The Trillion-Dollar Stack

AI has entered its industrial phase.

With SpaceX's record listing, and Anthropic and OpenAI moving towards their own, the next phase of artificial intelligence is shifting from model releases to power, capital expenditure and the physical limits of scale.

The public markets spent a few quiet years waiting for a new centre of gravity. This June, they received one record listing and two signals. SpaceX completed the largest initial public offering by proceeds in history, raising about $75bn at $135 a share, a debut that valued Elon Musk's rocket-and-satellite empire at roughly $1.75tn and, after a first-day jump of around 19%, pushed it above $2tn, briefly worth more than Tesla. Filings on either side did the rest. Anthropic confidentially submitted a draft registration statement on 1 June, only days after raising $65bn at a $965bn post-money valuation. OpenAI announced its own confidential S-1 a week later, with the careful note that it had "not decided on timing yet" because some things are simply easier to do as a private company. Whether the two model labs list this autumn or later, the paperwork has already changed the conversation.

The reason a rocket company belongs at the front of an article about artificial intelligence is that SpaceX, in early 2026 it absorbed xAI, Musk's AI lab, folding Grok and the former Twitter into the same balance sheet, and a growing share of its enormous capital budget now goes not to Mars but to compute. Its first-quarter capital expenditure more than doubled year on year, to around $10bn, most of it, the company says, on AI. The listing is therefore the first occasion on which public investors can buy frontier intelligence and the physical means of producing it in a single ticker, which is either the perfect expression of this moment or a warning about it, depending on one’s tolerance for capital intensity.

The Bill Arrives

For a few years the AI cycle has been a private affair, financed by investors willing to underwrite ambition on the strength of a demo and a curve. Public markets are less romantic. A listing forces narrative into disclosure: it turns ambition into risk factors, capital expenditure, customer concentration, gross margins, losses and the assumptions behind future growth, and it invites a few thousand strangers to disagree with the story in real time. Private markets can live with mythology for a long while. Public markets eventually ask for the bill.

The early readings of that bill are instructive. SpaceX posted $18.7bn of revenue for 2025 and a net loss near $5bn, which at the offer price worked out to more than ninety times trailing sales; one analyst opened coverage at "sell" with a target below the float, and Morningstar pegged fair value at less than half of it, describing most of the premium as an option on ambitions such as orbital data centres and Mars. OpenAI, by one widely circulated estimate, still loses around $1.22 for every dollar it earns. Anthropic has grown its revenue run-rate to about $47bn, from roughly $10bn a year earlier, which is the sort of figure that makes losses easier to forgive, though it too carries the compute bill that defines the category. None of these are software margins, they are the economics of an industry that has discovered it owns a great deal of very expensive plant.

For most users, AI arrived as something close to weightless: a prompt box, a chatbot, an image conjured from a sentence, its power apparently sitting entirely inside the interface. It was easy to forget the machinery, because the machinery belonged to someone else and never sent you an invoice. For anyone watching the stack, though, the binding constraint was always going to be industrial rather than magical. Scaling AI depends on chips, data centres, cooling systems, substations, grid connections, fibre, land, water, transformers, turbines and power-purchase agreements, and on a level of capital expenditure that belongs more naturally to the history of railways and electrification than to the mythology of software. The cloud was always a metaphor. Intelligence now has a balance sheet, and the IPO window is the moment it gets audited.

A Prophecy, With Footnotes

The more sophisticated reading of the AI cycle began to move away from the question of which model wins some time ago. The better question was where the bottlenecks would sit, who would control them, and which parts of the stack could turn scarcity into durable economics. The figure who made the shift most legible is Leopold Aschenbrenner, which is itself slightly odd, as he is a former OpenAI researcher who, two months after being dismissed from the company's Superalignment team in 2024, published a 165-page essay series called Situational Awareness: The Decade Ahead.

The argument, stripped of its length, runs like this. Progress in AI can be tracked by "counting the OOMs", the orders of magnitude of improvement delivered each year by three compounding forces: raw compute, algorithmic efficiency, and "unhobbling", the unlocking of latent capability through better tools and scaffolding. Extend that line and Aschenbrenner reaches artificial general intelligence by around 2027, followed by an "intelligence explosion" in which AI systems able to do the work of AI researchers begin improving themselves, compressing a decade of progress into a year or two. And here is the part that matters: none of it happens in the abstract. The race to AGI, he wrote, would not play out only in code and behind laptops; it would be a race to mobilise America's industrial might. He forecast individual training clusters costing first hundreds of billions and then trillions of dollars, each drawing power on the scale of a small American state, and framed the whole thing as a national-security contest that would eventually pull in the United States government.

You can disagree with the destination, and plenty of serious people do; 2027 strikes them as aggressive to the point of comedy, and Aschenbrenner himself has been conspicuously wrong before, most memorably when DeepSeek's cheap Chinese model briefly challenged the most aggressive version of his thesis in early 2025. But the structural claim under the prophecy is harder to dismiss, and it is the one the market has quietly started to price: that the binding constraint on advanced AI is physical. Compute, power and industrial capacity are what is actually scarce.

From Manifesto To Market Force

And then Aschenbrenner turned the thesis into a fund, and the fund has been, by any ordinary standard, a triumph. Situational Awareness LP launched in September 2024 with a few hundred million dollars and a manager who had never run money before, backed early by the sort of names that make other allocators pay attention, the Collison brothers, Nat Friedman and Daniel Gross, with the quantitative trading firm Jane Street later joining as an investor. It is reportedly up more than 1,000% net of fees since inception, around 270% in 2026 alone through May, and has grown its assets above $20bn in under two years, which puts a two-year-old shop run by a twenty-four-year-old in roughly the same weight class as Pershing Square. Its filings now move markets on their own: Nebius rose about 12% when the fund's stake surfaced, and the solar manufacturer T1 Energy jumped more than 20% on the same kind of news. When a manager’s quarterly filing becomes a tradable event, the disclosure is no longer just a record of conviction; it becomes part of the market’s own pricing mechanism.

For all the infrastructure preaching, Aschenbrenner's single largest disclosed position is reportedly Anthropic, a model lab, at around a fifth of the book. He bought in near a $60bn valuation in early 2025 and has watched it re-rate toward $965bn, which is to say the prophet of "it is the power plants, not the algorithms" has made his biggest paper fortune on, of all things, an algorithm company, and stands to make a great deal more if that company lists near a trillion dollars. Meanwhile his most recent 13F, the quarterly inventory of US holdings that the American regulator requires every large manager to file within forty-five days of each quarter's end, and which the whole market then dissects, disclosed billions of dollars in put-option exposure across marquee AI hardware names, including Nvidia and other semiconductor-linked exposures, sitting alongside long infrastructure bets. One should not over-read this. The filing is stale by the time it appears, captures only US-listed longs and the bare existence of options rather than the structure around them, and a put can be a hedge, a financing leg or one side of a spread quite as easily as a bearish wager. But the shape of it points at something more sophisticated than "buy AI infrastructure." It suggests a rotation within the stack itself: away from the most crowded chip expressions and towards cheaper owners of power, land, cooling and data-centre capacity.

The cleanest single emblem of the thesis is Nebius. In late May the fund disclosed a 5.6% stake, about $2.6bn, in the Amsterdam-listed neocloud that is itself the remnant of the old Yandex, and the stock jumped roughly 12% on the news, because that is now what happens when Aschenbrenner's name appears on a filing.

Beyond Unicorn

The numbers around all this have escaped the usual vocabulary of venture capital, which tends to top out around "unicorn". Goldman Sachs Research expects the four largest hyperscalers to spend a combined $5.3tn on AI and data centres between 2025 and 2030; its global institute, totting up chips, data centres and power across the whole industry, sketches something nearer $7.6tn cumulatively by 2031, while stressing that these are baseline estimates rather than forecasts, and so sensitive that changing a single assumption about how long an AI chip stays useful swings the total by hundreds of billions. McKinsey reckons global data-centre capital spending will approach $7tn by 2030, around $6.7tn, of which roughly $5.2tn is for the AI-specific kind. And the International Energy Agency expects data-centre electricity consumption to more than double by 2030, to around 945 terawatt-hours, slightly more than Japan's entire annual usage, with the AI-optimised slice projected to quadruple. These are the numbers of a heavy industry that happens to think in tokens, and they are now large enough to pull genuinely industrial capital into the centre of the AI story.

In June, KKR launched Helix Digital Infrastructure, a company seeded with more than $10bn of long-duration commitments and a founding shareholder list, KKR, the Kuwait Investment Authority, Nvidia and Vistra, that reads like a diagram of the entire problem. Nvidia supplies the design logic of accelerated computing and frets publicly about "tokens per watt"; Vistra, a generator with a fleet of roughly fifty gigawatts and existing hyperscaler contracts, supplies the electricity and the relationships; a Gulf sovereign fund and a buyout giant supply the patient balance sheet; and a former chief executive of Amazon Web Services, Adam Selipsky, runs it. The structure is the message. The line between technology and infrastructure has dissolved to the point where a chipmaker, a utility, a sovereign wealth fund and a private-equity firm now sit around one table, because none of them can build the future alone.

Jeff Bezos is walking up to the same wall from the opposite side. His startup Prometheus, run with the scientist Vik Bajaj, raised a $12bn round in June at a valuation near $41bn, on top of a $6.2bn launch only seven months earlier, taking total funding past $18bn before it has shipped a product or disclosed a dollar of revenue. Prometheus is not chasing another chatbot; it is building what Bezos calls an "artificial general engineer", AI aimed at the physical world, at jet engines and chips and drug compounds, treating the whole design-to-manufacturing pipeline as one end-to-end problem.

Where The Surplus Settles

All of which makes the trade more interesting and, awkwardly, harder. The first-order AI trade was conceptually clean: buy the obvious beneficiaries and let the narrative do the work. The second-order trade demands discrimination, because a company can be essential to the buildout and still be a poor investment at the price. A bottleneck can be real and temporary. A capacity provider can have all the demand it wants and watch its margins compress anyway. A chipmaker can remain extraordinary while its valuation already assumes it stays that way forever. A data-centre developer can hold exactly the right land with exactly the wrong power economics. This is the gap Aschenbrenner's puts seem to gesture at, and the IPO window will make it impossible to ignore, because a public price is a daily, unforgiving vote on precisely these questions.

History is generous with the precedent. Railways, telecoms, fibre, shale and cloud computing each carried a version of the same story, in which the infrastructure proved necessary and the returns to the people who financed it proved wildly uneven. Some of the rails became monopolies. A great deal of the late-1990s fibre was laid by companies that went bankrupt precisely so that later owners could buy the cable for cents and run the internet over it. The future was real; it was simply overfunded in the wrong places, by the wrong people, at the wrong moment. AI may well rhyme. Believing in the buildout and believing that every participant in it will earn an attractive return are entirely different propositions, and conflating them is how capital cycles tend to end.

So the question is no longer whether AI matters; the market has settled that. The harder question is where the economic surplus comes to rest once the capital has been spent. Does it accrue to the model owner, the chip supplier, the cloud platform, the neocloud, the power provider, the grid operator, the landowner, the private lender, or the hyperscaler with the cheapest cost of capital? Does it go to whoever controls the bottleneck, or to whoever turns compute into revenue before depreciation begins to bite? Capital expenditure can build empires and it can build overcapacity, frequently from the same blueprint, and the difference is rarely visible until afterwards.

AI As A Pressure, Not A Sector

For private markets, this reads less as a warning than an invitation, because AI infrastructure is almost perfectly suited to private capital. The assets are large, illiquid, capital-intensive and slow, and they reward patient financing, structured partnerships, development risk and real energy expertise rather than quick mark-ups. Goldman, unsurprisingly given who pays it, expects private markets to take an ever larger role in the financing, with infrastructure funds, which raised a record $221bn last year, potentially holding more than $3tn of assets by 2030. Helix is what that looks like once it is built.

It also quietly changes what "technology exposure" even means. A serious allocator may no longer reach AI mainly through a handful of mega-cap technology stocks or a venture fund backing application-layer startups. The exposure increasingly lives inside power, real estate, private credit, secondaries, equipment finance, energy transition, materials and industrial software, which makes AI less a sector than a pressure moving through the whole economy, surfacing in places that had nothing to do with computing five years ago. That is what makes it interesting, and what makes it messy. It competes for land, power and political tolerance; it reshapes local economies and electricity bills; it raises questions about emissions and affordability and national security and who gets priority when the grid is tight; and it invites the state back into the room, because infrastructure at this scale rarely stays a purely private affair for long.

The Factory Goes Off-Planet

The far edge of all this is, quite literally, off the planet. The idea of putting data centres in orbit, solar-powered, cooled by the cold of space, free of the day-night cycle and the planning objections, sounds like science fiction until you notice how many serious people are working on it. Musk has made the economic case directly, at Davos and elsewhere, and SpaceX has filed with the American regulator for a satellite "space cloud" to support AI compute.

Nvidia used its 2026 developer conference to launch hardware built for orbital data centres; Google is studying the same idea under the name Project Suncatcher; Blue Origin is reported to be working on space data-centre technology of its own; and a Nvidia-backed startup, Starcloud, has already run a model on a GPU in orbit and reached a billion-dollar valuation on the promise of building far more. The obstacles are not trivial, launch cost, radiation, thermal management, servicing, debris and latency among them, and none of it may pencil out. But the fact that the question is being asked at all is the tell. When the physical limits of terrestrial compute begin to bite, the market starts, quite seriously, to price moving the factory off-world. That is less a vision of the future than capital pressure looking for a release valve.

Food For thought

For two centuries, progress meant making the physical world cheaper, cheaper energy, cheaper materials, cheaper distance, so that human effort could move on to harder things. Artificial intelligence, the most abstract product humanity has built, may quietly reverse that bargain, turning intelligence into the abundant input and the physical world, power, land, water, time, into the scarce one again.

The next phase will be measured not only in parameters and tokens, but in gigawatts, and financing capacity and the patience of whoever underwrites the build.

When intelligence becomes abundant, what becomes scarce?

Disclaimer

Constellation is published by Nevv Capital for general information and educational purposes only. It is not investment, legal, tax or financial advice, and it is not an offer or solicitation to buy or sell any security, fund interest or other instrument. References to specific companies, funds, managers or transactions are illustrative of the themes discussed and are not recommendations. Figures and third-party statements are drawn from publicly available sources believed to be reliable as at the date of writing; they have not been independently verified and may be incomplete, dated or subsequently revised. Forward-looking estimates are inherently uncertain and should not be relied upon. The views expressed are those of the author at the time of writing and may change without notice. Past performance is not a guide to future results, and the value of investments can fall as well as rise. Readers should seek their own professional advice before acting on anything contained herein.

© 2026 Nevv Capital. All rights reserved.