AI bubble talk has become the background noise of every earnings call, every fund manager letter and every dinner-party argument about technology since late 2022. The question is almost always posed as a yes or no. New Cornell research says that framing is the problem, and that the honest answer is neither.
The paper behind the headline is a statistics paper, not a market commentary. It builds a test that looks at one company’s daily share price at a time and asks whether that specific price series is behaving explosively — and if so, on exactly which date the behaviour started and on which date it stopped. Run across twelve of the largest technology and chip firms, the test says the sector as a whole is not in a bubble, while a handful of named companies clearly are.
That distinction is the whole story, and it is the reason the Cornell Chronicle headline reads “Is there an AI bubble? No, but some tech companies show signs”. If you follow how AI infrastructure and tooling reach production, our AI models, tools and releases hub tracks the product side of the same boom; this article deals with the price side of it.
What follows is a full read of the method, the twelve verdicts, the date stamps the model produced, how the results compare with the standard test everyone else uses, what the Bank of England has said about the same risk, and what any of it should change for a business that is buying AI rather than trading it.
Table of contents
- What The Cornell AI Bubble Study Actually Found
- How The SV-ADF Method Date-Stamps An AI Bubble
- Which Companies Show AI Bubble Signs, And When
- The Selectivity Test That Separates The Two Methods
- What The Nasdaq And Crypto Results Add
- How This AI Bubble Compares With The Dot-Com Era
- What Regulators Say About AI Bubble Risk
- What An AI Bubble Would Mean For Ordinary Businesses
- How To Read AI Bubble Headlines Without Panicking
- The Limits Of This AI Bubble Study, Stated Plainly
- Frequently Asked Questions About The AI Bubble
- References
What The Cornell AI Bubble Study Actually Found
Start with the claim itself, because it is narrower and sharper than the headline suggests.
The headline finding, in one sentence
The AI sector is not in a bubble market, but the prices of some individual AI-exposed companies — Alphabet most prominently — behave with the price variability and exuberance characteristic of one. That is the finding, and every other number in the paper is in service of it. Treating “AI” as a single category, the authors argue, hides more than it reveals.
Who wrote it and where it was published
The authors are Abir Sarkar, a doctoral student in the field of statistics at Cornell, who is the paper’s lead author, and Martin Wells, the Charles A. Alexander Professor of Statistical Sciences in the Cornell Ann S. Bowers College of Computing and Information Science and in the ILR School. The paper appeared in July 2026 in Frontiers in Mathematical Finance, published by the American Institute of Mathematical Sciences.
The title is worth reading twice
The Cornell Chronicle renders it as “Is There an AI Bubble? Robust Data-stamping for Periods of Exuberance”. The paper itself, on arXiv since 13 April 2026 and revised on 9 May 2026, says Date-Stamping. That is not pedantry — date-stamping is the entire contribution. The paper is not primarily arguing about whether an AI bubble exists; it is arguing about how you would put a start date and an end date on one.
The number the press release led with
Alphabet’s share price has risen more than 70% in the year to publication, against roughly 20% for the whole NASDAQ Composite. That is about three and a half times the index move, from a company already among the largest in the world by market value.
Why one big number is not evidence on its own
A 70% year is not proof of an AI bubble. Plenty of companies have risen that much on earnings that justified it. The reason the paper matters is that it does not stop at the price change — it tests whether the path of the price is explosive in a statistical sense, which is a different and much harder question.
How The SV-ADF Method Date-Stamps An AI Bubble
The method is called SV-ADF: a stochastic-volatility-robust augmented Dickey–Fuller test. The name is ugly and the idea underneath it is simple.
What a bubble test is actually testing
A standard unit-root test asks whether a price series wanders like a random walk. A right-tailed version flips the question and asks whether the series is doing something stronger than wandering — growing at an accelerating rate, the statistical signature of a price detaching from what it is worth. That signature is what any AI bubble test is hunting for.
Why the standard tests over-call
The catch is that those tests are derived assuming the variance of daily returns is constant, or close to it. Real technology stocks are nothing like that. Volatility comes in long, persistent waves, and a wave of volatility looks, to a test built on the wrong assumption, exactly like an explosive price. So the standard procedure raises an alarm every time the market gets jumpy.
What SV-ADF changes
The Cornell test extends right-tailed Dickey–Fuller testing to models with highly persistent mean and volatility dynamics, without forcing a rigid parametric shape onto the variance. In plain terms, it lets the noise be as wild and as clustered as it really is, then asks whether there is still an explosive trend underneath. That is what makes it usable as an AI bubble detector rather than a volatility detector.
Two thresholds instead of one
The other change is structural. The standard procedure — Phillips, Wu and Yu, usually shortened to PWY — applies a single uniform threshold to both the start and the end of an episode. SV-ADF calibrates them separately: an origination threshold approximated by log(n)/10, and a collapse threshold approximated by log(n)/2. Starting and ending a bubble are different statistical events, so they get different critical values.
The persistence rule that filters out noise
Because the analysis runs on daily data, a single day above a threshold means very little. The authors require the recursive statistic to sit above the origination threshold for at least two consecutive calendar months before they will call a start, and to sit below the collapse threshold for a month before they will call an end. That rule is what converts a twitchy daily signal into a defensible date.
| Design choice | Standard PWY test | Cornell SV-ADF test |
|---|---|---|
| Assumption about volatility | Constant or weakly varying | Persistent and time-varying |
| Thresholds | One uniform threshold for both ends | Separate origination and collapse values |
| Threshold form | log log(ns)/100 | log(n)/10 start, log(n)/2 end |
| Behaviour in a volatility spike | Frequent transient false alarms | Alarms are stable across the spike |
| Unit of analysis | Often applied to whole indices | Individual stocks within one industry |
| Episodes supported | Multiple, in the PSY extension | One episode per stock by construction |
Which Companies Show AI Bubble Signs, And When
Twelve firms were tested on daily prices across 2018–2026, split into a pre-AI window and the post-2022 window that everyone actually cares about.
Alphabet is the live case
SV-ADF dates Alphabet’s episode from October 2025, and it was still unresolved when the sample ended. The paper reads it as a sustained upward revision in what investors expect from the company as an AI beneficiary, helped by stronger Google Cloud performance. It is the only one of the very largest technology firms the model flags in this cycle.
Nvidia’s episode has already opened and closed
Nvidia is the cleanest date-stamp in the paper: a bubble running from November 2023 to September 2024. The start lines up with the November 2023 earnings surprise and intensified after the February 2024 release. The collapse in September 2024 lines up with upward revisions moderating after the August 2024 earnings, despite operating performance that stayed strong.
TSMC and Micron are the ongoing chip cases
TSMC’s episode begins in September 2025 and remains open, which the authors tie to its bottleneck role in advanced AI chips. Micron’s begins in October 2025, also open, and with greater intensity — the paper describes high-bandwidth memory being re-rated from a cyclical product into a structurally scarce input. Micron is the standout: its estimated autoregressive coefficient interval lies entirely above one.
Tesla belongs to a different cycle
Tesla’s flagged episode runs June 2020 to March 2021 — profitability confirmation, the stock split, S&P 500 inclusion. In the post-2022 AI window Tesla’s estimated coefficient sits below one, which is the weakest evidence of exuberance in the whole recent sample. Whatever is happening to Tesla’s price now, this AI bubble test does not think it is that.
Where the model found nothing
For Microsoft, Apple and Amazon the evidence does not indicate a sustained exuberant regime in the post-2022 window. Meta shows a short-lived episode too brief to matter to the wider pattern. Palantir, Broadcom and ASML show shorter, less persistent bursts consistent with temporary repricing rather than a durable AI bubble.
| Company | Episode begins | Episode ends | State at end of sample |
|---|---|---|---|
| Alphabet | October 2025 | Not reached | Open and unresolved |
| TSMC | September 2025 | Not reached | Open, bottleneck re-rating |
| Micron | October 2025 | Not reached | Open, strongest intensity |
| Nvidia | November 2023 | September 2024 | Closed |
| Tesla | June 2020 | March 2021 | Closed, pre-AI cycle |
| Bitcoin and Ethereum | January 2021 | June 2021 | Closed |
| Apple, Microsoft, Amazon | Not detected | Not detected | No sustained episode |
The Selectivity Test That Separates The Two Methods
The most persuasive part of the paper is not the list of flagged firms. It is the comparison with what the standard method says about the same twelve companies.
PWY calls almost everything a bubble
Run over the post-2022 window, PWY rejects the null for eleven of the twelve firms — every name except Tesla. Over the pre-AI window it flags ten of the eleven that were testable. A test that labels essentially the entire large-cap technology complex as bubbly in every period is not giving you information; it is giving you a constant.
SV-ADF flags half as many
SV-ADF flags six of the twelve in the post-2022 window: Alphabet, Nvidia, TSMC, Broadcom, Palantir and Micron. In the earlier window it flags exactly one, Tesla. Six out of twelve is 50%; eleven out of twelve is roughly 92%. That gap is the paper’s real argument.
Reading the verdict table honestly
The table below is the paper’s own comparison, condensed. Read down the SV-ADF column and the pattern is obvious: the flags cluster in silicon and in one search company, not across large-cap technology as a class.
| Company | Segment | SV-ADF, 2022–2026 | PWY, 2022–2026 |
|---|---|---|---|
| Apple | Large-cap tech | No | Yes |
| Microsoft | Large-cap tech | No | Yes |
| Alphabet | Large-cap tech | Yes | Yes |
| Amazon | Large-cap tech | No | Yes |
| Meta | Large-cap tech | No | Yes |
| Tesla | Large-cap tech | No | No |
| Nvidia | Chipmaker | Yes | Yes |
| TSMC | Chipmaker | Yes | Yes |
| Broadcom | Chipmaker | Yes | Yes |
| Micron | Memory | Yes | Yes |
| ASML | Chip equipment | No | Yes |
| Palantir | Data platform | Yes | Yes |
Why the disagreement matters more than the verdicts
If two credible tests disagree on six of twelve names, then “there is an AI bubble” and “there is no AI bubble” can both be defended with real statistics. That is precisely why so much of the public argument goes nowhere. The useful question is not which side you are on but which test, over which window, produced the claim you just read.
What The Nasdaq And Crypto Results Add
The paper does not stop at individual firms, and the two extra experiments are the ones that keep it honest.
No exuberance in the index itself
Applied to the NASDAQ Composite over 2020–2026, SV-ADF finds no statistically significant exuberance at all, despite the entire AI-driven run. This is the single result that most directly supports the “no sector-wide AI bubble” headline, and it is also the one that will annoy the most people.
The 1990s replication is the credibility check
To show the test is not simply blind, the authors re-run the classic 1990s Nasdaq exercise. SV-ADF dates the origination to early April 1995 and the collapse to September 2000, against June 1995 to September 2000 from PWY. It catches the real thing two months earlier, under a stricter threshold — which is the whole point of a better test.
Bitcoin, Ethereum and one date that does not match
For cryptocurrency the test identifies significant bubble behaviour in Bitcoin and Ethereum during January 2021 to June 2021, an episode that the standard homoskedastic procedure fails to date accurately at all. Note that the Cornell Chronicle’s write-up describes crypto exuberance as beginning in December 2020; the paper’s own figure gives January 2021. Small gaps like that are worth checking before you quote a date.
How This AI Bubble Compares With The Dot-Com Era
Every AI bubble conversation eventually reaches for 1999. The comparison is useful in one direction and misleading in the other.
Cash flow is the difference
The dot-com cohort was full of companies with no earnings and, in many cases, no revenue. The firms at the centre of this cycle are among the most profitable enterprises in history, and much of the current build-out is funded from operating cash rather than debt. That is a materially different balance-sheet picture, and it is the strongest argument against a systemic AI bubble.
Concentration is the similarity
What does rhyme with 1999 is concentration. A small number of names now drive a large share of index returns, so a re-rating of those names is not a sector event — it is an index event, and through passive funds and pensions it reaches households that never chose to hold a chip stock.
What “the sector is fine” does not mean
A finding that the sector is not in a bubble is not a finding that no one will lose money. Nvidia’s flagged episode closed in September 2024 while the business kept performing. Prices can normalise without a single product failing, which is exactly the outcome most AI bubble commentary is not built to describe.
What Regulators Say About AI Bubble Risk
Central banks have been circling the same question with different tools, and their framing is worth putting next to the Cornell result.
The Bank of England’s framing
The Bank’s Financial Policy Committee has repeatedly flagged that equity prices have risen especially for AI-related stocks and that valuations look stretched on some measures. Its concern is conditional: those valuations rest on the infrastructure build-out succeeding, on continued access to financing, and on the pace at which AI is actually adopted across the wider economy.
Amplifiers, not just levels
The Committee’s sharper point is about amplification. A fall in AI-related equity prices would be magnified by high index concentration, by momentum-driven positioning and by rising leverage in equity markets. That is a mechanism argument, and it does not depend on resolving the AI bubble question at all.
Why a stock-level signal is useful to supervisors
This is where a date-stamping test earns its keep. As Robert Jarrow, the Ronald P. and Susan E. Lynch Professor of Investment Management at the Cornell SC Johnson College of Business, puts it: “Identifying stocks with bubbles is important because bubbles eventually burst, which result in dramatic price declines. Knowing which stocks contain bubbles enables investors to make better investment decisions.” Jarrow is joining Wells and Sarkar to extend the work.
What An AI Bubble Would Mean For Ordinary Businesses
Most organisations reading about an AI bubble do not hold these shares directly. They still have exposure, and it runs through procurement rather than portfolios.
Your budget is not a stock position, but your suppliers are
If you are buying AI-assisted tooling, your risk is not a share price. It is whether the vendor behind the tool is still funded, still supported and still charging the same amount in two years. That is a supplier-viability question, and it belongs in the same review as any other critical dependency in your digital strategy.
Pricing risk sits with the loss-making layer
The firms flagged by this test are mostly profitable chip and platform businesses. The pricing risk to buyers sits further up the stack, with application vendors selling below cost to win share. A product that wraps a large model to do document handling or natural language processing has a cost floor set by someone else’s silicon, and that floor does not care about the vendor’s growth plans.
Contract terms that survive a repricing
The practical protections are unglamorous: price-increase caps, notice periods, data-export rights and a written exit path. None of them require you to have an opinion on whether an AI bubble exists. They simply mean a supplier repricing is an inconvenience rather than a migration crisis.
| Decision | Exposed to a repricing | Resilient to a repricing |
|---|---|---|
| Model access | Single provider, no abstraction | Provider swappable behind one interface |
| Contract length | Multi-year, uncapped uplifts | Annual with a capped increase |
| Data | Locked in vendor format | Exportable on demand |
| Business case | Rests on future price cuts | Works at today’s list price |
| Vendor funding | Unprofitable, growth-funded | Cash-generative or well capitalised |
The cheap discipline is measurement
Every claim in this paper is a measured quantity with a stated window. The same standard applied internally — what did this tool change, over what period, measured how — is the fastest way to make your own AI spending immune to the argument. Our data analytics and AI strategy pages cover the mechanics of putting that measurement in place.
How To Read AI Bubble Headlines Without Panicking
Three questions separate a real AI bubble finding from a recycled opinion, and they take about a minute to apply.
Ask what was measured
Was the claim about prices, revenues, capital spending or adoption? These are four different debates that share one word. This paper measures prices only, and says so.
Ask which window
A bubble claim without a start and end date is not a claim, it is a mood. The Cornell result is unusual precisely because it names months. Nvidia’s episode is over; Alphabet’s was open at the cut-off in April 2026.
Ask which test, and how selective it is
If a method flags 92% of a sample, its next flag tells you almost nothing. Selectivity is the property that makes a signal worth acting on, and the paper’s simulations show SV-ADF detecting both the start and the end of a genuine episode more reliably than the standard alternative.
The Limits Of This AI Bubble Study, Stated Plainly
Good research is easier to trust when its boundaries are visible, and this paper’s are clear.
One episode per stock
The framework supports a single bubble episode per stock by construction. That is why the authors deliberately do not benchmark against PSY, the multiple-bubble extension. If a company has genuinely inflated, deflated and inflated again, this design will not show you all three.
A price statement, not a product verdict
The test reads prices. It has nothing to say about whether a model is good, whether a chip is fast, or whether an AI deployment pays back. A company can be flagged and still be building something durable — Nvidia in 2024 is the paper’s own example.
The sample ends in April 2026
Alphabet, TSMC and Micron are described as open episodes as of the cut-off. Open is not the same as permanent. Any of them could have resolved in the months since, and the honest reading of an ongoing AI bubble flag is “unresolved at the last measurement”, not “certain to burst”.
One quote worth keeping
Sarkar’s own summary is the most useful sentence in the coverage: “Our method is not saying that everything is a bubble. That is the key distinction from the standard bubble-detection methods.” Read alongside our write-up of the strategic fit study on AI winners, it points the same way — the interesting variation is between firms, not across the sector.
Frequently Asked Questions About The AI Bubble
Does the Cornell paper say there is an AI bubble or not?
It says there is no sector-wide AI bubble, but that several individual companies show bubble dynamics. Both halves of that sentence are load-bearing.
Which companies were flagged?
In the post-2022 window: Alphabet, Nvidia, TSMC, Broadcom, Palantir and Micron. Tesla was flagged only in the earlier 2018–2020 window.
Which were not flagged?
Apple, Microsoft, Amazon, Meta and ASML showed no sustained exuberant regime in the post-2022 window under the SV-ADF test.
Does an open episode mean the stock will crash?
No. It means the price path was still behaving explosively at the last measurement. Episodes close without dramatic falls, as Nvidia’s did in September 2024.
Why do other studies reach different conclusions?
Mostly because they use tests that assume steady volatility, which over-detect. The same twelve firms produce eleven flags under the standard procedure and six under this one.
Should this change my company’s AI spending?
Not the spending itself, but perhaps the contracts around it. Cap price increases, keep your data portable and make sure the business case works at today’s prices rather than hoped-for future ones.
Where can I read the paper?
The preprint is on arXiv under number 2604.12062, and the published version appeared in Frontiers in Mathematical Finance in July 2026.
References
Is there an AI bubble? No, but some tech companies show signs — Cornell Chronicle
Is There an AI Bubble? Robust Date-Stamping for Periods of Exuberance — arXiv
Sarkar and Wells, full paper (PDF)
Frontiers in Mathematical Finance — American Institute of Mathematical Sciences
Financial Stability Report, July 2026 — Bank of England
Financial Policy Committee Record, July 2026 — Bank of England
Financial Stability Report — Board of Governors of the Federal Reserve System
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.