Is There an AI Bubble? The State of Play as of October 2026

Disclaimer: This article is for general information only and is not financial or investment advice. Disclosure: This article was researched and drafted with help from an AI system built by Anthropic, one of the companies discussed below. We have tried to treat Anthropic exactly like every other company here, using only its public filings and independent reporting.

On Friday, October 30, 2026, the Nasdaq Composite fell 2.86% to close at 25,437. It was the index's worst day since early June. The trigger was not a war, a bank failure or a surprise rate hike. It was a run of quarterly reports from companies spending record sums on artificial intelligence, and from investors who were no longer sure they would get paid back. Intel fell about 28.6% in a single session. The broader chip index dropped 7.2%. One wealth manager summed up the mood: the question is "whether there will be a positive return on investment and if so, how quickly."

That sentence is the AI-bubble debate in miniature. Nobody serious argues that AI is fake. Revenue is real and growing very fast. Cloud divisions at Microsoft, Alphabet and Amazon are posting their best growth in years. Nvidia is reporting quarterly revenue that would have been a good annual figure for it not long ago. The real argument is about price, pace and plumbing. Are valuations running ahead of any plausible profits? Is capacity being built faster than demand can absorb it? And is the money flowing through financial structures (circular deals, off-balance-sheet vehicles, rising debt) that could turn an ordinary slowdown into something worse?

This guide sets out the evidence as of the end of October 2026: the spending, the revenue, the deals, the debt, the power grid, the market wobbles and the history. It gives both the bull and bear cases fairly. Where a figure comes from a company filing, we say so. Where it is an analyst's estimate or our own arithmetic, we label it.

Key takeaways

• The spending is unprecedented. The four largest US hyperscalers (Microsoft, Alphabet, Amazon and Meta) spent a combined $151 billion on capital expenditure in 2023 and $246 billion in 2024. Their guidance for 2026 now adds up to roughly $740–760 billion (our arithmetic from company guidance), and several have said 2027 will be higher still.
• The revenue is real but lags the spending. Google Cloud revenue more than doubled year on year to $30.7 billion in Q3 2026, Azure grew 48% and AWS grew 39%. But Alphabet, Meta and Amazon all reported negative free cash flow on heavy capex.
• The frontier labs are valued like mega-caps. OpenAI raised $122 billion at an $852 billion post-money valuation in March 2026 and is reportedly in talks at about $1.4 trillion. Anthropic raised $65 billion at $965 billion in mid-2026 and filed its S-1 publicly on October 26, 2026.
• Circular deals are the biggest red flag. Suppliers such as Nvidia and AMD are investing in, guaranteeing or issuing equity to their own customers, and the customers then spend that money on the suppliers' chips.
• Debt is replacing cash flow as the fuel. Off-balance-sheet structures like Meta's $27 billion Hyperion joint venture, Oracle's $125 billion debt load and record bond issuance worry the IMF and the Bank of England.
• Power is the hard limit. US data centres could use 9–17% of US electricity by 2030 (EPRI), up from about 4–5% today. Gas turbines are back-ordered for years, and capacity prices in the largest US grid have hit their cap.
• Verdict: contested. The evidence fits a capital-spending boom with bubble-like pockets better than a simple yes or no. The key unknown is whether AI revenue can grow into roughly $2 trillion a year by 2030, the figure Bain & Company estimates is needed.

What do we actually mean by "bubble"?

The word is used loosely, so it is worth being precise. Economists usually describe a bubble as a period when asset prices rise far above what fundamentals can justify, driven by expectations of further price rises rather than by underlying cash flows. Bubbles usually share several features:

1. A real underlying innovation. Railways, electricity, radio and the internet were all real, and they all had bubbles.
2. Extrapolation. Early growth rates get projected forward indefinitely.
3. Easy financing. Credit, new financial structures or retail enthusiasm supply the fuel.
4. Overbuilding. Physical capacity (track, fibre, data centres) runs ahead of paying demand.
5. A trigger and a repricing. Something shows that the cash flows will come later, or smaller, than priced in.

A key point from economic history is that a bubble and a transformative technology can coexist. Sam Altman, OpenAI's chief executive, said exactly this in August 2025. Asked whether investors as a whole were overexcited about AI, he said "my opinion is yes," and added that "when bubbles happen, smart people get overexcited about a kernel of truth." At the time, his company was negotiating a share sale at a $500 billion valuation.

So the useful question is not "is AI real?" It clearly is. The question is: are today's prices and today's build-out consistent with realistic future cash flows, and what happens to the financial system if they are not?

Part 1: The spending

Hyperscaler capex, quarter by quarter

The late-October 2026 earnings season gave the freshest data. Here is what the five biggest AI infrastructure spenders reported in their most recent quarters:

Sources: company releases and CNBC/Fortune coverage, October 21–29, 2026, and September 10, 2026 for Oracle. Microsoft's guidance periods were described inconsistently across outlets, so check the 10-Q for exact framing.

Three things stand out. First, the scale: Amazon's gross purchases of property and equipment over the past twelve months were $215.2 billion, more than the annual revenue of most Fortune 500 companies. Second, the direction: no major spender cut its outlook, and Alphabet, Amazon and Meta all signalled more spending in 2027. Fortune reported that analysts project Alphabet's 2027 capex at around $300 billion. Third, the cash flow has turned negative at companies that were, until recently, among the most cash-generative businesses ever built.

Capex by year

Combining the "big four" (Microsoft, Alphabet, Amazon, Meta) gives this picture:

Big-four hyperscaler capex, US$ billions  (each █ ≈ $20B)

2023        ████████                                  151
2024        ████████████                              246
2025 (est)  ████████████████████                     ~400
2026 Feb    █████████████████████████████████        ~650
2026 Oct    ██████████████████████████████████████   ~750

That is roughly a fivefold increase in three years. The International Energy Agency, in an April 2026 analysis, put 2025 capex across five large technology companies at more than $400 billion and said it was set to rise another 75% in 2026.

Is the spending earning anything?

The bull case starts here, because the revenue lines are strong:

• Google Cloud: $30.7 billion in Q3 2026, more than double a year earlier, with a 36.2% operating margin. Sundar Pichai said Alphabet is on "a half-trillion dollar revenue run rate."
• Microsoft Azure: $33.17 billion in the September quarter, up 48%, ahead of analyst expectations. It was the first time Microsoft disclosed Azure revenue in dollars.
• AWS: $46.0 billion, up 39%, its fastest growth in nearly five years. Amazon cited a $184 billion annualised run rate, and coverage of the call reported a $578 billion contracted backlog.
• Oracle: remaining performance obligations (contracted future revenue) of $664 billion, and cloud infrastructure revenue that "more than doubled" to $7.4 billion.

The bear case also starts here. Backlogs are promises, not cash, and much of the backlog is owed by a handful of AI labs that are themselves losing money. Investors' reaction to the October reports shows the tension. Alphabet's cloud business doubled, yet the shares fell more than 11% after hours because Search growth slowed to 12% and free cash flow went deeper into the red. One independent analyst told Fortune that "Alphabet is now funding the AI buildout with shareholders' money instead of its own cash." Fortune also reported that Alphabet announced an $84.8 billion equity raise in June, including $10 billion from Berkshire Hathaway.

Part 2: Nvidia, the arms dealer

No company captures the boom better than Nvidia. For the quarter ended July 26, 2026 (reported August 26), it posted:

• Revenue of $96.2 billion, up 106% year on year and 18% quarter on quarter
• Data-centre revenue of $89.0 billion, up 117%
• Gross margin of 75.0%
• Guidance of $108.0 billion (±2%) for the following quarter, assuming no data-centre compute revenue from China

That growth rate, more than doubling year on year at nearly $100 billion a quarter, has no real precedent for a company of that size. It is the strongest single argument that the boom is driven by real demand.

Nvidia's market value

Nvidia market capitalisation, US$ trillions  (each █ ≈ $0.2T)

Dec 2022   ██                          0.36
Dec 2023   ██████                      1.22
Dec 2024   ████████████████            3.29
Dec 2025   ███████████████████████     4.53
Oct 30 '26 █████████████████████████   4.94

Source: StockAnalysis.com year-end figures; October 30, 2026 close.

Note the shape. Most of the rerating happened in 2023–2024. Since then, Nvidia's market value has grown more slowly than its earnings, so on a price-to-earnings basis the stock has become cheaper even as the company got bigger. Bulls point to this as evidence that this is not 1999-style euphoria. Bears answer that the earnings themselves depend on customers whose spending is debt-funded and may not be sustainable, so a low multiple on peak earnings can be a trap.

Part 3: The frontier labs, from valuations to burn rates

The hyperscalers are rich enough to absorb a disappointment. The pure-play AI labs are where the bubble debate is sharpest, because their valuations depend almost entirely on the future.

OpenAI: funding and valuation history

OpenAI valuation, US$ billions  (each █ ≈ $25B)

Jan 2023   █                                          29
Oct 2024   ██████                                    157
Apr 2025   ████████████                              300
Oct 2025   ████████████████████                      500
Mar 2026   ██████████████████████████████████        852
Sep 2026*  ████████████████████████████████████████████████████████ ~1,400
                                          * reported talks, unconfirmed

Revenue has also grown quickly. OpenAI's reported revenue was $3.7 billion in 2024 and $13.1 billion in 2025. Reuters reported in late September 2026 that its annualised revenue run rate was approaching $70 billion. But OpenAI's commitments are far larger. In late October 2025, Altman said the company had committed to roughly 30 gigawatts of compute at a cost of about $1.4 trillion. Its known contracts include $300 billion of Oracle computing over five years, $250 billion of Microsoft Azure services and a 6-gigawatt AMD chip deal. OpenAI has said it does not expect to be cash-flow positive until 2029. It confirmed a confidential IPO filing in June 2026, and Altman later said the listing would slip to 2027.

Anthropic: funding and valuation history

Anthropic valuation, US$ billions  (each █ ≈ $25B)

Mar 2025   ██                                         61.5
Sep 2025   ███████                                   183
Feb 2026   ███████████████                           380
Jun 2026   ███████████████████████████████████████   965

Anthropic's public S-1 gives the most detailed look yet at a frontier lab's finances, and it supports both sides of the argument. According to Yahoo Finance's reading of the filing:

• Revenue of $16.3 billion in the first half of 2026, against $4.6 billion for all of 2025.
• A net loss of $80.7 billion for the half, which the company attributes mostly to non-cash accounting charges (secondary coverage cites about $72.4 billion from changes in the fair value of convertible notes and similar instruments).
• AI infrastructure spending of $9 billion in the half.
• Compute commitments of more than $430 billion, mainly with Amazon and Google, rising to as much as $637 billion including cancellable and smaller deals.
• Customer concentration: roughly one-fifth of revenue came from two customers.

Its "run-rate" revenue crossed $47 billion in May 2026, per the company. Run-rate figures annualise a recent month or quarter and are not the same as recognised revenue, which is one reason headline numbers for OpenAI and Anthropic should not be compared directly.

xAI and Mistral

Latest private valuations of major labs, US$ billions  (each █ ≈ $25B)

Anthropic (Jun 2026)  ███████████████████████████████████████   965
OpenAI    (Mar 2026)  ██████████████████████████████████        852
xAI       (Feb 2026)  ██████████                                250
Mistral   (Sep 2026)  █                                     ~24 (€21B+)

Mistral's dollar figure is an approximate conversion and an estimate.

The gap between the American giants and Europe's leading lab, about 40 to 1, says something about where capital is concentrated. It also shows how quickly valuations have compounded. Anthropic's valuation rose about 16-fold in roughly 15 months.

The revenue gap

The most widely cited framework for the "spend versus revenue" gap comes from Bain & Company's September 2025 Global Technology Report. Bain estimated that by 2030 the industry would need about $2 trillion in annual revenue to fund the computing that projected AI demand requires, and that even after counting AI-driven savings, revenue would fall about $800 billion short. It also put incremental global AI compute demand at about 200 gigawatts by 2030, with half of that in the United States.

To put that in perspective, the two leading labs' combined run rates, roughly $70 billion for OpenAI and $47 billion-plus for Anthropic, come to well over $100 billion. That is extraordinary growth from almost nothing in 2022, but it is still a small fraction of $2 trillion. Bulls argue that the hyperscalers' cloud revenue, advertising gains and enterprise software should be counted too, and that growth rates of 100% or more close gaps quickly. Bears argue that every year of 2026-scale capex makes the hole deeper, and that chips depreciate in a few years whether or not the revenue arrives.

Part 4: Circular deals, money that goes round in a loop

The feature of this cycle that most worries sceptics is what has come to be called circular financing: suppliers investing in, lending to or guaranteeing their own customers, who then use the money to buy the suppliers' products.

The main deals

Here is the basic loop, simplified:

        ┌──────────── equity / guarantees ─────────────┐
        │                                               ▼
   ┌─────────┐   chip orders    ┌──────────┐   cloud contracts   ┌──────────┐
   │ NVIDIA  │ ◄─────────────── │  CLOUD   │ ◄────────────────── │  AI LAB  │
   │  / AMD  │                  │ (Oracle, │                     │ (OpenAI, │
   └─────────┘                  │  Azure,  │                     │Anthropic)│
        ▲                       │   AWS)   │                     └──────────┘
        │                       └──────────┘                          │
        └───────────────── revenue booked by supplier ◄───────────────┘

Why it matters. When a supplier's investment helps fund its own sales, reported revenue can overstate independent, end-customer demand. If the lab at the centre of the loop struggles to raise its next round, the effect cascades: cloud backlogs shrink, chip orders slow and the supplier's investment loses value at the same time. The IMF's October 2026 Global Financial Stability Report explicitly flagged circular financing, in which AI firms invest in, finance and buy from one another, as a vulnerability.

The defence. Vendor financing is not new, and it is not inherently fraudulent. Nvidia's investments are small compared with its revenue, and the chips are going into data centres that are actually being built and used. Jensen Huang said Nvidia's March 2026 investment "might be the last time" it invests in OpenAI before the company goes public. That suggests at least some awareness of the optics.

The market's verdict. When news of the possible $250 billion Nvidia backstop broke on July 27, 2026, chip stocks sold off hard. CNBC calculated that 20 of the world's most valuable chip stocks lost $1.3 trillion in market value over the following sessions, including $238 billion from Nvidia alone.

Part 5: Debt, SPVs and the shift away from cash

The strongest "not a bubble" argument of 2024 and 2025 was that, unlike the dot-com telecoms, the hyperscalers were paying for AI out of their own enormous cash flows. In late 2025, Guinness Global Investors noted that, excluding Oracle, the other big hyperscalers could fund their forecast capex entirely from operating cash flow. By late 2026 that argument had weakened.

The evidence of rising leverage

• Bond issuance surged. Bank of America's Yuri Seliger counted $121 billion of debt issued in 2025 by Amazon, Google, Meta, Microsoft and Oracle, about four times their average over the previous five years. BofA strategist Savita Subramanian warned in December 2025 that "capex funded by operating cash flow is running out," while arguing that the boom is not a 2000-style bubble but that investors "should get ready for an air pocket."
• Forecasts of much more. Secondary reports in mid-2026 attributed to Morgan Stanley a forecast of roughly $570 billion of global AI-related debt issuance in 2026. We could not open the primary Morgan Stanley document, so treat this as a reported estimate.
• Oracle is the leveraged outlier. CNBC reported that Oracle carries $125 billion of debt, negative free cash flow and a lower credit rating than its rivals. Its shares were down 22% year to date in September even as its backlog hit $664 billion.
• Equity, not just debt. Alphabet's $84.8 billion equity raise in June 2026 shows that even the richest companies are tapping outside capital.

Off-balance-sheet structures

The best-known example is Meta's Hyperion campus in Richland Parish, Louisiana. In October 2025, Meta and funds managed by Blue Owl Capital formed a joint venture valued at about $27 billion to develop and own the campus. Blue Owl holds 80% and Meta 20%. The vehicle issued about $27 billion of investment-grade bonds, largely bought by PIMCO, backed by the project's assets. Meta leases the facilities on an initial four-year term with extension options and provides a residual value guarantee covering the first 16 years. If it walks away, it owes the JV a payment.

Hyperion is planned for 2 GW of AI computing at first, and Mark Zuckerberg has said it could eventually exceed 5 GW.

Why this worries people: the debt does not sit on Meta's balance sheet, yet Meta carries much of the economic risk through the guarantee. Analysts disagree about whether this is prudent project finance (similar to how pipelines and power plants are funded) or a way to make leverage look smaller than it is. The Bank of England's Financial Policy Committee said in October 2026 that AI-related investment is increasingly financed with debt, while valuations remain "heavily dependent" on future earnings expectations.

The depreciation question

A quieter debate concerns accounting lives. AI accelerators improve so fast that a chip bought in 2024 may be economically obsolete well before the five or six years over which some companies depreciate servers. Investor Michael Burry argued in November 2025 that longer useful-life assumptions could understate industry depreciation by roughly $176 billion between 2026 and 2028. That is his estimate, not an established fact. Goldman Sachs, in a September 2026 outlook summarised by secondary sources, said AI capex drove nearly half of this year's earnings growth and that rising depreciation would start to offset that boost in 2027. If chips wear out faster than the books assume, today's earnings are flattered.

Part 6: Power, the physical limit

Software scales without limit. Electricity does not. A full picture of the AI boom has to start with the grid.

How much electricity are we talking about?

US data-centre share of electricity  (each █ ≈ 1 percentage point)

2023 actual (LBNL)        ████▍                  4.4%
2030 low   (EPRI)         █████████              9%
2030 ref.  (LBNL)         ███████████▊           11.8%
2030 high  (LBNL)         ███████████████▎       15.3%
2030 high  (EPRI)         █████████████████      17%

For a sense of scale, CNBC's analysis of US Energy Information Administration data says 10 gigawatts, the size of OpenAI's proposed Ohio campus, is roughly the annual power use of 8 million US households.

Project by project, in gigawatts

Planned capacity of selected projects, gigawatts  (each █ ≈ 0.5 GW)

Pike County OH campus   ████████████████████   10.0
Stargate (7 US sites)   ██████████████████▌    ~9.3+
Hyperion (initial)      ████                    2.0
Doña Ana NM (Stargate)  ████▍                   2.2
Abilene TX (Stargate)   ██▍                     1.2

Bottleneck 1: the interconnection queue

Building a data centre takes about two years. Connecting new power plants to the grid takes much longer. Lawrence Berkeley National Laboratory's Queued Up 2026 edition, published in June 2026, found that about 2,061 GW of generation and storage capacity was waiting to connect to the US grid at the end of 2025 (roughly 1,312 GW of generation and 749 GW of storage), down from about 2,290 GW a year earlier. Most queued projects historically never get built, and the queue reflects a system struggling to plug in new supply fast enough.

Bottleneck 2: gas turbines

Because renewables and nuclear cannot be added fast enough, many AI campuses are turning to natural gas, often generated on site. That has caused a turbine shortage. In October 2025, GE Vernova said its gas-turbine capacity was largely committed through 2028 and that it was mainly booking 2029 orders. Its gas backlog and slot reservations had reached 62 GW, and hyperscalers accounted for about a third of paid reservations. CEO Scott Strazik said turbine prices were "accelerating." Later 2026 reports put GE Vernova's backlog and reservations above 100 GW and cited Mitsubishi deliveries stretching toward 2031, but we were not able to verify those figures in primary documents.

Bottleneck 3: nuclear takes time

Big Tech has become the nuclear industry's most important customer:

• Microsoft and Three Mile Island. Constellation is restarting the former Three Mile Island Unit 1, renamed the Crane Clean Energy Center, under a 20-year power purchase agreement with Microsoft. In March 2026, Constellation said the restart remained on track for the second half of 2027, after a Reuters report said grid operator PJM had pointed to 2031 for full deliverability. Constellation said that date concerned "certainty of full deliverability of its capacity injection rights."
• Meta's 6.6 GW of nuclear. On January 9, 2026, Meta announced deals with Vistra (20-year PPAs for 2,176 MW from the Perry and Davis-Besse plants, rising to 2,609 MW with uprates by 2034), TerraPower (up to eight Natrium reactors, the first two at up to 690 MW as early as 2032) and Oklo (a 1.2 GW campus in Pike County, Ohio, with the first phase possibly in 2030). Together they cover up to 6.6 GW by 2035.
• Small modular reactors. Google signed with Kairos Power for up to 500 MW by 2035, and Amazon led a $500 million round in X-energy, targeting more than 5 GW of new projects by 2039. These were announced in 2024. Most SMR power is due in the 2030s, not in time for today's build-out.

The pattern is clear. Nuclear is the long-term answer, gas is the short-term bridge, and the gap between the two is where the shortages are.

Bottleneck 4: who pays? Rising electricity bills

In PJM, the grid operator serving 13 US states and Washington, DC, the December 2025 capacity auction for 2027/2028 cleared at $333.44 per megawatt-day. That was the price cap and a record for the third straight auction. Total cost was a record $16.4 billion, and PJM ended up about 6,625 MW short of its target reserve margin. PJM said a 5,250 MW increase in its demand forecast was "almost entirely driven by data centers." The Citizens Utility Board's executive director said: "Our capacity market is breaking under the weight of data center demand and a dysfunctional interconnection queue."

At the national level, the EIA's May 2026 Short-Term Energy Outlook expected US residential customers to pay an average of 18.2 cents per kWh in 2026, nearly 5% more than in 2025, with increases of up to 7% a year on the East Coast. The EIA also projected that US commercial electricity sales (which include data centres) would overtake residential sales in 2027 for the first time on record. Utilities also cite fuel costs and storm-hardening as reasons for higher prices, so how much of the rise to blame on data centres remains disputed. Politically, though, "AI is raising my power bill" has become a real issue.

Why power matters for the bubble question: it cuts both ways. A power shortage means capacity cannot be overbuilt as easily as fibre was in 2000, because physics rations it. But it also means projects take longer and cost more, and debt-funded projects are especially sensitive to delays.

Part 7: Markets, the wobbles so far

There has been no crash, but there have been several sharp sell-offs, each linked to a different AI fear.

Single-day drops in AI-sensitive benchmarks (%)  (each █ ≈ 0.5 pt)

Jun 5 2026   Nasdaq 100    ██████████    -4.8
Jan 27 2025  Nasdaq Comp.  ██████▏       -3.1
Oct 30 2026  Nasdaq Comp.  █████▋        -2.86
Jun 5 2026   S&P 500       █████▏        -2.6
Oct 30 2026  S&P 500       ███▋          -1.82

Every sell-off so far has been followed by a recovery to new highs (the S&P 500 closed at 7,580 on October 30, 2026, even after that day's fall). Bulls read this as a sign that the market is resilient and absorbing scepticism in an orderly way. Bears point out that markets also recovered repeatedly in 1998 and 1999 before the 2000 peak.

Part 8: Is AI paying off for the companies that buy it?

The final link in the chain is the enterprise customer. If businesses that buy AI do not see returns, then cloud revenue, lab revenue and eventually chip revenue are all at risk.

The MIT "95%" report and its critics

In August 2025, a report titled The GenAI Divide: State of AI in Business 2025, produced with MIT's NANDA initiative, made headlines by finding that only about 5% of enterprise generative-AI pilots produce rapid revenue acceleration, while the large majority show little or no measurable impact on profit and loss. Fortune reported that the study drew on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments. The report found that purchasing tools from specialised vendors succeeded about 67% of the time, while internal builds succeeded about a third as often. Lead author Aditya Challapally said that "almost everywhere we went, enterprises were trying to build their own tool."

The headline became one of the most quoted statistics of the bubble debate. It also drew pointed criticism. Wharton professor Kevin Werbach, among others, argued that:

• the report does not show the data behind the "95% get zero return" claim;
• its narrowly defined "successful implementation" measure of custom tools is not the same as "zero return";
• the authors themselves describe the figures as "directionally accurate" and based on interviews rather than company reporting;
• the report was hard to obtain and the role of MIT faculty was unclear.

A fair reading is that the report captured something real (most pilots stall, and integration is hard) but that the "95% of companies get nothing" framing went further than its evidence.

McKinsey's 2026 survey

McKinsey's annual survey, The State of AI in 2026: On the Road to ROI (summarised by FM magazine in September 2026), gives a more measured picture:

McKinsey State of AI 2026  (each █ ≈ 5 percentage points)

Say AI boosts productivity  ████████████████   80%
Plan to raise AI spend      ████████████       60%
Scaling enterprise-wide     ████████▊          44%
See any EBIT impact         ███████▍           37%
High performers (≥5% EBIT)  █▏                  6%

The picture is consistent across studies: individuals feel more productive, companies keep spending, but bottom-line gains remain concentrated in a small group of leaders. That is typical of general-purpose technologies. Electricity took decades to show up in factory productivity, because factories had to be redesigned around it. But it is uncomfortable when capex is running at $700 billion-plus a year now.

Part 9: History, from dot-com to railroads

The dot-com comparison

The obvious parallel is Cisco in 2000, the "picks and shovels" supplier of the internet build-out, as Nvidia is for AI.

Cisco figures from Creative Planning (2025) and other analyses. Exact P/E figures vary by source and method.

The comparison cuts both ways. Bulls note that Nvidia's multiples are a fraction of Cisco's, that it is enormously profitable and that the Magnificent Seven (excluding Tesla) traded at a median forward P/E of about 26 in late 2025, according to Guinness, which is elevated but nowhere near 2000 levels. Bears point out that Cisco's business also kept growing after 2000. It was the valuation that collapsed, and the stock took more than 20 years to recover even as the company thrived. A great company can still be a bad investment at the wrong price.

Market-wide, the Bank of England noted in October 2025 that the five largest S&P 500 companies made up about 30% of the index, the highest concentration in 50 years, and that valuations based on past earnings were the most stretched since the dot-com era. It added that forward-looking measures looked less stretched.

The railroad comparison

The longer historical parallel is the railway manias. Economist Andrew Odlyzko has studied British railway investment in detail. His 2010 paper on the 1830s mania argued that it was a giant, wildly speculative and successful investment mania, involving real capital investment comparable, as a share of GDP, to about $2 trillion in the US economy of 2010. The 1840s mania that followed went far worse. Odlyzko describes it as by many measures the greatest technology mania in history, with one of the greatest busts. One popular account puts the decline in railway share prices at about 85% on average by 1850. Yet Britain was left with a national rail network that served it for a century.

How big is AI relative to the economy? Estimates vary hugely depending on what is counted:

The last two rows are secondary-sourced estimates.

On most measures, AI capex as a share of GDP is now in the same range as the 1990s telecom boom, which left behind tens of millions of miles of unused fibre, and below the peak railway years. But it is rising quickly, and GPUs wear out far faster than railway embankments or fibre-optic cable. Guinness noted that adjusting for this faster depreciation makes the current build-out look "slightly more aggressive."

The lesson of history is not that bubbles prevent progress. It is that the people who pay for the infrastructure are often not the people who profit from it. Railway shareholders of the 1840s and fibre investors of 2000 lost fortunes. Society, and later companies like Amazon and Google, benefited from what they built.

Part 10: The warnings and the rebuttals

Who is worried

• The IMF (Global Financial Stability Report, October 2026): AI investment "has helped keep markets and economic growth ticking along," but "the boom could quickly turn to bust if the promised profits fail to materialize." The report flagged concentration, rising tech borrowing and circular financing.
• The Bank of England (October 2025): "The risk of a sharp market correction has increased," with equity valuations looking stretched, "particularly for technology companies focused on AI." Its October 2026 record shifted the emphasis toward debt.
• Jefferies' Chris Wood warned of "massive capital destruction" from US AI spending, as quoted by The Register in October 2026.
• Sam Altman (August 2025): investors are "overexcited."
• Bain & Company (September 2025): an $800 billion annual revenue shortfall by 2030 even under generous assumptions.

Who says it isn't a bubble, or not yet

• Bank of America's Savita Subramanian (December 2025): "Is this 2000? Are we in a bubble? No," but expect "an air pocket."
• BlackRock's Jean Boivin: "We don't think the bubble framing is that useful at this stage for investors."
• Goldman Sachs (September 2026 outlook, as summarised by secondary sources): the feared "earnings bubble" is unfounded, though earnings are unusually high and market pricing does not assume a bust.
• Morningstar's Michael Field, on the July 2026 chip sell-off: the decline was "driven largely by sentiment rather than fundamentals."
• The demand data itself: Nvidia's guidance, cloud backlogs and AI labs' revenue growth all show buyers who are paying real money.

Bull case vs bear case

A bubble scorecard

We rated each classic bubble feature on what is visible in October 2026. This is editorial judgement, not a measurement.

On balance, the evidence describes a capex super-cycle with bubble-like pockets, especially in private lab valuations, vendor financing and the leverage now entering the system. It is not a pure mania detached from earnings, as parts of 1999 were. But it does contain the ingredients that turn a slowdown into a bust.

What to watch next

1. 2027 capex guidance in the January–February 2027 earnings season. A first-ever cut would be a regime change.
2. The Anthropic IPO pricing and any OpenAI listing timeline. Public-market pricing will test private valuations.
3. Free cash flow at Alphabet, Amazon and Meta. How long markets tolerate negative FCF.
4. Credit spreads and bond-deal demand for hyperscaler and data-centre debt.
5. Nvidia's guidance, especially any slowdown in sequential growth.
6. Power milestones: Abilene reaching 1.2 GW, the Crane restart in 2027 and the next PJM auction.
7. Enterprise ROI data: whether the 37% "any EBIT impact" figure finally moves.

Timeline: the AI boom and the bubble debate

• Nov 2022: ChatGPT launches.
• Jan 2023: Microsoft invests a further $10 billion in OpenAI at a ~$29 billion valuation.
• Dec 2023: Nvidia's market value reaches about $1.22 trillion.
• 2024: Big-four capex reaches $246 billion. Google signs Kairos (Oct 2024). Amazon backs X-energy.
• Jan 27, 2025: DeepSeek sell-off. Nvidia loses roughly $589–593 billion in a day.
• Apr 2025: IEA projects data-centre demand doubling to ~945 TWh by 2030. OpenAI raises $40 billion at $300 billion.
• Aug 2025: Altman says investors are "overexcited." The MIT NANDA "95%" report is published.
• Sep 2025: Nvidia announces up to $100 billion for OpenAI (Sep 22). Oracle signs a $300 billion OpenAI contract. Anthropic raises at $183 billion. Bain publishes the $2T/$800B analysis.
• Oct 2025: AMD–OpenAI 6 GW deal (Oct 6). BoE warns of correction risk (Oct 8). Meta–Blue Owl Hyperion JV. OpenAI's $1.4 trillion / 30 GW commitments.
• Nov 2025: Burry short disclosure; Nasdaq's worst week since April. Nvidia/Microsoft invest up to $15 billion in Anthropic.
• Dec 2025: PJM capacity auction clears at its cap ($333.44/MW-day).
• Jan 9, 2026: Meta announces 6.6 GW of nuclear deals.
• Feb 2, 2026: SpaceX acquires xAI in an all-stock deal ($250 billion for xAI).
• Feb 2026: Hyperscalers guide ~$650 billion of 2026 capex. Anthropic Series G at $380 billion. EPRI raises its 2030 range to 9–17%.
• Mar 2026: OpenAI closes $122 billion at $852 billion.
• Apr 2026: IEA reports 17% growth in data-centre electricity in 2025.
• May–Jun 2026: Anthropic Series H at $965 billion. OpenAI and Anthropic file confidentially for IPOs. Alphabet's $84.8 billion equity raise. SpaceX IPO (Jun 12). LBNL's queue and data-centre reports. Nasdaq 100 −4.8% on June 5.
• Jul 2026: Reports of a $250 billion Nvidia backstop talk; chip stocks shed $1.3 trillion. Alphabet raises 2026 capex guidance to $195–205 billion.
• Aug 26, 2026: Nvidia reports $96.2 billion quarterly revenue.
• Sep 2026: Oracle backlog hits $664 billion. Mistral raises €3 billion at >€21 billion. OpenAI reportedly seeks funding at ~$1.4 trillion.
• Oct 13, 2026: IMF warns that the AI boom "could quickly turn to bust."
• Oct 21–29, 2026: Meta, Alphabet, Microsoft and Amazon report; three of them show negative free cash flow.
• Oct 26, 2026: Anthropic files its S-1 publicly.
• Oct 30, 2026: Nasdaq −2.86%, its worst day since June.

Frequently asked questions

Is AI in a bubble right now? There is no consensus. Respected institutions such as the IMF and the Bank of England see bubble-like risks, especially in leverage and concentration. Banks such as Bank of America and Goldman Sachs argue it is not a 2000-style bubble. The most defensible view is that there is a genuine boom with speculative pockets, and that the financing has become riskier during 2026.

How much are tech companies spending on AI? The big four US hyperscalers' 2026 capex guidance adds up to roughly $740–760 billion (our arithmetic), up from $246 billion in 2024. Oracle, the AI labs and neocloud providers add more on top.

Is any of this money making a profit? Yes, for the suppliers. Nvidia's gross margin is about 75%, and Google Cloud's operating margin was 36.2% in Q3 2026. The AI labs are not yet profitable on a net basis, and several hyperscalers' free cash flow has turned negative because of the spending.

What is a "circular deal"? An arrangement where a supplier invests in, lends to or guarantees a customer that then buys the supplier's products. Examples include Nvidia's investments in OpenAI, and AMD's warrant giving OpenAI the right to up to 160 million AMD shares alongside a 6 GW chip purchase.

What would pop an AI bubble? Possible triggers include a failed funding round at a major lab, a hyperscaler cutting capex, a sharp rise in interest rates, a cheaper-model shock like DeepSeek, credit markets refusing to fund data-centre debt, or power and permitting delays that leave debt-funded projects without revenue.

How does this compare to the dot-com bubble? Valuation multiples at the leading companies are much lower than in 2000, and today's leaders are highly profitable. But market concentration is higher, and financing structures (SPVs, vendor financing) echo the telecom boom of the late 1990s.

Why does electricity matter so much? AI data centres need enormous, continuous power. US data centres could use 9–17% of the nation's electricity by 2030 (EPRI). Grid connections, turbines and nuclear plants take years to build, so power delays can push back revenue while debt costs keep accruing.

Are electricity bills going up because of AI? Partly, and the amount is disputed. The EIA expects US residential prices to rise nearly 5% in 2026. In PJM, data-centre demand drove capacity prices to their cap, though utilities also cite fuel and grid-hardening costs.

Did the MIT study really show 95% of AI projects fail? It reported that about 5% of enterprise generative-AI pilots achieve rapid revenue gains. Critics argue that the evidence behind the "95% get zero return" framing was thin. McKinsey's 2026 survey found 37% of firms see some EBIT impact.

What does history say? Transformative technologies often go through investment manias. Britain's 1840s railway mania and the 2000 dot-com bust destroyed investor wealth but left behind valuable infrastructure. A bust would not mean AI "fails," only that investors overpaid for it.

Should I sell my tech stocks? We cannot give financial advice. Anyone worried about exposure should consider their own time horizon and diversification, and consult a qualified adviser.

Sources and further reading

All URLs below were opened during research for this article (October 2026).

Company results and filings

• Nvidia Q2 FY2027 results: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027
• Meta Q3 2026 results: https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx
• Amazon Q3 2026 results: https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Third-Quarter-Results/default.aspx
• Alphabet Q4 2025 results (SEC 8-K exhibit): https://www.sec.gov/Archives/edgar/data/1652044/000165204426000012/googexhibit991q42025.htm
• Fortune, Alphabet Q3 2026: https://fortune.com/2026/10/28/alphabet-google-stock-drops-cloud-revenue-doubles-free-cash-flow-negative/
• CNBC, Microsoft FY27 Q1: https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html
• CNBC, Oracle FY27 Q1: https://www.cnbc.com/2026/09/10/oracle-orcl-q1-earnings-report-2027.html
• Market Chameleon, Meta full-year 2025: https://marketchameleon.com/articles/b/2026/01/29/meta-delivers-strong-growth-2025-expense-surge-and-bold-2026-plans
• EcomCrew, Amazon Q4 2025: https://www.ecomcrew.com/amazon-q4-2025-earnings-aws-ai-growth/

Capex totals

• Irish Times (FT), Big Tech capex 2023–2024: https://www.irishtimes.com/business/2025/02/07/big-tech-lines-up-over-300bn-in-ai-spending-for-2025
• Yahoo Finance, $650B 2026 guidance: https://finance.yahoo.com/news/big-tech-unveils-650-billion-121205995.html
• Yahoo Finance, $725B after Q1 2026: https://finance.yahoo.com/markets/article/magnificent-7-earnings-rush-reveals-ai-spending-surge-with-hyperscaler-capex-set-to-reach-725-billion-in-2026-224901707.html
• StockAnalysis, Nvidia market cap history: https://stockanalysis.com/stocks/nvda/market-cap/

AI labs

• Entrepreneur, OpenAI $122B round: https://www.entrepreneur.com/business-news/openai-just-closed-the-biggest-funding-round-in-history
• Pulse 2.0, OpenAI talks at $1.4T: https://pulse2.com/openai-reportedly-targets-at-least-30-billion-funding-at-1-4-trillion-valuation/
• Built In, Anthropic Series H: https://builtin.com/articles/anthropic-raises-65b-series-h-965b-valuation-20260601
• Yahoo Finance, Anthropic S-1: https://finance.yahoo.com/technology/article/anthropic-publicly-files-ipo-prospectus-noting-strong-revenue-growth-and-massive-spending-131807138.html
• Orrick, Mistral Series D: https://www.orrick.com/en/News/2026/09/Orrick-Advises-Mistral-in-its-3B-Series-D-at-21B-Post-Money-Valuation
• Wikipedia, OpenAI: https://en.wikipedia.org/wiki/OpenAI
• Wikipedia, Anthropic: https://en.wikipedia.org/wiki/Anthropic
• Wikipedia, SpaceXAI (xAI): https://en.wikipedia.org/wiki/SpaceXAI
• Bain & Company, Global Technology Report 2025 press release: https://www.bain.com/about/media-center/press-releases/20252/%25242-trillion-in-new-revenue-needed-to-fund-ais-scaling-trend---bain--companys-6th-annual-global-technology-report

Circular deals and financing

• CNBC, Nvidia–OpenAI $100B (Sep 2025): https://www.cnbc.com/2025/09/22/nvidia-openai-data-center.html
• CNBC, Nvidia–OpenAI $250B backstop talks (Jul 2026): https://www.cnbc.com/2026/07/27/nvidia-and-openai-in-talks-for-up-to-250-billion-dollar-ai-backstop.html
• The Register, AMD–OpenAI deal: https://www.theregister.com/2025/10/06/openai_and_amd_join_forces
• Bisnow, Meta–Blue Owl Hyperion JV: https://www.bisnow.com/news/national/data-center-capital-markets/meta-pushes-its-largest-data-center-project-off-its-books-with-27b-joint-venture-131490
• Fortune, BofA "air pocket" and AI debt: https://fortune.com/2025/12/03/is-ai-a-bubble-bofa-says-air-pocket-in-2026-data-center-debt
• Guinness Global Investors, "Are we in an AI bubble?": https://www.guinnessgi.com/insights/are-we-in-an-ai-bubble

Power and energy

• Epoch AI, Stargate sites: https://epoch.ai/publications/openai-stargate-where-the-us-sites-stand
• PV Tech, IEA Energy and AI: https://www.pv-tech.org/ai-data-centre-demand-to-more-than-double-by-2030/
• Rigzone, IEA 2026 update: https://www.rigzone.com/news/data_center_electricity_use_surges-20-apr-2026-183486-article/
• LBNL, US Data Center Energy Usage Report 2025 Update: https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025
• Data Center Knowledge, EPRI 2026: https://www.datacenterknowledge.com/build-design/epri-report-us-data-center-grid-strain-casts-cloud-over-ai-race
• Data Center Knowledge, GE Vernova turbines: https://www.datacenterknowledge.com/energy-power-supply/ge-vernova-sees-gas-turbine-sales-to-hyperscalers-surging
• World Nuclear News, Meta nuclear deals: https://world-nuclear-news.org/articles/meta-announces-landmark-agreements-for-new-nuclear
• Local 21 News, Crane/TMI restart: https://www.local21news.com/news/local/2027-restart-of-nuclear-reactor-grid-at-the-crane-still-on-track-three-mile-island-nuclear-power-crane-clean-energy-pa-pennsylvania
• Utility Dive, PJM capacity auction: https://www.utilitydive.com/news/pjm-interconnection-capacity-auction-data-center/808264/
• Utility Dive, EIA residential rates: https://www.utilitydive.com/news/data-centers-electricity-residential-rates-eia/820342/

Markets and warnings

• Yahoo Finance (Reuters), Oct 30, 2026 sell-off: https://finance.yahoo.com/markets/stocks/articles/nasdaq-futures-sink-amazon-intel-114737736.html
• CNBC, July 2026 chip sell-off: https://www.cnbc.com/2026/07/29/chip-selloff-sk-hynix-samsung-softbank.html
• The National, June 5, 2026 sell-off: https://www.thenationalnews.com/business/markets/2026/06/05/nasdaq-100-slides-48-as-traders-dump-tech-stocks/
• Kathmandu Post (Reuters), DeepSeek sell-off: https://kathmandupost.com/world/2025/01/28/deepseek-sparks-ai-stock-selloff-nvidia-posts-record-market-cap-loss
• The Register, IMF warning: https://www.theregister.com/ai-and-ml/2026/10/13/ai-bubble-could-burst-and-take-global-economy-with-it-imf-warns/5302781
• Yahoo Finance UK, Bank of England warning: https://uk.finance.yahoo.com/news/bank-of-england-ai-stock-market-bubble-risk-report-125937642.html

Enterprise ROI

• Fortune, MIT NANDA report: https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo
• Critique of the MIT report (Kevin Werbach, mirrored): https://neonvagabond.xyz/mirrors/mit-report-lacks-evidence.html
• FM magazine, McKinsey State of AI 2026: https://www.fm-magazine.com/news/2026/sep/companies-financial-value-from-ai-holds-firm-in-2026/

History

• Creative Planning, Cisco vs Nvidia: https://creativeplanning.com/insights/investment/cisco-nvidia-lessons-from-past/
• Andrew Odlyzko, "This Time Is Different" (railway mania): https://experts.umn.edu/en/publications/this-time-is-different-an-example-of-a-giant-wildly-speculative-a/
• A Wealth of Common Sense, railway bubble vs AI: https://awealthofcommonsense.com/2025/11/the-railway-bubble-vs-the-ai-bubble/