The trade clears at 09:30:01. A retired schoolteacher in Dundee owns a sliver of SpaceX now, bought at the open, at whatever the open was asking. She didn't choose it. She's never heard of the seasoning rule, couldn't tell you what an index provider is. Her pension sits in an index fund, and an index fund does what the index says the instant it's told, which this morning means buying into one of the most expensive listings in market history. The whole thing takes milliseconds.
She's not reckless. She's not even in the room. And she's just been put on the wrong side of a bet nobody showed her.
Let me separate two things first, because they keep getting welded together: the technology, and the bet. The AI you actually touch, the thing that drafts your emails and reads the contract and sits quietly inside your working day, is the most useful thing to happen to ordinary work in a generation. Genuinely. I build with it every morning. Hold onto that, because what follows isn't an argument against the machines. It's an argument about who's paying for them, and on what terms.
The Forced Bet
In poker you post the big blind before you've seen a single card. It's a wager you're made to place, blind, just for sitting at the table. This spring, without most people noticing, a very large blind got posted in their name.
The visible version landed in June, when SpaceX listed on Nasdaq in the largest flotation in history. That part's spectacle, and I'm an admirer of the rockets, much to my wife's despair over dinner. The part that matters is duller. In the weeks before the listing, FTSE Russell cut the seasoning period a new stock waits before it enters the indices, from around three months to five trading days. Nasdaq cut its own to fifteen. S&P looked at the same change and said no. The rules that did change are mechanical: index funds, the boring default under most workplace pensions, have to buy a stock the moment it joins the index, and analysts put the forced buying at ten to sixteen billion dollars in a single session.
So savers were bought in at the opening price, whether they wanted to be or not. And what they bought is a stock Morningstar reckons is worth 780 billion dollars, less than half the 1.77 trillion it listed at: a rocket company that, back in February, swallowed Elon Musk's cash-burning AI lab, xAI, which lost 6.4 billion dollars last year. Your pension might already hold a slice of it, and nobody asked. The rules got changed in the weeks before, so nobody had to.
That's the bit I can't get past. But sadly, it's only one hand at a much bigger table.
What Everyone Bet On
Pull back, and the whole table comes into view. The big technology companies have made one enormous, coordinated wager, and the scale of it is hard to hold in your head. Capital spending across Big Tech is running at around 725 billion dollars this year, up roughly 77 per cent. Morgan Stanley reckons another 800 billion dollars of the build-out will have to come from private credit. And a lot of the money just loops between the same few names. Nvidia invests in OpenAI. OpenAI buys compute that runs on Nvidia chips. The cloud providers underwrite the labs, and the chipmaker's revenue quietly counts everyone's spending twice. The economists who sat through the dot-com years have a word for money that circles back to where it started. Round-tripping.
The biggest stack belongs to Sam Altman. OpenAI's own leaked forecasts have it burning through something close to 665 billion dollars by 2030, and not breaking even until then. It raised 122 billion in the spring at a valuation near 852, and it'll need to raise again. None of this is illegal, and none of it is stupid. It's just an all-in bet: build a big enough furnace, and the intelligence, and the returns, will come.
And more and more, the furnace is paid for by someone other than the people making the bet. The data centres are financed off balance sheet, through special-purpose vehicles and private credit, so when a build runs over or sits half-lit, it's a pension fund or an insurer or a sovereign lender holding the paper, not the lab with its name on the door. The bet stays private. The risk gets laid off onto the public, one quiet structure at a time. The same public an index fund just marched into SpaceX.
Is the Bet Even Right?
That lay-off would be defensible if the bet were obviously right. It isn't yet obviously anything. Last summer a study out of MIT found 95 per cent of corporate AI pilots showed no measurable return to the bottom line. Goldman's chief economist, Jan Hatzius, reckons AI added "basically zero" to US growth in 2025, and called the spending FOMO, not ROI. I think these points aren't highlighting a case against AI, they're highlighting that the problem is with AI Adoption and Integration. Most companies have been hurling AI into every KPI without sitting down and properly understanding mapping where it truly adds value and where people have incentives to use it (indeed, too many forget the people side of the equation but that's for another essay - graph below illustrates this point well enough for now)
Meanwhile Michael Burry, who famously shorted the 2008 housing crash, says the cloud giants are flattering their profits by pretending their chips will last far longer than they will. He might be wrong, and the chipmakers say he is. Whatever the truth of the matter, every big player has a reason to keep the depreciation clock running slow, and Burry's not the first to ask, out loud, to see the cards.
The strongest reply is that the demand is real and the capacity is genuinely short. Satya Nadella says the bottleneck isn't a compute glut, it's power: he's got chips he can't plug in. I think that's true, and it's the part of the bull case I take most seriously. But a shortage today doesn't justify the spend if tomorrow needs a fraction of the compute.
You can see the tell in where the "smart money" says data centres have to eventually go. Not Arizona. Orbit. Musk says space-based compute will hit cost parity "within two years, maybe three." Google's already testing solar-powered chips in orbit, under the name Project Suncatcher. Wood Mackenzie ran the numbers: a one-gigawatt data centre in space would cost around 170 billion dollars, more than three times its earthbound twin, and cooling a single megawatt in a vacuum needs a radiator the size of an ice-hockey rink, because heat sheds a thousand times slower with no air to carry it off. When the answer to a power shortage on Earth is to leave the planet, the scale story has certainly outrun reality.
I'm not saying hockey rink-sized radiators in space are impossible. But apply Elon Musk's preferred 'first principles' thinking and, ironically, it's the option loaded with the most hidden assumptions, even though it sits at the centre of SpaceX's IPO-prospectus path to profitability.
The Wager Turns Over
This is where the wager goes from bad to worse. The bet assumes bigger is the only road, that better always runs through more: more chips, more power, more concrete. But scaling laws only ever described one phase, the early, brute-force one. They're not a law of nature that forbids the next phase. Those at the fore of LLM research do already understand this, as Andrej Karpathy recently talked about us entering into the post-Scale phase. Unfortunately for data centre investors, the next frontier is the one that cannibalises the scale investments. Efficiency.
Look at what efficiency has already done. DeepSeek trained a frontier-class model on 2,048 chips for around 5.6 million dollars, a rounding error next to the American budgets. Frontier models have quietly got an order of magnitude smaller since 2023, not bigger. Running a model as good as the one that stunned everyone in 2022 has dropped from about 20 dollars per million tokens to seven cents, nearly 280 times cheaper. Google's published a compression trick called TurboQuant that it says shrinks a model's working memory to about three bits, with no retraining. Gartner expects the cost of running even a trillion-parameter model to fall more than 90 per cent by 2030. We've barely started researching how little compute intelligence actually needs, and every month the answer comes back the same: less than we just spent.
Meanwhile the floor keeps rising under everyone. Open-weight models, free to download and run on your own machine, are closing on the closed frontier fast enough to frighten anyone whose business is selling access to a model. China's Zhipu released GLM-5.2 in June; on the labs' own benchmarks it trades blows with the best in the West. When a capable model is a free download, and running it gets 90 per cent cheaper in a decade, the model itself isn't a moat but instead it's a commodity.
So what's actually left? That's the question that matters, and it's where the value was hiding all along. The durable thing was never the model. It's the harness around it: the integration into real workflows, the proprietary data no download can hand you, the people who sit next to a customer and make the thing actually work. There's a name for those people now, the forward-deployed engineer, borrowed from Palantir, and right now the labs are hiring them as fast as they can find them. Sequoia put a number on it: for every dollar spent on AI software, about six go on the services around it. They went all in on the engine.
The labs know it, which is the tell inside the tell. Anthropic launched a 1.5 billion dollar enterprise-services arm in early May. OpenAI followed with its own a week or so later. The companies selling the model are quietly racing to become the companies selling the harness.
Where I Stand, and Who Survives
I should say where I stand, because I'm not a neutral observer here. I'm building CURN inside this same wave, warmed by the very fire I'm worried about, living off the same capital that props up the valuations I'm questioning. I don't have a clean answer for that. I just know the people writing breathless threads about the listing need to understand the unintended consequences of these companies reaching for the stars.
The sad thing is, I think these companies will come through largely unscathed. SpaceX, OpenAI and Anthropic's investors will mostly cash out through their IPOs, and the risk transfers onto the public markets. Between the three problems: model efficiency, open source catching up, and shady private market credit, it's almost certainly a question of when, not if, the reckoning comes.
Because the reckoning, when it comes, won't kill AI any more than 1873 killed the railways or the dot-com crash killed the internet. The track always gets inherited. Level 3 bought Global Crossing's transatlantic cables for three billion dollars in 2011, a fraction of what they cost to lay, and that fibre is still down there, still carrying your email. What a reckoning does is sort the people who built a business from the people left holding a forced bet.
And a forced bet landing on people who never agreed to it isn't new. It isn't even old. In 2007, the cleverest minds in finance took American mortgages, sliced them up, stamped them safe, and posted the pieces into pension funds and town halls and the savings of people an ocean away who'd never read the paperwork. When it blew, the people who built it were mostly fine. The losses landed on the ones who never chose to hold them. Burry was one of the few who saw it coming. It seems that private credit may have just filled the vacant role. He called it once except this time he's joined by many, many others calling it again.
Look at Your Cards
So before the next round's dealt, here's the one useful thing I'd urge: Know what you're holding. Most people you know have no idea their default pension or Stocks and Shares ISA / IRA index funds now hold a chunk of a loss-making rocket-and-chatbot conglomerate run, almost entirely, by one man. Whilst we have two further IPOs of enormous size to lay the risk on further.
None of this makes the machines any less real, or less wonderful. The technology is unquestionably great but it needs time for people to find its value, for adoption to truly take hold. The action worth sitting with is simple: know what you own, and help your loved ones know what they own too.