The Age of Reason
Making Sense of the AI bubble
Prisoners of Compute
Geography, faith, and the mortgage on the cathedral. The numbers stopped mattering
Musings of a madman. On Friday 17 July 2026, Taiwan Semiconductor Manufacturing Company reported a 77% jump in quarterly operating profit and its share price fell seven percent. See, it’s not just you Shopify.
Sit with that for a second, because everything else in this essay is downstream of it. I have been listening to Tim Marshall lately and he talked about the Age of Reason. I feel right in that moment last Friday. It spawned research to sit on top of the headlines, the side chats and the LinkedIn guff too. Markets are not my forte but they influence much of what we see and hear. More on this below.
Back to Taiwan SMC. The most strategically important manufacturing company on the planet, the firm that physically etches the future onto silicon at a resolution measured in atoms delivered results that would have triggered a champagne order in any other decade, and the market marked it down like a bakery with yesterday’s bread in the window. Kavanagh-esque IYKYK.
It was not alone. The Philadelphia Semiconductor Index fell 12.5% across the week, its worst run in more than fifteen months, closing more than twenty percent below its late-June peak and crossing, by the technical definition, into a bear market. Some $3.3 trillion in global semiconductor market value had evaporated since 22 June. SoftBank, which the market treats as a proxy for OpenAI the way a weather vane is treated as a proxy for the weather, dropped nine percent. Nvidia slipped enough to briefly hand the world’s-most-valuable-company crown back to Apple, which must have felt like being overtaken on the motorway by a man on a ride-on lawnmower. In Hong Kong, the Chinese AI names got it worse than the Americans: Z.ai down thirty percent, MiniMax down sixteen. FWIW, the NVIDIA change sparked Jensen Huang into action on Twitter - overshadowed only, by Bernard Arnault and his letter of perfection.
The trigger was a model. On 16 July, at the World Artificial Intelligence Conference in Shanghai, a Beijing startup called Moonshot AI released Kimi K3, 2.8 trillion parameters, a mixture-of-experts architecture, a million-token context window, and, critically, open weights. The largest open model ever released. It benchmarked competitively against the frontier, it cost a fraction of the frontier to run, and Moonshot announced the full weights would go public on 27 July. Richard Hendriks would be proud.
Anyone could have them. Anyone can still have them, as of this week. Independent testing put Kimi K3’s hallucination rate at roughly fifty-one percent.
51%. I thought I was hallucinating - Dad joke.
A model that is wrong about half the time moved three and a third trillion dollars of capital in four trading sessions, in the same week that the world’s most important chipmaker beat its numbers by seventy-seven percent and was punished for it.
The result is a loud expression of doubt.
And, this is the number I would tattoo on the inside of my eyelids, the Philadelphia Semiconductor Index, having lost a fifth of its value and crossed into a bear market and triggered a global panic, remained up more than sixty percent on the year.
As I mentioned, I am not even a novice on this field yet, but it is interesting.
The wrong revolution
I have been noodling on this idea for a number of months. And for months I had the wrong analogy, so let me correct myself in public.
The instinct is to reach for the Enlightenment.
The Age of Reason. Science versus religion, the scientific method arriving at the door of superstition, faith retreating in disarray, the priests scrambling for power as the ground shifted underneath them. It feels right. It has the correct shape, an old order of belief colliding with a new order of evidence, and everyone picking sides, and the argument getting nasty. Today, there is technology and there is government. Market capitalisations and the blurring of the lines between work and home life creating this semi cult like obsession with work culture being making our belief in the tech almost religion like. Governments, scrambling for relevance whilst maintaining power are all looking to create new wealth. This is, of course, how nations become great. New wealth.
It’s also, historians will tell you, largely a Victorian invention.
The idea that science and religion have been locked in perpetual warfare is known in the trade as the conflict thesis, and it was manufactured in the 1870s by two men, John William Draper and Andrew Dickson White, with axes to grind and books to sell. Modern historians of science have spent a century taking it apart. The picture that replaced it is far more interesting and far more useful to us. And if nothing else, history shows us men with double barreled names were always right, right? :)
Newton wrote more words on theology and alchemy than he ever did on physics. The Royal Society was thick with clergymen. Mendel was a friar. The universities that incubated the scientific revolution were, structurally and financially, religious institutions. Science did not arrive and evict religion from the building.
Science moved into religion’s house and kept the furniture.
It inherited the authority structure, a body of credentialled interpreters who could read the text when you could not. It inherited the canonical scripture, first the classical corpus, then the peer-reviewed literature. It inherited the eschatology, the promise that if we keep going, something enormous and redemptive waits at the end. It inherited the heresy trial, the schism, the excommunication, the reformation, the counter-reformation, the mission field, and the tithe.
And that is why the analogy stands up. Because when people say “AI has become a religion,” they usually mean it as an insult, a way of saying the believers are stupid. That’s not the argument. The argument is structural and it is not remotely an insult.
The argument is that when a system of knowledge becomes too large, too fast-moving, and too technically opaque for ordinary people to verify, it does not become irrational. It becomes ecclesiastical. It grows a priesthood, because it must. It grows a scripture, because it must. It grows a category of person who is trusted to interpret and a much larger category of person who is required to believe. My own belief too, is that, we need structure. In the absence of faith, where do we turn? We have self optimisation, but that requires discipline, faith, does not. Faith can be fickle, being faithful, even more so. The idea does not change but maybe the vehicle does.
Mark Coeckelbergh published a book this year called Artificial Religion: On AI, Myth, and Power, and his central move is exactly this: AI is not merely a technical project but a cultural and religious one, and those belief structures actively shape policy and product decisions in rooms where nobody thinks they are doing theology.
They are, though. Sam Altman once described artificial general intelligence as a magic intelligence in the sky. Ilya Sutskever’s rallying phrase inside OpenAI, according to former colleagues, was feel the AGI. Marc Andreessen wrote a manifesto. Yuval Noah Harari has argued the whole apparatus plays a role analogous to traditional faith. I am guessing they are all Star Wars fans too - the originals, just to point out, not the new series.
In May of this year, in New York, representatives of Anthropic and OpenAI sat down with leaders from the Hindu Temple Society of North America, the Bahá’í International Community, the Sikh Coalition, the Greek Orthodox Archdiocese and the Church of Jesus Christ of Latter-day Saints for the inaugural Faith-AI Covenant roundtable, to discuss how to build morality into the technology.
The new church, consulting the old church, about ethics. My Father in Law would shudder at the thought.
I didn’t invent that. I could not have invented that. If I had put it in a satirical script the notes would have come back saying too broad.
connectivity vel imperium
If the Enlightenment is the wrong frame, the Reformation is the right one, and it is right in a way that is almost embarrassingly literal.
The pre-Reformation Church held a monopoly on the text. The Bible existed in Latin. Latin was the language of a trained clergy. Access to God ran through an institution that controlled the interpretive layer, and that institution had a business model attached: indulgences. Salvation, sold forward, in instalments, against a future you could not audit.
The printing press broke it. Not because Gutenberg was a radical, he was a businessman, but because once the text existed in the vernacular and could be copied cheaply, the interpretive monopoly was structurally dead. It took a century of extremely unpleasant European history to work out what came next, but the moment the weights went public, the priesthood’s pricing power was finished. Stack em high, sell em, erm high?
On 27 July 2026, Moonshot released the weights.
The largest model ever built, handed to anyone who wants to download it, from a lab in Beijing, deliberately timed, deliberately open, and priced to make the closed frontier look like a cathedral charging admission.
And what did the market do? The market did what markets do when the interpretive monopoly is threatened: it repriced the priesthood, not the product. Nobody sold TSMC because chips got worse. They sold TSMC because if the frontier can be approximated for half the money by a model you can download, then the extraordinary premium attached to being the only one who can read the text comes off the price. Lots of conversations I am seeing today, probably because of my research centered around the new idea that a moat, is no longer a moat. The idea of a moat is foolish.
The tithe
There is something wonderful about the language of old. Here is what the faithful are being asked to pay.
JPMorgan’s midyear estimate puts global AI-related capital expenditure at $5.5 trillion through 2030, of which roughly $4.1 trillion is expected to be debt-financed. Hyperscaler capex alone is projected at around $650 billion in 2026, rising past $1.1 trillion in 2027. Morgan Stanley’s China equity desk has the top five cloud providers exceeding $800 billion this year and reaching between $1 trillion and $1.2 trillion next.
These are not corporate numbers any more. These are national accounts. This is the scale at which you build motorway networks and national grids and, historically, cathedrals. This is government level spending. It is insane.
Now the part that should make an operator’s blood run cold. Moody’s flagged earlier this year that hyperscalers carry approximately $662 billion in data centre lease commitments that have been signed but not yet commenced, obligations that, under the lease commencement standard, sit off the balance sheet. Six hundred and sixty-two billion dollars of promises to a future that has not arrived yet, structurally invisible to anyone reading the accounts casually.
If you are Irish and you have just felt a small cold hand on the back of your neck, that’s not superstition. That’s memory. We have seen what happens when an economy’s most important obligations are the ones that aren’t in the filing.
Against that: current AI-attributable revenue is estimated somewhere in the range of $50 to $150 billion a year, even when you generously credit every scrap of incremental cloud growth to AI. The gap between the tithe and the collection is not a rounding error. It is an order of magnitude.
And the money moves in a shape that anyone who has worked in ecommerce will recognise instantly, because we watched the DTC sector do a smaller version of it for a decade.
Investment capital flows into AI companies.
AI companies immediately spend it on compute from hyperscalers.
That spend books as revenue at the hyperscaler.
Revenue lifts the hyperscaler’s valuation. The lifted valuation justifies further investment into the sector.
Repeat.
It looks like growth from the outside. From the inside it’s a collection plate being passed down one pew and back up the other.
Multiple analyses have now noted that despite hundreds of billions deployed since 2022, there is still no measurable positive signal in US GDP attributable to AI. That is not proof of a bubble. Productivity gains from general-purpose technologies are famously slow to show up in the statistics, Solow said computers were everywhere except the productivity figures, and he said it in 1987, and he was eventually wrong. But it is the specific gap that historically shows up in the late phase of these things.
The precedent everyone reaches for is telecoms. The late-nineties fibre build-out laid an enormous quantity of glass into the ground on the assumption of demand that took a decade to arrive. Between eighty-five and ninety-five percent of it sat dark. The companies that laid it were substantially wiped out. There is much written about this. Remember Nokia.
The fibre was still there.
Hold that thought. We’ll come back to it.
AI wars = turf wars
Tim Marshall’s whole argument in Prisoners of Geography, and again in The Future of Geography when he takes the same lens off-planet, is that the choices available to a nation are pre-constrained by things that were decided before anyone was born. Russia’s behaviour makes sense when you notice the North European Plain is a doorway with no door. China’s makes sense when you notice how few of its rivers run the useful way. America’s makes sense when you notice it has two oceans, a navigable interior and no serious neighbours. The leaders change. The map doesn’t.
The AI industry has spent three years marketing itself as the first genuinely placeless technology. The cloud. Weights, tokens, inference, latent space, a vocabulary designed to sound like it exists nowhere in particular. Ask where a model lives and you’ll get an answer about parameters. Marketing, all marketing. It is ALL marketing.
It is a lie, and it is the most consequential lie in the sector, because AI is the most brutally geographic industry since oil.
Consider what the stack actually requires, physically:
Land, in large parcels, near fibre trunk routes. Power, in gigawatt quantities, delivered continuously. Water, for cooling, in volumes that put facilities into direct competition with agriculture and municipal supply. Chips, whose leading-edge fabrication is concentrated on an island 180 kilometres off the coast of a country that claims it. Lithography machines, made by exactly one company, in the Netherlands, in a supply chain that cannot be replicated within a decade at any price. Rare earths and processing capacity, dominated by China. Undersea cable, which follows the same routes the telegraph did, which followed the routes the shipping lanes did.
Every single one of those is a chokepoint on a map. Taiwan, Eindhoven, the Gulf, the Nordics, northern Virginia, and Dublin.
Because here’s where it stops being abstract for anyone reading this from Kerry.
Ireland’s data centres consumed 23% of the country’s total electricity in 2025, 7,663 gigawatt-hours, up ten percent in a single year while the rest of the country’s demand rose two percent. Over a decade, that’s a 360% increase. Total residential consumption, every home, every kettle, every immersion, for five and a bit million people, is around 28%. The server farms are within touching distance of overtaking the entire population.
Dublin is the second-largest data centre market in Europe. It got that way because of tax policy, English-language talent, EU membership and a transatlantic cable landing position, which is to say, because of geography and legislation, the two things Marshall says determine everything.
And then the map bit back. In November 2021 the Commission for Regulation of Utilities imposed what amounted to a moratorium on new grid connections in the Greater Dublin Area, because EirGrid could not guarantee the lights would stay on. That moratorium was lifted at the end of 2025, but with a condition that tells you everything about where the power actually sits now: any new data centre seeking a connection must install on-site generation or storage capable of meeting its own full demand, and must be able to give power back to the grid when the country needs it. This is arecent development and not yet set in stone. But it is an A typical Irish solution, from an A typical Irish government. Finger on the pulse as ever.
Read that again. The state has effectively told the hyperscalers: you may build your cathedral, but you must bring your own light, and when the village is cold you will heat the village.
That is a small country using the only leverage it has, physical connection to a constrained grid, against counterparties whose market capitalisations exceed its GDP by multiples. It is genuinely one of the more interesting regulatory manoeuvres in Europe and almost nobody outside the sector noticed. BUt maybe, there is a reason - it is so idiotic so as not to be believed.
This is the geo-technocratic layer, and it is the thing your average AI commentary misses entirely because it is written by people who have never had to ring ESB Networks. The geopolitical story is Washington versus Beijing. The geo-technocratic story is a planning application in Ennis, a substation upgrade in Clonee, a water abstraction licence, a Dutch export control list, and a cable landing station on a beach in Mayo.
Belief is weightless.
Two congregations
Now to the thing that I think is the actual insight in all of this, and it’s the one that comes from working on both sides of the Atlantic rather than reading about one side from the other. I mentioned this at the outset, ownership of 9//10ths of the law.
Americans are bullish about AI partly because they own it. Europeans are sceptical partly because we don’t.
That sounds glib. It isn’t. The numbers are stark and they explain a decade of mutual incomprehension.
In the United States, Gallup’s April 2026 survey found 58% of American adults own stock, through individual holdings, mutual funds, 401(k)s or IRAs. Federal Reserve data has household financial assets allocated to equities at around 45%, an all-time high. When the S&P moves, it moves through the retirement account of the man fixing your car and the woman teaching your children. (If this were an a movie, think The Big Short, the fun bits inter spliced with the rest of the footage. )
In the European Union, roughly €10 trillion of household savings sits in low-yield bank deposits, about seventy percent of the total. BCG puts the liquid cash-and-deposits figure closer to €12 trillion. The ECB has calculated that if EU households simply mirrored American allocation behaviour, an additional €8 trillion would flow into securities markets.
Roughly a third of European household financial wealth sits in cash. In the US the ratio is inverted.
BCG ran a digital forum with more than five thousand people across Germany, France, Italy and Spain to work out why, and buried in it is an useful anecdote. A respondent recalled being handed €100 by their father as a birthday present, with the instruction that they could do anything at all with it, buy video games, drink it, as long as they didn’t buy shares. Brewsters 100s we will call it.
Drink it, but don’t invest it. There is a thousand years of European history in that sentence. Two world wars fought across the continent, hyperinflation, currency collapse, bank failures, nationalisation, occupation, and, for the Irish specifically, a property crash so total it reorganised the national psyche around the conviction that anything which goes up that fast is lying to you.
Europeans don’t distrust the stock market because we’re stupid about money. We distrust it because we have institutional memory of what happens when everyone believes the same thing at the same time.
So watch what this does to the AI argument.
An American with a 401(k) heavy in the S&P 500, where the index’s returns have been dominated by a handful of AI-exposed names, is not merely observing the AI story. They are inside it. Their retirement is a leveraged position on the thesis being true. And human beings are not built to assess evidence neutrally when their own security depends on the conclusion. That’s not a character flaw. It is just how belief works. It’s the oldest finding in psychology and it’s why every religion in history has asked for money before it asked for anything else. Skin in the game is not a hedge against faith.
You put your money in, and then your mouth follows your money. This is my interpretation from a far. .
Meanwhile a European with the same wages and no equity exposure watches the same headlines from a position of total structural detachment. If it works, they get nothing. If it collapses, they lose nothing. That produces a scepticism which feels like superior analysis and is in fact mostly just the absence of a position.
Both sides are being shaped by their balance sheets and both sides believe they are being shaped by their reasoning. The American who thinks the European is a pessimist is wrong. The European who thinks the American is a mark is also wrong.
But, and here is where it gets uncomfortable for the American side of the argument, the ownership is far narrower than the participation. The wealthiest 1% of Americans hold 50% of all equities, worth roughly $27.6 trillion as of the first quarter of 2026. The top 1% hold more equity than the bottom 90% combined. The top 10% hold something like 87-93% of household stock market wealth. So while 58% participate, a majority of that 58% own an amount that would not cover a boiler replacement. These numbers will vary by a few points depending on where you get your info, but this is summary I am comfortable with.
Which means the American position is not really we all own it. It’s we are all in the congregation, and a very small number of us own the building.
And one more number, which I think is the most under-reported statistic of the year. That 58% ownership figure? It was 62% in 2025. It fell for the first time since 2016.
Attendance is down.
Not collapsed. Down. Four points. The first negative print in a decade. This was pretty mind blowing. In any other institution you’d call that the early tremor. If you’ve ever watched a parish notice its own numbers slipping, you know the sequence: nobody announces they’ve stopped believing. They just stop turning up.
The European position, meanwhile, has quietly become policy. The Savings and Investments Union is, stripped of its Brussels vocabulary, an attempt by the European Commission to get €10 trillion out of deposit accounts and into capital markets, because the Draghi report costed European competitiveness at an additional €750-800 billion of investment per year by 2030, and there is no other pot of money that size on the continent.
So set that up properly, because it’s genuinely funny and genuinely serious at once (like me): at the precise moment the American congregation records its first decline in a decade, Europe is running a formal, institutional, treaty-level campaign to recruit its population into the church.
We are trying to buy in at the top of the market, using the savings of a population whose grandparents’ entire life lesson was don’t.
Conversion therapy - govt wise
Governments do not usually join religions. That step is usually taken a long way time before that. They regulate them, tax them, occasionally establish them, and mostly try to survive them. What has happened over the last eighteen months is different: states have stopped regulating the AI story and started investing in it, which means they have moved from the magistrate’s bench into the pews.
The United Kingdom launched a £500 million Sovereign AI Fund in April 2026, a state-backed venture vehicle writing £1-10 million equity cheques into British AI companies, bundled with access to the national AI Research Resource and fast-tracked visas, alongside a £282 million Strategic Assets Grants Programme. The framing from Downing Street is that Britain must be an AI maker rather than an AI taker.
Macron used Davos in January to argue that Europe needs dramatically more investment in AI, quantum, cleantech and defence in the name of sovereignty and autonomy. The EU Chips Act is funding five AI gigafactories. Mistral has raised €1.7 billion at an €11.7 billion valuation with ASML as its largest shareholder at around eleven percent, which is to say, the company that makes the lithography machines now owns a piece of the company that makes the models, which is a vertical integration that would have got you an antitrust letter in 1975 and gets you a national champion designation in 2026.
And then, in March, the tell. Palantir and NVIDIA announced a Sovereign AI Operating System Reference Architecture, a turnkey national AI stack, hardware through applications, targeting what they described as a $600 billion market. They are pretty bullish about themselves. Opening up, if you can fathom this, a debate on transparency. Panantir?
Sovereignty. As a product. With a reference architecture and, presumably, a quote. This is not an overnight story, but it is a different approach from the company who usually title reports - What we do in the shadows.
You can sell a country the means of not depending on you. Rome ran the same model: you didn’t have to build your own cathedral from first principles, you could buy the plans, the relics, the trained staff and the liturgy, and the only condition was that the whole thing pointed back toward Rome. The word for that arrangement is not independence. The word for it is franchise.
Meanwhile Europe’s other instrument comes online on 2 August 2026, when the substantive obligations of the AI Act hit enforcement. And this is where the transatlantic asymmetry stops being cultural and starts being structural.
America is running revelation: build the frontier, discover what it does, apologise later, capture the value. Europe is running doctrine: define what is permitted, certify the compliant, tax the rest, protect the citizen.
Everyone frames this as innovation versus regulation, which is the argument the American side wants to have because they win it. The better frame is that Europe is not running a competing innovation strategy at all. Europe is running a counter-reformation, an institutional response by an older power structure to a disruptive new one, built on documentation, licensing, certification and the assertion of interpretive authority.
The counter-reformation, incidentally, worked. It did not stop Protestantism. It did preserve the institution, consolidate its remaining territory, professionalise its bureaucracy and keep it standing for another five centuries. If your goal is survival with dignity rather than victory, it is a perfectly rational strategy and it has an excellent track record.
Let’s stop pretending the AI Act is an innovation policy that went wrong. It’s a doctrinal instrument that is working precisely as designed.
Unclean, unclean: The heresy trial
By 24 July, nine days after the world’s largest open model went public, the White House had escalated its position on Moonshot from technological competition to allegations of illicitly obtaining Nvidia Blackwell chips and distilling US models at scale.
Note the shift in category. No longer, a trade dispute. This is a heresy trial, and it follows an age old pattern: the challenge appears; the challenge is initially dismissed as inferior; the challenge proves popular; the response reframes the challenger not as a rival interpreter but as a thief and a corrupter of the original.
Whether the allegations are true is genuinely material and I’m not equipped to adjudicate it, chip smuggling is real, distillation is real, and the fifty-one percent hallucination rate suggests something in K3’s training has gone sideways in an interesting way. There’s also a hard-nosed enterprise argument that has nothing to do with theology: China’s National Intelligence Law creates obligations for Chinese firms that no European or American compliance officer can wave away, and routing your business logic through Beijing-hosted inference is a decision with consequences beyond price per token.
But watch the structure of the argument rather than its content. Two weeks ago the US position was that America’s lead was insurmountable because of compute. Now the position is that the lead was breached by theft. Those are not compatible claims, and the speed of the switch tells you which one was load-bearing.
The tell is always the same. Even James Bond
Meanwhile, in the shop
Ecommerce hat back on. On 15 July 2026, the same week as everything above, which is either coincidence or the most on-the-nose scheduling in fintech history, Reuters reported that Stripe and Advent International had jointly bid $53.4 billion for PayPal. Sixty dollars fifty a share, a 28% premium, backed by around $50 billion in committed bank financing, with Stripe and Advent taking fifty percent each and PayPal kept intact rather than broken up.
It would be the largest fintech acquisition ever attempted. Axios flagged the structural oddity immediately: a venture-backed private company buying an S&P 500 constituent. Stripe, privately valued around $159 billion, bidding for a listed company that has spent four years being systematically dismantled by the market’s opinion of its future, a company that has already been through a CEO change this year, with HP’s Enrique Lores taking over from Alex Chriss.
Everyone read it as consolidation. Payments is a scale game, PayPal’s multiple was depressed, private equity smelled a carve-out. All true, all boring.
When everyone looks left, I look right, or wrong.
Stripe is not buying PayPal’s merchant business. Stripe has a merchant business; Braintree competes with it directly. What Stripe does not have, what Stripe has never had, and what it cannot build organically at any speed that matters, is the consumer side. More than 400 million active accounts. Venmo. Possibly, the single most recognised checkout button in the world. A direct relationship with the person holding the card, rather than with the shop taking it.
Why does that matter now, in the specific week that a Chinese lab released open weights and three trillion dollars fell off the chip complex?
Because if agentic commerce is real, if the buying decision migrates from a human on a product page to a model executing an instruction, then the entity that owns the authorisation and settlement layer owns the only piece of the transaction that cannot be abstracted away.
Just because it can does not mean it will.
The storefront can be disintermediated.
The product page can be disintermediated.
The search box has already been disintermediated.
Somebody still has to move the money, prove the identity, take the fraud risk and clear the payment.
All signs point to slowly slowly catchy monkey.
Rails are territory. In a world where the interface is a model that could be anyone’s, the rails are the only territory.
So read the Stripe bid as what it is: a ring-fencing manoeuvre. A pre-emptive land grab, executed at the precise moment the surface layer of commerce is being called into question, by a company that has decided the safest place to stand during a religious war is on the ground everybody has to cross.
And notice the shape it shares with everything else here, the sovereign funds, the Palantir architecture, the Irish grid conditions, the Dutch export controls. In every case, the response to a weightless, placeless, allegedly frictionless technology is somebody quietly building a fence around the physical or contractual chokepoint that the technology cannot route around.
Which brings me back to a thing I’ve been banging on about for two years. Friction is currency. Every era of technology sells itself on the removal of friction, and in every era the money ends up accruing to whoever owns the friction that could not be removed. Not the frictionless bit. The stubborn bit. The grid connection. The lithography machine. The card network. The customs declaration. The island 180 kilometres off the coast.
For the merchant on the ground, the practical translation is short and unglamorous:
Your model choice is depreciating fast, open weights at half the price with a hallucination problem is not a reason to switch today, but it is a reason not to sign a three-year commitment to anyone today either.
Your payments relationship may be about to change ownership, and the terms of a Stripe-PayPal combination will be set in a room you’re not in.
Your data sovereignty position is about to become a purchasing requirement rather than a legal footnote, from 2 August.
And your electricity bill, if you’re Irish, is being shaped by decisions about server farms that have nothing to do with you and every consequence for you.
None of that requires you to have an opinion about AGI. That’s rather the point.
The cathedral gets built anyway
I promised this wouldn’t be a sermon against sermons. The strongest argument against everything I’ve written is Carlota Perez’s, and it is very strong. I had a to look and I am in the getting the head out of the summary place, but I am learning. Perez’s work on technological revolutions describes a repeating structure: an installation phase, driven by financial capital, characterised by frenzy, overbuilding and spectacular capital destruction, followed by a turning point, and then a deployment phase in which the infrastructure built during the madness becomes the substrate of ordinary economic life for the next forty years.
Sounds familiar. Canal mania. Railway mania. The electrification boom. The telecom build-out. In every case the investors were slaughtered and the infrastructure remained. The British railway mania of the 1840s wiped out an enormous quantity of middle-class savings and left Britain with a rail network that carried its economy into the twentieth century. The dark fibre from 1999, eighty-five to ninety-five percent unlit for a decade, is the fibre that carried Netflix, YouTube and every ecommerce transaction any of us have ever processed.
Bubbles are, unfortunately, how societies finance infrastructure that no rational actor would fund at the right price. The mania is not a bug in the mechanism. On the available historical evidence, it is the mechanism.
So the sceptical position has to survive that, and mine does, but only just, and only by narrowing.
Because Kevin Kelly’s protopia idea does real work here and it cuts both ways. Kelly’s argument is that the future arrives not as utopia or catastrophe but as grinding, unglamorous, barely-perceptible improvement, a one percent a year that compounds into transformation you only notice in retrospect. Kaizen, wiuthout the glanmour. If that’s what AI is, then it’s real, it’s valuable, and it is also catastrophically overpriced, because nobody builds a cathedral to one percent a year. Protopia doesn’t need a $5.5 trillion capital programme. Protopia needs patience, and patience is the one thing an equity market allocating 45% of American household financial assets cannot supply.
And Tom Friedman’s acceleration point supplies the mechanism I couldn’t find on my own. His argument in Thank You for Being Late is that technology has begun outrunning the human and institutional capacity to adapt to it, that the machine is moving faster than the manual. Fine. But look at what actually fills that gap. When capability outstrips comprehension, and the people affected cannot personally verify the claims being made, the space between is not filled with analysis.
It is filled with faith. It has to be. There is nothing else that fits in there.
And Karen Hao’s contribution to my thinking has been the discipline of the question rather than any particular answer: what is the real cost of this, and who is paying it? Empire has always required a theology, because extraction at scale needs a story about why it is good. The story is never the mechanism.
So, It is a bubble, and it is real infrastructure, and both facts will be true simultaneously for years. The capital will be destroyed and the data centres will still be standing. The people who believed hardest will lose the most and will have been, in the long historical arithmetic, correct about the direction. The people who stood back and sneered will be proven right about the price and will have owned none of the upside, and will tell themselves that was wisdom.
The question that actually matters isn’t is it a bubble. That question is a way of avoiding a harder one.
The harder one is: when the cathedral is finished and the congregation has thinned, who holds the mortgage? Enough with the church analogies, sorry, not sorry.
The answer is already visible in the numbers. The tithe is being paid by American retirement accounts, by €10 trillion of European savings currently being coaxed out of deposit accounts by treaty-level policy, by Irish households paying network charges on a grid whose capacity is being consumed by facilities they don’t own and can’t visit, and increasingly by sovereign debt, $4.1 trillion of the build-out expected to be debt-financed, at a moment when governments have also decided to become equity investors in the same story.
The cathedral will get built. It always does. The fibre is still in the ground. The railway is still running. Where are we now in this debate. I don’t have a prediction, and if I offered one you should discount it, because the entire argument of this essay is that confident prediction in a period of low verifiability is just theology.
What I have is an orientation, and it’s this.
Start asking where the chokepoints are, who owns them, and what happens to your business when the price of the abundant thing collapses and the price of the scarce thing does not.
The abundant thing is intelligence. That’s what 27 July meant. Weights are going to the vernacular, capability is commoditising, and the models will end up costing roughly what electricity costs, which is to say: a lot in aggregate and almost nothing per unit.
The scarce things are the ones on the map. Grid connections. Water rights. Lithography. Cable landings. Card networks. Customs regimes. Regulatory permission. Trust.
Every single one of those is geographic, legal, or relational. Not one of them is a model.
Tim Marshall’s point was never that geography is destiny in a fatalistic sense. It was that leaders who understand the map make better decisions than leaders who believe they’ve transcended it. The AI industry has spent three years insisting it has transcended the map, and the map has spent the last three months quietly submitting its invoice, in a Dublin grid condition, a Dutch export licence, a Beijing weight release, a Washington indictment, and a $53 billion bid for the right to own the pipe rather than the promise.
We are not at the end of the Age of Reason. We never left it. We’re just discovering, as every generation does, that reason is expensive, slow, and much less comforting than the alternative, and that when the numbers get too large to check, human beings will reach for belief the way they always have, and then build something enormous on top of it.
The cathedral is going up. It is going up on land, with power, near water, on an island somebody else can blockade.
Say a prayer for whoever is left holding the paper.
Sources and figures verified as of 26 July 2026. Market data: Philadelphia Semiconductor Index performance and the $3.3tn semiconductor drawdown (multiple outlets, w/c 13 July 2026); TSMC results and share reaction (Fortune, 17 July); Kimi K3 specifications, release timing and independent hallucination testing (Moonshot AI announcements, Artificial Analysis, TechTimes). Capex: JPMorgan midyear outlook via Fortune; Moody’s lease commitment analysis; Morgan Stanley via Tencent Finance. Ownership data: Gallup Economy and Personal Finance survey April 2026; Federal Reserve Distribution of Financial Accounts Q1 2026; European Commission and Council SIU documentation; BCG household savings forum, June 2026. Irish energy data: CSO data centre metered consumption release 2026; CRU/EirGrid connection policy, December 2025. Stripe/PayPal: Reuters via TechCrunch, Axios, CNBC, 15 July 2026.