Are We in Another AI Bubble? A Beginner's Guide to the Debate

Update (September 8, 2026): Nvidia reported Q2 FY27 earnings on August 26, delivering $96.2 billion in total revenue and $89.0 billion in Data Center revenue (up 117% year-over-year), beating its $91 billion guidance. The S&P 500 closed at 7,718.60 on September 4, roughly flat from the August 31 close of 7,686. Palantir stock traded around $174-182 in early September, consistent with the mid-August range following its Q2 earnings beat. Burry expanded his AI shorts through late August, adding CoreWeave and increasing positions in Palantir, Micron, SOXX, and Oracle, with his short stock holdings now exceeding 21% of his portfolio per his August 26 Substack post. Anthropic's IPO timeline has shifted to mid-October at the earliest, with Reuters reporting on September 5 that the prospectus is now expected in late September rather than early September. The debate continues.

Before the long version, here's the debate at a glance, feel free to read this and skip ahead, or keep going for the named voices, current 2026 data, and the dot-com comparison.

What bears say: AI spending (over $700 billion in 2026 from four companies) has become detached from the cash AI is actually producing. The roughly $12 billion in consumer AI revenue today doesn't justify the buildout. (Voices: Burry, Grantham, Sequoia's Cahn, Zitron, Dalio, Sorkin's 1929 as historical mirror.)

What bulls say: That spending is matched by rapidly growing cloud-AI revenue (Azure AI at a $37B run-rate, +123% year-over-year; Google Cloud roughly $80B annualized at +63%), funded by profitable companies, and just the early innings of a real revolution. (Voices: Nadella, Pichai, Amodei, Morgan Stanley research.)

What both sides actually agree on:

• AI itself is real and transformative.
• The magnitude of hyperscaler spending is unprecedented.
• ROI timing (how long until revenue catches up to capex) is genuinely uncertain.

What separates them: Bears think current AI stock prices don't yet reflect the revenue-catchup risk. Bulls think they already do. That's the entire fight in one sentence.

The surprising thing while researching this piece: bulls and bears were almost always looking at the same numbers (Nvidia's $75 billion quarter, the $700 billion capex figure, the 35% Mag-7 concentration) and reaching opposite conclusions about what those numbers will mean three years from now. Neither side can prove what happens next; they're arguing about the consequences of the same facts, not the facts themselves.

The rest of this article walks through each side carefully (what they actually argue, with current 2026 data and specific named voices) so you can decide for yourself whose framing you find more convincing.

Where the debate stands, late July 2026: Bears have sharpened. Michael Burry disclosed fresh short positions across Nvidia, Tesla, Micron and the SOXX semiconductor ETF, warning of a possible drawdown of around 30 percent, and the Magnificent 7 shed roughly $2.2 trillion of market value in June. Bulls still point to Nvidia's audited $75.2 billion data-center quarter, note the stock is up roughly 13 percent year to date, and cite that about 95 percent of covering analysts rate it a "Strong Buy." Palantir reported Q2 2026 earnings on August 4, 2026 showing revenue of approximately $1.81 billion (up roughly 81 percent year-over-year) and strong U.S. commercial growth of 150 percent, and the stock rebounded sharply, trading near $172-174 in mid-August after the earnings beat, and it still trades near 67 times trailing-12-month sales, a price-to-sales level that history has found hard to sustain above roughly 30 times. A recent Nvidia partnership added about $23 billion to Palantir's value. Historical data shown; past performance does not indicate future results.

If you've been online recently, you've seen the headlines. The New York Review of Books ran a long essay in March 2026 about Andrew Ross Sorkin's book 1929, drawing parallels between the run-up to the Great Crash and today's AI boom. Burry's Substack (yes, Substack; we'll get to that) has been calling the market overheated for months. Bloomberg and Fortune publish a new "AI stock bubble" piece about every other week. The question keeps coming back because the underlying AI stocks keep going up.

There are real numbers behind the chatter. The S&P 500 reached a record high of 7,798.99 on August 13, 2026, then pulled back to around 7,652 on August 24 and closed at 7,686 on August 31. The seven largest U.S. stocks, known as the "Magnificent 7" (Apple, Microsoft, Alphabet, Amazon, Meta, Nvidia, and Tesla), now make up roughly 35% of the entire S&P 500's market capitalization. That's up from about 12.5% a decade ago.

Then there's the spending. The four largest cloud companies, Microsoft, Alphabet/Google, Amazon, and Meta, collectively guided to roughly $700–725 billion in capital expenditure for 2026, almost all of it on AI infrastructure: data centers, Nvidia GPUs, networking, and the electricity to power them. That's up roughly 77% from 2025, when the same four spent around $410 billion. We have a glossary entry for capex if the term is new.

And Nvidia, the company that sells most of the chips inside those data centers, reported on May 20, 2026 that its Q1 FY27 revenue hit $81.6 billion in a single quarter, with $75.2 billion of that coming from its Data Center segment alone, up 92% year-over-year. Our earlier breakdown of Nvidia and Palantir covers how the company got here.

So: an unprecedented chunk of the stock market is concentrated in a handful of names, those names are spending unprecedented amounts on a single technology, and serious investors are publicly arguing about whether the whole picture is brilliant or about to break. That's why the question is everywhere.

Before we wade into who's right, it helps to know what we're arguing about. "Bubble" gets thrown around so loosely that most people use it to mean "prices went up a lot." That's not what it means in finance.

A bubble is a specific pattern. It's when asset prices become disconnected from the cash those assets are likely to produce, when prices are sustained mostly by narrative and the expectation that someone else will pay more for the same asset tomorrow, rather than by valuation tied to underlying earnings, revenue, or cash flow. If you'd like to try one on a real company, our P/E ratio calculator is a quick way to compare a stock's price to its earnings.

The economist Hyman Minsky described five stages every bubble tends to move through. In plain English:

1. Displacement: a real innovation appears that genuinely changes the economy (the internet, the railroads, electricity… AI).
2. Boom: prices start rising as investors recognize the innovation's importance. This part is usually rational.
3. Euphoria: prices keep rising, but now the rise is driven less by fundamentals and more by the story. Late-stage investors enter believing prices will keep rising because they have been.
4. Distress: early skeptics or smart-money insiders start to sell. Volatility increases. The narrative starts to fray.
5. Revulsion: prices collapse, often well below their pre-boom levels. Confidence shatters. The narrative is rewritten.

For the AI debate: by this definition, asking whether AI stocks are overvalued is partly a question about what investors are buying and why, not just about what prices are doing.

• If you believe Nvidia's $75 billion of quarterly Data Center revenue is real and growing fast, and that hyperscaler capex will generate cash flows that justify today's prices, you're saying we're in the boom phase, rational, productive, fundamentals-driven.
• If you believe the prices already imply revenue that hasn't shown up yet (and may never show up at the levels assumed), you're saying we've drifted into euphoria, narrative-driven, vulnerable.

The bears and bulls in the next two sections aren't really arguing about whether prices are high. They're arguing about which Minsky stage we're in.

Multiple investors and analysts have reached similar bear conclusions from different angles. Here's the strongest version of what they collectively argue, in their own words where possible.

Michael Burry is the most famous bear of the moment. The investor who shorted the 2008 housing market shut down his fund Scion Asset Management in late 2025 and now publishes warnings on his Substack "Cassandra Unchained," which has more than 200,000 subscribers as of mid-2026. His final Scion 13F filing in late 2025 showed large notional short positions on Palantir, Nvidia, Oracle, the SOXX semiconductor ETF, and the QQQ Nasdaq ETF, mostly through long-dated put options expiring in 2027. In May 2026 he declared the market had "jumped the shark" and added a leveraged short via January 2027 SOXX puts. Burry has publicly argued Palantir is significantly overvalued; we won't reproduce specific price targets here because they're his opinion, not StockCram's recommendation.

Jeremy Grantham, co-founder of asset manager GMO and the long-running superbubble watcher, co-wrote a paper in January 2026 titled "Valuing AI: Extreme Bubble, New Golden Era, or Both?" His verdict, paraphrased: mostly the first. The paper cited OpenAI's 2025 numbers (roughly $12 billion in revenue against $8 billion in operating losses) and projected losses scaling to $35 billion by 2027. In an April 2026 ThinkAdvisor interview, Grantham assigned what he described as "slim to none" odds that the AI bubble doesn't burst. That's one analyst's view, and a famous bear's view at that; we cite it as a data point in the debate, not a forecast we endorse.

Sequoia Capital's David Cahn has been writing for over a year about what he calls the AI "revenue gap." His updated 2026 estimate: roughly $600 billion of annual AI revenue would need to materialize to justify the current pace of capex, and the gap is widening, not closing. Sequoia is a Silicon Valley venture firm with deep skin in the AI game, so the fact that they're flagging the AI capex bubble matters.

Ed Zitron, an independent industry analyst whose newsletter Where's Your Ed At has become widely read in finance and tech circles, published "What If We're In An AI Bubble? (Part 2)" on May 22, 2026. His sharp data point: roughly $178.5 billion of U.S. data-center debt is currently financing the buildout, often through private credit. If AI revenue lags, that debt becomes a problem far beyond the tech sector.

Ray Dalio, the founder of Bridgewater Associates, joined the bear chorus publicly on Bloomberg TV on June 3, 2026. He said AI shows bubble signs that "will eventually burst" and that his bubble indicators are at "1929 and 2000 levels", a direct echo of the Sorkin parallel. Bridgewater pegs Mag-4 2026 capex at around $650 billion (slightly below the $700B+ figure the broader counts produce). Like Burry and Grantham, Dalio isn't predicting timing; he's describing what he sees.

And finally, Andrew Ross Sorkin's book 1929 (Viking, October 2025) has become a touchstone for bears not because Sorkin himself predicts a crash but because the book documents how confidently the smartest people of 1928 dismissed bubble warnings, until they couldn't. Sorkin's recurring observation, paraphrased: at some point in any mania, the math may stop adding up.

Bears don't claim AI is fake. They claim the spending is wildly ahead of the cash AI is producing today, and history suggests that gap has to close.

If bears worry about the gap between spend and revenue, bulls answer: the spend is funded by profit, not debt; the revenue is arriving fast and from sources you can verify; and the technology is still in early innings. Here are their voices.

Satya Nadella, CEO of Microsoft, has had to defend his company's roughly $190 billion 2026 capex budget repeatedly. At Morgan Stanley's TMT conference earlier this year, he argued that software-driven improvements in total cost of ownership and server utilization will produce real returns on invested capital, not in some distant future but throughout 2026 and 2027. He also said something that doesn't usually come from CEOs justifying massive spend: Microsoft has been capacity-constrained through 2026. Demand for Azure AI services exceeds supply. The Azure AI run-rate alone is reported at roughly $37 billion, up 123% year-over-year.

Sundar Pichai, CEO of Alphabet, raised Google's 2026 capex guidance on the April 29 Q1 earnings call to between $180 and $190 billion. He told staff that supply constraints, not demand worries, "keep me up at night." Google Cloud, which carries much of Alphabet's AI investing exposure on the revenue side, grew 63% year-over-year to a roughly $80 billion annualized run-rate.

Dario Amodei, CEO of Anthropic (one of the leading AI model labs), said at a Morgan Stanley conference in February 2026 that the scaling laws (the relationship between model size, training compute, and capability) "have not hit a wall." He expects "radical acceleration" in 2026. Anthropic's confidential IPO filing on June 1, 2026 put its current run-rate near $47 billion at a roughly $965 billion valuation, substantially above the $30 billion figure cited earlier in 2026, and an order of magnitude above 2025 levels. By mid-August 2026, investor speculation (per Financial Times reporting citing six backers) pointed to a potential $2 trillion IPO valuation for a September or early October listing, though the company has not confirmed this target. Crucially, Amodei has also been refreshingly honest about the risk: "If I'm just off by a year in that rate of growth… then you go bankrupt." That's a bull acknowledging the bear case, useful for understanding why even AI insiders take the question seriously.

Morgan Stanley's equity research team has watched its own framing evolve. The earlier line, that bubble fears were "misplaced or premature", has shifted in 2026 to something more careful: "We don't see this risk as a 2026 story, but vigilance is a 2026 responsibility." Their analysts estimate roughly $3 trillion of total data-center spend across 2025–2028, with about half flowing through private credit markets. They flag data-center debt potentially exceeding $1 trillion by 2028, which is closer to the bear case than the 2024-era "don't worry about it" framing.

And then there's the profitability argument that bulls cite most often: unlike the dot-com era, today's index leaders are profitable. The Mag-7 generate substantial free cash flow and fund AI capex from earnings rather than debt. That profitability gap is the strongest single difference between 2000 and 2026, examined more closely below. The bull framing of the valuation story: today's prices reflect earnings that exist, not earnings that might.

Bulls also push back on the dot-com comparison with what we'll call the Cisco caveat: Cisco was a real, profitable company at the center of a real internet revolution, and its stock still fell about 85% from the 2000 peak. Bulls' point: yes, that's true. But Cisco's P/E ratio at the peak was in the triple digits. Today's Nvidia, even after the run, isn't priced like that. Different fundamentals can justify different multiples.

Bulls don't claim capex doesn't matter. They claim the cash flows backing it are real, growing fast, and (mostly) funded from profits rather than debt.

If there's one ticker symbol the AI stocks debate revolves around, it's NVDA. Nvidia is the largest single beneficiary of hyperscaler capex (it sells most of the GPUs that go into the data centers) and so the "Nvidia bubble" question is, in many ways, the broader AI stocks question in miniature.

The bull case for Nvidia. The numbers are extraordinary and they're audited. Q1 FY27 (reported May 20, 2026): $81.6 billion in total revenue, $75.2 billion of that from the Data Center segment alone, +92% year-over-year. That single quarter generated more Data Center revenue than the entire company produced in all of fiscal 2024. The company is profitable; gross margins remain north of 70%; cash from operations is enormous. Beyond raw numbers, Nvidia has a software moat in CUDA (the developer ecosystem that makes its chips the default choice for AI training) which competitors haven't been able to crack despite years of trying. Bulls argue that even a "normalized" Nvidia, growing at half its current pace, would justify the current valuation.

The bear case for Nvidia. Three things make bears nervous, and only one of them is the P/E ratio.

Customer concentration is one concern. Roughly half of Nvidia's Data Center revenue comes from a handful of hyperscaler customers: Microsoft, Google, Amazon, Meta, and Oracle. If any one of them slows orders, the revenue line moves immediately. Bears point out that this is exactly the dynamic that hit Cisco in 2000: when the telecom buildout paused, Cisco's revenue collapsed even though the underlying technology kept growing.

Circular financing makes bears uneasy. Nvidia has disclosed substantial investments in OpenAI and Anthropic during recent funding rounds (the precise dollar figures vary by source), and some of that capital flows back to Nvidia as GPU purchases. The arrangement is legal and not unprecedented in tech (Microsoft does something similar with OpenAI), but it makes bears uneasy: the revenue isn't fully "organic" if a chunk of the cash funding the customers comes from the supplier.

The inference shift is another risk. Stanley Druckenmiller's recent positioning is instructive: he exited Nvidia in late 2024 (a move he later publicly called a "big mistake") and has rebuilt his AI exposure through Broadcom, Intel, and Arm Holdings. His thesis: the next leg of AI spend will be dominated by inference (running AI models for users) rather than training (building the models). Inference workloads may favor different chips and architectures. If Druckenmiller is right, Nvidia's revenue mix has to evolve, and there's execution risk in that pivot.

The honest take on Nvidia specifically: at $75 billion in a single quarter of Data Center revenue, you cannot reasonably claim Nvidia is a fake business. The bear case requires belief that the growth rate breaks, not that the business does. The bull case requires belief that hyperscaler customers stay patient and that the inference transition doesn't open the door to credible competitors. Both are arguable. Neither is obviously right.

The other name that comes up constantly in any AI investing conversation is Palantir Technologies (PLTR). If Nvidia is the picks-and-shovels play, Palantir is the software-layer play: the company selling AI-powered analytics and decision-making platforms to governments and large enterprises. And it is, by some measures, the most expensive large-cap stock in the U.S. market.

The bull case for Palantir. Real revenue, growing fast. The company is profitable. Its AIP (Artificial Intelligence Platform) product has been adopted by a growing roster of large commercial customers in addition to its long-standing government work. U.S. commercial revenue growth has remained in the high double digits year-over-year for multiple recent quarters, well above the company's pre-AIP baseline. Total company revenue reached $1.63 billion in Q1 2026, up about 84.7 percent year-over-year. The bulls' argument: Palantir is the rare software company that has actually figured out how to monetize AI at the application layer, while most others are still in pilot programs. If you believe enterprise AI software becomes a category as large as cloud infrastructure, Palantir has positioned itself well.

The bear case for Palantir. Almost entirely about the P/E ratio and, just as much, the price-to-sales ratio. Even bulls acknowledge Palantir's valuation is rich: the stock trades near 67 times trailing-12-month sales, a level that history has found hard to sustain above roughly 30 times, and it was down about 29 percent year to date through late July 2026, but rebounded sharply following a strong Q2 2026 earnings report on August 4. A recent Nvidia partnership added about $23 billion to its value. Burry's Substack has been particularly focused on Palantir; his final Scion 13F before winding down showed Palantir as the largest single short position by notional value, and he has continued to maintain and expand bearish AI positions through early August 2026, including fresh Oracle and Nebius shorts disclosed August 6-7. We won't reproduce his price target because that's his opinion, not ours, but it's substantially below the current price. The bear case in three words: growth-priced-to-perfection. Bears argue that even if Palantir continues to execute, the current price already reflects several years of flawless growth, leaving no margin for any stumble.

The Nvidia/Palantir comparison. It's useful to see why these two stocks are bucketed together when they're really different bets:

• Nvidia bears worry about customer concentration and the inference transition. The business is enormous and profitable; the question is durability of the growth rate.
• Palantir bears worry primarily about multiple compression. The business is smaller but still growing fast; the question is what investors are willing to pay for that growth.

For our broader breakdown of Palantir and Nvidia's recent performance, the existing article has more company-specific detail. The takeaway here: "AI stocks overvalued" doesn't mean the same thing for these two names, and treating them as one trade is a mistake even Burry doesn't make.

Every AI stock bubble debate ends here, because the dot-com bubble is the only modern bubble most readers have a vivid mental image of. So let's compare them honestly: what's the same, what's different, and what each comparison actually implies.

The similarities are real and not flattering. Like dot-com, today's market is narrative-driven and top-heavy. A single technology story is justifying valuations across a wide swath of stocks. The press is breathless. The IPO market is reopening for AI-adjacent names. Smart people are loudly disagreeing in public. Most importantly: there's an undeniable real underlying revolution. The internet was real in 1999; AI is real in 2026. That part isn't in dispute.

The differences are also real, and they matter:

• Index leaders are profitable. All seven of the Magnificent 7 generate substantial free cash flow today. Many dot-com leaders did not. Some were losing money quarterly.
• Capex is funded by earnings, not debt. Microsoft, Google, Amazon, and Meta are spending hundreds of billions on AI infrastructure largely from their own profits. There's a wrinkle: private credit and data-center debt are growing fast as a side channel, and Morgan Stanley estimates this could exceed $1 trillion by 2028. But the core capex is still earnings-funded.
• Valuation multiples are lower. The Nasdaq's P/E at the March 2000 peak was somewhere around 175×. Today it's roughly 32×. That's not cheap, but it's not the same animal.
• Revenue at the top is verifiable. Nvidia's $75 billion of Data Center revenue in a single quarter shows up in audited financials. Many dot-com companies had revenue that was, charitably, speculative.
• Speculative IPOs are limited so far. The dot-com era featured no-revenue companies IPO'ing at billion-dollar valuations. So far in 2026 the AI market is still dominated by incumbents, not pre-revenue startups.

The table below summarizes the comparison.

The honest caveat that bulls and bears both share: differences from dot-com don't mean immunity. Cisco was profitable in March 2000 and still fell about 85% from its peak. It took 24 years for the stock to recover its dot-com high. A stock can be backed by a real revolution, generate real cash, and still be priced for an outcome that doesn't arrive. Different fundamentals shift the odds; they don't eliminate the possibility.

Most bubble articles describe the bear case in vibes. We'll do it in specifics. Bears are watching for specific events that would signal the AI capex bubble has started to deflate.

1. Hyperscaler capex cuts. This is the single biggest signal. If Microsoft, Google, Amazon, or Meta lowers its capex guidance, not because of a one-time write-down but because the company itself sees lower ROI, bears would treat that as game-on. So far, every hyperscaler guidance revision in 2026 has gone up. Bears watch the quarterly earnings calls of all four carefully for the first downward revision.

A related early-warning signal is downstream chip demand. On June 3, 2026, Broadcom posted a record Q2 ($10.8B AI semiconductor revenue, +48% YoY) but guided Q3 AI revenue to $16B against a $17.2B consensus, a textbook "beat-and-miss" that sent AVGO down roughly 14% and dragged Nvidia, AMD, and Intel in sympathy. One quarter doesn't make a trend, but this is exactly the kind of "expectations have outrun reality" moment bears watch for. Nvidia reports Q2 FY27 earnings on August 26, 2026 against guidance of approximately $91 billion in revenue, and any AI-revenue qualitative commentary from that report and the next round of hyperscaler earnings will be more telling than any single number.

2. Cloud-AI revenue growth slowing materially. Azure AI is at a $37B run-rate growing 123% YoY. Google Cloud is ~$80B annualized at 63% growth. AWS is the largest at ~$150B annualized at 28% growth. Bears watch for deceleration. Azure AI dropping below 50% year-over-year growth, or Google Cloud below 30%, would signal that enterprise AI demand isn't keeping pace with infrastructure spend.

3. OpenAI or Anthropic stumbling. The two leading independent labs are increasingly central to the AI revenue story. If OpenAI's reported revenue trajectory breaks, particularly if it can't justify another massive funding round, that would undercut the entire framing of "the revenue is coming." Anthropic's $80–100B run-rate target by year-end is a number bears will scrutinize closely against actual reported figures.

4. Data-center financing stress. The Ed Zitron and Morgan Stanley framing both center on private credit funding the data-center buildout. Credit spreads on data-center-backed debt widening, defaults at smaller data-center operators, or institutional investors publicly walking away from the asset class would all signal financing stress. Any of those would suggest the financing assumptions under the buildout are getting tested.

5. Better, cheaper models. A "DeepSeek moment" 2.0, a breakthrough that delivers competitive AI capabilities with materially less compute, would directly undercut the case for $700B/yr of GPU spend. Any open-source release or architectural advance that suggests scaling laws can be sidestepped would directly undercut the case for $700B/yr of GPU spend. (This is the risk Amodei himself flagged with his "if I'm off by a year" quote.)

6. Power and supply constraints binding. Even bulls acknowledge electricity is the new bottleneck. If new data centers can't get grid hookups, or if power prices spike enough to compress hyperscaler margins, the capex math gets harder. Utility filings, transmission queue data, and any announcement of capex deferral due to power constraints would all signal binding supply limits.

7. Regulatory action. Either antitrust action against any of the Mag-7 (limiting their ability to deploy AI vertically) or AI-safety regulation that significantly constrains deployment timelines could change the ROI picture. EU AI Act enforcement, U.S. state-level AI laws, and any FTC/DOJ activity around big-tech AI integration could all change the ROI picture.

One-quarter Nvidia revenue misses, single hedge fund losses, or individual Tweets from famous bears don't make this list. Structural signals matter more than headlines.

Bulls are watching for specific events that would tell them the AI investing thesis is playing out and the bear case is fading.

1. Enterprise AI adoption accelerating with measurable productivity gains. The bear-side critique that "90% of firms report no productivity impact from AI" (a February 2026 study cited by Grantham) is the most damaging single data point in the bear case. Bulls are watching for that to flip, specifically large-cap corporate reports starting to attribute concrete revenue or cost savings to AI tools. The first earnings call where a non-tech S&P 500 company says "AI delivered X basis points of margin expansion" would be a major signal.

2. AI agents going mainstream. The leap from chatbots to autonomous AI agents that handle real workflows (booking travel, processing claims, writing code end-to-end) is the bull case's killer-app moment. Bulls are watching for measurable agent-driven workflow adoption inside Fortune 500 companies. If agents become as common in 2026–27 as chatbots became in 2024, that would significantly expand the addressable revenue base.

3. The $700B vs. $12B revenue gap closing. Bulls don't need the gap to close immediately. They need it to start closing measurably. The headline gap is mostly consumer AI revenue. Cloud-AI revenue (Azure AI, Google Cloud's AI services, AWS's AI offerings) is rapidly approaching combined $200+ billion annualized. If that line shows another year of 50%+ growth, the gap narrows fast.

4. Margins improving on AI workloads. A recurring critique of cloud-AI economics is that it's harder to make high gross margins on GPU-heavy workloads than on traditional cloud. If Microsoft and Google start showing that AI is actually accretive to gross margin, not just additive to revenue, that would defuse one of the most credible bear concerns. Watch the cloud segment margin lines on Q1 and Q3 2026 reports.

5. New AI product categories emerging. Beyond chatbots and code completion, bulls are looking for genuine new product categories: vertical AI for healthcare diagnostics, AI-driven scientific discovery, consumer hardware integration. The bull case strengthens dramatically if AI starts producing categories of value that didn't exist before, not just better versions of existing software.

6. Scaling laws continuing to hold. Amodei's framing is that scaling laws have not hit a wall. If the next generation of frontier models (GPT-5 / Claude 4 / Gemini 2.5 successors) shows clear capability jumps proportional to compute increases, the bull case for continued infrastructure spend strengthens. If those models disappoint, the case weakens.

7. Geopolitical AI competition driving more, not less, spend. The U.S.-China AI race has so far been a tailwind for hyperscaler capex. If U.S. policy continues to view AI infrastructure as a strategic priority, and if European and Asian competitors accelerate their own spend, that creates a competitive dynamic that supports the bull case regardless of near-term ROI.

Validators on both sides take time to play out. The bear case wouldn't be fully validated by a single bad quarter, and the bull case wouldn't be fully validated by a single strong product launch. The market spends most of its time in the in-between, which is exactly why the debate stays alive.

Reading the bull and bear research side by side for this article, one observation kept jumping out: both sides cite the same numbers (Nvidia's $75 billion quarter, the $700 billion capex figure, the 35% Mag-7 concentration) and reach opposite conclusions about what those numbers mean. That's the unusual feature of this debate, and it's worth its own section.

If you read Morgan Stanley's 2026 update, Grantham's GMO paper, Burry's Substack, and Nadella's earnings calls in sequence, three areas of quiet agreement become obvious.

1. Concentration risk is real. When seven companies make up ~35% of the S&P 500, the index goes where those seven go. Bulls don't dispute this. They argue the concentration reflects deserved success, but they don't argue it isn't there. Concentration is not the same as overvaluation, but it does mean a diversified index isn't as diversified as it used to be. Our note on market capitalization gives the technical definition.

2. Capex magnitude is unprecedented. Whether you think $700 billion of annual hyperscaler spend is brilliant or reckless, it's genuinely new in scale. The number itself isn't contested. Even Burry and Nadella would agree on the dollar figures, they'd just frame what those figures mean very differently. The AI capex bubble framing, popularized by Sequoia's Cahn, doesn't get pushback on the magnitude, only on the interpretation.

3. ROI timing is uncertain. Bulls expect the AI revenue ramp to validate the spend within a few years. Bears expect a stumble: either a revenue shortfall, a hyperscaler pulling back on capex, or stress in the private credit financing the data centers. Both expect some lag between investment and full return. Amodei's quote, "If I'm just off by a year in that rate of growth… then you go bankrupt", lives precisely in that overlap. He's saying "I expect to be right, but the margin for error is thinner than it looks."

When you set the agreements next to the disagreements, the actual question gets clearer. It's not "is AI real?" Both sides say yes. It's not "is the spending unprecedented?" Both sides say yes. The real question is: do current stock prices already reflect the risks that both sides see? Bears say no, prices reflect a smooth ramp that may not arrive. Bulls say yes, prices reflect realistic earnings growth from companies actually delivering. The disagreement is about whether current prices already reflect these risks, not about the risks themselves.

Here's a possibility most AI stocks coverage skips: history suggests both sides may end up being right at the same time, just measuring different things on different time horizons.

The internet was real. The dot-com bubble was also real. Both descriptions of 1995–2002 are true. There's no contradiction between them. Cisco was a profitable, dominant company at the center of an actual technological revolution that permanently reshaped the global economy. And Cisco's stock still fell about 85% from the March 2000 peak, and didn't recover that high until 2024. Investors who said "the internet will change everything" were correct. Investors who said "Cisco at $80 is overpriced" were also correct. They weren't actually in conflict. They were measuring different things.

The pattern shows up earlier. The railroad reshaped 19th-century America in ways that still echo today. The 1840s railroad investment bubble in Britain wiped out enormous amounts of investor capital. Both statements describe the same period. Real revolution; real losses for many of the people who funded it. The market sorts winners and losers within a transformative industry, not just transformative industries themselves.

If this turns out to be the AI scenario, the honest framing is: technology revolution = yes; AI stock returns = maybe. A real revolution can produce real long-term winners and still produce dramatic losses for investors who paid too much, were positioned in the wrong names, or simply got the timing wrong. Even after the dot-com pop, the eventual large internet winners (Amazon, Google, Meta) delivered enormous returns. They just had to be held through the bust by people who didn't already know they'd survive.

In that scenario, both Burry and Nadella could be partly right. Burry could be right that a meaningful repricing event is coming in some AI stocks, particularly those priced for flawless growth. Nadella could be right that AI capex compounds into durable cash flows for the patient long-term investor. The bears would be right about prices; the bulls would be right about the technology. That's not a comforting answer. It's just the historically most common one when a real innovation meets an enthusiastic market.

Most articles dodge this part. After laying out both sides, they pretend to render a verdict. We won't, and the reason is practical, not evasive.

Bubbles are obvious in hindsight. In real time they're contested by very smart people on both sides, and roughly half of them will turn out to have been wrong. In 1999, every panel discussion about "is this a bubble?" had multiple credentialed economists on each side. The bears who turned out to be right were, by 2000, treated as geniuses. They had been called premature, hysterical, or simply mistaken for the four prior years.

Second: timing is famously hard. John Maynard Keynes (economist, not exactly a paragon of patience) is supposed to have said "the markets can stay irrational longer than you can stay solvent." Even if you correctly identify an AI stock market bubble in 2026, the pop could be tomorrow, in 2027, in 2029, or never. The Nasdaq peaked on March 10, 2000. It didn't return to that peak until 2015. Bubbles deflate in waves, sideways, slowly, or sometimes not at all.

"Is it a bubble?" turns out to be less useful than people think. A more useful one, and the one we tried to set you up to answer for yourself in the last two sections, is what specifically would have to be true for each side to be right? Watching those specific things (Azure AI run-rate, Google Cloud growth, hyperscaler capex revisions, signs of stress in private credit, productivity attribution from non-tech S&P 500 companies) is more useful than watching the daily Nvidia close.

For a beginner: the most valuable thing isn't picking a side. It's learning to hold uncertainty without panicking. Smart investors disagree because the future is genuinely unknowable, not because one of them is dumb. Your reaction to reading this article (relief, anxiety, conviction in either direction) is itself information about your own risk tolerance and time horizon. That's information worth paying attention to. What you do with it is a separate question, and one this article can't and won't answer for you.

For a 2026-specific expression of this debate, see our coverage of the trillion-dollar AI IPO wave: the same hyperscaler-capex-funding dynamics and many of the same named investors now show up in the SpaceX, Anthropic, and OpenAI public listings, with their own specific validators and timing risks. The IPO wave has itself become a data point in the debate. SpaceX went public on June 12, 2026 at $135, popped to about $161 and peaked near $225 on June 16, then fell below its IPO price to roughly $113 by late July, a post-IPO slump that bubble skeptics cite. OpenAI has delayed its own IPO to 2027, from an earlier 2026 track, citing market volatility, valuation concerns (CEO Sam Altman reportedly insists on a $1 trillion valuation), and SpaceX's rocky June debut, while Anthropic is targeting a mid-October 2026 listing at the earliest, with the prospectus now expected in late September per September 5 Reuters reporting. Historical data shown; past performance does not indicate future results.

If you'd like to go deeper on how stocks are valued in the first place, the Foundations course is the starting point. If you're newer than that, Start Investing walks through the basics of building toward a long-term portfolio.

Four common misreads we hear from beginners. None of these are dumb questions. They're the questions the headlines invite.

"If it's a bubble, I should sell everything." Maybe. (More on the urge to sell everything when a scary headline hits.) But historically, the investors who correctly identified bubbles often called them years too early, and during those years, the market kept going up. The investors who sold the Nasdaq in 1997 because it "looked frothy" missed three more years of gains before they were right. Acting on a bubble call requires being right about both the bubble and the timing, and the second is much harder than the first.

"If smart people disagree, one of them must be wrong." This is a deep misread of how markets work. Bull and bear can both be looking at the same data and weighting it differently: different assumptions about discount rates, growth durability, technological adoption, geopolitical stability. Disagreement among informed people is the norm for a healthy market, not the exception. A market where everyone agreed would already have priced in that agreement; the price wouldn't move, and there'd be no opportunity at all.

"Crashes are predictable from headlines." By the time "AI bubble" is on the front page of The New York Review of Books, the question is already widely discussed and at least partly priced in. Sometimes the warning has been on the cover for months while prices kept rising, because the warning itself was already in the price. "Cover indicator" is folklore, but the underlying point is real: public consensus is often a lagging signal.

"Bubbles always pop suddenly." Sometimes. Often not. The dot-com Nasdaq peaked on March 10, 2000 and was below that level for fifteen years. It declined in waves: sharp drops, partial recoveries, then more drops. The Japanese stock market peak of 1989 wasn't fully reclaimed until 2024, three and a half decades. A bubble pop can be a multi-year, sideways, painful slog rather than a single cinematic crash.

For a more current-market read on how confusing things have gotten, the rest of the Market Explainers series takes the same both-sides approach to other questions: why stocks keep rising despite high yields, why the housing market feels frozen, why markets seem numb to war headlines. Coming up over the next few weeks.

Selected primary sources for the figures and quotes used in this article. We encourage readers to check the originals. Financial debates are easier to follow when you read the underlying documents yourself, not just the takes about them.

Earnings & company disclosures

• Nvidia Q1 FY27 CFO Commentary (filed May 2026 with the SEC): primary source for the $75.2B Data Center revenue figure and growth metrics. SEC filing
• Alphabet Q1 2026 earnings press release (April 29, 2026): primary source for Google's capex guidance raise to $180–190B. blog.google
• Microsoft FY26 earnings calls and Morgan Stanley TMT conference fireside chat (2026): primary source for Nadella's TCO/utilization framing and Azure AI run-rate. Morgan Stanley interview

Bear research

• Jeremy Grantham & Edward Chancellor (GMO), "Valuing AI: Extreme Bubble, New Golden Era, or Both?" (January 2026): primary source for OpenAI revenue/loss figures and Grantham's framing. GMO
• Michael Burry, Substack ("Cassandra Unchained"), 2026 posts: primary source for Burry's current positioning and framing.
• Ed Zitron, "What If We're In An AI Bubble? (Part 2)" (May 22, 2026): primary source for the $178.5B data-center debt figure. Where's Your Ed At
• Ray Dalio interview, Bloomberg TV (June 3, 2026): primary source for Bridgewater's "1929 and 2000-era" bubble framing and ~$650B Mag-4 capex figure. Bloomberg
• Anthropic confidential IPO filing (June 1, 2026): primary source for the $47B run-rate figure and $965B valuation. Fortune
• Broadcom Q2 FY26 earnings and Q3 guidance (June 3, 2026): primary source for AI semiconductor revenue and the guidance miss. Bloomberg

Bull research

• Morgan Stanley AI Market Trends 2026: primary source for the evolved "vigilance, not panic" framing and $3T 2025–28 data-center spend estimate. Morgan Stanley
• Dario Amodei (Anthropic), Morgan Stanley conference (February 2026) and Fortune interview (February 14, 2026): primary source for the scaling-laws and "off by a year… bankrupt" quotes. Fortune

Historical framing

• Andrew Ross Sorkin, 1929: Inside the Greatest Crash in Wall Street History, and How It Shattered a Nation (Viking, October 14, 2025). Penguin Random House
• Jacob Weisberg, "Tick, Tick…Boom!": review of Sorkin's 1929 (New York Review of Books, March 26, 2026). NYRB

Market data

• MacroMicro Magnificent 7 share of S&P 500 (May 2026): primary source for the 34.8% concentration figure. MacroMicro
• Federal Reserve H.15 (June 1, 2026) and Trading Economics: primary source for current 10-year Treasury yield and Fed funds rate.

Last reviewed: September 8, 2026.