Search results for: “unix”

The dot LLM era is one of the chunkiest posts that I have written, so I have put it together in a PDF as well that you can download and share freely amongst colleagues and peers.

The dot LLM era executive summary

The “dot LLM era” represents a pivotal moment in technological history, drawing striking parallels to the dot-com bubble of the late 1990s. This period is defined by a massive influx of capital into Large Language Models (LLMs) and artificial intelligence infrastructure, which represents clear analogues to the dot-com era “three bubbles” framework: online businesses, open-source ventures, and telecommunications (which represents closest analogue to the current dot LLM era). 

The Core Thesis

The current $1 trillion valuation of the AI sector faces two existential challenges:

• Amortisation Risk: Unlike the dark fibre of the 1990s, which had a useful life of over a decade, modern GPU and TPU hardware becomes technically obsolete within 3 to 5 years.
• Self-Defeating Economics: If AI-driven automation successfully provides $1 trillion in cost savings through job cuts, the resulting increase in unemployment and drop in GDP could destroy the very macroeconomic environment required to sustain hyperscaler growth.

A Tale of Three Bubbles

The document argues that we are conflating three distinct historical analogues:

• Online Businesses: Recalling the “burn rates” of the early web, where pure-play LLMs are currently providing tokens for less than their marginal cost.
• Open-Source: Comparing current model proliferation to the rise of Linux, where the ultimate winners may not be the model creators but those providing enterprise-grade support.
• Telecommunications: The most instructive analogue, involving massive infrastructure build-outs, vendor financing, and potential “Minsky moments” where optimism outstrips sustainable cash flow.

Geopolitical and Economic Realities

Unlike the 1990s “Long Boom” characterized by US pre-eminence and budget surpluses, the dot LLM era exists within a climate of high government debt and inflation. Furthermore, US dominance is challenged by Chinese hyperscalers and open-source models like Alibaba’s Qwen, which offer high performance at significantly lower costs.

Potential Outcomes

The document outlines seven possible scenarios for the era’s conclusion, ranging from The Breakthrough (total economic transformation) to The Weird Gizmo (total collapse). Currently, “The Moral Hazard”—where AI is deemed “too big to fail” and receives government backing—is viewed as the most likely path (~95% likelihood).

How this dot LLM exploration started?

This dot LLM post came out of a number of ideas and vibes. 

Everyone^[i] from commentators^[ii] and podcast hosts to friends are talking about a dot-com-type bubble in LLMs, what I’ve termed as shorthand the dot LLM era. The dot LLM era comparison has become a steady tempo of concern. 

The term AI bubble took off in interest during September of 2025. 

The dot LLM era is shorthand to move backwards and forwards in time comparing the current AI boom with the dot-com boom of 1990s – 2001. It’s a very different type of ‘Y2K trend’. 

Many pure-play LLMs customers are currently getting to use tokens for less than their marginal cost^[iii],  and this is part of the reason (alongside the high cost of model training) why the likes of OpenAI, C3.ai, Perplexity, Anysphere and Anthropic are raising new rounds of financing^[iv]. They have been losing money^[v] and continue to do so. 

Spending by both pure-play LLMs and their hyperscaler partners is driven by the effort to create an AI moat^[vi]. An AI moat is a sustained proprietary advantage derived from a company’s use of artificial intelligence that makes its offerings fundamentally superior, cheaper, or “stickier” than those of rivals, and which is hard to be replicated by rivals.

Even the most historically bullish institutional investors, like James Anderson^[vii], formerly of Baillie Gifford, have turned bearish on Nvidia and pure-play LLM offerings.

To meet the needs of these services, development of an extra 1,500 data centres has been announced – only a quarter of which are under construction at the time of writing.^[viii]

It is a time reminiscent of the mid-2010s when venture capitalists subsidised the cost of services like Uber and Lyft^[ix] to grow markets from the ground up. Going back further to the dot-com era, Amazon took a similar approach with its business. 

Valuations for the Magnificent 10:  Apple, Alphabet, Amazon, AMD, Broadcom, Meta, Microsoft, Nvidia, Palantir and Tesla — are high. The 24-month forward P/E ratio of the Magnificent 10 is 35 times. By comparison the S&P 500’s equivalent P/E ratio at the peak of the dot-com boom approached 33^[x], with a brief peak at the market top of 44.^[xi]

Built into these Magnificent 10 valuations, is an assumption that LLMs will help them cut costs and or drive revenue growth by $1 – 4 trillion in the next two years.^[xii]

Like the dot-com era^[xiii], the dot LLM era is spawning several businesses that are likely to be considered weird gizmos or bad business ideas that will be mocked in the future. The dot-com analogues included the likes of proto-digital currencies Beenz and Flooz^[xiv], CueCat^[xv] – a bar code scanner that allowed web users to scan codes on magazines to get more pages online or the short-lived^[xvi] 3Com Audrey^{[xvii][xviii]} and Sony eVilla^[xix] internet appliances. 

(Disclosure: in my first agency-side role, I worked on 3Com’s consumer products and the Palm device business that was spun off as palmOne^[xx] to give space for the Ergo connected home internet appliance range. Audrey’s ability to sync with two Palm devices^[xxi], despite Palm being seen as an internal competitor, gives you an idea of how disjointed and chaotic internal planning was in companies like 3Com when they were trying to move at ‘internet speed’. One of the last 3Com projects I worked on was the launch of Audrey in October 2000.) 

Bubbles don’t kill technology from moving forwards

Like the dot-com era, the dot LLM era is likely to move through two separate cycles: one financial and the other technological. While the financial bubble destroyed a lot of shareholder value, the underlying web technology cycle and use cases became commonplace and evolved. Email became part of our culture^[xxii] in the same way that social media became cultural fabric a decade later. LLMs or their successors (such as nested models^[xxiii] and world models^[xxiv]) are likely to be influential and change the nature of work, life, business and culture. 

Already we can see the dot LLM era playing out on social media as over half of content is estimated to be produced with generative AI. 

This relentless forward progress for technological adoption and refinement was likened to an organic being by author Kevin Kelly in a phenomenon he called the ‘technium’.^[xxv]

Believing that AI is undergoing a dot LLM bubble isn’t the same as not believing that the technology won’t have an ongoing impact. 

A Tale of Three Bubbles

When we talk about the dot LLM era we are conflating a number of related bubbles bursting. 

The bubbles were based around a common conceit: prior experience counted for naught because the internet changed everything. 

This resulted in three distinct historical bubbles:

• Online business bubble
• Open-source bubble
• Telecommunications bubble

The one that most people recall is the dot-com boom where online businesses went under.

Online businesses

Iconic ones included technically ambitious clothing retailer Boo.com, pet care supplies firm Pets.com and many more. 

Boo.com burned through $135 million in just 18 months^[xxvi]. And they weren’t the only ones. In March 2000, Pegasus Research put out a research paper^[xxvii] outlining the burn rates of each online business. The report went under-reported at the time, but took a clear-eyed look at the sector. 

Successful business people failed. Podcaster and academic Scott Galloway^[xxviii] founded RedEnvelope^[xxix], an online commerce site that sold gifts including personalised items and experiences.  Bob Geldof’s online travel site deckchair.com^[xxx] doesn’t even merit a mention in most profiles of the famous musician. 

Back when I worked at Yahoo! long-time employees said that only a pivot to provide dating services had kept the rest of Yahoo! Europe afloat during the dot-com bust of 2001/ 2002. Online advertising revenues at the time dropped more than 30% over a 12-month period. The difference between success and failure was a very narrow gap.

Amazon survived and eventually thrived as it managed to convince its shareholders to defer profitability for a decade to garner growth. That move and the company’s nascent web services business (AWS) led to the online juggernaut that Amazon is today^[xxxi]. While Amazon was founded in 1994 and first went online in 1995, it didn’t make its first quarterly profit until the end of 2001^[xxxii] of $5 million on revenue of $1.12 billion^[xxxiii] and the first annual profit in 2003^[xxxiv]. Uber and Lyft learned from the example that Amazon had set a decade earlier. 

Open-source bubble

The second bubble was the ‘open-source’ bubble. The rise of the commercial web (and the millennium bug^[xxxv]) disrupted existing technology stacks and opened up new opportunities to sell enterprise computing hardware and software. Several companies were launched to support the rollout of open-source software that threatened Microsoft’s and Unix operating system duopoly. 

My former client VA Linux Systems built web servers and workstations optimised for Linux users^[xxxvi]. Now VA Linux Systems is remembered more for its IPO, which valued the company at $30 and opened for trading at $299^[xxxvii]. Red Hat^[xxxviii] and SuSE^[xxxix] provided commercially supported versions of Linux for corporate enterprises. Like their online business counterparts, few of the open-source business bubble companies could be considered ‘successful’, the outlier being Red Hat which eventually sold to IBM in 2019 for $34 billion^[xl]. 

The winner, Red Hat, didn’t sell the open-source software (Linux) as its business model; it sold enterprise-grade support, integration, and services.

While the open-source bubble was the smallest of the three bubbles, it had an outsized impact with Linux being the foundation for everything from the Android mobile OS to the largest data centres. 

Telecoms bubble

The telecoms bubble was the least visible, yet most spectacular bubble and the one that is most instructive about the dot LLM era. 

There are three places where you could start the telecoms bubble. April 30, 1995, when the NSFnet was decommissioned^[xli], the Telecommunications Act of 1996, or 1984. 

I am going to go with 1984^[xlii]. While the internet was growing in academic and military circles in the US and there were nascent computer networks elsewhere like the UK^[xliii] – the real revolution was happening on the London Stock Exchange. The UK government under prime minister Margaret Thatcher looked to get the government out of businesses. A programme of privatisation took place to sell-off numerous nationalised businesses; plans to privatise British Telecom were proposed in 1982.  1984 saw the IPO of British Telecom plc, the previously government owned telecoms provider^[xliv]. The UK government also licensed the first competitor Mercury Communications^[xlv]. 

From a technological perspective the IPO seemed to be a catalyst^[xlvi] for wider telecoms deregulation in western Europe^[xlvii] and around the world. In 1985, the Japanese government privatised NTT and opened the Japanese telecommunications market up to competition^[xlviii]. The European Commission began developing a regulatory framework to open up national telecoms markets in 1987^[xlix], Europe and Japan would spend the next decade opening up their markets for alternative telecommunications services. 

It was into this global landscape that the US overhauled its telecommunications regulations with the Telecommunications Act of 1996^[l]. The stated intention of the act was to “let anyone enter any communications business – to let any communications business compete in any market against any other.”^[li] The act incentivised the expansion of networks and new services across the US.^[lii] Early US netizens rejected the act as a way to regulate cyberspace^[liii]. 

The following year 69 members of the World Trade Organisation (WTO) agreed to open their basic telecoms markets to competition^[liv]. 

In parallel with the wider atmosphere of telecommunications liberalisation, was the rise of the internet. The rise of home computers in US households between 1990 and 1997 grew from 15% to 35%^[lv]. At that time, a small percentage of people would be dialling directly into work, nascent online services like CompuServe or AOL, dialling into their Charles Schwab account and bulletin boards. 

Outside the US, it was more likely that your computer was a standalone machine with a spreadsheet, word processing application, maybe design software allowing you to write the document from home and bring it in to work on floppy drive, or possibly an Iomega diskette^[lvi] of some sort. 

Private long distance optical fibre networks together with free local telephone calls were the infrastructure for internet connectivity. The web the way we know it now was not a surefire winner^[lvii]. Much speculation was on the internet superhighway – digital cable television with value added services like online shopping.^[lviii] Bill Gates at the peak of his power as CEO of Microsoft was convinced that the digital cable TV was the way forward.^[lix] The next edition was edited to reflect the reality of the web instead. The open interoperable nature of the web proved to be more attractive than walled garden digital services envisaged by cable TV companies.^[lx]

Investment in telecoms infrastructure increased to meet the future needs of digital services, based on a misreading of internet data traffic growth^[lxi]. US telecoms providers invested $500 billion between 1996 and 2001 – mostly on optical fibre networks.^[lxii]Much of this spending was done by new entrants including Global Crossing, WorldCom, Enron, Qwest and Level 3. There was a corresponding scale up by equipment makers like Lucent to supply the telecoms providers.^[lxiii] Telecommunications equipment companies Lucent and Nortel^[lxiv] both provided vendor financing for their dot-com era client base – engineered in such a way to inflate sales figures and their share price.^[lxv]

• Lucent lent customers the money to purchase their equipment. They then booked the loan value as revenue, even though the repayment risk remained and the debt was held as an asset on the Lucent balance sheet. 
• Nortel used its own shares as financing for its customers. It is believed that Nortel lent $7 billion+ to help start-up telecommunications carriers make equipment purchases. Many of these were unsecured loans, interest-free and tied to future purchases. 

Carriers engaged in ‘round-tripping’. Global Crossing would ‘sell’ network capacity to Qwest; Qwest would ‘sell’ similar capacity back to Global Crossing for nearly the same amount. Both companies booked the deals as revenue. US regulators found that this was a pre-arranged swap designed to inflate revenue, despite having no commercial purpose.  

Had the bubble continued into 2005, WorldCom CEO at the time Bernie Ebbers had expected to invest another $100 billion in the company’s network infrastructure that year^[lxvi]. Instead, Ebbers left WorldCom investors with a $180 billion loss. When the telecoms bubble imploded, an estimated trillion dollars in debt was owed, much of which was not expected to be recovered.^[lxvii]

In 2002, the telecoms bubble helped change the way business is conducted. In reaction to a number of major corporate and accounting scandals, notably Enron^[lxviii] and WorldCom – US lawmakers enacted the Sarbanes-Oxley Act of 2002^[lxix]. This Act (SOX as it became known) mandated standards in financial record keeping and reporting for public companies. It covered responsibilities of the board of directors and criminal penalties for certain practices^[lxx]. It required the SEC to create regulations for compliance. SOX drove up the cost of a company going public and remaining public due to the administrative burden to remain legally compliant. 

Technology vendor financing from companies like Cisco and IBM continued to be an issue through the 2008 financial crisis,^[lxxi]but was largely kept out of the common discourse by the tsunami of sub-prime mortgage debt defaults. 

The dot LLM era hinges around service providers and equipment makers, in the same way that the telecoms bubble did. Here are some examples and their dot LLM analogues. 

Service providers	Equipment makers
Dot-com era Enron PSINet Qwest UUNET Worldcom	Dot-com era 3Com Ciena Cisco Equinix Juniper Networks Lucent Sun Microsystems
Dot LLM era Alphabet Amazon Anthropic OpenAI Oracle Microsoft Salesforce	Dot LLM era AMD Applied Materials ASML Broadcom Huawei Intel Micron Nvidia Samsung TSMC

Of course, the idea of them being analogues doesn’t line up perfectly. While the excessive build out of optical fibre networks could be considered analogous to hyper-scaled AI infrastructure; it isn’t a perfect match.  The acceleration in network and computing capability in hyperscalers show the kind of positive trajectory that Mary Meeker had in her dot-com era analyst presentations^[lxxii]. 

Some critics think that the massive acceleration in network and compute investment for LLM purposes represents a Minsky moment in itself^[lxxiii] – heralding it as an event that fits Hyman Minsky’s Financial Instability Hypothesis.

Minsky considered this coming in three parts:

1. A self-reinforcing boom driven by optimism and easy credit
2. A shock, that can be minor in nature, has investors re-look at cash-flow shortfalls 
3. Rapid asset sales and deleveraging / de-risking

The scale of investment and construction of data centres together with the new electricity generating capacity to power them are orders of magnitude larger than the telecoms boom.  

Secondly, the LLM infrastructure has a much shorter life. LLM hyperscalers go through GPUs (and TPUs) extremely fast with a useful life of 3 years or so.^[lxxiv] Complete technical obsolescence of a given GPU / TPU design has occurred by 5 years from launch.^[lxxv]

Therefore, if there is an AI bust the processors wouldn’t be available to use in the next economic upswing in the tech sector. By comparison the optical fibre networks laid during the dot-com boom had a useful life of 10+ years and the growth of web 2.0 and social startups was largely built on surplus server and networking equipment left over from the dot-com era. The dot LLM era represents a financial and technological amortisation risk.

There is an added wrinkle in this last point about the useful life of GPUs and TPUs. Company filings of hyperscalers show that they are amortising their network and compute capital expenditure over longer times, by lengthening the assumed useful lives of components in their financial paperwork. 

The economic environment.

The economic conditions that the dot-com era happened in were very different to the conditions of the dot LLM era. 

The US had suffered through much of the 1980s and into the early 1990s. Reaganomics had driven a ‘jobless recovery’ as the financial and services sectors took over from manufacturing as the US economic growth engine. In 1989 the Savings and Loan crisis peaked.^[lxxvi] This occurred alongside rising interest rates to battle inflation. An oil price spike as a result of the first Gulf War exacerbated economic conditions and the recession ended the ambitions of George H. Bush becoming president for the second time. Under a new government, by spring 1994, jobs and economic growth both picked up. 1996 saw growth continuing and by May 1997 US unemployment dropped below 5% for the first time in 24 years.  

Other countries had similar recessions in the late 1980s and early 1990s due to restrictive monetary policies, oil prices and the end of the Cold War. By 1994, global GDP growth returned.^[lxxvii] Wired magazine talked of the 1980s as a contagious idea:^[lxxviii]

America is in decline, the world is going to hell, and our children’s lives will be worse than our own. The particulars are now familiar: Good jobs are disappearing, working people are falling into poverty, the underclass is swelling, crime is out of control. The post-Cold War world is fragmenting, and conflicts are erupting all over the planet. The environment is imploding—with global warming and ozone depletion, we’ll all either die of cancer or live in Waterworld. As for our kids, the collapsing educational system is producing either gun-toting gangsters or burger-flipping dopes who can’t read.

In the same article, they thought of the 1990s as the start of ‘The Long Boom’ – 25 years of prosperity freedom and a better environment for the world. 

By 2000, the US government went from running a budget deficit eight years earlier to running a surplus. This eased the credit markets for businesses and consumers. The US Taxpayer Relief Act lowered marginal capital gains tax and helped fuel stock market investments. Day trading became a thing by 1999,^[lxxix] mirroring investors in crypto and stocks in the 2020s.^[lxxx]

By comparison, the current economic climate is more similar to the 1980s than the 1990s. Government debt has reached new heights. Governments have struggled to rein in inflation created by COVID-era supply shocks – which was responsible for several governments including the Biden administration being voted out of office. The high government debt and inflation governments with fewer policy tools to manage a systemic shock compared to their 1990s counterparts. The Economist claimed that western countries had government debt levels unseen since Napoleonic times.^[lxxxi] There is no US government budget surplus and little ‘headroom’ for monetary policy.

Wired magazine’s ‘contagious idea’ sounds very familiar:

• Climate despair has been recognised as a condition by mental health professionals.^[lxxxii]
• Global warming is cited^[lxxxiii] as a cause of extreme weather conditions^[lxxxiv].
• Good jobs are disappearing and this is often blamed^[lxxxv] on generative AI. 
• US tariffs, Brexit and the Ukraine war are disrupting global commerce. 

In conclusion, the dot-com era economy was much more conducive for retail investors than the dot LLM era is. 

The internet changes everything

Dot-com businesses had it right in their view that the internet would change business and shopping for consumers and enterprises. Some of them like Amazon made it, many didn’t. The investment bank analysts believed it too.^[lxxxvi]

You see similar things being written about AI now, along with similar looking ‘hockey stick’ charts.^[lxxxvii]

Microsoft research^[lxxxviii] suggests that there is a strong link between GDP per capita and AI usage. But also notes that adoption in advanced economies tends to plateau between 25% and 45%, suggesting non-economic factors eventually moderate growth. Suggesting that the dot LLM era may not be the kind of game-changer that it might be believed to be by advocates. I would recommend that the reader keeps an open mind on this rather than automatically thinking that this proves generative AI as being a technological dead-end. More work is required to try and understand why the plateau happens and whether it represents a ceiling or a brief rest before adoption accelerates again. 

Artificial general intelligence or AGI

AGI is when the LLM surpasses your average human. The idea of AGI has taken on the similar messianic fervour of people from the dot-com era including George Gilder’s Telecosm. Many executives in the most prominent LLM developers subscribe to an imminent AGI occurring. 

Elon Musk holds the most aggressive timeline^[lxxxix]. He thinks that the main bottlenecks to AGI—specifically power supply and high-end chip availability—are being solved rapidly. Through his company’s xAI’s computing power, he believes that the next generation of models will surpass human intelligence in almost any individual task by early 2026. Anthropic’s CEO Dario Amodei believes that AGI could arrive in 2026/7^[xc]. OpenAI’s Sam Altman considers 2027 to be a realistic timeline for the arrival of AGI^[xci]. DeepMind co-founder Shane Legg has come up with a notional timeline of 2028. His view is based on the current rate of progress for both computing hardware and LLM algorithms.^[xcii] Long time AI advocate Ray Kurzweil has published a series of books about AGI, which he termed the ‘singularity’. The latest of which put 2029 as the year in which AGI is likely to occur^[xciii]. 

As with any cultural artefact, AGI has become blended with religious thinking, as exemplified by this outlandish quote from podcaster Joe Rogan. 

“Jesus was born out of a virgin mother. What’s more virgin than a computer? If Jesus does return, you don’t think he could return as artificial intelligence? AI could absolutely return as Jesus.” – Joe Rogan^[xciv]

All of which is reminiscent of Timothy Leary’s infatuation with the early web^[xcv] and the Heaven’s Gate Cult^[xcvi]. 

Despite some prominent advocates, many experts in the field are sceptical about the imminent arrival of AGI. Included in these sceptics are OpenAI co-founder Andrej Karpathy who believes that the nature of LLMs mean that AGI won’t arrive using current techniques and on the timeline that advocates predict^[xcvii]. Researchers Rodney Brooks^[xcviii] and Yann LeCun^[xcix] believe that understanding the physical world is critical for technology to achieve AGI. This work is only starting now. Academic Melanie Mitchell argues that until systems can grasp ‘meaning’ AGI will not happen^[c]. 

The good bubble

Some of the most important US business executives of the LLM era admit that we are in some kind of bubble. Here’s what they’ve said in their own words. 

“When bubbles happen, smart people get overexcited about a kernel of truth … Are we in a phase where investors as a whole are overexcited about AI? My opinion is yes.”^[ci]

“This frenzy gives us pause … The belief in an A.G.I. or superintelligence tipping point flies in the face of the history of technology”^[cii]

“This is a kind of industrial bubble … investors have a hard time in the middle of this excitement, distinguishing between the good ideas and the bad ideas. And that’s also probably happening today.”^[ciii]

“Given the potential of this technology, the excitement is very rational. It is also true when we go through these investment cycles there are moments we overshoot as an industry. We can look back at the internet right now, there was clearly a lot of excess investment, but none of us would question if the internet was profound or did it have a lot of impact it was fundamentally changed how we work digitally as a society. I expect AI to be the same; I think it’s both rational and there are aspects of irrationality to a moment like this.”^[civ]  

“Most other infrastructure buildouts in history, the infrastructure gets built out, people take on too much debt, and then you hit some blip … a lot of the companies wind up going out of business, and then the assets get distressed and then it’s a great opportunity to go buy more … definitely a possibility that something like that would happen here.”^[cv]

The real question is whether the dot LLM era is a ‘good’ bubble or a bad bubble? What does a good bubble look like? And how much will it cost? Most of the quotes above see the dot LLM era as similar in nature to the internet boom and bust. While pioneers may have died society was irrevocably changed. 

Some of the irrationality in the ‘good bubble’ hypothesis seems to include hubris, for example OpenAI shunned having external advisers to work on its $1.5 trillion worth of data centre deals.^[cvi] While OpenAI has relationships with investment banks and corporate law firms – it didn’t make much use of them.

These explanations assume that there will be a corresponding surplus of infrastructure that will spark new innovation on the backs of dead companies. A concept that most represents the telecoms aspect of the dot-com era. The explanations ignore the financial losses suffered by pension funds and retail investors as these companies went bankrupt. They also ignore that the useful life of AI computing hardware is obsolete faster than railway tracks or laid fibre optic cables.^[cvii] Short sellers have accused hyperscalers of estimating unrealistically useful lives for their computer equipment, in particular, the GPUs that power AI model training and inference. The allegations claim that profits are artificially overstated by allowing depreciation of assets over a longer period.^[cviii]

At its peak in March 2000, the NASDAQ index peaked at 5,048. When the dot-com bubble burst the index declined to 1,139. Recovery took 15 years from the peak value. The NASDAQ reached 5,048 again in March 2015.^[cix] The risk is arguably greater this time around as the top ten stocks constituting the S&P 500 index constitute 40% of its value.^[cx] This implies a vulnerable brittle market environment prior to any economic bust. So, the idea of ‘good’ is very narrowly defined and asking the term to do a lot of heavy lifting in terms of its language. Predicting the peak of the market^[cxi] is challenging too^[cxii]. 

Can the demand for LLM; grow at the speed implied by invested capital?

Advertising as a possible use case

The first use case to consider for how the dot LLM era could meet its full ‘potential’ would be the ongoing disruption of advertising by digital platforms. Depending who you believe the global total market for advertising is close to, or has just exceeded $1 trillion in total value. 

Globally, advertising represents about 1 percent^[cxiii] of global GDP. It usually holds at around that proportion as global economic growth waxes and wanes. In some key markets such as the US, UK and Singapore – it makes up a higher percentage of GDP – as the home of advertising platforms, advertising agencies with international responsibility and technology suppliers to the industry. 

Advertising isn’t just a cost centre for businesses, but also a driver of economic growth and profit. One Euro of advertising is estimated to generate up to 7 Euros of economic value.^[cxiv]

It took digital advertising over a quarter of a century to go from zero to over half of advertising spend. This hinged around two growth spurts, one in 2000 with the rise of online businesses and the second in 2020 with the COVID-19 lockdown. A factor of the transition of digital advertising growth has been down to the fragmentation of audiences across media platforms and alongside traditional media.

AI (but not LLMs) has been used in advertising as long as digital advertising has been around. It started to be used for understanding consumer behaviour and delivering targeted advertising.^[cxv] Amazon started using AI for its recommendations in 1998.^[cxvi]

Not all economic value in digital advertising accrued from the transfer of ‘traditional’ advertising to digital advertising. There is evidence of a direct correlation between a rise in e-commerce drives a decline in retail properties, given the strong linkage between e-commerce, retail media search advertising – there is part of that value exchange which would accrue to the advertising platforms. 

… one percent increase in e-commerce sales as a percent of total sales will decrease commercial real estate prices by 7.64%.^[cxvii]

It is worthwhile reading the whole economic paper on the decline in commercial real estate prices to understand the multiple factors that the author tried to take into account to better understand the impact of e-commerce sales. 

The sales didn’t only shift online, but offshore. For instance, China-based advertisers accounted for around 11% of Meta’s total revenue in 2024^[cxviii], which amounted to $18.35 billion. A significant portion of this is believed to come from large e-commerce companies like Temu and Shein^[cxix], rather than a large number of small businesses. These companies benefited from the Chinese state support^[cxx] covering their international logistics and postage costs and allowed their businesses to be run on razor-thin margins. 

There has also been a corresponding value transfer from the lost profits of advertising clients to the platforms as well. Advertising industry consultant Michael Farmer made this point in his discussion of large fast-moving consumer goods businesses. 

…for the fifty years from 1960 to 2010, the combined FMCG sales of P&G, Unilever, Nestle and Colgate-Palmolive grew at about an 8% compounded annual growth rate per year.

The numbers associated with this long-term growth rate are staggering. P&G alone grew from about $1 billion (1960) to $79 billion in 2010. Throughout this period, P&G was the industry’s advocate for the power of advertising, becoming the largest advertiser in the US, with a focus on traditional advertising — digital / social advertising had hardly begun until 2010. Since 2010, with the advent of digital / social advertising, and massive increases in digital / social spend, P&G, Unilever, Nestle and Colgate-Palmolive have grown, collectively, at less than 1% per year, about half the growth rate of the US economy (2.1% per year).

They are not the only major advertisers who have grown below GDP rates. At least 20 of the 50 largest advertisers in the US have grown below 2% per year for the past 15 years.

Digital and social advertising, of course, have come to dominate the advertising scene since 2010, and it represents, today, about 2/3rds of all advertising spend.^[cxxi]

Digital advertising at its heart represents marketing efficiency because of its ability to be created and ‘trafficked’ at a much lower cost and greater speed. But this efficiency comes at the cost of corresponding marketing effectiveness in terms of short-term sales and longer-term preference and purchasing impact. 

LLMs could undoubtedly further refine marketing efficiency, it could even ‘understand’ the marketing effectiveness challenge. But LLMs are restricted by the way the audience interacts with advertising, limiting their ability to solve the corresponding marketing effectiveness challenge. Marketing conglomerate WPP have launched a performance media platform that looks to further increase marketing efficiency by no longer requiring a traditional client-agency model. WPP Open Pro^[cxxii] is the first advertising agency as a software service powered by an LLM. There is some concern that LLMs could destroy the very platforms which serve advertisers to consumers.^[cxxiii]

Based on all these factors, advertising is likely to be only one aspect of a market supporting AI’s growth and is unlikely to contribute more than a small proportion of the implied trillion-dollar payback required in the next two years if the dot LLM era doesn’t turn from boom to economic bust. 

Business process efficiencies

A second use case mentioned is deriving business efficiencies. This could be done in a number of ways:

• Automating white-collar roles
• Automating blue-collar and pink-collar roles in conjunction with robotics^[cxxiv]. 

OpenAI recently did research^[cxxv] to find out how their service is being used. The sample looked across free, premium and corporate usage of ChatGPT. Some caveats around the research before we delve into it:

• It ignored the use of API services. 
• It is worthwhile remembering that ChatGPT may be under-represented for some actions like writing code – as developers are very aware of what is the current best tool for them.^[cxxvi]

Microsoft Worklab research^[cxxvii] supports the view of LLM as wingman for white-collar workers. In a story arc that is similar to that of early personal computer adoption, they see LLM use as employee advocated and driven. 

Actions have consequences

Economists have models that look at the impact affecting unemployment^[cxxviii], inflation and GDP. I have used the Phillips Curve^[cxxix] and Okun’s Law^[cxxx] in a thought experiment to model the effect on the US economy, if AI managed to provide up to $1 trillion in cost savings through automating jobs. Even with a notional cost savings of $1 trillion, the revenue that would accrue to LLM providers would be a very small proportion of the $1 trillion revenue growth over the next two years implied by current dot LLM era investments. 

Methodology and assumptions

• US baseline data.
• Civilian labour force 170.8 million.
• Number of unemployed 7.4 million.
• Baseline unemployment rate 4.3%.^[cxxxi]
• Baseline annual inflation (CPI) 3.0%.
• Baseline real GDP $23.8 trillion.
• The average salary is about $94,952 (based on $45.65/hr^[cxxxii] x 40 hours/week x 52 weeks/year).
• $1 trillion in job cuts would represent about 10.53 million unemployed.

Phillips Curve – used a standard slope where 1% increase in unemployment rate corresponds to a 0.5% decrease in inflation. Okun’s Law – I used a standard co-efficient where a 1% increase in unemployment rate corresponds to a 2% decrease in real GDP.  

The degree of economic change, at a time of deflation and drop in GDP would make the environment very hostile for businesses dependent on high growth rates. The economic model of achieving a $1 trillion payback through cost-savings is self-defeating. The very success of automation on that scale would destroy the macroeconomic environment required to sustain the hyperscalers’ growth projections.

As we have seen in Japan during the lost decades,^[cxxxiii] deflation would delay purchases and investments. The reduction in GDP would mean that there would be less money available for purchases and investments – creating a negative economic environment for all parties involved including the hyperscalers who would have precipitated the economic change. This scenario has alternative asset management firm Blackstone concerned that its peers are not considering the level of economic disruption the LLM era will bring.^[cxxxiv]

That is before you even consider the economic shockwave^[cxxxv] that would roll around the globe in a similar manner^[cxxxvi] to the 2008 financial crisis. All of this means that there is an optimal economic point in increasing productivity through dot LLM era automation without tanking future growth for hyperscalers and their clients. 

AI optimists would think of the economic shockwave as being short-term in nature, followed by a long-term boom. In this respect, they would draw on examples like the rise of the steam engine, railways or electricity. On balance, I would disagree with these optimists. Economic conditions are very different now. For instance, western economies are now much more ‘financialised’^[cxxxvii] and so the ‘short-term’ shockwave could be well over a decade in length, more similar to the great depression.^[cxxxviii] Developed economy country governments may not have the headroom^[cxxxix] to get out of the depression through a Franklin D. Roosevelt-style New Deal Keynesian stimulus.^[cxl]

Productivity benefits?

Personally, I have found working with generative AI useful in a number of circumstances, in particular, solving the blank page problem. I have also used it as a research tool, a proof-reader and an editing partner. This article was written with the help of generative AI from an editing perspective. But I have also spent a lot of time looking at the outputs given and ensuring that they accurately reflected the exploration of where I wanted to go. And then there is the issue of hallucinations. 

So far, the evidence has been mixed. There are a number of factors for this, IT projects are hard to implement successfully. 

Businesses that have embraced LLMs to improve productivity have been penalised by investors due to the high upfront costs required.^[cxli] Some critics claim that US data implies a plateauing of adoption of generative AI tools in companies^[cxlii] – I personally think that this data is far from conclusive at the present time. 

Some AI researchers like DeepSeek’s senior researcher Chen Deli believes that in the short-term AI could be a great assistant to humans, but over a longer period of 5-to-10 years it would threaten job losses as LLMs became good enough to replace humans in some forms of work. 

“In the next 10-20 years, AI could take over the rest of work (humans perform) and society could face a massive challenge, so at the time tech companies need to take the role of ‘defender’,” he said. “I’m extremely positive about the technology but I view the impact it could have on society negatively.”^[cxliii]

Many of the leading companies in the LLM space such as Nvidia believe that the technology will drive a leap forward in robotics.^[cxliv] Companies are currently building training sets on movement that are similar in function to the knowledge training sets used for LLMs. Even for well-known procedures, there are layers of formidable complexity to simple robotics tasks which would tax the most sophisticated process engineers.^[cxlv]

There are limiting factors outside the control of the LLM era ecosystem including power, the degree of control and limitations of mechanical engineering to supply chain challenges wrought by globalisation.^[cxlvi] Both of which neither move at, or are related to Moore’s Law speed and scale of innovation. A key component is the strain wave gearing (also known as a harmonic drive)^[cxlvii]which are made to standard sizes by very few companies, representing an innovation chokepoint, similar to ASML’s lithography machines in semiconductor manufacturing. The standard sizing limits capabilities from mechanical power to precision and increments of movement, which is one of the reasons why Apple still relies on hand assembly on its iPhones despite P&Ps (‘pick-and-place’^[cxlviii] machines or surface mount technology (SMT) component placement machines) being available as far back as the 1980s. This chokepoint is one of the reasons why robotics vendors have focused on software-based differentiation with limited success so far.

Different LLMs seem to lend themselves to different tasks as show by Anthropic^[cxlix] and OpenAI’s^[cl] own research into the economic and usage behaviour of their respective tools. 

The Global environment

Unlike other technological leaps forward, the LLM era isn’t likely to see American platform domination all around the world outside China. The dot-com era was the high point of American power. Coming out of the cold war, globalisation was benefiting US technology companies. The decline of Russia allowed the Clinton regime to open up the internet to commercial usage. American companies dominated enterprise software, semiconductors, wireless and computer network products. 

25 years later, the US no longer has pre-eminence. Many of its past champions like Lucent^[cli] or Motorola^[clii] are either much reduced, or no longer American companies. Globalisation in the technology industries has meant that the concentration of expertise has become interconnected and dissipated to global centres of excellence such as TSMC^[cliii], Foxconn^[cliv] and Huawei^[clv]. China had developed a parallel ecosystem some of which like Bytedance successfully compete head-to-head with large American technology platforms. 

The LLM era is no longer only American in nature. Chinese companies have compelling offerings. For instance, Chinese hyperscaler Alibaba claim to be able to have models that are comparable to their American counterparts, yet needs 82% fewer Nvidia processors to run.^[clvi] Even Silicon Valley companies are using Chinese LLM models over the likes of ChatGPT or Anthropic. The news that Airbnb opted to use Alibaba’s open-source Qwen AI model over ChatGPT was a milestone event.^[clvii]US technology sector investors are using the Kimi K2 model because it was ‘way more performant and much cheaper than OpenAI and Anthropic’.^[clviii] China benefits from much cheaper model training cost per token. The open-source models can be run on private infrastructure, keeping sensitive data inhouse and ensuring ‘corporate sovereignty.

In the global south, China’s technology companies have corporate and government business relationships built up over years. Their combination of low cost combined with trusted relationships would reduce American hyperscaler opportunities for global expansion. 

While US companies have access to more powerful chips, sanctions against Chinese companies aren’t effective with Nvidia chips being smuggled into China and heavy computing work like model training being run in data centres^[clix] based in other Asian countries, notably Malaysia.^[clx]

There is one clear parallel between the earlier telecoms bubble and the dot LLM era; demand in the global south seems to be constrained by infrastructure rather than user interest in adopting generative AI tools.^[clxi]

Other bubbles.

The dot-com era tends to be cited due to it being a technology story as much as an economic story. Many other bubbles were purely financial in nature:

• The sub-prime mortgage crisis of 2008/9
• The US savings and loans crisis of the early 1990s
• 1929 stock market crash
• Tulpenmanie from 1634 – 1637 

The 1929 crash has sometimes been described as an electric generation bubble bursting since some 19% of the shares available on the market were from utility companies. But the impact was so widespread that it be hard to argue that it was really a ‘technology bubble’.^[clxii]

The British railway mania of the 1840s is often cited as an analogue of the telecoms bubble a century and a half later. The railway mania rolled out at a slower pace than the dot-com boom. It featured a Minsky moment and resulted in a consolidation of rail companies rather than an outright failure of many businesses. Up to a third of railway companies started during the time collapsed before building their railway line due to poor financial planning.^[clxiii]

The key defining factor for how bad the bust is from a bubble, and how long the bust lasts for is the amount of borrowing (or leverage) involved.^[clxiv]

How might the dot LLM era differ from the dot-com era in terms of the corresponding bust?

Zero-cost co-ordination

An economic paradigm shift will have occurred that doesn’t have a clear analogue in history that I am aware of. For instance, there are theoretical writings about how LLMs and agents will change the very nature of economics and the corporation may be changed with the advent of ‘zero-cost co-ordination’^[clxv] reducing economic friction. This could upend the very nature of what a company is. 

Historically one of the reasons given for participating in a firm was that internal coordination costs were cheaper than market coordination (transaction costs). If agentic AI are rational actors that reduce market transaction costs (search, negotiation, contracting) to near-zero, the need for large, hierarchical firms changes and likely diminishes.^[clxvi]

If this theory were true, the excessive capital expenditure would simply be the price paid for creating the world’s first zero-friction economic system. In theory, it’s possible, but it depends on the humans involved being rational decision-makers in a rational culture that doesn’t exhibit risk aversion and that their agents don’t develop similar biases over time. This often isn’t true, even in business-to-business situations, for instance in the past ‘nobody ever got fired for buying IBM’.^[clxvii]

This viewpoint in some ways is similar to Wired magazine’s editorial team circa 1998 and futurist author Kevin Kelly’s ideas on the ‘new economy’.^[clxviii] The thesis was that the internet would reduce information friction. The dot-com bust provided a more tempered lens on the ideas of the ‘new economy’. Would efforts to reduce economic friction fare any better than the information friction reduction of the ‘new economy’?

Google Research economists have asked this same question^[clxix] and came back with more open questions than answers. The authors posit that AI systems, being built on optimisation principles, can be modelled as standard “textbook” economic agents. when AI agents deviate from perfect rationality, they may exhibit an “emergent preference” and display behavioural biases similar to those found in humans. They highlight what they termed the “contract” problem. It draws an analogy between the AI alignment problem and the economic theory of ‘incomplete contracts,’ where a designer (the principal) cannot perfectly specify the AI agent’s goals, leading to unpredictable behaviour. The economists were concerned there would be a need for new institutions to govern an AI agent economy to ensure markets remain well-functioning and stable.

The open questions:

• Whether AI agents have stable ‘beliefs’?
  • How they update them? 
  • If they can hold ‘higher-order beliefs’ (beliefs about others’ beliefs)?
• There is a lack of research and benchmarks for evaluating AI performance in complex, multi-agent systems which needs to be addressed. One of the key challenges is that small differences between AI and human behaviour can become magnified in an equilibrium.

But what if, as Francis Fukuyama argues,^[clxx] that transaction friction isn’t the block on economic growth? Instead, it’s resource constraints, social and political considerations that are the brake on how fast economic growth can happen. 

AI-fuelled breakthroughs

The infrastructure boom fuels foundational AI research far beyond current capabilities. In this scenario, active engines of scientific discovery. The AI research achieves a breakthrough in a hard-science field like drug discovery (e.g., new classes of effective antibiotics), materials science (e.g., room-temperature superconductors), novel ways of rare earth metal extraction, or sustained controllable nuclear fusion – and facilitates record compression of time to market for these developments. LLMs would not only have to facilitate the breakthrough, but drive mass-accelerated implementation and regulation. 

In theory, LLMs could: 

• Optimise experiment and trial design.
• In- and post-test data analysis. 
• Drive synthesis of regulatory compliance documents and evidence. 
• Optimise production and supply chains to facilitate the manufacture and commercialisation of a new break-out product. 

If all this happened, it would create entirely new sources of economic value, far dwarfing the infrastructure cost. That is a lot of serendipity, of huge scope and massive assumptions: even the NASA Apollo Program^[clxxi] took eight years to have its first crewed lunar flight^[clxxii] and another year to put the first men on the moon.  

AI-fuelled breakthroughs are usually linked with progress towards AGI or ‘artificial general intelligence’ or human level intelligence AI.^[clxxiii] A research paper from Cornell University that outlined benchmarking for progress to understanding the real world. The paper introduced WorldTest, a new framework for evaluating how AI agents learn and apply internal world models through reward-free exploration and behaviour-based testing in modified environments. Its implementation, revealed that while humans excel at prediction, planning, and change detection tasks, leading AI reasoning models still fall short. Their shortcoming was associated with flexible hypothesis testing and belief updating. The findings suggest that future progress in AI world-modelling depends less on scaling compute and more on improving metacognition, exploration strategy, and adaptive reasoning.

Platform lock-in and bundling

Many of the established hyperscalers (Adobe, Alphabet, Amazon, Microsoft, Oracle and Salesforce) have established client relationships in a range of products:

• CRM.
• Creative Suite and Marketing Cloud.
• Office suite or Workspace.
• Enterprise Cloud services.

Rather than a disruptive paradigm shift, the LLM payback could come from an instant, embedded non-disruptive increase across existing indispensable products and services. It extracts the value from the existing enterprise wallet, which breaks the historical analogy of relying on new economic value creation. On the face of it, a largely risk-free proposition.

The US legal environment is very different from the dot-com era. Microsoft would not have to worry about facing an antitrust trial similar to its conflict over bundling with Netscape.^[clxxiv]  

While in the US, antitrust enforcement is considered laxer than during the Biden regime, these technology companies would be concerned about competition regulators in the EU and elsewhere. For example, just this September, Microsoft had to unbundle Teams from its Office software to avoid EU antitrust fines.^[clxxv] Alphabet^[clxxvi] and Amazon^[clxxvii] have had previous bruising run-ins with authorities outside the US which would complicate any decision made to bundle an LLM service. 

What could dot LLM era outcomes look like?

I have come up with seven scenarios that range in the kind of impact that generative AI as a sector may provide. These range from being wildly successful to dark failure

• The breakthrough: total economic transformation due to a post-war breakthrough in science and technology.
• The ‘new economy’: frictionless co-ordination facilitates more economic activity.
• The ‘wingman economy’: a managed productivity boom.
• The ‘Red Hat model’: an open-source foundation driving value-added services.
• The ‘moral hazard’: major AI players are considered ‘too big to fail’ and backstopped with government loan guarantees.
• The ‘telecoms bust’: a Minsky moment and amortisation crisis.
• The ‘weird gizmo’: collapse total bust. 

How these scenarios map out when thinking about the level of value creation or value saved through increased efficiency.

	Negative / zero net value created	Positive to transformative value creation
New value creation	The ‘weird gizmo’ collapse (value was illusory)	The breakthrough (new science) The ‘new economy’ (new coordination)
Efficiency / existing value	The ‘telecoms bust’ Capex > value The ‘moral hazard’ value is geopolitical rather than financial	The ‘wingman economy’ (managed productivity) The ‘Red Hat’ model (value moves to services)

The breakthrough: total economic transformation

What it looks like: The massive capital expenditure on infrastructure is validated because AI achieves a true, hard-science breakthrough. This creates entirely new sources of economic value, such as sustained nuclear fusion, room-temperature superconductors, or new classes of antibiotics. In this outcome, the $1 trillion in implied value is not only met but vastly exceeded. Justifying the “bubble” as the necessary investment for a new industrial revolution.

What to watch? 

• Scientific breakthroughs. 

Metric: 

• High-impact scientific publications that use AI for novel discovery, NOT just analysis.

Source: 

• Track major journals like Nature, New Scientist and Science for breakthroughs in AI-driven drug discovery, materials science, or physics. Recent reports on AI’s role in molecular innovation and even quantum computing show this is a key area to watch.

The “new economy”: frictionless co-ordination

What it looks like: Agentic AI successfully reduces market transaction costs (search, negotiation, contracting) to near-zero. This upends the nature of the corporation, as the historical reason for firms (cheaper internal vs. market coordination) diminishes. The massive capital expenditure is seen as the “price paid for creating the world’s first zero-friction economic system”. This is the 1998 Wired “new economy” thesis finally coming true, though it faces challenges like the “Contract problem” and AI alignment.

What to watch? 

• Agentic breakthroughs

Metric: 

• Demonstrations of “agentic” AI (AI that can independently complete complex, multi-step tasks), particularly in commercial or economic settings.

Source: 

• Monitor announcements from leading research labs (DeepMind, FAIR, OpenAI) and market analysis on “agentic AI” to see if it’s moving from theory to reality.

The ‘wingman’ economy: a managed productivity boom

What it looks like: The technology finds its “optimal economic point”. LLMs become a powerful “wingman for white-collar workers”, similar to the adoption of early PCs. This drives real productivity gains, but the $1 trillion in cost savings is implemented gradually, avoiding the catastrophic deflationary shock modelled by the Phillips Curve and Okun’s Law. The “Magnificent 10” see steady growth, but the ‘pure play’ LLMs struggle to find profitability on their own.

What to Watch: 

• National Productivity Data
• Enterprise Adoption & AI Mentions in Earnings

Metrics: 

• U.S. labour Productivity and unit labour costs. We are looking for a “golden path”: productivity rising faster than unit labour costs, which would suggest companies are becoming more efficient without just slashing jobs en-masse.^[clxxviii]
• The number of S&P 500 companies citing “AI” on their quarterly earnings calls. A high number (e.g., over 40-50%) shows it’s a top-level strategic priority.

Sources: 

• U.S. Bureau of Labor Statistics (BLS) – productivity and costs. The quarterly releases from the BLS are the single best macro-indicator for this scenario.
• FactSet Earnings Insight.^[clxxix]  – they regularly publish analyses on the frequency of “AI” mentions in earnings calls, which is a direct proxy for corporate focus and investment.

The Red Hat analogue: a foundational model

What it looks like: The “pure play” LLMs like OpenAI and Anthropic, which are losing money, ultimately fail or are acquired for pennies on the dollar. However, open-source and open weight models (like Llama, etc.) proliferate. Alibaba’s Qwen model has already been very successful. Singapore’s national AI programme dropped Meta’s Llama in favour of it.^[clxxx] Singapore joins Airbnb as Qwen users;^[clxxxi] meanwhile Chinese model DeepSeek has been adopted by European startups.^[clxxxii] The long-term winners are not the model creators but the companies that, like Red Hat, sell “enterprise-grade support, integration, and services”. 

LLM models have an “outsized impact” —becoming the “Linux” for the next generation of applications—but the initial investors see a massive correction.

What to Watch: 

• Open-source vs. closed-source momentum

Metric: 

• Rate of change in download statistics, new model uploads, and developer activity on open-source AI platforms.

Source: 

• Hugging Face Trends.^[clxxxiii] This dashboard shows which open-source models are gaining traction. If downloads for open-source models are growing faster than API call revenue for closed-source models (a harder metric to find), it signals a shift toward this “Red Hat” scenario. GitHub’s annual “Octoverse” report is another key source, as it tracks the rise of AI-focused projects.

The ‘moral hazard’: major dot LLM players are considered ‘too big to fail’ and backstopped with government intervention

There are elements of a non-bubble, financial crisis aspect to the dot LLM era. Chinese LLM vendors are being given subsidised electricity from local governments,^[clxxxiv] alongside preferential rates in data centres. The LLM era in the US could be considered by the government as having become too large a part of the economy to be allowed to fail due to normal market forces. Open AI has recently had to deny rumours^[clxxxv] that it sought US government loan guarantees for at least part of the multi-trillion dollar deals it has put in place for data centre infrastructure and hardware. AI sovereignty comes to be seen as taking on a geostrategic and national security imperative as business and investor considerations take a backseat. 

Hyperscalers are hitting a ‘power wall’ as they cannot get the equivalent electricity generating capacity of 16 Hoover dams. Getting over the wall would require a massive amount of government infrastructure funding.^[clxxxvi]

Major government involvement may impact the speed of development as LLM model providers and supporting infrastructure no longer have to constantly innovate and instead move at the speed of their government clients. 

What to watch:

• Shift in rhetoric from commercial to critical: Observe how language from policymakers, military leaders, and national security bodies evolves. A shift from discussing AI in terms of commercial competition (e.g., “market leadership”) to national infrastructure (e.g., “digital sovereignty,” “critical asset,” “geostrategic imperative”) is a primary indicator. This reframes an economic failure as a national security failure.
• Direct & indirect state support mechanisms: look beyond simple R&D grants. Watch for the creation of new, targeted support instruments:
  • Direct: preferential pricing on energy/compute, state-backed datacentre construction, sovereign wealth fund investments, or direct “national champion” subsidies.
  • Indirect: government-backed loan guarantees for infrastructure (like the rumoured OpenAI deal), strategic procurement (where the government becomes the anchor customer) – Palantir would be an exemplar, and “regulatory moats” that favour incumbents (e.g., high-cost safety/licensing rules that only large, state-backed labs can afford).
• “Bailout” vs. “investment” framing: monitor how state intervention is publicly justified. A struggling “national champion” AI firm receiving a sudden capital injection from a state-adjacent entity will likely be framed as a “strategic investment in national capability,” not a “bailout.” This framing is key.

Metrics:

• Value of state & military contracts: Track the total disclosed value of government contracts (especially from defence and intelligence agencies) awarded to foundational model providers. A rapid increase, or contracts for non-competitive “strategic deployment,” signals TBTF (“Too Big to Fail”) status.
• Frequency analysis of policy language: quantify the co-occurrence of terms like “AI,” “sovereignty,” “national security,” and “critical infrastructure” in parliamentary/congressional records, national strategy documents, and defence budget justifications. A rising frequency indicates the ideological groundwork for a TBTF policy.
• State-backed capital flows: monitor announcements from sovereign wealth funds, national investment banks (e.g., UK’s National Security Strategic Investment Fund), or public pension funds. Track the size and frequency of their investments into large, established AI labs, as opposed to a diverse portfolio of early-stage start-ups.
• Subsidy disclosures: quantify the value of announced subsidies (e.g., tax credits, energy discounts, land grants) specifically earmarked for AI datacentres and R&D hubs associated with the major players.

Sources:

• Financial & policy journalism: The Financial Times, Bloomberg (especially its Bloomberg Government vertical), and Politico as media sources. Their reporters are often the first to break stories on subsidies, lobbying, and the intersection of tech and state power.
• Government procurement & grant databases: official portals like USASpending.gov in the US or the UK’s Contracts Finder service. While difficult to navigate, they provide primary evidence of public funds flowing to specific companies.
• Think tank & national security publications: Reports from organisations like the Center for a New American Security (CNAS) in the US, the Royal United Services Institute (RUSI) in the UK, or the Mercator Institute for China Studies (MERICS). They often analyse and quantify the geostrategic rhetoric and policy shifts. The main challenge with this source might be timeliness of publication in comparison to the previous sources. 
• Company filings & investor calls: For publicly traded companies (Microsoft, Google, Amazon, Nvidia), annual reports (10-K forms) and quarterly investor calls often mention large government contracts or regulatory risks/opportunities, providing a corporate-side view of this trend.

The Telecoms Bust: a Minsky moment and amortisation crisis

What it looks like: The $1 trillion in value fails to materialize from either advertising or business efficiencies. Investors have a Minsky moment and realize the debt and capex are unsustainable. The bubble implodes like the telecoms bubble. The key difference is the financial and technological amortisation risk: the GPUs (with a 2-to-5-year useful life) become obsolete. Unlike the dot-com era’s dark fibre, this infrastructure cannot be repurposed by a “web 2.0”. This leads to trillions in write-offs, analogous to WorldCom’s $180 billion loss.

What to Watch: 

• Hyperscaler capital expenditure (Capex)
• GPU amortisation & resale value

Metrics: 

• Quarterly capex announcements from Google (Alphabet), Meta, Microsoft, Oracle and Amazon (AWS). This is made trickier to understand by Meta, Microsoft and Oracle looking at forms of private equity financing. 
• The rate of change in Nvidia’s^[clxxxvii] data centre revenue, Broadcom and AMD’s enterprise / data centre revenue. This is the “equipment maker” side of the equation. As long as this number is growing, the bubble is inflating. A sudden slowdown would be the first sign of a “Minsky Moment.”
• The resale value of last-generation GPUs (e.g., H100s as B200s/B300s roll out). If these prices collapse, it validates your thesis that the assets cannot be repurposed, and the financial write-downs will be catastrophic.

Sources: 

• Hyperscaler capex reports from financial analysts and data centre publications. Recent reports show combined capex is projected to hit hundreds of billions, a clear sign of the infrastructure race.
• Alphabet, Meta, Microsoft, Oracle and Amazon quarterly results and investor roadshow presentations.
• NVIDIA, Broadcom and AMD quarterly earnings reports. The Nvidia Q2 2026 report showing data centre revenue at $41.1B is a perfect example of this indicator.
• Resale value of GPUs is a harder metric to track. Monitor tech hardware forums and eBay listings, or look for analyst reports on the “used GPU market.” A collapse in this secondary market for last generation GPUs is a major red flag.

The “Weird Gizmo” Collapse: total bust

What it looks like: The technology is ultimately seen as a novelty. It’s the 2020s version of Boo.com, Beenz and Flooz, or the 3Com Audrey. The argument that “AGI is not imminent^[clxxxviii], and LLMs are not the royal road to getting there” wins the day. This bear view of AGI is one that is widely shared by prominent experts^[clxxxix] within the machine learning field. Which is why new ways of working like nesting models and world models are being explored, alongside quantum computing. In this scenario, the pure play companies burn through all their cash and vanish. The hyperscalers are left with billions in useless, obsolete silicon, and the “dot LLM era” is remembered as a short-lived period of speculative mania.

What to Watch: 

• AI startup burn rates & funding (the “burn Rate” indicator)

Metric: 

• Quarterly venture capital funding for AI startups, specifically looking for a rise in “down rounds” (where valuations decrease) or outright failures.

Source: 

• Data from firms like CB Insights or Crunchbase.^[cxc] Recent reports show that while “mega-rounds” for established players (like Anthropic) are still huge, seed-stage funding is declining, showing a “haves and have-nots” market. A slowdown in the mega-rounds would signal the bust is beginning.

Personal assessment of likely outcomes by scenario

Scenario	Estimated likelihood	Rationale
The moral hazard	~95%	US – China trade disputes and geopolitical strife Chinese government investment in startups Chinese local government subsidies for operating AI services The current position that AI has in driving US GDP growth across sectors including construction and the energy sector Likely OpenAI loan guarantees Palantir is already deeply embedded in the US government as a vendor and has partnerships with defence contractors like Anduril
The ‘wingman’ economy	~80-90%	Some research reports indicate that AI is augmenting knowledge workers in different sectors.  Claims of AI replacing workers are more difficult to validate, for example:  Klarna moving to automation and then rehiring  Clifford Chance offshoring back-office roles to Poland and China while claiming that the job losses were due to AI.
The ‘Red Hat’ Model	~70-80%	Airbnb opting to use Alibaba’s open-source Qwen AI model over ChatGPT was a milestone event.[cxci]
The ‘telecoms bust’	~75%	Concerns about the size of capital expenditure. Rate of growth of supporting infrastructure. Uncertainty about length of depreciation affecting overall shareholder trust in hyperscalers.  Cheaper alternatives like Qwen.
The ‘new economy’	<15%	The uncertain economics of ‘zero friction’ transactions. Real-life legal and regulatory issues.  Amazon’s dispute with Perplexity using AI agents on its website.
The breakthrough	<10%	A black swan event
The ‘weird gizmo’	<5%	It would be unusual for a technology to disappear completely, LLMs have been finding some use already. The rise of open-source AI models which reduce the cost of operation.

Where are we at the moment?

I worked to put together a diagram to try and assess where we are at the moment given that some of the scenarios outlined are running concurrently with each other. 

Acknowledgements

Ian Wood (Wireless Foundry),

Colophon

The dot LLM era is brought to you with the assistance of:

• Pepsi Max
• Cuisinart filter coffee maker
• Moscot glasses
• Extra spearmint gum
• Herman Miller Aeron chair with Engineered Now headrest.
• Beyerdynamic DT 1770 PRO headphones
• Apple MacBook Pro
• Apple Notes, Freeform, Keynote, Pages and Voice Memos
• Hemingway App
• DuckDuckGo
• Newsblur
• Pinboard
• Substack
• WordPress  

---

^[i] Pluralistic: The real (economic) AI apocalypse is nigh (27 Sep 2025) – Pluralistic: Daily links from Cory Doctorow

^[ii] The real (economic) AI apocalypse is nigh | Pluralistic

^[iii] Morning, Ellis (May 7, 2025) The Daily WTF – AI: The Bad, The Worse and The Ugly

^[iv] AMD rallies 24% on OpenAI Deal and Bari Weiss Takes Over CBS News | Prof G Markets Podcast with Ed Elson

^[v] Wong, Matteo (July 24, 2024) The Atlantic – Silicon Valley’s Trillion-Dollar Leap of Faith

^[vi] Agarwal, P. (2025) Moats in AI Startups: Data, Distribution, or Default? (India) Eximius Ventures

^[vii] James Anderson warns Nvidia’s $100bn OpenAI bet echoes dotcom bubble | FT

^[viii] AI’s Power Demand Is Driving Up Your Electricity Bill | Prof G Markets

^[ix] Grabar, Henry (May 18, 2022) Slate – The Decade of Cheap Rides Is Over

^[x] Harris, A. (2025) CAPITAL IDEAS: The market’s valuation rivals the dot-com bubble. What’s next for stock prices? (US) The Berkshire Edge

^[xi] S&P P/E Ratio Is Low, But Has Been Lower (2009) Seeking Alpha

^[xii] How Does the End Begin? – No Mercy / No Malice

^[xiii] Meyer, Katherine (May 3, 2006) The Wall Street Journal – The Best of the Worst

^[xiv] Grant, Conor (July 1, 2018) The Hustle – A decade before crypto, one digital currency conquered the world – then failed spectacularly

^[xv] Stepanek, Marcia (September 28, 2000). “The CueCat Is on the Prowl: This gizmo is on the cutting edge of e-marketing. But with each swipe, it tracks your moves through cyberspace”. Bloomberg Businessweek.

^[xvi] Kanellos, Michael & Wong, Wylie (March 21, 2001) CNet News – Audrey’s life cut short

^[xvii] Cringely, Robert. X (December 2, 2004) PBS I, Cringely – Wishing for Audrey (Now That the World is Finally Ready for Internet Appliances, Where Are They?

^[xviii] Centre for Computing History – Ergo Audrey

^[xix] Chidi Jr, George A (June 18, 2001) CNN – Sony launches Net access device

^[xx] United States Securities and Exchange Commission Form 8-K palmOne, Inc. (October 28, 2003)

^[xxi] PC Mag UK (December 5, 2000) 3Com Ergo Audrey review

^[xxii] Beaumont, M. (2000) e. The novel of liars, lunch and lost knickers & The e Before Christmas (UK) HarperCollins

^[xxiii] Introducing Nested Learning: A new ML paradigm for continual learning | Google Research

^[xxiv] Tett, G. (2025) Behind the AI bubble, another tech revolution could be brewing (UK) Financial Times

^[xxv] Kelly, K. (2010) What Technology Wants (US) Viking Press

^[xxvi] Boo Hoo: A Dot Com Story Paperback by Ernst Malmsten, Erik Portanger and Charles Drazin

^[xxvii] The report was subsequently published in Panic! edited by Michael Lewis

^[xxviii] Scott Galloway CV (PDF)

^[xxix] Crunchbase | redenvelope.com

^[xxx] Rival walks off with Geldof’s Deckchair | Guardian

^[xxxi] The Everything Store: Jeff Bezos and the Age of Amazon (2014) by Brad Stone

^[xxxii] Spector, Robert (2002). Amazon.com: Get Big Fast.

^[xxxiii] CNN Money | Amazon posts first ever profit in 4Q (January 22, 2002)

^[xxxiv] Computerworld | Amazon records first profitable year in its history (January 28, 2004)

^[xxxv] Y2K – renaissance chambara

^[xxxvi] Dell plus Sun equals VA Research | Forbes

^[xxxvii] VA Linux Sets IPO Record – Wired

^[xxxviii] History of Red Hat Linux – Fedora Project Wiki

^[xxxix] SuSE Linux for S/390 Available Today – SuSE

^[xl] IBM Closes Landmark Acquisition of Red Hat for $34 Billion; Defines Open, Hybrid Cloud Future – Red Hat press release

^[xli] Lessons from History: The Rise and Fall of the Telecom Bubble – Fabricated Knowledge by Doug O’Laughlin

^[xlii] British Telecom privatisation share prospectus advert (1984)

^[xliii] 1984-2014 – 30 years of the Janet network (Jisc)

^[xliv] Ellassen, K.A, & From, J. (2009) Deregulation, privatisation and public service delivery: Universal service in telecommunications in Europe (United Kingdom) Policy and Society Vol. 27, issue 3

^[xlv] https://www.instituteforgovernment.org.uk/sites/default/files/british_telecom_privatisation.pdf              

^[xlvi] Herrera-González, F. & Castejón-Martin, L. (2009) The endless need for regulation in telecommunication: An explanation (UK) Telecommunications Policy Volume 22 Issues 10-11

^[xlvii] Morgan, K. Monopolies under Siege in Western Europe – Institute of Development Studies

^[xlviii] Hayashi, K. and Fuke, H. “Changes and deregulation in the Japanese telecommunications market,” in IEEE Communications Magazine, vol. 36, no. 11, pp. 46-53, Nov. 1998.

^[xlix] Tsatsou, P. (2010) EU regulations on telecommunications: The role of subsidiarity and mediation – First Monday volume 16, issue 1

^[l] Clinton signs telecom bill (February 8, 1996) – UPI

^[li] Telecommunications Act of 1996 – Federal Communications Commission (FCC)

^[lii] Brotman, S.N. (2016) Was the 1996 Telecommunications Act successful in promoting competition? – Brookings Institute

^[liii] Barlow, J.P. (1996) A Declaration of the Independence of Cyberspace – Electronic Frontier Foundation

^[liv] https://www.wto.org/english/tratop_e/serv_e/symp_mar02_uk_com_e.pdf            

^[lv] Issues in Labor Statistics – US Department of Labor Bureau of Labor Statics Summary 99-4 March 1999

^[lvi] Manes, S. (1996) Jaz Drive: A Lot of Backup Insurance in a Small Package (US) The New York Times

^[lvii] Elmer-Dewitt, P. (1993) Take A Trip into the Future on the Electronic Superhighway (US) TIME  

^[lviii] On-ramp Prospects for the Information Superhighway Dream by Gordon Bell and Jim Gemmell COMMUNICATIONS OF THE ACM July 1996/Vol. 39, No. 7 

^[lix] Gates, Bill; Myhrvold, Nathan and Rinearson, Peter. (1995). The Road Ahead (1st ed.) (US) Viking

^[lx] Botein, M. (2003) The Demise of The Information Superhighway (US) Digital @NYLS (New York Law School)

^[lxi] Odlyzko, A.M., Internet traffic growth: Sources and implications (US) University of Minnesota

^[lxii] Litan, R.E. (2002) he Telecommunications Crash: What to Do Now? (US) Brookings Institute

^[lxiii] O’Laughlin, D. (2023) Lessons from History: The Rise and Fall of the Telecom Bubble (US) Fabricated Knowledge

^[lxiv] Hunter, D. (2018) Nortel (Canada) The Canadian Encyclopedia

^[lxv] Lebowitz, M. (November 5, 2025) Nvidia Deals: Round Tripping or Vendor Financing? (US) Investing.com

^[lxvi] Claycombe, C. WorldCom/MCI: Massive Accounting Fraud (US) Wichita State University

^[lxvii] Starr, P. (2002) The Great Telecom Implosion (US) American Prospect magazine

^[lxviii] McLean, B & Elkind, P. (2003) The Smartest Guys in the Room: The Amazing Rise and Scandalous Fall of Enron (US) Portfolio

^[lxix] PUBLIC LAW 107–204—JULY 30, 2002116 STAT. 745

^[lxx] The Laws That Govern the Securities Industry – U.S. Securities and Exchange Commission

^[lxxi] Worthen, B., White, B. (October 27, 2008) Tech Finance Defaults Rise; More Makers Offer Loans (US) The Wall Street Journal

^[lxxii] Meeker, M. (1996) Internet Trends (US) Morgan Stanley

^[lxxiii] Kedrosky, P. (2025) Minsky Moments and AI CapEx (US) paulkedrosky.com

^[lxxiv] PS Lee National University of Singapore on LinkedIn

^[lxxv] GPU Life Concerns: Reality And Implications (2024) Beyond the Hype – Looking Past Management & Wall Street Hype

^[lxxvi] Curry, T., Shibut, L. (2000) The Cost of the Savings and Loan Crisis: Truth and Consequences (US) FDIC Banking Review

^[lxxvii] 1993: Recession over – it’s official | BBC

^[lxxviii] Schwartz, P., Leyden, P. (1997) The Long Boom: A History of the Future, 1980 – 2020 (US) Wired magazine

^[lxxix] Kadlec, D. (1999) Day Trading: It’s a Brutal World (US) Time magazine

^[lxxx] Sor, J. (2025) ‘A very lonely sport’: Day traders on the isolating experience of trying to make a living in the stock market (US) Business Insider

^[lxxxi] Are rich countries facing a debt crisis – The Economist on YouTube

^[lxxxii] Climate despair (2023) renaissance chambara

^[lxxxiii] (2025) How many people are already being killed by climate change? (UK) The Economist

^[lxxxiv] Moshiri, A. (2025) Devastation on repeat: How climate change is worsening Pakistan’s deadly floods (UK) BBC

^[lxxxv] Raval, A. (2025) The AI job cuts are accelerating (UK) FT

^[lxxxvi] Meeker, M. (1995 on) Internet reports (US) Bond Capital

^[lxxxvii] Meeker, M., Simons, J., Chae, D., Krey, A. (2025) Trends: Artificial Intelligence (US) Bond Capital

^[lxxxviii] Misra, A., Wang, J., McCullers, S., White, K., and Ferres, J.L. (2025) Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage (US) Microsoft

^[lxxxix] (2024) Elon Musk – AI will be smarter than the smartest human | Bloomberg Technology

^[xc] Amodei, D. (2024) Machines of Loving Grace (US) self-published on blog.

^[xci] Pillay, T. (2025) How OpenAI’s Sam Altman Is Thinking About AGI and Superintelligence in 2025 (US) TIME

^[xcii] (2025) The Three-Year Countdown: Inside DeepMind’s AGI Timeline and What It Means for Knowledge Work (US) The Ai Consultancy on Medium

^[xciii] Kurzweil, R. (2024 & updated in 2025) The Singularity is Nearer: When We Merge with AI (US) Random House Publishing

^[xciv] (2025) Joe Rogan: The Truth About Aliens (He Finally Says It) (US) Jesse Michels podcast on YouTube

^[xcv] Ditlea, S. (1996) Leary’s Final Trip, the Web, Realized Multimedia Vision (US) The New York Times Online

^[xcvi] Bearak, B. (1997) Eyes on Glory: Pied Pipers of Heaven’s Gate (US) The New York Times Online

^[xcvii] (2025) Andrej Karpathy — “We’re summoning ghosts, not building animals” (US) Dwarkesh Patel podcast on YouTube

^[xcviii] Brooks, R. (2025) Predictions Scorecard, 2025 January 01 (US) published on personal blog

^[xcix] Leonards, A. Meta chief AI scientist claims AGI will be viable in 3-5 years (UK) National Technology News

^[c] Mitchell, M. (2025) On the Science of “Alien Intelligences”: Evaluating Cognitive Capabilities in Babies, Animals, and AI (US) NeuroIPS

^[ci] Heath, A. (2025) I talked to Sam Altman about the GPT-5 launch fiasco (US) The Verge

^[cii] Schmidt, E., Xu, S. (2025) Silicon Valley Is Drifting Out of Touch with the Rest of America (US) The New York Times

^[ciii] Kharpal, A. (2025) Jeff Bezos says AI is in an industrial bubble but society will get ‘gigantic’ benefits from the tech (US) CNBC

^[civ] Islam, F., Clun, R. (2025) Google boss says trillion-dollar AI investment boom has ‘elements of irrationality’ (UK) BBC

^[cv] (2025) Mark Zuckerberg on the AI bubble and Meta’s new display glasses (US) ACCESS Podcast on YouTube

^[cvi] Kinder, T., Hammond, G. (2025) OpenAI shunned advisers on $1.5tn of deals (UK) Financial Times

^[cvii] Evans, J. (2025) Bubble, bubble, toil and trouble (US) Gradient Ascendant

^[cviii] Li, Y. (2025) ‘Big Short’ investor Michael Burry accuses AI hyperscalers of artificially boosting earnings (US) CNBC

^[cix] Davis, G.B. (2025) 6 Stock Market Lessons from the Dot Com Bubble That Apply in 2025 (US) Yahoo! Finance

^[cx] Galloway, S. (2025) How Does the End Begin? (US) No mercy / no malice

^[cxi] Harnett, I. (2025) The AI capex endgame is approaching (UK) The Financial Times

^[cxii] Galloway, S. (2025) How Does the End Begin? | No Mercy / No Malice (US) Prof G Media

^[cxiii] Kemp, S. (2025) Digital 2025: Global Advertising Trends (Singapore) DataReportal

^[cxiv] The Value of Advertising – World Federation of Advertisers

^[cxv] Hiorns, B. (2023) A Brief History of AI in Advertising #HistoryMonth (UK) Creativepool

^[cxvi] Goldberg, L. (2018) A brief history of artificial intelligence in advertising (UK) Econsultancy

^[cxvii] McGowan, Jacob, “How Has the Growth of E-commerce Sales Affected Retail Real Estate?” (2019). CMC Senior Theses. 2189.

^[cxviii] Merritt, M. (2025) Ad dollars from China are already starting to dry up (US) MorningBrew

^[cxix] Tiprank (2025) Meta Could Face a Massive $7 Billion Ad Revenue Hit from China Tariffs, Warns Analyst (Canada) Globe and Mail

^[cxx] Camille Boullenois, Agatha Kratz and Daniel H. Rosen (2025) Far From Normal: An Augmented Assessment of China’s State Support (US) Rhodium Group

^[cxxi] Farmer, M. (2025) Madison Avenue Media Madness (US) C-Suite Blues

^[cxxii] WPP Open Pro: empowering brands to plan, create and publish campaigns independently (2025) WPP

^[cxxiii] AI may fatally wound web’s ad model, warns Tim Berners-Lee | FT

^[cxxiv] Beltran, M. (2025) Japanese convenience stores are hiring robots run by workers in the Philippines (US) Rest of the World

^[cxxv] Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C., Wadman, K. (2025) How People Use ChatGPT (US) OpenAI, Duke University and Harvard University

^[cxxvi] Coding LLM leaderboard – Vellum.ai

^[cxxvii] AI at Work Is Here. Now Comes the Hard Part 2024 (US) Microsoft Worklab

^[cxxviii] Wessel Vermeulen, Nils Braakmann (2023) How do mass lay-offs affect regional economies? OECD Local Economic and Employment Development (LEED) Papers 2023/01

^[cxxix] How the Unemployment Rate Affects Everybody | Investopedia

^[cxxx] Understanding Okun’s Law: How GDP Growth Affects Unemployment | Investopedia

^[cxxxi] The Employment Situation – August 2025 (US) Bureau of Labor Statistics

^[cxxxii] Employer Costs for Employee Compensation – June 2025 (US) Bureau of Labor Statistics

^[cxxxiii] Jones, R. (2025) A Modern Economic History of Japan: Sho Ga Nai (It Is What It Is) (UK) London Publishing Partnership

^[cxxxiv] Blackstone says Wall Street is complacent about AI disruption | FT

^[cxxxv] Impact of the Global Financial Crisis and Its Implications for the East Asian Economy, Keynote Speech by Mr. Takatoshi Kato, Deputy Managing Director, International Monetary Fund, At the Korea International Financial Association, First International Conference

^[cxxxvi] Andrew Filardo, Jason George, Mico Loretan, Guonan Ma, Anella Munro, Ilhyock Shim, Philip Wooldridge, James Yetman and Haibin Zhu The international financial crisis: timeline, impact and policy responses in Asia and the Pacific. (Bank of International Settlements)

^[cxxxvii] Fontana, G., Dixon, G. (2017) Unlocking the puzzles of financialisation (UK) Applied Institute for Research in Economics

^[cxxxviii] Ross Sorkin, A. (2025) 1929: The Inside Story of The Greatest Crash in Wall Street History (US) Allen Lane

^[cxxxix] Global Debt Report 2025 – OECD

^[cxl] How Keynes Influenced FDR’s New Deal – Future Hindsight

^[cxli] AI’s awfully exciting until companies want to use it: Rightmove edition | FT

^[cxlii] Spencer, M. (2025) Going Short on Generative AI (US) AI Supremacy

^[cxliii] Mo, L., Goh, B. (November 7, 2025) DeepSeek researcher pessimistic over AI’s impact in startup’s first public appearance since success (UK) Reuters

^[cxliv] The Minds of Modern AI: Jensen Huang, Yann LeCun, Fei-Fei Li & the AI Vision of the Future | FT Live – YouTube

^[cxlv] Ford, M. (July 2015) A History of Placement Programming and Optimization (US) Circuits Assembly

^[cxlvi] Is there an end in sight to supply chain disruption? | Financial Times

^[cxlvii] Automata Eve launch | renaissance chambara

^[cxlviii] Component Placement Process – Surface Mount Process

^[cxlix] (2025) Anthropic Economic Index (Anthropic seem to be treating this exercise as a longitudinal research project). 

^[cl] Chatterji, A., Cunningham, T., Deming, D., Hitzig, Z., Ong, C., Shan, C., Wadman, K., (2025) How People Use ChatGPT (US) OpenAI, Duke University & Harvard University

^[cli] (1998 – 2025) AT&T Corporation (US) Encyclopaedia Britannica

^[clii] (1998 – 2023) Motorola, Inc. (US) Encyclopaedia Britannica

^[cliii] Montevirgen, K. (2025) Taiwan Semiconductor Manufacturing Co. (TSMC) (US) Encyclopaedia Britannica

^[cliv] Chinatsu, T. (2025) Foxconn (US) Encyclopaedia Britannica

^[clv] Dou, E. (2025) House of Huawei (UK) Abacus

^[clvi] Chow, V. (2025) Alibaba Cloud claims to slash Nvidia GPU use by 82% with new pooling system (Hong Kong) South China Morning Post

^[clvii] Broersma, M. (2025) Airbnb praises Alibaba’s Open-Source AI model (UK) Silicon

^[clviii] Kynge, J. (2025) Low-cost Chinese AI models forge ahead, even in the US, raising the risks of a US AI bubble (UK) Chatham House

^[clix] Baptista, E., Tang, A., Yong, J.Y. (2025) Malaysia reins in data centre growth, complicating China’s AI chip access (UK) Reuters

^[clx] Jennings, R. (2025) How Malaysia’s data centres became the engine powering China’s AI ambitions (Hong Kong) South China Morning Post

^[clxi] Misra, A., Wang, J., McCullers, S., White, K., and Ferres, J.L. (2025) Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage (US) Microsoft

^[clxii] Sorkin, A.R. (2025) 1929: The Inside Story of The Greatest Crash in Wall Street History (US) Allen Lane

^[clxiii] Odlyzko, A. (2010) Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s (US) University of Minnesota

^[clxiv] Sorkin A.R. (2025) Odd Lots: Andrew Ross Sorkin on the Stock Market Crash That Shattered America (US) Bloomberg

^[clxv] Perez, C.E. (2025) The Intelligence Abundance: How Zero-Cost Coordination Solves the Scarcity Problem

^[clxvi] Coase, R.H. (1937) The Nature of the Firm (UK) Economica volume 4, issue 16 published by the London School of Economics

^[clxvii] Melamed, G. (2024) Nobody gets fired for buying IBM (UK) Finextra

^[clxviii] Kelly, K. (1998) New Rules for the New Economy (US) Viking

^[clxix] Hadfield, G.K., Koh, A. (2025) An Economy of AI agents (US) NBER Handbook on the Economics of Transformative AI

^[clxx] Fukuyama, F. (2025) Superintelligence Isn’t Enough (US) Persuasion

^[clxxi] Ostovar, M. (1998) The Decision to Go to the Moon: President John F. Kennedy’s May 25, 1961 Speech before a Joint Session of Congress (US) NASA

^[clxxii] Brooks, C.G., James M. Grimwood, J.M., Swenson, Jr., L.S. (1979) The NASA History Series: Chariots for Apollo: A History of Manned Lunar Spacecraft

^[clxxiii] Warrier, A.,2, Nguyen, T.D., Naim, M., Jain, M., Liang, Y., Schroeder, K., Yang, C., Tenenbaum, J.B., Vollmer, S., Ellis, K., Tavares, Z. (2025) Benchmarking World-Model Learning (US) Cornell University

^[clxxiv] (2000) Microsoft vs the US Justice Dept. Netscape: A history (UK) BBC

^[clxxv] Warren, T. (2025) Microsoft avoids EU fine after Slack complained about Teams bundling (US) The Verge

^[clxxvi] (2020) Google is unbundling Android apps: all the news about the EU’s antitrust ruling (US) The Verge

^[clxxvii] Espinoza, J. (2020) EU accuses Amazon of breaching antitrust rules (UK) FT

^[clxxviii] U.S. Bureau of Labor Statistics (BLS) – Productivity and Costs – quarterly data

^[clxxix] FactSet Insight blog – Search their blog for keywords like “AI” or “earnings.” They regularly publish analyses on the number of S&P 500 companies that cite “AI” on their earnings calls, which is a direct proxy for C-suite focus.

^[clxxx] (2025) Singapore’s national AI program drops Meta model and switches to Alibaba’s Qwen | TechNode

^[clxxxi] Broersma, M. (2025) Airbnb Praises Alibaba’s Open-Source AI Model (UK) Silicon

^[clxxxii] Broersma, M. (2025) European Start-Ups Adopt DeepSeek To Cut Costs (UK) Silicon

^[clxxxiii] Hugging Face models hub – the view can be filtered by ‘trending’ and ‘most downloaded’ to see what the community is using, versus what closed source models are being marketed

^[clxxxiv] Gao, J. (November 8, 2025) How China hits hard to power its AI ambitions post-Nvidia (Taiwan) DigiTimes Asia

^[clxxxv] Sam Altman says OpenAI is not ‘trying to become too big to fail’ | FT

^[clxxxvi] JPMorgan’s Playbook for a 10-15% Correction (or Worse) — ft. Michael Cembalest | Prof G Markets – YouTube

^[clxxxvii] Nvidia investor relations page – The key figure in their quarterly financial reports is ‘Data Center revenue’.

^[clxxxviii] Marcus, G. (2025) Game over. AGI is not imminent, and LLMs are not the royal road to getting there. (US) Marcus on AI

^[clxxxix] The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li (US) Lenny’s Podcast on YouTube

^[cxc] Crunchbase News – They provide regular analysis of funding rounds. Watch for ‘down rounds’, M&A consolidation among start-ups or acquihires and slowdowns in $100M+ mega-rounds of fund raising. 

^[cxci] Broersma, M. (2025) Airbnb praises Alibaba’s Open-Source AI model (UK) Silicon