP. Web properties in 1996 were overwhelmingly garbage too. The five percent that work, coding, support, ads, search, are carrying nine-figure and ten-figure revenue lines. The correct read is that enterprises are still in the experimentation phase of a real technology, which is an argument about 2027, not about the technology.
72. "Depreciation schedules hide the truth." The claim: six-year server lives understate how fast GPUs obsolesce, inflating earnings. The rebuttal is empirical rather than theoretical: hyperscalers extended useful lives based on observed failure and utilization data, and last-generation GPUs demonstrably keep earning in inference, with 2020-vintage A100s still running production traffic across the industry. There is a genuine debate about pace as the Blackwell-to-Rubin cadence accelerates. But an accounting timing disagreement is a valuation risk, not a fake-revenue scandal. Enron this is not.
73. "The deals are circular." Nvidia invests in OpenAI, which buys Nvidia chips, alongside deals with AMD, Broadcom and Oracle that all interlock. The pattern is real and it deserves scrutiny, because vendor financing greased 2000's collapse too. Scale is the answer: the circular slice is tens of billions against an annual buildout north of $700 billion funded overwhelmingly by cash-rich operators. In 2000, vendor financing was the funding. In 2026 it is garnish on a cash-funded meal.
74. "Private marks are absurd." OpenAI at reported marks approaching $850 billion, round after round at rising valuations. Agreed: private-market enthusiasm is the frothiest corner of this cycle. It is also exactly where froth does the least damage, in structured preferred paper with liquidation stacks and capped returns, held by investors who underwrote the downside in writing, which is the honest half of why AI startups continue attracting investment at marks like these; the other half is that the revenue lines being bought grew 70 to 100 percent this year. Private losses will hurt their holders. They do not leap through bank balance sheets the way public telecom debt did in 2002.
75. "OpenAI missed its own forecast by $20 billion." The strongest single datapoint of the fall, and I will not spin it. A $50 billion run rate growing around 70 percent annually, with the shortfall driven partly by consumer churn and compute supply, while the nearest competitor ran its pace up to $65 billion over the same period. Frontier-lab revenue in aggregate climbed the entire time. If OpenAI stalls and Anthropic stalls with it, this reason dies and I will say so publicly. Today it reads as share shift inside an exploding market.
76. "Margins are getting squeezed. Look at Alphabet." Alphabet's October miss was real: AI costs compressed profits, and the market removed half a trillion dollars of value in a day. Here is the other side of that same ledger: $514 billion of contracted cloud backlog growing double digits per quarter, cloud revenue up 82 percent, and margins that remain among the best in corporate history. Building annuities costs money up front, and depreciation is the accounting cost of a backlog converting to revenue. This is the utility-construction phase of a platform build, and it was always going to look exactly like this on the income statement.
77. "Data center debt is cracking at the edges." A bank shopping around $500 million of data center exposure, Australian operators under pressure, a central bank flagging the sector's financing. All true, all worth monitoring quarterly. And all of it is fringe-of-the-fringe relative to the core: hyperscaler balance sheets, investment-grade issuance, and $80 billion equity raises that got oversubscribed. In 2000 the cracks were in the core of the funding stack. In 2026 the core is overcapitalized and the cracks are at the edge. Different animal, different blast radius.
78. "Stargate UK got paused." It did, in April 2026, over energy costs and regulatory conditions. Also true in the same period: the US flagship campus is live, UAE capacity is energizing, the Malaysia anchor deal was signed, and independent tracking has the US sites above nine gigawatts by 2029. Every infrastructure program in history pauses its weakest sites and finishes its strongest. The transcontinental railroad and the Panama Canal both had bankruptcies in the middle of them. Construction reality is not demand fiction.
79. "The midterms mean regulation." Maybe, and siting rules, power interconnects and safety regimes will slow things at the margin. But the data-center states compete for these projects the way they compete for factories, the security establishment in Washington treats compute as a strategic asset to be expanded rather than taxed, and Europe regulates AI while pouring 200 billion euros into it. Regulation raises the cost of a buildout. It does not delete the demand behind it, and power demand did not drop the morning after the election.
80. "There is still no killer app." This one expired in 2025 and has not been renewed since. AI coding is a multi-billion-dollar product category with auditable revenue, and Claude Code alone reached a $2.5 billion annualized pace five months in, with rivals at similar scales. It is a whole aisle now, and we priced the main coding CLIs, Kimi's included, so you can see the menu for yourself. The honest critique in late 2026 is no longer "no killer app." It is "not every app is killer yet," which is a maturity complaint about the second wave, not a bubble indictment of the first.
9. The Flywheel Beyond Chatbots (Reasons 81-90)
The buildout is not just serving chat windows. It is dragging energy, chips, science, and labor markets into its orbit.
81. It ended two decades of flat electricity demand. US power consumption was essentially flat from the mid-2000s through 2020, and data centers flipped the curve. A reactor at Three Mile Island is being restarted for compute, small modular reactors are on order, gas turbines are booked toward 2028, and grid engineering stopped being a legacy career. Whatever happens to any single AI stock, the power infrastructure being built serves the economy for fifty years, and that is how AI infrastructure supports future growth in the most literal sense: substations, turbines and transmission that outlive any model cycle. That was the railroads' silver lining, and it is real here too.
82. Chips came home. TSMC runs fabs in Arizona, the advanced packaging chain is being duplicated onshore, and the whole effort enjoys support across both parties. Whatever happens to AI valuations, the reshoring of leading-edge semiconductor manufacturing is a decades-long industrial project with national security attached. It does not unwind because the Nasdaq has a bad quarter, which is precisely why it keeps getting funded through bad quarters.
83. Demographics make automation structural, not cyclical. Japan, much of Europe, and soon China are running out of working-age people. Care economies alone face shortages measured in millions of workers this decade. When the 2030s labor gap arrives in full, the question will not be whether to automate. It will be why anyone waited. Demand for automation that is driven by demographics does not care about stock prices, and it does not negotiate. That is why AI-powered automation will continue expanding whether the Nasdaq behaves or not: the workforce curve is the customer, and it is shrinking on schedule.
84. Defense is a non-discretionary customer. ISR data floods, drone autonomy, logistics under fire, targeting pipelines: modern defense has become an AI procurement story, and defense budgets are the most recession-proof spending lines on earth. Ukraine's battlefield made autonomous systems a standing requirement for every NATO procurement office. That demand does not participate in drawdowns. It is exempt from them by law.
85. Science is compounding on machine time. AI systems performed at gold-medal level at the International Mathematical Olympiad in 2025. One DeepMind materials model surfaced 2.2 million new inorganic crystal candidates, centuries of traditional discovery in a single run. AI-designed molecules are in clinics. Fusion experiments are steered by learned controllers. None of this is revenue yet. It is the raw material of the next fifty years of GDP, arriving weekly.
86. Pharma signed the milestone checks. Multi-billion-dollar AI drug discovery partnerships, Isomorphic's pharma deals, Recursion's pipeline, Insilico's trials, have put the most cash-rich industry in the world on the demand side of AI compute. Biopharma does not fund hype. It funds molecules, and it signs milestone schedules with its eyes open, because a molecule that survives Phase 2 is worth more than the entire compute bill.
87. The robots are in the warehouse, not the keynote. Amazon's million-plus robots, humanoid pilots at automakers and logistics operators, vision models migrating from screens into machines. Physical AI in 2026 is where chatbots were in 2023: early, uneven, obviously on a slope. The addressable market is the global wage bill, and the constraint is engineering time, which reason 88 is busy fixing.
88. AI is building AI. Models write a third of the code at Google and Microsoft. The labs use their own models to design experiments, build data pipelines, and accelerate chip design itself. The agent tooling joined in: our OpenCode with Ollama walkthrough has a coding loop you can run on a laptop with no API key. A technology that improves its own production function is the textbook definition of compounding, and it is the engine behind the second scaling axis in Section 6. Every efficiency gain the bears cite as deflationary for chip demand is simultaneously a speed-up for capability delivery. That loop is how AI innovation is changing the technology industry from the inside: the tool chain now improves the tool maker.
89. The education surplus is real and measured. A randomized Harvard physics course found students learned more with an AI tutor than in the active-learning control section. Tutoring at near-zero marginal cost is a civilizational upgrade for the majority of humanity that cannot afford a human tutor, and it is arriving through the cheapest distribution device ever built, the phone. Adoption there is bounded by access and electricity. It is not bounded by demand.
90. The cyber arms race made defensive AI mandatory. Attackers got AI, and phishing, deepfakes and automated exploitation got dramatically cheaper. This year's breach catalog makes the point without my help. The security industry, a market on the order of $200 billion a year, now carries an AI line item that is not optional, because the alternative is losing to adversaries who automated first. Fear-driven demand is the stickiest demand there is, and it renews every year as the attacks evolve.
10. The Long Game (Reasons 91-100)
Zoom out. The last ten reasons are about time horizons, the one dimension where every bubble comparison quietly cheats.
91. General-purpose technologies take decades, and that is the bull case. A general-purpose technology, in the economics sense, is one that reshapes every industry it touches, like steam, electricity, and computing. Electricity needed roughly forty years to show up in factory productivity, because factories had to be physically redesigned around the electric motor. The internet needed fifteen years to hit the productivity statistics. AI is three years old as a mass product, which is the entire answer to why the artificial intelligence revolution is still in its early stages: mass adoption is here, mass restructuring is not, and the base rate says that part takes another decade. "It has not transformed the economy yet" is not a bear argument. It is the standard timetable, and the capex is arriving early relative to it, not late.
92. Platform shifts arrive on a schedule, and AI is on it. Mainframes to PCs in the 1980s, PCs to web in the 1990s, web to mobile in the late 2000s, mobile to AI in the 2020s. Roughly every fifteen years the stack resets, and the decade after each reset minted the largest companies on earth. The middle game of every platform shift looks like this: heavy infrastructure spending, thin-looking applications, and endless essays about the promised revolution being late. This is what the middle game is supposed to look like, and it is why AI technology will continue growing straight through it: every prior platform compounded through its own "where are the apps" years.
93. IT spend absorbs every wave, then grows. Global IT spending runs on the order of five to six trillion dollars a year, and it has absorbed client-server, the web, cloud, and mobile, growing through each transition. AI is still a single-digit share of that envelope. The runway for AI inside existing enterprise budgets is measured in trillions before anyone invents a new budget line, and history says the envelope itself expands with each wave rather than getting cannibalized. That envelope is the long-term growth potential of artificial intelligence in one sentence: trillions in existing spend, a technology still in single digits inside it.
94. Boards remember the cloud decade. The penalty for under-investing in cloud, ask IBM or Oracle about their 2013-era positioning, was a lost generation of enterprise relevance. Every board internalized that lesson. The downside of over-investing is a write-down and an awkward earnings call. The downside of under-investing is irrelevance. That regret asymmetry, not hype, is what keeps capex lines growing straight through selloffs, and it is rational. Keep it in mind whenever you read AI investment trends and future opportunities pieces, including this one: the check signers are optimizing for survival, not narrative.
95. Compute policy survives elections. Export controls, chip funding, research budgets: both US parties compete on AI dominance, and the direction has been constant across administrations, treat compute as strategic. Europe regulates the technology while investing 200 billion euros in it. You can argue about the wisdom of any particular policy, and I often do. But there is no electoral scenario in any major economy where the winning platform is "stop building compute."
96. China guarantees the floor. DeepSeek, Qwen, and a domestic accelerator program mean the compute race is two-sided, and neither side can afford to stop, for the same reason neither superpower stopped launching rockets after failures. Great-power competition is a demand floor with a defense budget attached. It is the least sentimental buyer in the market, and it does not read bubble essays.
97. Compliance is now a permanent line item. The EU AI Act, audit regimes, safety evaluations, model cards, red-teaming requirements: an entire compliance industry is forming around AI the way one formed around privacy after GDPR. Compliance spend is non-discretionary, boring, and utterly immune to hype cycles. An industry that regulators are building a permanent framework around is an institutionalized industry. Fads do not get their own statute books.
98. Human capital keeps compounding. AI research output has been roughly doubling every couple of years, and the field is absorbing mathematicians, physicists, and biologists at rates the profession has never seen. Talent flows are the slowest-moving and most predictive variable in technology. There is no historical scenario in which hundreds of thousands of the world's most technical people pile into a field for a decade and it produces nothing.
99. The stress tests keep getting passed. DeepSeek day: $600 billion of market value gone in a session, mostly recovered within weeks. The spring 2025 tariff shock: Nvidia down roughly 40 percent, at new highs by summer. March 2026: multiple compression, earnings kept arriving. And now November: a $3.2 trillion day. Each shock crushed prices. None has dented the revenue, the tokens, or the backlogs. Falling multiples alongside rising fundamentals is the market squeezing froth out while the underlying compounds, which is the opposite of 2000, where the fundamentals were the froth.
100. The meta-reason: you cannot burst what has already been delivered. Bubbles pop when the promised cash flows turn out to be fictional: the revenue was projected, the users were bought, the demand lived in a spreadsheet. Here the revenue is billed, the tokens are metered, the backlog is signed, the rides are ridden, the molecules are in trials, the code is in production. A crash can still take valuations down a lot, because that is what markets do. But "burst" means the underlying was hollow, and three years in, the receipts are printed. The 2000 crash revealed the internet's revenue was mostly a promise. A 2026 crash would reveal what, exactly? That we only built seven gigawatts instead of twelve?
The AI Bubble FAQ: 9 Questions
The questions readers and search boxes keep asking, with reason numbers where the long answers live.
Will the AI bubble burst in 2026 or 2027?
On the evidence above, no. A burst happens when the market learns the cash flows were fiction, and here the cash flows are billed, metered and contracted, and they grew straight through every selloff of the last three years. What 2026 and 2027 will contain is more volatility: $3.2 trillion days, missed forecasts, punished stocks. That is a correction doing its job. The five tripwires at the bottom of this post are the specific things that would flip my answer, and none is close to firing as I type this.
Is artificial intelligence a bubble or a revolution?
A false choice, but if you force it: a revolution wearing a bubble's price tag at the edges. The core buildout, meaning the chips, the clouds and the models with real revenue, is financed by cash and billed to customers. The froth, meaning some private marks and fringe data center credit, is real and contained. The railway age had manias in it and still rewired the economy. The honest read in 2026: the technology is compounding like a revolution, parts of the paper trade like a bubble, and those are not the same thing, which is the entire reason this post keeps them separate.
What makes the AI boom different from the dot-com bubble?
Four line items, all checkable in the Section 7 scorecard. Funding: operating cash from the most profitable companies in history, versus borrowed money. Demand: clouds sold out with $664 billion of Oracle RPO on the books, versus about 3 percent of the country's fiber actually lit. The flagship stock: Nvidia near 23 times earnings and doubling revenue, versus Cisco above 100 times with orders decelerating. Monetization: five working revenue rails versus one unproven banner-ad bet. The chart shape rhymes. The balance sheets do not.
Is the AI industry sustainable in the long run?
The demand side is structural: shrinking workforces, token consumption up roughly 7x in a year, sovereigns treating compute as a strategic asset, defense treating it as mandatory. The supply side is booked years ahead: turbines toward 2028, high-bandwidth memory sold out, advanced packaging allocation-constrained. The honest risks sit at the edges, data center credit and private marks, and they are worth watching quarterly, which is what the tripwires are for. Sustainable does not mean smooth. It means the bills keep getting paid, and they are.
Is artificial intelligence investment sustainable?
For the industry, the previous answer. For your portfolio, the question is different and the answer is more boring: valuations can compress a long way without the technology losing. Nvidia spent all of 2026 de-rating to a seven-year-low multiple while its revenue doubled. If you cannot stomach 40 percent drawdowns in the builders, the index-fund math from our Money School series is the sane way to own the theme. The buildout does not need your conviction. Your plan needs your discipline.
Why AI demand could remain strong for years
Because the biggest drivers are not sentiment. Inference is now the majority of compute, and live traffic renews daily (reason 56). The installed GPU base enters a five-to-six-year refresh cycle worth hundreds of billions a year before any new workload exists (reason 58). Demographics force automation regardless of stock prices (reason 83). Sovereigns buy compute the way they buy oil reserves (reason 47). And every price cut so far has been answered by a volume explosion, sevenfold on Google's token curve alone. Demand built on contracts, demographics and elasticity does not expire with a headline.
What will drive the next phase of AI growth?
Agents with receipts: coding agents are already a multi-billion-dollar category, with Claude Code at a $2.5 billion pace five months in and rivals at similar scale. Agentic commerce, where models transact instead of chat. Physical AI: a million-plus robots at Amazon, humanoid pilots at the automakers, vision models moving into machines. And science: AlphaFold's structures, 2.2 million material candidates, AI-designed molecules already in trials. The artificial intelligence business opportunities in 2026 cluster where the last wave proved the money: code, support, ads and search. The next wave builds on intelligence that gets cheaper every quarter.
Could artificial intelligence become the next industrial revolution?
The pattern fits so far. General-purpose technologies, steam, electricity, computing, reshaped every industry they touched, took decades to do it, and were called bubbles repeatedly along the way. Electricity needed roughly forty years to show up in factory productivity; the internet needed fifteen to reach the productivity statistics. AI is three years old as a mass product. If this is an industrial revolution, we are in year three of forty, which is the actual bull case, and the reason "it has not transformed everything yet" is a timetable observation rather than a verdict.
What is the future outlook for the global artificial intelligence market?
Every serious number points one way. Global IT spending runs five to six trillion a year, with AI still a single-digit share of it. Goldman's forecast has monthly token processing climbing another order of magnitude by 2028. The contracted backlogs at Oracle and Google Cloud together approach $1.2 trillion. Artificial intelligence market growth prospects are not a vibes question; they are a backlog-and-elasticity question, and both inputs are rising. Expect the path to stay loud, corrections, misses and failures included, because that is what scaling looks like from the inside.
The Verdict, and What Would Change My Mind
That is the case. The revenue is real and compounding. The funding comes from the most profitable balance sheets in history, topped up by investors who underwrote the downside in writing. The demand is contracted, scarce-supplied, and partially prepaid. The unit economics improve every quarter, and usage explodes every time prices fall. The dot-com analogy fails on every line item that matters, and the best bear arguments, the $20 billion miss, the margin squeeze, the private marks, the fringe debt, are arguments about pace and price. Not one of them is an argument that the cash flows are fiction. If you want the one-line version of why artificial intelligence is here to stay: it already survived three years of stress tests with its revenue curve intact.
None of this means AI stocks are cheap, or that November 5 was the bottom. Valuations can compress a long way without breaking the thesis: Nvidia's multiple spent all of 2026 compressing to seven-year lows while its revenue doubled, which is the opposite of a mania's mechanics. If you are trading weeks, none of these hundred reasons matter. If you are investing years, the boring answer in our index fund lesson still beats stock-picking heroics. And if you are asking the structural question, whether this is 2000 redux, a hollow buildout destined to vaporize and take the infrastructure with it, the answer keeps coming back no, in a hundred checkable ways.
And because a post like this is worthless without a falsification clause, here is mine. Not "the Nasdaq falls 20 percent," because that is a price, and prices are the one thing corrections and bubbles have in common. Watch these five instead:
As I write this, none of those five is close to triggering. The capex line keeps being raised, twice at Alphabet this year alone, with Meta lifting its plan to $125-145 billion and Amazon to around $200 billion. Google's token volume grew 7x last year. The frontier labs are growing 70 to 100 percent. The clouds are supply-constrained. Renewals expand. Pause for a second and think about the Companies laying off employees in the name of AI? do you think they might have done calculation wrong! You know the answer and they know the answer!
The bubble crowd has a chart that looks like a chart. This side has the cash flows. One of those ages better than the other, and if you have read this far you already know which one I am betting on. Check every number in this post, that is what they are there for, and if one of my five tripwires trips, I will be one of the first to write the obituary. The future of the artificial intelligence industry will contain drawdowns, because everything does. It will also keep billing, building and compounding, because that is what the last three years show. Until then: correction, yes. Burst, no. The buildout pays its own bills, and that has never once been true of a bubble on the day it popped. Don't Panic due to headlines! As always think !!