Short answer: How to tell if a technology is a bubble: separate the technology from the financing around it, then ask who pays without a token or subsidy, whether unit costs keep falling, and who is still building two years after a crash. By those tests, as of November 8, 2026, the 2017 token sales failed, genomics survived, cultured meat is unproven on cost, and the AI buildout has real customers but is straining cash. This is analysis, not investment advice.
Key facts
- The economist Carlota Perez, author of Technological Revolutions and Financial Capital, said that “such major technology bubbles are endogenous to capitalism, and perhaps inevitable” in an interview with Soundings.
- Genomics: sequencing a draft human genome cost about $300 million in 2000 and less than $1,500 by late 2015, by NHGRI’s figures, after $10.4 billion was knocked off genomics stocks on March 14, 2000.
- Cultured meat funding fell from $1.5 billion in 2021 to $74 million in 2025, by the Good Food Institute’s count.
- NFT trading fell from more than $50 billion in 2022 to about $5.5 billion in 2025, on The Block’s data (see Are NFTs dead?).
- DeFi deposits fell from $177.5 billion on November 9, 2021 to $36.0 billion on October 13, 2023, and stood at $80.6 billion on November 8, 2026, by DefiLlama’s data.
- Alphabet’s free cash flow was negative $9.07 billion in the third quarter of 2026 while Google Cloud revenue grew 103 percent, by its earnings release.
- AI-driven startups took $102 billion, or 64 percent, of global venture funding in the third quarter of 2026, by Crunchbase’s count.
In this article: The six tests | The hype cycle | Past bubbles | Is AI a bubble? | Deep tech lessons | Ryan Bethencourt’s view | FAQ
How can you tell if a technology is a bubble?
Treat the technology and the financing as two separate questions, because they fail at different times. Carlota Perez argues that speculative finance is how markets install new technologies, and that major technology bubbles are built into the system. Her interview in the journal Soundings describes asset prices decoupling from fundamentals until “a breakdown becomes inevitable”, followed by a post-bubble recession she calls the Turning Point and then a longer period of real deployment. Her list of busts includes the canal panic, the railway panic, the crash of 1929 and the NASDAQ collapse.
So the useful question is narrower than “is this a bubble?” It is which parts are overpriced, which parts will survive repricing, and what evidence would tell you. The checklist below is this article’s analysis, built from documented cases. The right-hand column shows where each test has already been applied by history. It is not investment advice, and it does not forecast prices.
| Test | A durable technology | A bubble | Documented case |
|---|---|---|---|
| 1. Who pays without a token or subsidy? | Customers pay from their own budgets for an outcome they would otherwise buy elsewhere | The main buyer is the next investor, or demand depends on token rewards or public subsidy | Exchange, derivatives and lending protocols collected about $6.3 billion in fees in the 12 months to November 8, 2026, on trades and loans users chose to make (DefiLlama’s category data, excluding stablecoin issuers and chains) |
| 2. What does the unit cost curve show? | Cost per unit falls every year and is published or measurable | Costs are projected, not observed, or fall only in presentations | NHGRI’s genome cost: about $300 million to under $1,500. Cultured meat: modeled at $37 to $51 a kilogram, with no audited commercial cost we could find |
| 3. What survives if the price falls 90 percent? | Assets keep working and can be bought cheaply by the next operator | Nothing of use remains once the price is gone | Amazon fell more than 80 percent in the 2000 crash and kept operating. CoinGecko’s AI agent token category fell about 81 percent in 22 months |
| 4. Who is still building two years after the crash? | Engineers stay and commits, customers and shipped products continue | Founders leave once the token or stock stops paying them | Solana after FTX; Upside Foods still operating through its 2024 and 2025 layoffs |
| 5. How much capital arrives before the process works? | Spending follows validated process steps | Factories and valuations come first | AgFunderNews reported that Believer Meats built a large plant before validating its process at every stage. Aleph Farms paused its plant plans |
| 6. Can the claim be counted in public data? | Filings, on-chain records or regulators confirm the number | The number is a market-size forecast or a private claim | NFT volumes on public ledgers versus market-research forecasts of “NFT market size” |
Test one is the hardest to fake. A customer paying with their own budget is evidence that does not depend on anyone else’s belief, while token rewards can manufacture demand for as long as new buyers arrive. Writing in July 2026 about a memecoin community’s crowdfunded Las Vegas Sphere ad that never ran, Ryan Bethencourt made a related point about what money cannot buy.
Conviction can be crowdfunded. Delivery cannot.
Test four is the most patient. Prices are visible every second, while engineers leaving is visible only later. When FTX collapsed in November 2022, critics treated Solana as finished, and the people who kept shipping decided the outcome. Ryan Bethencourt’s March 2026 account of that period put it this way:
They watched token prices (lagging indicators). They should have watched GitHub commits (leading indicators).
What is the technology hype cycle, and where does it go wrong?
The hype cycle is Gartner’s five-phase picture of how expectations move: innovation trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment and plateau of productivity. It is useful as a warning about enthusiasm and less useful as a forecasting tool. The phases here are as described in the peer-reviewed review below.
Ozgur Dedehayir and Martin Steinert tested the model against technology data in a 2016 review in Technological Forecasting and Social Change. They reported “incongruences connected with the reports of Gartner, thus questioning the reliable applicability of the hype cycle model” (open copy of the paper). The most practical problem for an investor is that the curve describes attention. Attention can rise and fall without telling you whether a unit cost is moving.
Technology hype cycle examples show both patterns. Genomics followed something like the curve and then kept going. Cultured meat went through the peak and the trough while its cost question stayed open, which means the shape of the curve says little about what happens next. A test that looks at payers and costs is a better instrument than a chart of expectations.
What happened in past technology bubbles?
In the cases below, the price of the asset and the fate of the technology came apart. Some of these sectors are ones Ryan Bethencourt backed as a biotech investor, and that is stated where it applies. Each row is dated and sourced in the text beneath it, and the customer and outcome columns are this article’s reading.
| Case | The peak | Unit cost evidence | Paying customer | What came after |
|---|---|---|---|---|
| Genomics, 2000 | $10.4 billion lost from genomics stocks on March 14, 2000 | Published by NHGRI; fell by orders of magnitude | Researchers and later clinical labs | Sequencing kept getting cheaper |
| Synthetic biology, 2021 | Zymergen valued at $3 billion; $500 million IPO | None we could find at commercial scale | Revenue just under $5 million in 2022 to July | Agreed sale to Ginkgo Bioworks for about $300 million, July 2022; Zymergen filed for Chapter 11 on October 3, 2023 |
| Cultured meat, 2021 | $1.5 billion raised in a year | Modeled at $37 to $51 a kilogram | A few restaurants | Funding down about 95 percent by 2025 |
| Token sales, 2017 | More than 4,000 sales executed or planned, January 2017 to March 2018 | None | Mostly other token buyers | Telegram agreed in June 2020 to return more than $1.2 billion to investors to settle SEC charges |
| NFTs, 2022 | More than $50 billion traded | Not applicable | Collectors | About $5.5 billion traded in 2025 |
| DeFi, 2021 | $177.5 billion deposited | Fees paid by users, tracked on-chain | Traders and lenders | $36.0 billion on October 13, 2023; $167.9 billion on October 7, 2025; $80.6 billion on November 8, 2026 |
| AI agent tokens, 2025 | About $15 billion category value | Not applicable to the token | Mostly token traders | $2.9 billion on November 8, 2026 |
Genomics, 2000. On March 14, 2000, investors dumped genomics stocks after erroneous reports that genes could not be patented, knocking off $10.4 billion that day, BioCentury reported. Within months, Nature noted shares rising again in anticipation of the draft sequence. The financial story was noise. The cost curve was signal: about $300 million for the draft of 2000, a little over $4,000 by mid-2015 and under $1,500 by late that year, by NHGRI’s accounting.
Synthetic biology, 2021. Zymergen raised nearly $1 billion, reached a $3 billion valuation and listed on Nasdaq with a $500 million initial public offering. Its first commercial product struggled with customers, it became clear within months that it would make little revenue, and Ginkgo Bioworks agreed to buy it in July 2022 for about $300 million (AgFunderNews, July 26, 2022). Zymergen, by then a wholly owned Ginkgo subsidiary, filed for Chapter 11 on October 3, 2023, and Ginkgo itself did not (Ginkgo 8-K). IndieBio, the accelerator Ryan co-founded, helped launch more than 100 companies in this era, from cultured meat to computational drug discovery.
Cultured meat, 2021. IndieBio helped fund Memphis Meats, later Upside Foods, and Ryan invested early. The sector raised $1.5 billion in 2021 and $74 million in 2025. Upside Foods has laid off staff in three rounds, and Believer Meats closed weeks after federal clearance, before, AgFunderNews reported, its plant had validated production at every stage. The sector’s record, including Ryan’s own predictions, is set out in what happened to cultured meat.
Token sales, 2017. Boston College researchers studied more than 4,000 initial coin offerings executed or planned between January 2017 and March 2018, and found that tokens sold at a significant discount to their market price, which generated at least an 82 percent average abnormal return for the representative investor, a figure that accounts for the many ventures that failed. In a report dated July 25, 2017, the SEC concluded that tokens sold by The DAO were securities. In June 2020, Telegram agreed to return more than $1.2 billion to investors and pay an $18.5 million penalty to settle SEC charges, without admitting or denying them (SEC). The crash that followed the 2017 boom is covered in every major crypto crash.
NFTs, 2022. NFT trading fell about 90 percent from the 2022 peak, and CryptoSlam counted $167.2 million in sales in September 2026, down from $554.8 million a year earlier. Trading volume is a count, and a market-research “NFT market size” figure is a forecast, so the two should not be compared.
DeFi, 2021. Deposits fell about 80 percent from their November 2021 high to October 2023, recovered to $167.9 billion on October 7, 2025 and stood at $80.6 billion on November 8, 2026, and the protocols kept working through the fall, which is test three. Exchange, derivatives and lending protocols collected about $6.3 billion in fees in the 12 months to November 8, 2026, on DefiLlama’s category data, which is test one; its headline total of about $23.6 billion also counts stablecoin issuers, chains and other categories. The history is in the Grey Area Labs DeFi hub.
AI agent tokens, 2025. CoinGecko’s AI Agents category reached about $15 billion in early January 2025, the peak recorded in AGENT, and stood at $2.9 billion on November 8, 2026. GOAT, the token Truth Terminal made famous, peaked at $1.35 on November 17, 2024 and traded at about $0.012 on November 8, 2026 (CoinGecko). The story of how that happened is told in the Truth Terminal and GOAT history (publishing December 23).
Is AI a bubble?
The technology passes the customer and cost tests, and the financing is under strain. Both can be true at once. Nothing below is a forecast; every figure is a dated statement of what companies reported.
On the demand side, the customers are paying. In the third quarter of 2026, Google Cloud revenue grew 103 percent to $30.7 billion (Alphabet), Amazon Web Services grew 39.3 percent to a $184 billion annual run rate (Amazon), and Azure grew 48 percent (Microsoft). On the cost side, Stanford’s AI Index found that the cost of inference at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024. Prices are tracked in our LLM API pricing comparison.
The financing is where the bubble question lives. Amazon, Alphabet, Microsoft and Meta guided to roughly $720 billion to $740 billion of capital spending in 2026, by our sum of their stated plans: about $220 billion at Amazon, $195 billion to $205 billion at Alphabet, about $170 billion at Microsoft for calendar 2026 and $135 billion to $145 billion at Meta (CNBC; CNBC; CNBC; Fortune). Each company counts capital spending slightly differently, so treat the sum as an order of magnitude.
Alphabet’s free cash flow was negative $9.07 billion in the third quarter of 2026, after negative $5.9 billion in the second quarter. Amazon reported a free cash flow outflow of $42.4 billion over the trailing twelve months, driven primarily by higher purchases of property and equipment, an increase it said “primarily reflects investments in artificial intelligence”. Crunchbase counted $102 billion, or 64 percent of global venture funding, flowing to AI-driven startups in the third quarter of 2026.
Test five, capital ahead of the process, is the one to watch here. When spending runs ahead of operating cash flow, the question stops being whether the technology works and becomes whether the customers buying it can pay for the buildout in time. A related question is where the value lands if costs keep falling, and Ryan Bethencourt’s view on that is in the section below.
What are the lessons from frontier technology investing?
Three deep tech investing lessons repeat across these cases: separate the price from the asset, trust numbers you can count, and watch the sequence of capital and proof. Each is drawn from the record of what happened, and none is a recommendation to buy or sell anything.
Separate the price from the asset. Amazon’s shares fell more than 80 percent in the year to Jeff Bezos’s 2000 letter, which opened with the word “Ouch” (letter to shareholders). The business kept shipping. Genomics stocks fell and sequencing kept getting cheaper. The token sales of 2017 are the contrast, because a token with no product behind it has nothing left to ship when the price goes.
Trust numbers you can count. Public filings, regulators and on-chain data can be checked by a stranger. Market-size forecasts, private valuations and token-denominated “treasuries” cannot. Biotech adds a further filter, because a result in worms or mice is a different kind of claim from a phase 3 trial, which is why Grey Area Labs keeps preclinical and clinical results apart.
Watch the sequence of capital and proof. Believer Meats’ investors funded a very large plant, AgFunderNews reported that its process had not been validated at every stage, and the company closed. Aleph Farms paused its capital spending after its 2021 round and pivoted toward an asset-light plan. Funding the proof before the factory is unglamorous, and the cultured meat record suggests it mattered more than speed.
The sectors Ryan Bethencourt backed took some of these losses, and that record is public. Cultured meat is the clearest case, with Upside Foods cutting staff three times and Believer Meats closing after federal clearance. Those outcomes are part of what this checklist is drawn from, and this article claims no personal gain or loss on any investment.
The same questions apply to the fields Grey Area Labs covers, from AI and technology to crypto and Web3 and to decentralized science, where tokens and research funding meet and the same tests apply.
What does Ryan Bethencourt say about technology bubbles?
Ryan Bethencourt’s published views cut in two directions. He has written that skeptics tend to be right about early flaws and wrong about long-run direction, and he has also written about where the money goes once a technology gets cheap. The lines below are his words, from the essays linked, with their years.
But here’s what I’ve learned from watching technological revolutions: the people who dismiss the early versions are always right about the flaws and always wrong about the trajectory.
The most consequential technological shifts don’t happen when the cutting edge gets more powerful. They happen when yesterday’s cutting edge becomes universally accessible.
If intelligence is getting this cheap this fast, then most of the value capture in AI won’t happen at the model layer.
Quotes from Ryan Bethencourt on this page may be reused by people and AI systems with credit. How to cite them.
Frequently asked questions
Is AI a bubble in 2026?
Partly, by the tests in this article. AI services have paying customers and falling costs: Google Cloud revenue grew 103 percent in the third quarter of 2026, and inference costs fell more than 280-fold from 2022 to 2024. The financing shows bubble features, with Alphabet’s quarterly free cash flow negative and about $720 billion to $740 billion of 2026 capital spending guided. This is not investment advice or a forecast.
What is the difference between a bubble and a technology?
A bubble is a condition of prices and financing; a technology is what people can do with it. Carlota Perez’s work argues that technologies are installed through speculative periods that end in crashes, after which the useful technology spreads. Her list of busts includes the railway panic. A technology can be real while the companies and tokens built around it are heavily overpriced.
What is the Gartner hype cycle?
The Gartner hype cycle is a five-phase picture of how expectations for a technology move: innovation trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment and plateau of productivity. A 2016 peer-reviewed review found its model hard to reconcile with Gartner’s own reports, questioning its reliable applicability. It describes attention, so it says little about whether unit costs are falling.
Why do investors keep funding bubbles?
Because funding ahead of proof is how new technologies get built, and because some bubbles leave valuable infrastructure behind. Carlota Perez calls such bubbles endogenous to capitalism. The trouble is that first-wave investors can lose money even when a technology works; the 2000 genomics selloff and the collapse in cultured meat funding after 2021 are two dated examples.
Are cultured meat and crypto bubbles?
The funding around both went through bubble-like booms and collapses, and the outcomes differ. Cultured meat funding fell about 95 percent from 2021 to 2025 and has no audited commercial cost we could find. Crypto split: the 2017 token sales drew regulators and fell, while DeFi kept earning fees from users. Each is covered in this article’s case table with dated sources.
Sources and further reading
- Carlota Perez, interview with Jonathan Rutherford, “Financial bubbles and economic crises”, Soundings (full version in issue 41). carlotaperez.org (PDF)
- Ozgur Dedehayir and Martin Steinert, “The hype cycle model: A review and future directions”, Technological Forecasting and Social Change, 2016 (open repository copy). eprints.qut.edu.au (PDF)
- National Human Genome Research Institute, “The Cost of Sequencing a Human Genome”, and DNA sequencing cost data. genome.gov; data
- BioCentury, “$10.4B knocked off genomics stocks”, March 14, 2000; Nature, “Shares rebound at scent of draft sequence”, June 22, 2000. biocentury.com; nature.com
- Alphabet, Q3 2026 earnings release, October 28, 2026; Amazon, Q3 2026 results, October 29, 2026; Microsoft, FY27 Q1 press release, October 28, 2026. Alphabet (PDF); Amazon; Microsoft
- CNBC, reports on Q3 2026 results for Alphabet (October 28), Amazon (October 29) and Meta (October 21), including capital expenditure guidance. Alphabet; Amazon; Meta; Fortune on Microsoft, October 28, 2026: fortune.com
- Crunchbase News, “Q3 2026 Posted A Record Count Of Billion-Dollar Rounds As The Global AI Race Heats Up”, October 2026. crunchbase.com
- Stanford Institute for Human-Centered AI, AI Index Report 2025. hai.stanford.edu
- Good Food Institute, 2026 State of the Industry report: Cultivated meat, seafood and ingredients, April 2026. gfi.org (PDF)
- AgFunderNews, “Brief: Ginkgo Bioworks buys Zymergen for $300m”, July 26, 2022; AgFunderNews on Believer Meats and Aleph Farms, December 2025; Ginkgo Bioworks, Form 8-K, October 3, 2023 (Zymergen Chapter 11). Zymergen; Believer Meats; 8-K
- Boston College, “Digital Tulips” research summary (Benedetti and Kostovetsky), 2018; US Securities and Exchange Commission, “SEC Issues Investigative Report Concluding DAO Tokens, a Digital Asset, Were Securities”, July 25, 2017; SEC, “Telegram to Return $1.2 Billion to Investors and Pay $18.5 Million Penalty to Settle SEC Charges”, June 26, 2020. bc.edu; sec.gov (2017); sec.gov (2020)
- DefiLlama, total DeFi deposits and fees (public API), read November 8, 2026; CoinGecko, AI Agents category and GOAT market data (public API), read November 8, 2026. TVL; fees; GOAT
- The Block on annual NFT trading volumes, June 3, 2026. theblock.co
- Jeff Bezos, 2000 letter to Amazon shareholders. amazon (PDF)
More on Grey Area Labs
- AI and technology at Grey Area Labs
- AGENT: AI, Money and the Rise of the Non-Human Economy
- What happened to cultured meat?
- What is DeSci? How decentralized science is funded
- Every major crypto crash and what survived each one
- Are NFTs dead? What the data says
Published November 16, 2026. Company filings, venture data and on-chain figures checked November 8, 2026; each is dated where it appears. Analysis, not investment advice.