Is AI another dot com bubble waiting to burst?

Two loud camps, both partly right

Everyone has an opinion on this one, and most of them are loud. Half the internet is convinced it is 1999 all over again, that the whole thing is a house of cards, and that when it falls it will take the pension funds with it. The other half thinks anyone who doubts AI will be out of a job by Christmas. Both camps are exhausting, and both are partly right, which is the annoying bit.

So let us think about it properly, without hype and without doom: where the money is going, where the value actually is, and what the last great technology bubble teaches a business deciding where to spend on AI.

This is not investment advice. I say nothing about which shares or funds to buy, sell or hold. The question is narrower and more useful: if you run a business, what does the history mean for your own AI budget? Every market figure is dated, because this was written in June 2026, and comes from the original source.

What the dot com crash actually looked like

The Nasdaq Composite, the American index packed with technology shares, closed at 5,048.62 on 10 March 2000. On 9 October 2002 it closed at 1,114.11. It did not close above its 2000 peak again until 23 April 2015, more than 15 years later.¹

The horizontal axis shows each year from 1995 to 2006, the vertical axis the index level at the last close of that year.¹ The takeaway is the shape: four years up, three years down, and a recovery so slow that someone who bought at the top waited a decade and a half to get back to even.

The fall is easiest to grasp as a percentage change:

$$\Delta\% = \frac{x_{\text{new}} - x_{\text{old}}}{x_{\text{old}}} \times 100$$

Here $\Delta\%$ is the percentage change, $x_{\text{old}}$ is the starting value and $x_{\text{new}}$ is the value you compare it with. If percentages feel rusty, Foundations: percentages builds them from scratch.

Worked example: why a 78% fall needs a 353% rise

Step one, the fall from the March 2000 close to the October 2002 close:

$$\Delta\% = \frac{1{,}114.11 - 5{,}048.62}{5{,}048.62} \times 100 = -77.9$$

So $x_{\text{old}} = 5{,}048.62$, $x_{\text{new}} = 1{,}114.11$, and the index lost 77.9% of its value.

Step two, the rise needed to get back. Now the starting point is the bottom:

$$\Delta\% = \frac{5{,}048.62 - 1{,}114.11}{1{,}114.11} \times 100 = 353.2$$

This time $x_{\text{old}} = 1{,}114.11$ and $x_{\text{new}} = 5{,}048.62$. A fall of 78% needs a rise of 353% to undo it. That asymmetry is why bubbles hurt so much: the arithmetic of recovery is brutal, whatever the technology underneath is worth.

How stretched are prices now, compared with then

The usual yardstick for "are shares expensive" is the cyclically adjusted price to earnings ratio (CAPE), published by the economist Robert Shiller. It divides the price of the S&P 500, the broad American share index, by the average of ten years of earnings after inflation. In December 1999 it reached 44.2, the highest value in his data, which runs back to 1881. In May 2026 it stood at 40.1. Across the whole record the average is about 18.²

An example: a CAPE of 40 means buyers pay €40 for every €1 of average yearly profit over the past decade. At 18, they pay €18. Neither number says when prices fall, or whether they will. It only says that, as of May 2026, the broad American market was priced closer to the 1999 peak than to its own long run average. That is the strongest evidence on the bubble side, and it is a market wide number, not an AI number.

The money question, because that is what people actually mean

When someone asks if AI is a bubble, they are asking a financial question: is the money going in justified by the money coming out?

The four largest spenders publish their numbers every quarter. In the three months to 31 March 2026, measured by each company's own cash flow statement:³⁴⁵⁶

The horizontal axis names the company, the vertical axis shows what it spent on property and equipment (data centres, servers, chips, buildings) in that one quarter, in billions of US dollars. Amazon's and Alphabet's figures are purchases of property and equipment, Microsoft's are additions to property and equipment, and Meta's include principal payments on finance leases, each as the company itself reports it.³⁴⁵⁶ The takeaway: together, about 130.6 billion dollars in three months. The same four spent about 72.7 billion dollars in the first quarter of 2025, so the pace rose by about 80% in a year. At the March 2026 pace, that is about 520 billion dollars a year.

Two more facts from the same releases. Meta raised its own forecast for 2026 spending to between 125 billion dollars and 145 billion dollars.⁶ And Amazon's free cash flow, the cash left after running the business and paying for the build out, fell to 1.2 billion dollars for the twelve months to March 2026, from 25.9 billion dollars a year earlier, because spending on property and equipment, net of proceeds from sales and incentives, rose by 59.3 billion dollars.⁴

Now hold that next to the revenue

Few companies report AI revenue separately. Microsoft gives a number: in April 2026 it said its AI business had passed an annual revenue run rate of 37 billion dollars, up 123% on the year before.⁵ A run rate is the current month's or quarter's revenue scaled up to a year.

An example, using only Microsoft's own figures. Its additions to property and equipment were 30.9 billion dollars in one quarter.⁵ At that pace, the yearly spend is:

$$30.9 \times 4 = 123.6$$

Here $30.9$ is one quarter's spending in billions of dollars and $4$ is the number of quarters in a year. So about 123.6 billion dollars a year of spending against about 37 billion dollars a year of AI revenue, a ratio of roughly 30%. Two caveats keep this fair. Not all of that spending is for AI, and revenue growing at 123% a year closes gaps quickly if it keeps going. But it shows the shape of the bet: the spending runs well ahead of the revenue, and the industry is betting that the revenue catches up, the way broadband eventually used the fibre laid in the late nineties.

The machines wear out quickly

There is a detail in the annual reports that rarely makes the headlines. Microsoft depreciates computer equipment over two to six years.⁷ A data centre building lasts decades. The servers and chips inside it do not.

An example: a business buys €120,000 of servers and writes them off in equal parts over four years. Each year, €30,000 of their value disappears from the books:

$$120{,}000 \div 4 = 30{,}000$$

Here $120{,}000$ is the purchase price in euro and $4$ is the years of use. To be worth it, those servers must earn back €120,000, plus running costs, before they are replaced, and then the business buys the next set. The bet has to pay off within the life of the machines, which is why the timing of the revenue matters so much.

The case that it is a bubble

The pessimists deserve a proper hearing. First, the spending is justified by projections, not by profits, as the figures above show. Second, some of the money goes round in a circle.

An example, with invented companies. A chip maker, Alpha, invests €100 million in an AI lab, Beta. Beta uses the money to buy €100 million of Alpha's chips. Alpha books €100 million of sales. Beta shows €100 million of investment in computing power. And if Beta's next funding round values it higher, Alpha can book a gain on its stake. On paper, three good news stories. In cash, no new money arrived from outside the two companies. Nothing in this is improper, and the chips are real. But the growth figures are telling you less than they seem to.

Real deals have this shape, and they are public. In September 2025 NVIDIA announced a letter of intent, not yet a final contract, under which it intends to invest up to 100 billion dollars in OpenAI as at least 10 gigawatts of NVIDIA systems are deployed.⁸ In October 2025 AMD announced a 6 gigawatt agreement to supply OpenAI with its chips, and issued OpenAI a warrant for up to 160 million AMD shares that vests as milestones are reached.⁹ And Amazon's net income for the first quarter of 2026 included 16.8 billion dollars of pre tax gains from its investments in Anthropic.⁴ All of it is properly disclosed. It is also exactly the kind of loop to watch when you read a growth number.

Third, the cost of running these models never stops. Every answer costs electricity and hardware time. In the dot com days, a website, once built, served its next visitor for almost nothing. A large AI model is the opposite shape: expensive to build, and expensive again every time someone uses it. Where the price a customer pays does not cover the cost of serving them, growth makes the loss bigger, not smaller.

And the hype has detached from reality in places. When a pitch deck gains value by adding the letters AI, a market is pricing a story, not a product. That is the smell of 1999.

The case that it is not that kind of bubble

Now the other side, because it is stronger than the doomers like to admit.

In 1999, many of the companies with silly valuations had little revenue and no profit, just a domain name and a plan. The largest AI spenders today fund the build out from businesses that already make money: search, advertising, cloud, software and retail. Alphabet's revenue for the first quarter of 2026 was 109.9 billion dollars and Amazon's 181.5 billion dollars.³⁴ If AI spending were cut in half tomorrow, these companies would still have their day jobs.

And the technology is used, by people and by businesses. Alphabet reports that its own models were processing more than 16 billion tokens a minute through direct API use by customers, up 60% on the previous quarter.³ (A token is a small piece of text, roughly part of a word.) In the UK, the Office for National Statistics found that about a quarter (25%) of businesses were using some form of AI in late December 2025, rising to 44% among businesses with 250 or more employees.¹⁰ Ireland is on a similar curve. The Central Statistics Office found that 20.2% of Irish enterprises with ten or more people used some form of AI in 2025, up from 15% in 2024, and 57.7% of large enterprises with 250 or more did.¹¹ The two surveys ask in different ways, so the figures are a direction, not a like for like comparison.

A bubble built on something nobody wants pops at once. A bubble built on something people use can deflate slowly and still leave an industry standing, which is exactly what the internet did after 2000.

Are we actually using AI correctly?

Here is where I start giving out, because I think the answer is mostly no. The industry has fallen in love with the wrong use cases.

The instinct across the industry has been to cram a new kind of tool into familiar shapes. A chatbot gets bolted onto everything. The bank has one, the energy supplier has one, and often it is worse than the button it replaced. The most flexible technology in a generation, mostly used to answer emails slightly faster and summarise documents nobody was going to read.

Electricity went through the same phase. Factories at the end of the nineteenth century ran on one central steam engine that drove every machine through shafts and belts. The first step with electricity was to put a big electric motor where the engine had been. The large gains came later, when factories gave each machine its own motor and redesigned the whole building around that idea. Between 1880 and 1930, American manufacturing moved from shaft and belt drive to motors on individual machines, and that change, not the motor alone, raised output per worker and per unit of capital.¹² AI is in the big central motor phase: a new thing doing the old thing slightly better.

The useful question is not "how do I add AI to my product". It is "what could I never do before that is possible now". Almost everyone is asking the first.

An example: an invented wholesaler in Waterford, Suir Valley Supplies (suirvalleysupplies.example), receives 200 orders a week as PDF attachments, and two people type them into the order system. The first idea on the table is a chatbot on the website. It answers questions nobody was asking, and the 200 PDFs still get typed. The second idea asks what was blocked before: reading every PDF, filling in the order, and sending only the unclear ones to a person. Same technology, a different question, and only the second one changes the business.

Most AI startups recycle old ideas with a fresh coat

There is a whole category of company that people unkindly call a wrapper. The product is a thin interface on top of somebody else's model, doing something you could largely do yourself if you knew the prompt. Note takers, email writers, generic chatbots for this or that industry. They recycle ideas that already existed, and the only new ingredient is the model underneath, which they do not own and which their competitor can rent just as easily.

The problem for these companies is that they have no moat. If the whole product is a clever prompt and a nice login screen, then the moment the underlying model improves, the product either becomes redundant or the model provider ships it as a free feature. Founders call this getting steamrolled, and it is the defining risk of building on a platform that improves faster than you do.

But the wrapper sneer is a bit smug. A shop sits on top of a wholesaler, and a restaurant is a wrapper around ingredients you could buy yourself. Sometimes the value is in the trust, the specific workflow, the dull integration work nobody else wants to do. A wrapper that understands one narrow industry, speaks its language and slots into its existing tools can survive, because the moat was never the model. The moat was knowing the customer better than the model provider ever will.

An example, with two invented Limerick startups. Quillwise (quillwise.example) sells a general "AI email writer" for €12 a month, and has nothing left to sell once the big email clients add the feature for free. Harbourline (harbourline.example) builds software for independent letting agents: it reads tenancy paperwork, fills in tenancy registration forms and flags clauses for a person. Same rented model underneath. What Harbourline owns is the knowledge of how a letting agency works, and that does not ship in a model update.

So the honest split: most AI startups recycle old ideas and will not make it. A minority do something that could not be done before, or wrap the model in so much hard won domain knowledge that the wrapper becomes the valuable thing. On the surface they look identical, which is what makes a bubble hard to read in the moment.

Where AI is actually, undeniably useful

This is the part I actually enjoy, because under the hype there is a solid core of things that work, and they share a shape. They are specific, unglamorous solutions to problems that used to be blocked by one of three things: they were too small to be worth building, they needed more thinking power than people have, or they cost too much to do by hand.

Small custom solutions that never needed to scale

Countless small problems inside real businesses were never worth solving, because a custom tool for a dozen users cost more than doing the job by hand forever. AI changes that sum. A small firm can now build the internal tool that reads its particular messy spreadsheets, sorts its particular kind of customer email, or drafts the one odd report its regulator wants every quarter. It does not need to scale. It needs to save eleven people two hours a week each.

Worked example: is the small tool worth building?

An invented lettings agency in Kilkenny, John's Quay Lettings (johnsquaylettings.example), has 11 staff who each spend two hours a week copying details from emailed documents into their property system. The figures below are the example's assumptions, chosen to be plausible, not quotes.

Step one, the hours saved. 11 people times two hours is 22 hours a week. Over 46 working weeks:

$$22 \times 46 = 1{,}012 \text{ hours a year}$$

Here $22$ is the hours saved each week and $46$ is the number of working weeks.

Step two, the value of those hours. At an assumed €22 an hour, including employer costs:

$$1{,}012 \times 22 = 22{,}264$$

So $1{,}012$ hours at $22$ euro an hour is about €22,264 a year of time back.

Step three, the costs, all of them. Building the tool: €6,000, once. Running it (the model usage, called inference): €40 a month, €480 a year. Someone checking the output and fixing what it gets wrong: two hours a week, which over 46 weeks at €22 is €2,024 a year.

Step four, the first year:

$$22{,}264 - 6{,}000 - 480 - 2{,}024 = 13{,}760$$

So the first year leaves $13{,}760$ euro.

Step five, every year after, without the build cost:

$$22{,}264 - 480 - 2{,}024 = 19{,}760$$

So every later year leaves $19{,}760$ euro. That is a sound investment, and a dull one. Note that the checking and the running costs are in the sum. Leave them out and the case looks better than it is.

Problems that were blocked by thinking power

Some problems were never blocked by money or ambition. They were blocked by the fact that a human brain, or a large team of them, cannot hold enough variables at once.

Protein folding is the poster child. Working out the shape a protein folds into could take a researcher years for a single structure. You could not win by hiring more graduate students, because it was a problem of pattern recognition across a search space far bigger than any mind can grip. In 2024 half of the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper "for protein structure prediction", the work behind AlphaFold.¹³ The AlphaFold Protein Structure Database, run by EMBL's European Bioinformatics Institute, now gives open access to more than 200 million predicted structures.¹⁴

That is not a chatbot writing emails. It is a new key to a locked door, and where the bottleneck is pattern rather than effort, AI is a new capability, not a slightly better tool.

Problems that were blocked by cost

Then there is the large middle ground of things that were always possible but never worth the money. Translating a small company's website and help pages into many languages. Giving every customer a personal reply instead of a template. Reading every contract a mid sized firm signs, instead of spot checking a few and hoping.

When a task becomes much less expensive, you do not just save money, you unlock things that were off the table. A small company can behave, in some respects, like a large one. That shift of capability down to the small players is, to me, one of the most exciting and underrated stories in the field, and it gets drowned out by arguments about whether a chatbot is conscious.

An example: an invented Cork shop, Blackrock Ceramics (blackrockceramics.example), never translated its 400 product pages because the likely orders did not justify a translator. Machine translation with a native speaker reviewing it changes the answer from "never" to "test one market and measure".

Why the companies building AI still employ so many people

The story in 2025 was clean and frightening: AI writes code now, so fewer people are needed to write code, so the layoffs. Some of that was real. But the filed numbers of the largest AI spenders tell a calmer story.

Alphabet had 194,668 employees at the end of March 2026, against 185,719 a year earlier.³ Meta's headcount was 77,986, up 1% on the year.⁶ Amazon reported 1,575,000 full and part time employees, also up 1%.⁴ Microsoft employed about 228,000 people full time at the end of June 2025, 80,000 of them in product research and development.⁷

Worked example: Alphabet's headcount change

Using the percentage change formula from the first section, with $x_{\text{old}} = 185{,}719$ and $x_{\text{new}} = 194{,}668$:

$$\Delta\% = \frac{194{,}668 - 185{,}719}{185{,}719} \times 100 = 4.8$$

So $\Delta\% = 4.8$: the company spending tens of billions a quarter on AI grew its workforce by 4.8% in the same year.

Two things matter here for the bubble question.

The first is that headcount staying flat does not mean the jobs stayed the same. Some routine work can now be assisted or automated, and it is fair to suspect that some cuts were a correction of pandemic era hiring, dressed in fashionable AI language. What stays in demand is the senior work: designing systems, fixing what AI breaks and building the actual products. In the UK, the ONS found that only 4% of businesses using AI reported that their overall headcount had fallen because of it.¹⁰

The second is the quiet admission inside all of this. If AI could replace software engineers wholesale, the companies that build the models would be the first to stop employing them. They have every incentive and the best tools. Instead, Alphabet, Amazon and Meta each employ more people than a year earlier, and Microsoft reports 80,000 in product research and development. The tools are very good at assisting a skilled person and nowhere near replacing one. If you want one honest signal of how mature the technology is, ignore the press releases and watch who the model builders keep on the payroll.

Where the true costs are hidden

This is the part the glossy demos never show, and where a lot of the risk lives, for the market and for your own budget.

The first hidden cost is inference, the cost of running the model every time someone uses it. Training makes the headlines, but inference is forever. A product priced below its cost to serve is betting that chips get less expensive, or that prices can rise later without customers leaving. That is the same bet many dot com companies made right before the music stopped.

The second hidden cost is the one nobody puts on a slide: the cost of being wrong. An AI that is right 95% of the time sounds fantastic, until you remember that the wrong 5% arrives in exactly the same confident tone. So a person has to check the output, which eats part of the promised saving. The useful way to price this is the expected cost of the errors:

$$E[C] = n \times p \times c$$

Here $E[C]$ is the expected cost of the errors, $n$ is the number of items the AI produces, $p$ is the share of them that is wrong, and $c$ is the average cost of one error that reaches a customer. The idea of an expected value, a probability times a cost, is set out in Foundations: expected value.

Worked example: pricing the wrong 5%

An invented online tea shop in Galway, Mutton Island Tea Co (muttonislandtea.example), wants AI to write 2,000 product descriptions. All figures are the example's assumptions.

Step one, without any checking. With $n = 2{,}000$, $p = 0.05$ and an assumed $c = 15$ euro per wrong description (a return, a refund, a complaint about a caffeine claim):

$$E[C] = 2{,}000 \times 0.05 \times 15 = 1{,}500$$

So $E[C] = 1{,}500$ euro of expected damage.

Step two, with checking. A person spends two minutes on each description, which is 4,000 minutes, about 66.7 hours. At €25 an hour:

$$66.7 \times 25 = 1{,}667.5$$

So $66.7$ hours at $25$ euro an hour is about €1,668 to check everything, which removes most of that €1,500 of risk.

Step three, against writing by hand. At 15 minutes each, 2,000 descriptions take 500 hours, which at €25 is €12,500. AI plus checking still wins comfortably here.

Step four, change one number. Now the AI drafts contract clauses for a law firm, and an error that slips through costs $c = 20{,}000$ euro. Even one wrong clause in a thousand ($p = 0.001$, $n = 1{,}000$) gives:

$$E[C] = 1{,}000 \times 0.001 \times 20{,}000 = 20{,}000$$

So $E[C] = 20{,}000$ euro, from a single expected error. In law, medicine, accounting or engineering, the checking is not optional, and its cost decides whether AI saves anything at all.

The third hidden cost is people and process. Bringing AI into a company is not plug and play, whatever the vendor says. Someone has to clean the data, connect the systems, retrain the staff, rewrite the workflows and watch the thing when it drifts. The licence fee is often the smallest number in the project.

And there is the cost that lands on no single company's books: electricity. The International Energy Agency estimates that data centres used about 415 terawatt hours of electricity in 2024, around 1.5% of the world's consumption, and projects that this will more than double to around 945 terawatt hours by 2030.¹⁵ Part of that bill is paid by everyone, through energy prices and strained local grids, whether they use AI or not.

Why tokenmaxxing is quietly rotting the leaderboards

This one is more technical, so let me build it up gently. The scoreboards used to decide which AI is best can be gamed, and one way of gaming them makes models slower and more expensive for reasons that have nothing to do with being more useful to you.

Labs prove a model is the best with standard tests called benchmarks: maths problems, coding challenges, exam questions. The score goes in a chart, the chart goes in the marketing. The trouble is an old observation, usually called Goodhart's law: once a number becomes the target, people improve the number rather than the thing it was meant to measure. The anthropologist Marilyn Strathern traced the same effect in the rating of British universities, where assessment, in her words, "has a life of its own".¹⁶

Contamination: the exam paper was in the study notes

Models learn from vast sweeps of the internet, and benchmark questions and answers often sit on the internet too. So a model may have seen the exam paper before sitting the exam. One study masked a wrong option in multiple choice questions from the MMLU benchmark and asked models to fill the gap. ChatGPT and GPT 4 guessed the exact missing option 52% and 57% of the time, which is hard to explain unless the questions were in their training data.¹⁷ You have measured memory and called it intelligence.

Tokenmaxxing: thinking out loud for the scoreboard

The second problem is subtler. Models often score higher when they are allowed to reason through a long chain of intermediate text before answering. Research on this "test time compute" shows that spending more computation per question can lift results enough for a smaller model to beat one 14 times its size on some problems.¹⁸ The name for pushing that as far as it goes is tokenmaxxing: maximising the tokens a model chews through to win a benchmark point.

Researchers have a politer name for the side effect, overthinking: models that spend a great deal of computation on simple problems for little or no gain in accuracy.¹⁹ The title of that paper says it best: "Do NOT Think That Much for 2+3=?".

Here is why that should annoy a user rather than impress them. Every extra token costs money and time. The cost of a month of answers is:

$$K = q \times t \times \frac{r}{1{,}000{,}000}$$

Here $K$ is the monthly cost, $q$ is the number of questions a month, $t$ is the average number of tokens the model produces per answer, including its reasoning, and $r$ is the price per million tokens. It is a plain rate multiplied out, as in Foundations: ratios and rates.

Worked example: the verbose model on the support desk

The invented Mutton Island Tea Co puts an AI assistant on its customer service, answering 20,000 questions a month. It compares two models, both assumed to cost €8 per million output tokens (an assumed price for the example, not a quote). Model A, top of the leaderboard, produces 3,000 tokens per answer. Model B produces 800, and gives answers that are just as good for questions like "where is my order".

Model A:

$$K = 20{,}000 \times 3{,}000 \times \frac{8}{1{,}000{,}000} = 480$$

So $K = 480$ euro a month for model A.

Model B:

$$K = 20{,}000 \times 800 \times \frac{8}{1{,}000{,}000} = 128$$

So $K = 128$ euro a month for model B. That is €480 against €128, a difference of €352 a month or €4,224 a year, and model A's customers also wait longer for each answer. The benchmark says A is better. The bill and the queue say otherwise.

It is also a bad incentive for the whole field: models pushed towards being more expensive to run, just when the open question is whether anyone can run them profitably. When the next chart shows model X beating model Y by two points, the right response is a shrug and three questions: better at what, measured by whom, and at what cost per answer?

So, bubble or not? A grown up verdict

Yes, there is very likely a bubble in parts of the market. Prices across the American market sat, as of May 2026, closer to the 1999 peak than to their long run average. The spending runs far ahead of the reported AI revenue, some of the money goes round in circles, and a crowd of wrapper companies will not survive a correction. When it comes, a lot of people who said AI could not fail will discover they were describing a share price, not a technology.

And no, it is not only a bubble. Underneath the froth is a technology that businesses and people use every day and will not give back, a core of problems in science and in thousands of ordinary businesses that AI has made solvable for the first time. That core does not disappear when share prices fall. It gets quietly more valuable while everyone watches the crash.

The dot com era is the right guide precisely because both halves were true. The crash was real and it ruined people. The internet was also real and it went on to reshape the economy. The bubble bursting and the technology mattering are not opposing predictions. They are the same story told at two different speeds.

For a business deciding where to spend, the history turns into one practical filter.

The diagram reads from the top. The first question separates ideas that change something from ideas that decorate an old process. The second prices the hidden costs before you spend. An idea that passes both is worth building whatever the share prices do next, and one that fails either was never worth building, bubble or no bubble.

The winners will be the ones who stopped asking how to sprinkle AI on top of what already exists, and started asking what became possible that never was before. Everyone else is laying fibre nobody will light, and buying an expensive front row seat for the pop.

Footnotes

1. Nasdaq OMX Group, NASDAQ Composite Index (NASDAQCOM), daily closes 1995 to 2016, retrieved from FRED, Federal Reserve Bank of St. Louis, https://fred.stlouisfed.org/series/NASDAQCOM

2. Robert J. Shiller, Online data: US stock market prices, earnings and the cyclically adjusted price earnings ratio (CAPE), monthly since 1871, data file ie_data.xls, values for December 1999 and May 2026, https://shillerdata.com/

3. Alphabet Inc., Alphabet Announces First Quarter 2026 Results, earnings release, 29 April 2026, https://s206.q4cdn.com/479360582/files/doc_financials/2026/q1/2026q1-alphabet-earnings-release.pdf

4. Amazon.com, Inc., Amazon.com Announces First Quarter Results, earnings release, 29 April 2026, https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-First-Quarter-Results/

5. Microsoft Corporation, Earnings Release FY26 Q3, press release, 29 April 2026, https://www.microsoft.com/en-us/investor/earnings/fy-2026-q3/press-release-webcast

6. Meta Platforms, Inc., Meta Reports First Quarter 2026 Results, press release, 29 April 2026, https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-First-Quarter-2026-Results/default.aspx

7. Microsoft Corporation, 2025 Annual Report (Form 10 K for the fiscal year ended 30 June 2025), sections Human capital resources and Property and equipment, https://www.microsoft.com/investor/reports/ar25/index.html

8. NVIDIA, OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems, press release, 22 September 2025, https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems

9. AMD, AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs, press release, 6 October 2025, https://ir.amd.com/news-events/press-releases/detail/1260/amd-and-openai-announce-strategic-partnership-to-deploy-6-gigawatts-of-amd-gpus

10. Office for National Statistics, Business insights and impact on the UK economy: 8 January 2026, Business Insights and Conditions Survey wave 147, 15 to 28 December 2025, https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/bulletins/businessinsightsandimpactontheukeconomy/8january2026

11. Central Statistics Office, Information Society Statistics, Enterprises 2025, chapter Artificial Intelligence, CSO statistical release, 6 February 2026, https://www.cso.ie/en/releasesandpublications/ep/p-isse/informationsocietystatistics-enterprises2025/artificialintelligence/

12. Warren D. Devine Jr., From Shafts to Wires: Historical Perspective on Electrification, The Journal of Economic History 43(2), Cambridge University Press, 1983, https://www.cambridge.org/core/journals/journal-of-economic-history/article/from-shafts-to-wires-historical-perspective-on-electrification/500078D9B4764BA1109A7967437CF226

13. The Nobel Prize, The Nobel Prize in Chemistry 2024, summary, NobelPrize.org, 2024, https://www.nobelprize.org/prizes/chemistry/2024/summary/

14. EMBL European Bioinformatics Institute and Google DeepMind, AlphaFold Protein Structure Database, https://alphafold.ebi.ac.uk/

15. International Energy Agency, Energy and AI, executive summary, IEA, 2025, https://www.iea.org/reports/energy-and-ai/executive-summary

16. Marilyn Strathern, Improving ratings: audit in the British University system, European Review 5(3), Cambridge University Press, 1997, pages 305 to 321, https://www.cambridge.org/core/journals/european-review/article/improving-ratings-audit-in-the-british-university-system/FC2EE640C0C44E3DB87C29FB666E9AAB

17. Chunyuan Deng, Yilun Zhao, Xiangru Tang and others, Investigating Data Contamination in Modern Benchmarks for Large Language Models, arXiv:2311.09783, 2023, https://arxiv.org/abs/2311.09783

18. Charlie Snell, Jaehoon Lee, Kelvin Xu and Aviral Kumar, Scaling LLM Test Time Compute Optimally can be More Effective than Scaling Model Parameters, arXiv:2408.03314, 2024, https://arxiv.org/abs/2408.03314

19. Xingyu Chen, Jiahao Xu, Tian Liang and others, Do NOT Think That Much for 2+3=? On the Overthinking of o1 Like LLMs, arXiv:2412.21187, 2024, https://arxiv.org/abs/2412.21187