1. Transition

The overall theme of what we are going through now is the second great transition of the 21st century economy. Through 2008 we were at the tail end of high industrialism; from 2008 to 2022 we saw the rise of the bits-based economy; and since 2022 we have seen the beginnings of the token economy — where the token, the unit of model output priced per million, becomes the basic commodity of computing, as the bit was before it.

The easiest way to picture this is in the ten most valuable public companies. In 2008 there was one tech company (Microsoft) alongside a mix of oil and gas majors, state banks, telecoms, and industrials. By early 2022 the list was roughly 70% tech (depending on how you count Tesla). In 2026, nine of the ten most valuable companies in the world — Aramco is the only holdout — are not just tech companies but companies that design, manufacture, or are existentially dependent on accelerated compute, whether custom for themselves or as part of the supply chain for others.

Because many of those are household names (Google, Apple, Amazon) whose main business appears to be something else, the effect is less jarring than it should be. But the theme is one of transition: from an industrial economy in 2008, to a bits-based economy in 2022, to the beginnings of a token-based economy in 2026.

2. Value chain

It is now usual to talk about the AI value chain in terms of power; infrastructure (datacenters and the things that connect datacenters to power); semiconductors (particularly accelerated compute in datacenters, though as inference becomes more important, presumably at the edge too); the model layer; and the application layer.

Three key things to think about across that value chain at the moment:

(a) Transition

Placing yourself, and the assets you are interested in, on the transition spectrum is crucial at every layer of the AI value chain. To date there has been a significant mismatch between the speed at which supply of energy and compute can grow and the speed at which AI demand is rising. TAM is expanding as the models improve and we get better at finding things to do with them, and products can be delivered at software scale — but energy and compute are subject to often severely constrained physical supply chains.

This has produced the bottleneck dynamic that has been a popular trade for the last twelve months (though it is suffering as I write). For a long-term-oriented investor there are probably three types of trade in this market:

(i) Pure bottleneck rents. I have written before about the relay-race dynamics you see in constrained supply chains: capital floods to the binding constraint until it is resolved, then the baton passes to the next constraint. This can make for very attractive returns right up to the point the constraint is resolved, but it becomes a trap if you assume constraint-level pricing power or margins will persist, or if you get the timing wrong.

Given the background of the recent bottleneck blow-up, I should add that my own research indicates multi-year and multi-layer bottlenecks, especially in power (see Figure 1). At no point on the timeline is the whole stack cleared at once. The identity of the binding layer is predictable; the binding itself is permanent.

(ii) Bottleneck re-rates. These are businesses that enjoy particular pricing power while they are the bottleneck, but which after the transition become staples of the new economy and benefit from long-term re-rating or increased volume. The pattern match here is to GE in the electrification era. GE did not sail through the Great Depression — its revenues collapsed with everyone else’s — but it survived the transition, remained structurally necessary as electrification became a staple of the economy, and compounded for decades afterwards (albeit ultimately through diversification rather than a core electricity business). The instructive contrast is RCA, the bottleneck-rent stock of the same era, which took decades to reclaim its 1929 high. The lesson: re-rate candidates do not need to dodge the bust; they need to be structurally necessary on the other side of it.

There seem to be strong candidates for that kind of re-rating in power and in the semiconductor supply chain — particularly TSMC and, if they remain the sole provider at the leading edge, ASML.

(iii) Transition winners. These are currently pure growth plays that need to harness the excitement around AI today to fund the growth required to be winners on the other side. This most obviously applies to the model and application layers, but in energy, infrastructure, and semiconductors there are growth and venture opportunities — from modular nuclear to alternative inference chip design — that are able to raise at scale specifically (and perhaps exclusively) because of the capital flowing into the bottlenecks.

I find this transition spectrum useful both for thinking about individual opportunities and for thinking about how an overall portfolio is weighted.

(b) The application layer — the dog that hasn’t barked (yet?)

Currently something like 90% of AI revenue is concentrated in the model layer and below, and something like 70% of it is going to semiconductors, particularly NVIDIA. In mature SaaS, by contrast, something like 70% of revenue went to the application layer.

The VC and growth community is torn between two beliefs: either a similarly proportioned application layer will emerge for AI (which would be several orders of magnitude larger than the software layer), or it is all over for the application layer because the model layer companies will simply eat everything.

The coding experience lends significant support to the second belief. Coding has so far been the standout application for AI, at least in revenue terms, and it produced the standout application-layer winner in Cursor — acquired by SpaceX for $60bn, the largest venture-backed exit ever. But despite a great product and a head start, Cursor’s revenues at the time of acquisition were a fraction of OpenAI’s and Anthropic’s coding-related revenues, and much of the deal commentary framed Cursor as caught in a margin trap — paying rent to the model layer — that made it a seller rather than an independent survivor. So far, the standout application-layer use case has been won by the model layer. It is worth keeping a close eye on OpenAI and Anthropic in finance and law, both of which have produced early breakout application-layer companies.

(c) Emerging opportunities

A way of synthesizing the fear and greed around the application layer:

The foundation models have shown themselves to be very good at generating code. Almost all of what we currently think of as the application layer is code, so we should expect that the models will either win there outright, or render the production of code so cheap that there is not much left to win.

But the availability of abundant cheap software, and the capabilities associated with probabilistic compute, also enable genuinely new things. Obvious examples include doing interesting things with proprietary or otherwise differentiated data sources, robotics, and materials and biological sciences.

Perhaps the analogy is the internet’s effect on media. The free availability of content was disastrous for the traditional media industry — but it made possible a whole new set of media opportunities in social networks, user-generated and curated video at YouTube, and streaming. Notably, the durable winners were not the producers of content but the owners of distribution, curation, and network effects. The application-layer analogue: where the code itself is no longer the moat, the durable opportunities lie in what the code cannot commoditize — the data, the physical interface, the network.

3. Dispersion of outcomes

Any portfolio that touches AI at the moment — and given the scope of the AI value chain, that is almost any portfolio — has to contend with a core confounding factor:

We are currently short watts, wafers, and (as a result) frontier tokens.

As well as producing the bottleneck dynamic, this means we do not really know the true extent of demand for AI (even at today’s level of capability, let alone with continued scaling), or the true margins that AI providers can command. We can see the effect of this in the reported gross margins of the large players. Anthropic is reportedly commanding gross margins around 70% — at this stage of the company’s growth we would expect it to trade margin for yet more growth if it could. Similarly, Google could in theory price-bomb the whole market, and it would make commercial sense for them to do so, but they simply do not have the compute for that kind of strategy.

This is one of the things that makes assessing the impact of open source so difficult: in a regime of rationing, it is hard to tell whether open source is genuinely taking share or merely absorbing demand that closed source cannot serve anyway. And beyond that, it means there is a very wide dispersion of possible outcomes for every layer of the AI value chain.

To make that concrete: it is plausible that by the end of 2028 the leading frontier lab is generating a $1tn revenue run rate at software gross margins (and when I say plausible, that implies gross margins holding and growth slowing considerably from year-to-date 2026); it is also plausible that open source, intense inter-lab competition, and emergent application winners have rendered the labs commodity players doing lots of low-margin revenue selling tokens. The cascading effects differ accordingly: in the first outcome, the leading lab is the GE of the token economy — a bottleneck re-rate at the model layer, with the layers below as suppliers to a king; in the second, value drains out of the model layer in both directions — down to compute and power, which keep scarcity pricing, and up to whichever applications own the customer. The point is that we should all have a mental model, if not a spreadsheet, that weighs how plausible those outcomes are, and their cascading effects through the AI value chain, and whether our portfolios reflect that weighting. And we should all be watching the resolution of the watts, wafers, and tokens shortages for the early indicators of which outcome will in fact occur.

Which brings us back to where we started, and the two-part test this was all building towards — the components of analysis for AI opportunities, whether at the individual or portfolio level:

Begin by placing the opportunity on the transition spectrum, and then weight that against your view of the dispersion of outcomes for AI demand and margins as the supply shortages are resolved.

The distillation debate has produced two camps that agree on almost nothing except the stakes.

One holds that AI is a technology of near-existential importance, that America must lead, and that the frontier labs (together with Nvidia) are how it leads. Chinese labs that distill American models in breach of their terms of service and release the results as near-frontier open source are therefore a threat to be regulated away, up to and including the entities list. The other holds that AI is a technology of near-existential importance, that it must diffuse through the American economy as fast and as cheaply as possible, Chinese open source helps and the labs are not owed a business model. Pursued to its logical end, each camp’s preferred path arrives, by a different road, at the same closed frontier.

The lines against protectionist policies are well rehearsed. I want to focus here on the second camp, whose openness-focused policies risk being self-defeating. Faced with a monetisation window compressed toward zero, it will be most logical for the labs to stop releasing the frontier, using it internally to build the next model, and selling closed access to a handful of deep-pocketed counterparties. In a world of unrestricted distillation the best option is to close things down and the mechanisms of that are already in place.

Instead, I propose a statutory bargain borrowed from pharmaceutical exclusivity: protect frontier models from distillation for a few months after release. Once the window lapses, distillation is deemed lawful, and terms of service cannot say otherwise. Inside the window enforcement runs on provenance certification rather than unwinnable infringement suits. And the design deliberately gives Chinese labs a legal route into the American market, instead of a reason to strip-mine it from outside.

What the labs are actually for

The labs’ long-term goal, broadly stated, is to build an artificial superintelligence, and one as general as possible. The medium-term goal that appears necessary to get there is recursive self-improvement: the point at which models are good enough to contribute meaningfully to the generation of the next model, improving everything from energy efficiency to chips and algorithms.

Short-term revenue should be understood as instrumental to those goals rather than a goal in itself. It demonstrates the size of the prize to external funders, and it will increasingly let the labs fund their own development. But nothing in that logic requires general public releases.

To date, the labs have tolerated a bargain in which a frontier release earns a shortish window of strong monetisation followed by a steep drop-off as capability commoditises. Distillation-powered open source compresses that window. If it becomes too short, the rational response is not to stop training frontier models, it is to stop releasing them. Instead they will use the frontier internally, for their own recursive-self-improvement efforts, and sell closed access to a small number of trusted, deep-pocketed counterparties. It is entirely reasonable to think that an Nvidia or a Jane Street would pay billions for exclusive access to the leading model for a period of time, and that others would pay tens or hundreds of millions to be first in their industry.

The machinery already exists

The infrastructure for closed frontier deployment has already been built, tested, and normalized. Currently it just wears safety clothing.

Anthropic’s Project Glasswing already shows the shape: an unreleased frontier model, a vetted coalition that launched with around fifty partners and now numbers roughly two hundred organisations across more than fifteen countries, tiered access, delivery to approved organisations only. The stated rationale is misuse risk, and it is a serious one; the claim here is not that Glasswing is a covert auction. The claim is that the machinery now exists and has been accepted, whatever the motive. Repurposing this machinery for commercial reasons is a pricing decision.

Absent a response to distillation, that is the direction of travel: a frontier that is more closed, seen by fewer people, and where the leading open-source options come only from those willing to break terms of service.

A statutory window

The better solution is to take distillation out of the terms of service altogether and put it on a statutory footing. A regime borrowed from pharmaceutical exclusivity, forced-labour supply-chain enforcement, and SOC-style compliance.

The window. Leading-edge models should receive statutory protection against distillation for a fixed period from release. This should be longer than distillation turnaround and shorter than the next generation model release: probably around nine months, though the right figure is an empirical question this piece doesn’t settle.

After the window. Once protection lapses, distillation of the model should be deemed lawful and, critically, the statute must pre-empt contract. All companies and labs should have legal certainty that they can merrily distill.

Inside the window. Enforcement should never depend on proving infringement, which is close to impossible, it should depend on the absence of provenance. A government agency (BIS is the natural home) would maintain a list of models and labs presumed to be built on in-window distillation. The presumption is rebuttable, and certification is how it is rebutted. Certification attests to the model’s lineage (base models past the window, documented data manifests, auditable training logs) and should be scoped to frontier capabilities within the window.

Who gets sued. The cause of action should run against hosts and distributors (marketplaces, hubs, API resellers carrying uncertified in-window models) not end users, for the same reason copyright law learned to sue intermediaries rather than ten million downloaders. And it needs statutory damages per deployment, since actual distillation damages are unprovable and an unquantified suit is worthless. The federal government should then anchor demand by refusing to buy uncertified models, on the FedRAMP model. From there the market polices itself. Enterprise procurement will come to demand certification the way it came to demand SOC 2 reports, insurers will reprice uncertified deployment, and it will become very hard to sell uncertified models or products built upon them.

Getting the Chinese to play

I waved off the protectionist route earlier on. But it is worth acknowledging that a nationality ban and a statutory window are both restrictive, but not equally protectionist. And we should also acknowledge the success of restrictive US policy to date: there has been no frontier Chinese model; and the very fact of distillation is to some extent an acceptance that they will not be able to get there without leading edge semiconductors.

But a nationality ban gives every Chinese lab identical incentives: nothing to gain from compliance, nothing to lose from infringement, and thereby legislates them into behaving as exactly the coordinated bloc the protectionists fear. A statutory regime creates a reason for individual exit. Certification, not the window itself, is the lever: any single lab can certify and change its own legal status, independent of what its peers do.

While we must assume at least some state level co-ordination of Chinese open-source there is serious diversity among the players. Perhaps eight labs are at or near the frontier and they are making real, novel contributions, particularly on efficiency, because being GPU-poor forced them to. The talent pool capable of frontier and near-frontier work is very small, and a significant portion of it is in China. The techniques that pool has developed should be broadly shared, and the US should give it every incentive to contribute to the US system rather than to serve as the foundation of a parallel Chinese one.

Under this statutory regime defection is a real option with real payoffs: certify, gain lawful access to the US and likely broader Western market, and distill legally the moment the window opens. The labs most likely to take the exit are those whose parents have hostage assets: cloud businesses, Western revenue, listing exposure. The pure-play independents may never come. That is fine. The regime does not need universal compliance; it needs enough defection that the lagging edge stays contested and to make state coordination harder. Every legal distiller at the trailing edge also disciplines pricing at the frontier, which is good for everyone.

Conclusion

This system would not be perfect. It leaks at the edges (synthetic-data laundering is genuinely hard to audit) and would be gamed. But it does not need to be watertight; it needs to bind the enterprise revenue layer, because that is where the money is, and the defection incentive with it.

This is what gives an incentive for the labs to keep releasing broadly, and something closer to a level playing field for US open source than exists today. Both camps believe the fight is over who captures the value of AI. The fight that matters is whether the frontier stays visible at all.

Two readings of AI infrastructure dominate the current debate. Both rest on category errors. The first is the glut thesis: the $700 billion gap between projected hyperscaler capex and any defensible revenue path, and therefore a commodity-style clearing event within the next two or three years. The arithmetic is coherent, but the category is not. The buildout would resemble prior commodity cycles such as fibre, DRAM, and shale only if the two sides of the market respond at similar speeds. They do not: demand for AI infrastructure responds at software speed; supply responds at physical speed; and that asymmetry has no close analogue in the cycles the glut thesis pattern-matches to. The cost of being wrong on the glut thesis is stepping out of a general-purpose-technology deployment in its early innings.

The second reading is bottleneck commentary: the steady weekly diet of analyst notes on the next constraint to clear, with power this month, transformers next quarter, HBM by 2028, treating each binding layer as the trade. This is closer to the truth than the glut thesis, but it has its own characteristic error. It treats each bottleneck as a milestone to be cleared rather than as one phase in a moving system, and it tends to pull investors into the most crowded position at the moment the rent is most visible. The cost of being wrong here is getting caught long the wrong layer when the binding migrates, or, more dangerously, getting caught at the one layer where the supply response can actually arrive (as may well occur at the neocloud or other ‘financing moat’ layers).

The asymmetry between demand and supply response times produces neither a clearing market nor a sequence of milestones, but a relay race: a succession of binding constraints, each ceding to the next, with the locus of pricing power moving through the stack on a timescale measured in decades rather than years, as in electricity and oil and gas. The race is, within reasonable bounds, predictable, because the response functions of the physical inputs are knowable. It also produces three categories of winner whose advantages differ in durability, a distinction that matters most for long-duration capital. Some positions earn rent only while their specific constraint binds. Others carry an advantage that outlasts the binding period, re-rating during their turn and keeping the rating against a much larger market once the buildout matures.

1. The asymmetry

Demand for AI infrastructure responds at software speed while supply responds at physical speed. When inference prices fall, the market for applications using that inference expands within weeks, because the decision to deploy more is a procurement decision made by software teams inside the deployment windows of their engineering cycles. When more capacity is needed, the decision to build it runs into gas turbine backlogs, FERC timelines, electrical steel prequalification cycles, ASML throughput, and TSMC yield curves, none of which compress to the speed capital would like, regardless of how much capital is deployed.

The aggregate data is already consistent with this. The price of a GPT-4-class model has fallen by roughly a factor of 1,000 in two years (Demirer et al., 2025), and rather than producing saturation this has been accompanied by continued exponential growth in usage. Enterprise spend on generative AI has risen more than threefold over the same period, inference spend has overtaken training, and computing capacity devoted to AI has continued to double every 7 months. Electricity behaved this way across the twentieth century: unit costs fell by roughly two orders of magnitude between 1900 and 1960, and electricity did not find a new equilibrium at the lower price but became the substrate for industrial motors, urban rail, refrigeration, appliances, electrochemistry, and eventually the computing infrastructure on which this is being read. Intelligence is behaving the same way, and the category it belongs to is general purpose technology deployment, not commodity buildout.

Both errors named in the opening, the glut frame and the bottleneck frame, show up clearly in recent pricing cycles. H100 spot rates fell from roughly $8 per hour in early 2024 toward the $2–$3 range through 2025, and the glut commentators read this as the demand crack they had been forecasting. The bottleneck thinkers who had been chasing GPUs as the trade were simultaneously caught flat-footed: by the time the H100 spot dip was visible, the binding constraint had already migrated to power, and Microsoft acknowledged publicly that it now held GPUs it could not plug into the grid. The H100 trough was a product-cycle handoff inside a generational transition. By early 2026, H100 contract pricing had rebounded roughly 40% as the Hopper-to-Blackwell transition completed and the secondary market cleared. DRAM showed the same pattern on a different clock: the 2023 spot-price trough was read as structural oversupply at the time, and pricing had fully reset within 12 to 18 months as AI memory demand absorbed the slack. In both cases, the apparent glut was the asymmetry operating exactly as it should, and the bottleneck thinkers who had been long the layer in question were the ones who paid for the misreading.

What the commodity-buildout template misses is that under this response asymmetry, supply and demand do not clear in the usual sense. Each capacity addition activates a use case that the previous cost structure could not support, and the frontier of unactivated use cases recedes roughly as fast as new capacity arrives. The buildout continues until either the physical inputs commoditise at the rate demand expands, which in electricity took 80 years, or the GPT itself matures into a different consumption pattern. In the meantime, the system is permanently running against a constraint somewhere in the stack. Capital deployed at the constrained layer earns rent for as long as the constraint binds. What the bottleneck template misses is that the constraint moves, and the rents move with it. Understanding both the mechanism and the sequence is the entire analytical problem.

2. The relay race

Because the constraints are physical, they do not all bind at once. One constraint binds until the capacity response catches up, at which point demand, having grown into the earlier capacity, binds the next. The locus of scarcity is a moving target, and the question for any asset is where in the sequence it sits.

The historical case most often cited as a cautionary tale against this thesis is the internet fibre buildout of 1996 to 2002. The received reading is that telecoms overbuilt, fibre became worthless, and the lesson is that infrastructure cycles end in gluts. The data tells a different story. At the peak of the overbuild, only between 2% and 5% of laid fibre was carrying traffic, and long-haul transit prices collapsed from roughly $1,200 per megabit per second in 1998 to under $1 by 2015. The equity loss was genuine. But the dark fibre did not stay dark. By 2010 it was fully utilised. Demand grew into the capacity on a timescale of about a decade, and what looked like durable oversupply in 2002 was, in retrospect, early delivery of a capacity layer whose commoditisation was the prerequisite for the next layer to emerge.

The next layer was the content delivery network. Akamai’s revenue went from $4 million in 1999 to $163 million by 2003, to $636 million at the YouTube streaming inflection in 2007, and to $1.37 billion by 2010. Fibre had commoditised; CDN had not, because its binding constraint was operational knowledge about edge placement and caching rather than capital-responsive physical capacity. The handoff then occurred again, from CDN to colocation, colocation to cloud, and cloud to the current generation of AI compute. At no point in the 25 years that followed did a system-wide glut materialise. The fibre overbuild, read correctly, is not the cautionary tale against the infrastructure thesis. It is the clearest available precedent for it.

Electrification tells the same story over a longer arc and with higher amplitude. The buildout from 1880 to 1960 proceeded through overlapping constraint phases. Generation was binding from roughly 1880 to 1910, and the Insull-era utilities that built the first central stations earned extraordinary returns. By the 1920s generation had commoditised (the apparent glut), and the binding constraint had migrated to transmission and distribution. Copper consumption for T&D alone rose from 60 million pounds in 1920 to 260 million pounds in 1930, a 400% increase even as generation growth slowed. Firms indexed to the new binding layer compounded through the Depression. General Electric’s sales doubled between 1921 and 1929; its earnings-to-sales ratio held steady at 13–14% from 1920 all the way through 1940. Sangamo Electric, whose business was meter manufacturing for the expanding utility market, grew employment from 1,200 in 1929 to 3,000 by 1943, through the Depression. By the 1940s, the binding constraint had migrated to distribution equipment and end-use. By the 1960s, to load management. 80 years of compounding returns were distributed across a succession of layers, and the investors who fared best were those who understood that the binding constraint was a moving target rather than a permanent address.

The useful property of the current cycle for an investor is that the sequence is visible in advance. As of early 2026, the near-term binding constraints, in approximate order of severity and duration, are grid power and gas turbine capacity (5–10 year lead times, structurally the longest); large power transformers and grid interconnection (3–7 years, Wood Mackenzie’s 30% supply deficit holding through the decade); HBM memory and CoWoS advanced packaging (18–24 months, easing by 2028 as Micron’s Idaho fab comes online); indium phosphide substrates for optical networking (2–3 years, worsening as co-packaged optics compound demand); data centre construction labour (a 439,000-worker shortfall against a workforce with 41% retiring by 2031); EUV lithography and advanced-node logic (3–5 year fab cycles, stable but concentrated); and frontier model training talent (fewer than 500 senior practitioners globally). Each constraint has a knowable response function. Which layer binds at any given moment determines where pricing power sits.

Microsoft’s acknowledgement that its binding constraint has shifted from GPUs, which it now holds in inventory, to power, which it cannot plug them into fast enough, is the relay race in real time, disclosed by the most data-rich operator in the stack. The bottleneck thinker who was long GPUs in mid-2025 is the trade most exposed to handoffs of this kind. The relay-race thinker is positioned to anticipate them.

3. Three kinds of winners, two kinds of rent

If the relay race is the right description of the buildout, the next question is which positions earn rent only while their specific constraint binds, and which earn rent during their binding period and then keep the re-rating against a larger post-buildout market. The relay race produces three categories of winner. Two extract rents that are temporary, tied specifically to their constraint being the binding one at the moment. The third earns the same temporary rent while its constraint binds but carries a durable advantage that outlasts the binding period. This is the payoff pattern that matters most for long-duration capital, and the one the market mis-prices. One distinction is worth making first. A competitive advantage is a structural feature that cannot be replicated within the investment horizon. A rent is what that advantage extracts given the current demand environment. TSMC had its process-knowledge advantage in 2019, but the rent was smaller than it is now because HPC demand was not yet stressing the leading-edge node. The advantage persists across cycles; the rent is the advantage multiplied by the scarcity conditions of the moment.

The first category holds scarce physical or regulatory positions that cannot be replicated within the cycle. The natural gas pipeline network connecting producing basins to the load centres emerging around hyperscale campuses is Ricardian in this sense: the rights-of-way were established decades ago, FERC certification of a new interstate line takes years, and the long-term (10–20 year) supply and power agreements pipeline operators are signing with hyperscalers, including the Williams-Meta agreement and the Energy Transfer-Oracle deal, price at reservation rates that do not respond to incremental capex elsewhere in the system. Generation sites with existing grid interconnection are Ricardian because the queue to obtain new interconnection runs at medians above 4 years against a national backlog of 2,300 gigawatts. Cleveland-Cliffs’ Butler Works is Ricardian as the sole North American source of grain-oriented electrical steel, feeding a transformer supply chain in which demand is up 116% since 2019 and the physics floor on new GOES capacity is 5 to 7 years.

Ricardian advantages are defensible against technological change because the scarcity is physical or legal rather than technical. A better algorithm does not compress a FERC timeline or create new electrical steel mills. The rent, however, is conditional on the specific output remaining in demand and on the regulatory regime that creates the barrier holding in place. A shift in fuel mix, routing, or regulation can reset the rent without eroding the underlying advantage. The position earns extraordinary returns while its layer is binding and reverts toward competitive economics when adjacent capacity eventually arrives.

The second category holds a technology-frontier lead. Coherent’s architectural lead in 800-gigabit and 1.6-terabit optical transceivers, with book-to-bill above 4× and most of 2026 booked, reflects a technology lead the field has not yet closed. Vertiv’s liquid-cooling architecture and Eaton’s electrical balance-of-plant for high-density AI racks (dollar content per megawatt up from $1.2 million in the cloud era to $3.4 million in AI) earn from being first to the required technical specification. These are Schumpeterian rents. The rent flows while the lead holds, and the lead holds only as long as the incumbent keeps running. The field catches up because the constraint driving the rent is technical rather than physical; an architectural discontinuity can reset the curve altogether. These are real earnings, but an investor pricing them as durable is pricing them wrong.

The third category, the one most consequential for a long-duration book, holds accumulated operational know-how that competitors cannot buy in factor markets, cannot recruit out of any single labour pool, and cannot compress the clock on. The canonical case is TSMC. At the 3-nanometre node it achieves yields approaching 90%. Samsung, having deployed more than $200 billion in capital, still struggles to hit 50%. On a 300-millimetre wafer costing $15,000 to $20,000 at the leading edge, that yield gap is nearly a 2:1 effective cost per good die, and the gap has widened rather than narrowed across successive nodes. China’s two-decade, $150 billion semiconductor subsidy programme has left SMIC approximately two generations behind, with gross margins of 19% against TSMC’s 60%. Capital does not buy this. Morris Chang’s characterisation of TSMC Arizona as a “very expensive exercise in futility” was a statement about what operational capability requires, namely sustained co-presence and accumulated practice, neither of which capital can compress. The AI cycle is stressing TSMC’s advantage now; the advantage will persist after the cycle moves on, and TSMC will be making substantially more leading-edge logic at higher margins than it was pre-cycle, because leading-edge logic serves mobile, automotive, and industrial markets whose demand will have stepped up in the meantime.

ASML sits in the same category through a different mechanism: EUV systems depend on Zeiss optics produced under a craft tradition running back to 1846, and every leading-edge fab in the world depends on that single supply chain. The Western gas turbine majors, GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries, are another instance. Their current backlog through 2028 is rent. But a heavy-duty gas turbine requires 50 years of metallurgical and combustion engineering to build reliably at scale, and that knowledge base does not transfer to a new entrant on any timescale capital cares about. Chinese competitors have been making serious turbines for two decades and still cannot compete at the heavy-duty frontier. When the AI-driven power cycle ends, the Western three will still be producing substantially more turbines than they did pre-cycle, because electrification of heating and transport, reshoring, and coal retirement will have become live demand drivers, and the market will still be paying these firms for the output.

NVIDIA looks like a classic Schumpeterian play, but the sheer complexity of what they ship (silicon, NVLink, rack-level thermal and power design, the CUDA-X library stack, co-design with hyperscaler customers) compounds into something closer to operational capability than to a technology lead. Trainium and TPUs have not materially bitten into NVIDIA’s margins despite years of hyperscaler investment, which is evidence for this read. They are more brittle than TSMC or ASML, because the paradigm they serve is younger and could shift, but they are not to be dismissed as a cyclical Schumpeterian position either.

4. What would change the view

Three categories of risk are worth naming. None delivers the commodity-style clearing event the glut thesis requires, but each changes parts of the allocation map in specific ways.

Technical change sits on two margins. The familiar one is a further architectural breakthrough in model efficiency, with DeepSeek V3’s $5.6 million pre-training run (the reported GPU rental cost, not the all-in research and capital figure) as the existence proof, which on a Jevons reading accelerates the thesis rather than breaks it, by activating use cases the current cost structure cannot support. Every prior compute-era efficiency jump has produced an order-of-magnitude expansion in total compute demand rather than a contraction. The less familiar risk is a substrate shift that resets operational-capability advantages specifically: neuromorphic, optical compute at scale, or a model architecture that makes current kernel libraries obsolete. This is the paradigm risk for the NVIDIA call, and to a lesser extent for the advanced-packaging layer. TSMC, ASML, and the turbine majors are less exposed because their capability base transfers across substrate generations. If an investor wants to hedge the NVIDIA call without stepping away from the broader thesis, the hedge is an overweight in the positions that survive the paradigm shift.

Geopolitics operates differently than it appears. The first-order effect, supply chain fragmentation, export controls on Chinese advanced computing, and tariffs on specialty materials, typically entrenches the operational-capability positions rather than eroding them, because substitution attempts fail on exactly the tacit-knowledge dimensions the taxonomy identifies. TSMC’s position has strengthened through the export control regime, not weakened. The Western gas turbine majors have benefited from restrictions on Chinese competitors’ access to Western markets. The second-order effect, a durable Taiwan contingency or a rupture in the Netherlands-Taiwan-US litho axis, is a genuine tail risk that would reset the taxonomy at the leading edge. This is a supply collapse risk, a different shape of loss from a glut, and one an allocator should hedge through geographic diversification within the operational-capability layer rather than by under-weighting the layer itself.

Macro compresses timing rather than changing structure. A recession delays the demand curve without bending it; the AI capex cycle is currently funded largely out of hyperscaler free cash flow, which is itself durable. The relay race runs slower in a downturn. The positions identified still pay, and the operational-capability compounders in particular tend to use downturns to widen their competitive position while capital-constrained competitors fall behind. Macro is a timing risk, not a thesis risk.

The scenario worth ending on, because it is the most plausible one that genuinely changes the analysis and the one that intensifies rather than weakens the thesis, is a robotics breakthrough. The current application of AI is, by historical GPT standards, narrow, and largely confined to productivity software. Embodied AI at commercial scale, general-purpose robotic labour, autonomous industrial systems at volume would extend the demand vector into a second physical stack that the software-era system does not use: battery cells, rare earth magnets, precision actuators, sensors, industrial metals, high-volume precision manufacturing capacity. Those inputs sit on a different supply geography (the lithium triangle, Chinese rare earth processing, Korean and Japanese battery chemistry, German and Japanese and Chinese precision manufacturing) with different constraint types and different geopolitical overlays. The software-era bottlenecks persist and a second set opens alongside them. The taxonomy does not break. It extends. Ricardian advantages emerge in new materials geographies. Schumpeterian advantages emerge at the battery, actuator, and sensor frontiers. Operational-capability advantages emerge in high-volume precision manufacturing, the kind Denso, Bosch, and the tier-one Chinese industrial integrators have spent decades accumulating. The stress test for this piece is a demand expansion into categories the current stack is not built for.

Demand responds at software speed; supply responds at physical speed; the gap between those response times is where institutional capital should sit. The layers where capital can respond freely are the glut layers, and the bottleneck-thinker’s most exposed trade. The layers where it cannot are the layers where durable returns accrue. The task is to be positioned at the second.

The AI buildout to date has been driven by training compute, large, episodic, datacenter-concentrated workloads. The next phase is different: agentic AI creates continuous, latency-sensitive, geographically distributed inference demand that runs around the clock. That shift has direct physical consequences that follow a clear thread from where AI demand is going to where it lands in the physical stack. This note follows that thread, from the nature of agentic workloads through to the classes of infrastructure that will be under sustained utilisation pressure for the next decade, and stress-tests that picture against the most plausible technology disruption scenarios.

1. The Demand Shift: From Training to Agentic Inference

To understand what is coming for physical infrastructure, it helps to be precise about the three distinct types of AI workload that have existed in rough chronological order, and why the third is categorically different from the first two.

Training: where the buildout began

The AI infrastructure boom that began in earnest in 2022–2023 was driven almost entirely by training compute. Training a large language model is an episodic, intensive event: a cluster of thousands of GPUs runs at near-100% utilisation for weeks or months, then stops. The cluster might then sit relatively idle while teams evaluate the model before the next run begins. The infrastructure profile looks like a series of peaks: very high demand, then relative quiet.

This episodic character shaped how the entire supply chain was built. Datacenters were sized for peak training loads. Power procurement was structured around the assumption that demand would cycle. Gas turbines were dispatched as peaking or intermediate resources rather than baseload. Grid interconnections were sized on the assumption that load would vary. The whole stack was optimised for a demand signal that looked like bursts separated by troughs.

Standard inference: the first evolution

As trained models were deployed publicly (ChatGPT being the most visible example), inference workloads emerged alongside training. Inference is the process of running a model to generate a response. It scales with users rather than with model development cycles: a hundred million daily users generates enormous aggregate compute demand, even though each individual query is relatively light.

Standard inference has a recognisable rhythm tied to human activity. Mornings and evenings in major time zones see higher traffic; overnight hours see lower traffic. The infrastructure profile is smoother than training but still cycles with the working day. The underlying assumption baked into the physical stack, that demand varies, that there are quiet periods, that infrastructure can be sized for average plus a reasonable peak, remains broadly intact.

Agentic AI: why this phase is different

Agentic AI describes systems where models do not simply respond to a single query but take sequences of actions, use tools, maintain context across multiple steps, and operate autonomously toward a defined goal. An AI agent tasked with researching a topic, drafting a document, and scheduling a follow-up is not answering one question; it is executing a persistent, multi-step workflow that may run for minutes, hours, or continuously in the background.

Several properties of agentic workloads combine to make the infrastructure implications qualitatively different from what came before.

Agentic workloads are always-on. An enterprise deploying agents to handle customer interactions, monitor systems, or execute ongoing workflows is not generating episodic compute demand. It is generating baseload demand around the clock. The concept of off-peak hours largely ceases to apply.

They carry persistent state. An agent mid-task maintains a record of everything it has done and learned, stored as its context window. Architectural approaches can externalise some of this into databases and task queues, allowing limited resumption after interruption, and these techniques are being actively developed. But even with external memory, agentic workloads carry fundamentally higher curtailment costs than training runs: partially completed multi-step tasks must be re-executed, real-time user-facing workflows cannot be paused without consequence, and FERC’s curtailability classification for grid interconnection purposes is binary. The practical effect is that agentic datacenters are poor candidates for curtailable-load treatment, which matters directly for how they interact with the grid interconnection queue, as discussed below.

They are far more energy-intensive per unit of output. Reasoning models, the class used in agentic systems, consume roughly 30 times more power per query than standard inference alternatives. A full agentic session consumes in the range of 10 to 15 kilowatt-hours; a standard chatbot query consumes approximately 0.3 watt-hours. That difference, roughly ten thousand times, reflects the additional computation involved in multi-step reasoning, tool use, and context maintenance.

They generate fundamentally different network traffic. Training generates intense east-west traffic within a campus; GPU clusters communicate internally. Agentic workloads generate persistent north-south connectivity: continuous links between datacenters, between datacenters and users, between agents co-ordinating across different sites. The geographic footprint of network demand changes entirely.

The shift from episodic training to always-on agentic inference converts every infrastructure constraint from a peak-capacity problem into a continuous-baseload problem.

The load factor argument

The most direct way to quantify the physical consequences is through load factor, the percentage of a facility’s capacity being used on an ongoing basis. A datacenter campus primarily running training workloads might operate at around 65% average load factor: high during runs, lower between them. A campus dominated by agentic workloads is likely to operate above 85%, because the demand never meaningfully stops.

A 500 megawatt campus at 65% average load factor consumes around 19.9 billion cubic feet of natural gas per year to power it. The same campus at 85% consumes around 26.1 billion cubic feet, a 31% increase in annual gas demand from the load factor shift alone, with no change in the size of the facility. Across the more than 100 gigawatts of on-site gas-fired power currently being planned by US datacenter operators, the aggregate numbers are very large.

But the load factor argument matters beyond gas. Every layer of the physical stack, including pipelines, generators, grid connections, network capacity and cooling, was sized, financed, and economically underwritten around assumptions that embed peaks and troughs. Agentic AI, at scale, replaces that cycling demand profile with something that looks much more like heavy industrial baseload. The assets positioned to serve continuous demand are not always the same ones that served the episodic training era well.

For context: hyperscale cloud datacenters, the generation before AI training dominated, typically operated at 50 to 65% average load factor. AI training pushed toward the upper end of that range. The shift to 85%+ for agentic workloads is the second step-change in a decade, and unlike the cloud transition, which expanded into abundant spare capacity, it arrives against a physical infrastructure stack that was built for cycling demand and has not been reconfigured for continuous baseload.

The direction of travel

It is worth being explicit about what is established and what remains uncertain. The direction of travel, from training-dominated demand toward always-on agentic workloads, is not speculative. The largest AI labs are all building and deploying agentic systems. Enterprise adoption is early but accelerating. The compute procurement decisions being made today by hyperscalers are being made on the explicit assumption that agentic demand is coming and that it will be substantial.

What is genuinely uncertain is the pace. Agentic AI at hyperscaler scale could be the dominant workload type by 2027 or by 2030. The difference matters for which constraints bind first and how severely. But the pace uncertainty does not change the destination; the question is less whether it arrives than whether it arrives in three years or in five. The final section of this note addresses those scenarios directly.

2. What That Does to the Physical Stack

Following the thread from agentic demand into physical infrastructure requires going layer by layer. The five constraint layers that matter, namely gas-fired power, gas transmission, grid infrastructure, datacenter physical and networking, each respond to the demand shift in a specific way. In every case, the direction is toward higher and more continuous utilisation.

Gas-fired power generation

Gas turbines are the primary power source for the AI buildout. Renewable energy plays a growing role in datacenter power procurement, and the largest hyperscalers have meaningful commitments to carbon-free energy. But the firm, dispatchable power that keeps GPU clusters running regardless of weather or time of day comes overwhelmingly from gas. For large H-class combined-cycle turbines, the most efficient units available, the physics-constrained supply chain already has lead times extending to 2028–2030 from the three dominant OEMs: GE Vernova, Siemens Energy, and Mitsubishi Power.

The demand shift to agentic workloads does not add pressure to turbine procurement in the first instance; it changes the character of what those turbines are asked to do. GE Vernova’s CEO confirmed this directly at the end of 2025: hyperscaler customers are now underwriting 20-year business cases where turbines run continuously at baseload, not as peakers they can dispatch when needed. Hyperscalers have moved from roughly 10% to approximately one-third of GE Vernova’s new orders, and they are buying the largest H-class units specifically for continuous operation.

For the physical infrastructure of power generation, this means the maintenance and availability profile changes. Baseload operation accumulates fired hours faster than peaking operation, moving forward combustion inspection intervals. But it also means fewer cold starts, each of which adds the equivalent of 150 to 200 operating hours of thermal fatigue to critical components. The economic character of a turbine running at continuous baseload is more favourable than one cycling repeatedly; the maintenance planning requirements are simply different from what gas-fired operators accustomed to cycling dispatch have historically needed to manage.

Natural gas transmission pipelines

In several key geographies, pipeline capacity will bind before turbine delivery. This is the constraint that the public discussion has most consistently underweighted.

In the Southeast corridor, where Georgia Power has approved nearly 10 gigawatts of gas-fired expansion, the Transco mainline currently has only 200 to 300 million cubic feet per day of spare southbound capacity into North Carolina. The approved expansion requires gas delivery at roughly ten times that rate. The price signal of this tightness is unmistakable: the Transco Zone 4 basis differential, which reflects what Southeast buyers pay above the Henry Hub benchmark, blew out from $0.28 per million BTU in 2025 to $2.70 in early 2026. The infrastructure is already strained before the bulk of the AI-driven demand has arrived.

The critical path project is the Transco Southeast Supply Enhancement: 1.6 billion cubic feet per day of new southbound capacity, approved by FERC in January 2026, with construction beginning in Q4 2026 and targeted in-service in Q4 2027. In ERCOT Texas, the binding constraint is upstream of transmission: intra-basin Permian gathering capacity, where planned additions of 1.5 billion cubic feet per day are running against 5 billion cubic feet of processing growth. Interstate relief arrives with Energy Transfer’s Hugh Brinson pipeline in late 2026; the gathering constraint is less tractable.

There is a second-order constraint here that has received almost no public analysis. Compressor stations, the facilities required at regular intervals along any pipeline to maintain gas pressure, are powered by industrial gas turbines. The same OEMs with backlogs extending to 2028–2030 for power generation turbines supply the turbines that make new pipeline capacity operational. A pipeline that has received regulatory approval and completed civil construction cannot serve its full rated capacity without its compressor stations, and those compressor stations require the same constrained equipment as the power plants the pipeline is designed to serve.

For pipeline operators with existing contracted capacity in the right corridors, the baseload transition is unambiguously positive. Higher and more continuous utilisation of contracted infrastructure means more throughput revenue against the same fixed asset base, with long-term take-or-pay contracts replacing the variable utilisation of the cycling era. Williams’ 20-year Meta contract and Energy Transfer’s 10-year Oracle arrangement are early examples of what the contracted structure looks like when the customer is a baseload AI operator rather than an interruptible industrial load.

Grid infrastructure: transformers and interconnection

The grid layer presents two distinct constraints operating on different timelines, both of which are made more severe by the demand shift to agentic workloads.

Large power transformers, the equipment that steps voltage up and down at substations and connects generators to the grid, have lead times of up to 143 weeks for generator step-up units and 128 weeks for power transformers. This is the commonly cited constraint. What is less commonly understood is that the delivery lead time for a specific unit is entirely separate from the lead time to build new manufacturing capacity. New large power transformer manufacturing capacity requires five to seven years from investment decision to first commercial shipment, gated by four sequential physical constraints: grain-oriented electrical steel supply (where a single facility in Butler, Pennsylvania produces the only North American output), specialised winding equipment, high-voltage testing infrastructure, and workforce training for craft skills that cannot be compressed.

The United States is 80% import-dependent on power transformers, and demand for generator step-up units has increased by 274% since 2019. Two datacenters in Silicon Valley have been physically built but cannot operate because the large power transformers needed to connect them to the grid have not arrived. This is not a theoretical constraint; it is already determining which facilities can and cannot turn on.

The second grid constraint, the FERC interconnection queue, has an unusual interaction with agentic AI specifically. The queue currently holds over 2,600 gigawatts of pending projects, with a median wait time exceeding five years. The only regulatory mechanism designed to accelerate this process, FERC’s proposal that curtailable loads receive expedited treatment, is structurally unavailable to agentic AI datacenters. A training campus can be designed to pause runs when grid conditions require it, qualifying as a curtailable load. An agent mid-task cannot pause without losing its entire session state. The workload that most needs fast grid access is the workload that cannot, by construction, qualify for the fast lane.

Datacenter physical infrastructure

At the level of the datacenter itself, the demand shift to agentic and reasoning workloads changes two things: the thermal environment inside the building, and the labour required to build and commission the buildings.

NVIDIA’s Blackwell architecture, the current generation of GPU hardware, requires 142 kilowatts of cooling per rack. This is beyond the thermal envelope of air cooling. The next generation is expected to reach 240 kilowatts per rack, with NVIDIA’s published roadmap targeting one megawatt per rack. These are not future projections to plan around cautiously; they are the hardware that hyperscalers are deploying now, and they require liquid cooling infrastructure rather than the raised-floor air conditioning that characterised the previous datacenter generation.

Liquid cooling has a counterintuitive consequence for water consumption. Closed-loop direct-to-chip systems circulate coolant in sealed piping and release heat through air-cooled radiators with negligible ongoing water use, compared to evaporative cooling towers that can consume up to 1.8 billion gallons per year for a large AI datacenter. The workloads that drive the highest power density are also the workloads that make liquid cooling mandatory, and liquid cooling largely eliminates the ongoing water consumption risk that has been cited as a constraint on datacenter growth in water-stressed regions.

Construction labour is a different story. A projected shortfall of roughly 500,000 workers in the US construction sector by 2026, with 41% of the existing workforce approaching retirement age and construction unemployment at a record low, represents a constraint that capital cannot simply buy away. A 30% wage premium over standard construction rates has not cleared the market. The consequence is that fully-constructed, permit-approved, equipment-delivered datacenter facilities are sitting idle waiting for the specialised electricians and commissioning engineers needed to make them operational, at a cost, for a typical facility, of roughly $14 million per month of delay.

Fiber and optical networking

The networking constraint is frequently conflated with the chip supply chain, but it operates on an independent logic. The bottleneck in 800-gigabit optical transceivers is indium phosphide laser wafer capacity, a compound semiconductor that cannot be substituted with conventional silicon processes, grown on small wafers at a handful of specialist fabs globally. This is not a TSMC problem; it is an InP problem. Even if chip supply constraints were entirely resolved, the networking layer would remain independently bottlenecked.

The scale of that bottleneck is stark. Global demand for indium phosphide runs at roughly two million pieces per year; current production capacity is around 600,000 pieces. Coherent, one of the two dominant transceiver manufacturers, has a book-to-bill ratio above four times, with most of calendar 2026 already booked and 2027 filling rapidly. The CEO has stated publicly that supply-demand balance will not return in 2026 or 2027. Indium prices have roughly doubled since January 2023, and China’s control of approximately 70% of refined indium, subject to export controls introduced in early 2025, introduces geopolitical supply risk that cannot be engineered around.

The demand shift from training to agentic inference changes the character of networking demand in a way that intensifies this constraint. Training generates east-west traffic that stays within a GPU cluster. Inference and agentic workloads require persistent north-south connections: between datacenters, between datacenters and users, and between agents co-ordinating across sites in real time. The shift means more transceivers are needed per unit of compute deployed, and the traffic those transceivers carry runs continuously rather than in the bursts that characterised training. The fiber route miles needed to carry that traffic are estimated to need to double, from 95,000 to 187,000 US route miles, by 2029.

3. The Asset Class Map

Following the utilisation thread from agentic demand to physical infrastructure produces a reasonably specific map of which classes of asset will be under the most sustained pressure over the next decade.

Four asset classes stand out. A fifth is noted separately as significant but with important qualifications.

1. Gas transmission pipelines

In the Southeast and ERCOT Texas corridors, pipeline capacity binds before turbine delivery in the AI power build-out. The baseload demand profile of agentic AI converts pipeline throughput from variable to near-continuous, improving the economics of contracted capacity. The take-or-pay contract structures that pipeline operators are establishing with AI datacenter developers, with terms of 10 to 20 years and minimum volume commitments, are suited to infrastructure assets with long-lived physical assets and regulated or contracted returns. The constraint is most acute in existing mainline systems with established right-of-way in corridors where new pipeline development faces multi-year permitting timelines: the Transco mainline serving the Southeast and Mid-Atlantic, and the interstate pipeline network serving ERCOT from the Permian. The utilisation thesis does not depend on AI demand growing faster than expected; it holds even under conservative demand scenarios, because the pipeline capacity needed to serve already-approved utility gas expansion programmes is already insufficient.

2. High-voltage grid hardware and its upstream supply chain

Large power transformers and generator step-up units are the equipment through which almost all grid-connected electricity flows, and their supply chain is the most severely constrained single bottleneck in the grid layer. The five-to-seven-year floor on new manufacturing capacity, gated by grain-oriented electrical steel, specialised winding equipment, high-voltage testing infrastructure, and craft workforce development, means the constraint does not respond to capital in any conventional timeframe. The most durable chokepoint within this asset class is the upstream supply of grain-oriented electrical steel, where a single facility in Butler, Pennsylvania produces the only North American output. US energy efficiency standards requiring higher-grade electrical steel across the installed transformer base by April 2029 create a replacement cycle demand on top of the AI-driven new-build demand. The asset class includes the transformer manufacturing facilities themselves, the electrical steel mills that supply them, and the substation and interconnection infrastructure where transformers sit.

3. Optical networking infrastructure: dark fiber routes and InP supply chain

Two distinct assets within the networking layer face sustained utilisation pressure. The first is dark fiber, specifically fiber route miles along the corridors connecting major hyperscaler datacenter clusters in secondary markets: West Texas, the Phoenix metro, the Carolinas, and the Northern Virginia-to-Ashburn corridor. The always-on, north-south traffic profile of agentic workloads means fiber pairs serving inference and agentic traffic run at near-continuous utilisation rather than the episodic peaks of training; anchor-tenant contracts with hyperscalers are the relevant investment structure. The second asset is further upstream: indium phosphide substrate production capacity and the epitaxial reactor supply chain (where a single equipment manufacturer holds roughly 90% market share). Every optical transceiver pathway, whether conventional pluggable modules, co-packaged optics, or next-generation photonic interconnects, requires InP-based laser sources. This constraint intensifies with each bandwidth generation, as the InP content per module increases while module volumes grow substantially.

4. Dispatchable baseload power generation infrastructure

The physical infrastructure of gas-fired power generation, specifically combined-cycle plants sized and contracted for baseload rather than peaking operation, faces a decade of high utilisation as the primary firm power source for AI datacenters. The relevant assets are not the turbines themselves, which are the subject of the equity discussion around OEMs, but the physical generation infrastructure: the plant sites with grid connections, fuel supply agreements, and operating permits already in place. These assets benefit from the baseload conversion because their economics improve with higher utilisation: fewer cold starts, more revenue hours against the same capital base, and a demand profile that supports long-term power purchase agreements. Behind-the-meter generation, turbines sited at or adjacent to datacenters that bypass the FERC interconnection queue, represents a growing share of new capacity additions and creates direct contractual links between the generation infrastructure and the datacenter load it serves.

5. Liquid cooling infrastructure (noted with qualifications)

The shift to high-density agentic workloads makes liquid cooling not merely preferable but physically necessary for new AI-native datacenter builds. The relevant infrastructure includes cooling distribution unit networks, heat exchange equipment, and the piping and secondary cooling loops that route coolant to and from server racks. The qualification is that this is a market where supply constraints are actively being addressed by capital investment; the major vendors are expanding capacity 6 to 12 months ahead of demand curves, and no single physics-constrained upstream input creates the kind of durable chokepoint that characterises the other four asset classes. The utilisation thesis is sound; the scarcity premium is less durable.

4. The Technology ‘To Be Sure’

Any honest analysis of decade-horizon infrastructure utilisation must reckon with the technology scenarios that could change the picture. There are several plausible developments in AI architecture, power generation, and cooling technology that are actively being worked on. The question is less whether any of these succeed (some will, some will not) and more which asset classes remain highly utilised when they do.

The answer that emerges from a systematic review is more reassuring than the technology change narrative might suggest. Most of the technologies in active development do not eliminate constraints; they shift them. The asset classes that look most durable are the ones that sit at the point where constraints converge regardless of which technical pathway dominates.

Model efficiency: the Jevons effect

The most frequently cited reason for caution on AI infrastructure demand is the efficiency trajectory of AI models. DeepSeek’s demonstration in early 2025 that competitive performance could be achieved at a fraction of the training cost of frontier models drove a sharp short-term reassessment of datacenter capex. The efficiency improvement thesis holds that each successive generation of models will do more with less, eventually reducing per-query energy consumption enough to limit infrastructure demand growth.

The historical pattern in technology markets suggests caution here. Jevons’ Paradox, the observation that efficiency improvements in the use of a resource tend to increase total consumption of that resource by expanding the scope of what is economically worth doing, has held across every prior computing era. The PC became more energy-efficient per unit of compute, and total PC electricity consumption rose for decades. Mobile became vastly more efficient than desktop computing, and total mobile energy consumption dwarfed what it replaced. The relevant question is not whether per-query AI energy consumption will fall (it will), but whether total AI compute demand grows faster than efficiency gains. The evidence from token consumption data (growing at roughly 15% per quarter while AI revenue grows at 15% per year) suggests it is.

The efficiency scenario is also less disruptive to the specific asset classes identified above than it might appear. Gas transmission pipelines, grid hardware, and dark fiber routes are throughput assets: their utilisation depends on total AI activity, not on energy per query. A world where models become ten times more efficient but are used one hundred times more still produces higher pipeline throughput, more transformer stress, and more fiber traffic than today.

Alternative compute architectures

Several alternative compute architectures, including wafer-scale integration (Cerebras), optical computing (Lightmatter’s Envise platform), and processing-in-memory (Samsung and SK Hynix’s PIM designs), are frequently presented as potential disruptors to the NVIDIA-dominated GPU infrastructure model.

The infrastructure implications of these alternatives are more limited than commonly assumed. Cerebras’ wafer-scale approach eliminates the advanced packaging bottleneck (CoWoS) but transfers demand to TSMC wafer area: each unit consumes an entire 300mm wafer. Power and cooling requirements are unchanged; at 23 kilowatts per system, Cerebras hardware is not a power reduction story. Optical computing addresses only the multiply-accumulate operations in inference, not the memory operations that dominate power consumption in agentic workloads; the power reduction at the datacenter level is 10 to 30%, not the orders-of-magnitude improvement suggested by the physics of optical computation alone. Processing-in-memory improves throughput efficiency but does not reduce the memory footprint set by model parameter counts.

More importantly for the infrastructure utilisation thesis: none of these architectures reduces the demand for gas-fired power, pipeline capacity, transformer supply, or fiber. They are, in the main, architectural changes within the datacenter that leave the external infrastructure stack unchanged.

Behind-the-meter generation and its trade-off

Behind-the-meter gas generation, turbines sited at or adjacent to datacenter campuses that bypass the FERC interconnection queue entirely, is the fastest-scaling near-term solution to the grid interconnection constraint. Roughly 46 gigawatts of behind-the-meter turbine capacity is in the development pipeline. Microsoft, Amazon, and others have made explicit commitments to this approach.

The infrastructure implication is important: BTM generation does not reduce gas pipeline demand. A datacenter powering itself with on-site turbines still requires a gas supply pipeline to those turbines. In some cases, for example remote sites in West Texas, the pipeline infrastructure needed to fuel BTM generation at scale does not yet exist. BTM addresses the grid interconnection constraint by removing the facility from the interconnection queue; it redirects, rather than eliminates, the pressure on the gas transmission infrastructure identified above.

BTM also concentrates procurement pressure on turbine OEMs, whose order books are already extending to 2028–2030. The approximately 46 gigawatt BTM pipeline competes for the same manufacturing slots as utility-scale gas-fired generation. The infrastructure bottleneck shifts from the FERC queue to the turbine factory; the turbine OEM constraint is, if anything, harder to resolve quickly, given the physics of blade foundry qualification and rotor forging press installation.

Small modular reactors

Small modular reactors are the technology that generates the most interest in discussions of long-term AI power supply. The logic is appealing: firm, carbon-free, always-on power with no fuel supply chain vulnerability. Google, Microsoft, and Amazon have all made SMR commitments.

The timeline is unambiguous: no purpose-built SMR delivers commercial power before 2031, and that is the optimistic scenario. The pathway from application to commercial operation requires NRC construction permit review (17 to 18 months based on recent experience), first-of-a-kind construction (four to six years), and a separate operating licence. X-energy’s most advanced project targets the early 2030s. For the asset utilisation thesis covering the next decade, through approximately 2035, SMRs provide zero relief for the constraints already identified. They are a genuine long-term hedge on gas turbine fuel cost exposure; they are not a near-term competitor to the infrastructure described in this note.

What survives all the scenarios

Running each of the plausible technology scenarios against the four asset classes in the preceding section produces a clear pattern.

Gas transmission pipelines survive the efficiency scenario (total demand grows faster than per-query efficiency), the BTM scenario (BTM requires gas delivery to the point of generation), the SMR scenario (SMRs arrive after the decade under analysis), and the alternative compute architecture scenario (none of which changes gas consumption). The only scenario that significantly changes the picture is a demand collapse scenario, a fundamental failure of agentic AI to reach enterprise scale; that is a demand view, not a technology view.

High-voltage grid hardware faces rising demand in all scenarios except a complete shift to BTM generation. Even partial BTM adoption leaves the majority of new datacenter capacity grid-connected. The efficiency standard replacement cycle for transformers, driven by DOE requirements independent of AI demand, creates a baseline demand for GOES steel and transformer manufacturing regardless of the AI trajectory. The five-to-seven-year floor on new manufacturing capacity means the constraint is durable even if demand growth moderates.

Dark fiber and InP supply chain assets are the most technology-sensitive of the four. Co-packaged optics, where optical components are integrated directly into switching hardware rather than using pluggable transceivers, is a genuine architectural evolution that begins to ship in networking equipment in 2026 and 2027. The important detail is what co-packaged optics does to InP demand: it does not reduce it. Silicon cannot generate light; every optical platform requires InP-based laser sources. CPO eliminates the pluggable transceiver module but intensifies the InP dependency at the substrate level, shifting the constraint from transceiver assembly to laser wafer supply. The bottleneck changes form; it does not disappear. Fiber route miles are needed under every scenario in which AI inference reaches users via terrestrial networks.

Baseload gas generation infrastructure is the asset class most exposed to the long-run SMR scenario: a world after 2031 where SMRs offer firm, carbon-free power at competitive levelised cost would represent genuine competition for new gas-fired generation. For existing gas-fired assets contracted at baseload rates through the 2030s, the exposure depends on contract terms and the pace of SMR commercialisation. For the decade under immediate analysis the exposure is limited.

Technology risk in AI infrastructure is not primarily about whether individual technologies succeed. It is about the persistent upstream and regulatory constraints that determine deployment pacing regardless of which technical pathway wins.

5. The Durable Utilisation Case

Following the thread from agentic AI demand to physical infrastructure, and stress-testing the picture against the most plausible technology scenarios, produces a reasonably crisp conclusion.

Three things are happening simultaneously. AI workloads are shifting from episodic training to always-on agentic inference. That shift converts the demand signal reaching physical infrastructure from cycling to continuous baseload. And the physical infrastructure stack, built for a world of peaks and troughs, was not dimensioned for this conversion.

The asset classes that sit at the points of highest and most durable utilisation pressure are the ones where the constraint is not primarily one of capital or technology, but of regulatory process, physical materials, or physics-constrained supply chains that do not respond quickly to demand signals. These are the characteristics of genuine infrastructure assets: long-lived, irreplaceable within relevant timescales, earning returns from utilisation rather than from growth.

Gas transmission pipelines in the key AI buildout corridors benefit from being the first constraint to bind in the critical path, before turbines are delivered, before datacenters are operational, and from the contractual evolution toward long-term take-or-pay structures that the shift to baseload demand makes commercially rational for both sides. The physical right-of-way and regulatory certificate held by existing systems in these corridors cannot be replicated quickly.

High-voltage grid hardware, and specifically the upstream supply of grain-oriented electrical steel, benefits from a constraint floor that is measured in years, not months, and that is reinforced rather than relieved by technology change. Every grid-connected facility, regardless of its power source or compute architecture, passes its electricity through large power transformers. Demand for the equipment and its inputs is growing from multiple directions simultaneously: the AI buildout, electrification more broadly, and the replacement cycle driven by regulatory efficiency standards; supply is growing slowly, from a concentrated base, with no credible path to rapid expansion.

Dark fiber routes in secondary hyperscaler markets and the InP substrate supply chain sit at the point where the traffic demands of agentic AI land in the physical network. The always-on, north-south traffic of agentic workloads runs continuously on fiber pairs that were sized for more variable demand. InP substrate is required by every optical connectivity pathway, including pluggable, co-packaged, or photonic interposer, and faces a demand-to-capacity ratio that worsens with each bandwidth generation.

Baseload gas generation infrastructure completes the picture for the decade under analysis, providing the firm dispatchable power that renewables alone cannot; its long-run exposure to the SMR scenario is real but post-2030 in timeline, and assets contracted at baseload rates through the 2030s are largely insulated from that risk within the investment horizon that matters for decisions being made now.

Summary: The Four Durable Asset Classes

These four asset classes share a common characteristic: they are throughput infrastructure. Their returns come from physical utilisation, gas flowing through pipe, current through transformers, light through fiber, rather than from technological leadership or market position that could be disrupted by a successor product. The AI demand shift is increasing the throughput demand on each of them. That is the simple version of the thesis. It does not require a precise view on which AI architecture wins, which efficiency gains materialise, or which energy technology eventually displaces gas; it requires only the view that agentic AI reaches meaningful scale, a view that the current trajectory of investment, deployment, and enterprise adoption makes increasingly difficult to contest.

I. The Argument

The Financial Times recently published a careful summary of the case for why the Iran conflict and the potential closure of the Strait of Hormuz pose a material threat to the AI supply chain. It is worth engaging with seriously because it expresses, with admirable clarity, an argument that many investors are currently working through. The intuition: the chip supply chain is long and intricate, crossing more than seventy borders before a finished semiconductor reaches its destination; it depends on energy, and much of East Asia’s energy comes through the strait; it depends on specialty chemicals, several of which originate in or transit the Gulf; and, beyond the chips datacenters themselves are energy intensive and exposed to price increases. The longer the disruption, the deeper the fallout.

It is a coherent argument. The problem is not with the logic, it is with the underlying facts about how global commodity markets actually work. Each link in the causal chain, East Asian energy, specialty chemicals, US datacenter energy, breaks on contact with the structural realities of the relevant market. Below we work through each in turn.

II. East Asian Energy: Real Dependency, Wrong Conclusion

The FT is right that South Korea and Taiwan are heavily dependent on imported fossil fuels, and that a meaningful share of those imports transits the Strait of Hormuz. Taiwan sources over a third of its LNG from the Middle East. This is a genuine exposure.

The question is whether that exposure transmits into chip output in a way that materially affects the AI buildout. Here the arithmetic is clarifying. Energy accounts for approximately 3–5% of total semiconductor manufacturing cost at a leading-edge fab. A 3nm wafer at TSMC costs approximately $19,500 to produce. The energy component of that is perhaps $600–975. A doubling of energy costs, an extreme scenario by historical standards, adds at most that same amount again to a wafer that already sells for multiples of its production cost.

The magnitude does not move the needle. But there is a deeper point about how constrained inputs are allocated that sits above the cost arithmetic, and we will come to it in the final section. For now it is sufficient to note that even under a severe and sustained energy price shock, the direct cost impact on chip manufacturing is small, and that the chips being made are among the highest-value manufactured objects on earth, with the pricing power to absorb cost increases that would be existential for lower-margin industries.

The FT also notes that Taiwan keeps only around eleven days of LNG in reserve and depends on just-in-time deliveries. This is a real operational vulnerability for the island’s energy system. Whether it translates into fab shutdowns depends on how Taiwan’s government manages allocation priorities under stress. Advanced semiconductor manufacturing, which contributes disproportionately to GDP and sits at the centre of Taiwan’s strategic identity, would likely be the last sector to face enforced curtailment.

III. US Datacenter Energy: The Transmission Mechanism Does Not Exist

The FT’s second energy argument is distinct, and it deserves a separate answer. The claim is that US LNG exporters, seeing higher prices in Europe and Asia driven by Gulf disruption, will rush to sell their supplies to those markets. This will pull up Henry Hub and with it the electricity costs of US datacenters, at which close to 75% of planned on-site power comes from natural gas. Since electricity accounts for roughly half of a datacenter’s operating expenses, the argument runs, elevated US gas prices make the AI infrastructure buildout less viable.

The mechanism this describes (arbitrage between Henry Hub and international gas prices) is real in principle but is blocked in practice by a hard physical constraint: LNG export capacity. The United States produces roughly seven times what it exports as LNG. The ability to send that gas to international markets is limited by the number and throughput of liquefaction terminals, which cannot be expanded at short notice regardless of the price signal. You cannot liquefy and ship more gas than your terminals can process. The arbitrage is capped by infrastructure, not by price.

The empirical record on this is unambiguous. The EIA’s own current outlook notes that since the Hormuz closure, European and Asian gas prices have spiked by 62% and 83% respectively, while Henry Hub has stayed ‘relatively flat’. In 2021–22, the European energy crisis drove TTF, the European gas benchmark, to levels more than ten times Henry Hub at peak. US LNG exports increased over that period, but only as fast as terminal capacity allowed. Henry Hub barely moved. The structural decoupling between US domestic gas prices and global gas markets has now been tested twice in four years , and has held both times, for precisely the same reason: the transmission mechanism runs straight into the wall of fixed export infrastructure.

The AI buildout is concentrated in the United States, and its primary energy input is priced at Henry Hub. The Gulf energy channel, as it applies to US datacenters, is broken at the structural level, not because there is no incentive to arbitrage, but because the infrastructure through which that arbitrage would flow does not exist at the scale required.

IV. Chemicals: Real Dependencies, Frequently Misattributed

The FT identifies three chemical inputs — helium, sulphur, and bromine — as sources of vulnerability. On the underlying dependencies, it is largely correct. But the supply chain implications require more precision.

The systematic error in the chemicals argument is the conflation of raw material supply chains with semiconductor-grade chemical supply chains. These are different supply chains, with different geographies, different intermediaries, and different substitution options. The Gulf’s share of a raw material’s global production does not translate directly into fab exposure if fabs are not buying that raw material, they are buying a processed derivative manufactured elsewhere.

Consider sulphur. The FT notes that roughly half of global seaborne sulphur transits the strait, a striking statistic. But fabs do not buy elemental sulphur. They buy ultra-high-purity sulphuric acid (H₂SO₄) for use in Piranha cleaning solutions. That acid is manufactured by chemical companies in Japan, Germany, and Korea from sulphur sourced across global refinery byproduct streams, many of which are not Gulf-derived. The transit statistic is accurate at the raw material level and essentially irrelevant at the fab input level.

Helium is a more substantive case. Qatar does supply approximately 30% of global helium, and the FT is right that South Korea and Taiwan source heavily from there. Ras Laffan, the world’s largest LNG complex, which is also a primary helium processing facility, sustained damage from drone strikes and reportedly ceased production in early March 2026. Phil Kornbluth, founder of the leading helium consultancy, estimates it will take four to five weeks to resume gas and helium production, and a further two to three months to restore the supply chain to pre-crisis condition. Against this, South Korean chipmakers are reported to hold approximately six months of helium inventory. The cushion holds, but narrowly, and a prolonged closure would begin to bite.

What the cushion argument reveals, however, is the relevant magnitude: helium is approximately 0.5–1% of total fab operating costs. Even a complete and extended helium disruption is not an existential threat to chip output, it is an operational headache, and an expensive one, but it compresses margin rather than halting production.

The one case where the FT’s dependency claim and the supply chain geography are both correctly identified is bromine. ICL Group, headquartered in Israel and drawing from the Dead Sea and Negev Desert, controls approximately 35–40% of global bromine supply. South Korea imports the great majority of its bromine from ICL, and bromine compounds are genuine inputs into photoresists and packaging materials in leading-edge semiconductor manufacturing. This is a real concentration in a strategically important supply chain.

Three factors limit the risk. Jordan, which accounts for approximately 25–30% of global bromine production through Arab Potash Company, is uninvolved in the conflict and provides a meaningful alternative source. Bromine inventories in the supply chain typically run three to six months. And the constraint that a sustained bromine disruption would compound is High Bandwidth Memory production, which is already sold out through 2026 in its entirety, with SK Hynix stating explicitly that it cannot fully meet demand even at full utilisation. A bromine supply tightening would intensify a constraint that is already fully bound.

V. The Structural Point: How Scarce Inputs Are Allocated

Each of the preceding arguments can be made independently and each independently undermines the FT’s conclusion. But there is a structural observation that sits above all of them,

When inputs are scarce, whether those inputs are shipping slots, specialty gases, or chemical feedstocks , they are allocated by the price that competing users are willing and able to pay. The relevant question for the AI supply chain is not whether disruption would affect the availability of helium or sulphuric acid or bromine in the abstract. It is whether the manufacturers of advanced semiconductors would be outbid for those inputs by competing users.

The answer is almost certainly not. A GB200 NVL72 rack is worth $3–4 million and is subject to an order backlog measured in millions of units from the world’s largest technology companies. The helium required to manufacture it, the bromine compounds, the ultra-pure acids, each represents an infinitesimal fraction of the end-product value. The purchasing desks at TSMC, SK Hynix, and Samsung are, in any genuine input scarcity, among the highest-value buyers in every relevant market.

The lesson from Covid-19 and the Ever Given container ship blocking the Suez Canal in March 2021 is instructive here. The disruption cascaded most severely through low-margin, high-volume consumer goods (furniture, lower-end electronics, fast fashion) because those products could not outbid other cargo for scarce shipping capacity. Advanced semiconductor supply chains were disrupted less and recovered faster than virtually every other manufactured product category, precisely because they could pay whatever was necessary to move to the front of any queue that formed. A prolonged strait closure would replicate the dynamic.

Conclusion

The FT has done something useful: it has articulated, with clarity and precision, the intuitive case for why a Gulf disruption threatens the AI supply chain. The intuition is not unreasonable. It draws on real dependencies such as East Asian energy imports, Qatari helium, Israeli bromine — and traces a plausible transmission mechanism from commodity shock to chip output to AI buildout.

The problem is that the plausibility of the mechanism depends on assumptions about how global commodity markets work that do not survive contact with their actual structure. East Asian energy dependency is real, but energy is a small fraction of chip manufacturing cost, and the hierarchy of value means that chipmakers are at the front of any rationing queue. The US datacenter energy argument requires a transmission mechanism (arbitrage between Henry Hub and international gas prices) that is physically blocked by LNG export infrastructure, as Ukraine war experience demonstrated. The chemicals argument correctly identifies several raw material dependencies but misattributes them to supply chains that fabs do not actually use; with the partial exception of bromine, where the dependency is real but the margin for disruption is bounded by inventory and alternative supply.

The AI buildout does face genuine constraints. They are not in the strait. They are in the fabs: in the High Bandwidth Memory that is sold out through 2026, in the advanced packaging capacity that NVIDIA has substantially locked up, in the logic wafers where AI demand is absorbing a rising share of the world’s most advanced production node. These are the constraints that will determine the pace of the buildout over the next two years, after which power becomes the binding constraint and attention will shift to turbines and interconnect queues. They have nothing to do with Hormuz.

I once spent close to 48 hours in an office once without going home. Our client was selling a copper mine in South America and by that point there were perhaps a hundred highly paid professionals working through successive nights to make it happen. Deal counsel, local counsel, company counsel, bankers, accountants, consultants, executive teams. Two sides, multiple jurisdictions, and so on.

We were not working all night because that was the optimal number of hours to spend on the problem. Quite the opposite. I was a corporate lawyer, and I would estimate that a third of my time during those 48 hours was spent waiting for comments from someone else. The rest was work that an engaged 16-year-old could have done with sufficient instruction. The documents were not complicated because they required rare genius. They were complicated because the transaction had no right answers, only wrong ones, and then below the wrong ones a vast contested space where every provision was a negotiation. We were not trying to finish. There was no finishing. We were trying to extract the maximum possible advantage from a situation that would remain contested until the moment it closed and immediately reconstitute itself in the next transaction.

The reason our client hired a hundred professionals was not that a hundred professionals were the minimum required to execute the task. It was that this was an important transaction and important transactions reward firepower. I am certain that if the professionals had been cheaper, they would have bought more of them, not fewer.

I thought about that transaction recently when Dario Amodei used the Centaur analogy from chess: human plus AI outperforming either alone, but only as a transitional phase before AI pulls ahead entirely. It’s a compelling frame but not one that I recognize from those hours on the copper mine.

Chess ends. There is a winner and the winner is verifiable. Deep Blue doesn’t need to persuade Kasparov that he’s lost because the board tells him. The Centaur phase was real: for a period, the best human with AI could beat the best AI alone. That period is over. The best AI now plays better chess than any combination of human and machine, because every position has a fact of the matter and the machine finds it. Dario’s contention is that software engineering is in the Centaur phase now, and that the rest of knowledge work will follow. He is probably right about software engineering, for the same reason he’s right about chess. Code compiles or it doesn’t. Tests pass or they don’t. The signal is outside the players.

The warranty schedule I was working on was not like that. It had been drafted weeks before closing, commented upon, and disclosed against. Standard work, the kind AI handles capably already, probably better than the junior associate who first drafted it. By the time we were on those late-night calls, all of that was done and we were arguing about the final points.

Warranties are risk apportionment. These final points weren’t wrong answers waiting to be corrected, they were positions. Each side wanted the other to take more risk. The stakes weren’t high enough to escalate, so it fell to junior lawyers to grind it out, call by call, until we were told to stop. There was no external signal to consult. The answer was being produced by the people on the call, and it would only exist once they’d made it.

The bots will agree the technical points faster and to a higher standard than the junior lawyers ever did. That is unambiguously good. It means getting to the warranty schedule’s final points sooner, with cleaner drafts, with fewer errors to unpick. And then the junior lawyers will be on the late-night call anyway. Because those final points aren’t verifiable. There is no signal outside the room to optimize against, only the signal the players produce in real time.

You could build a negotiator bot one that haggles over warranty provisions faster and more strategically than any junior lawyer. But if both sides have one, the contested ground shifts to whoever instructed the bot and what mandate they were given. The haggling doesn’t disappear. It moves. And if the verifiable work gets cheaper, the clients don’t spend less. They spend the savings on more contested ground.

Keynes made the same mistake in 1930. Industrial inputs were getting cheaper, machines were replacing labour, and he extrapolated forward: within a generation, people would work fifteen hours a week and spend the rest in leisure. He was wrong by such a margin that the prediction reads as comedy now.

The obvious response is that this time is different, not because society won’t advance, but because all the advancing is verifiable and so will be done by AI. Fine, you might not be able to turn the haggling over copper mine warranties to a bot, but who cares, because it will be out here curing cancer, building space stations, finding supermaterials to replace copper. The contested work is a rounding error on the upside.

But that compounds the Keynes error rather than escaping it. There is no theoretical limit on the number of contestable tasks human beings can dream up. And their relative value won’t just hold steady as AI does the verifiable work, it will rise, because the productivity surplus from an AI boom has to go somewhere, and contested ground is where it goes. This is what happened to services in the industrial economy Keynes was imagining. He saw the substitution and missed the reconstitution.

And beyond that, technological progress tends to create problems that the same generation of technology cannot solve. Tell an agronomist in 1850 that by the 2010s overeating would be one of America’s greatest health crises and that reengineering human metabolism would seem the answer. Caloric abundance didn’t solve the problem of scarcity and leave nothing in its place. It created a new category of problem, one that couldn’t have existed before the abundance arrived. AI-generated abundance will do the same. We don’t know what the cognitive equivalent of obesity is. We know it follows from the structure of what’s happening.

Dario has given us the appealing image of a nation of geniuses working in data centres on the hard problems: cancer, poverty, the scientific frontiers that have been stalled for decades. That future seems both attractive and likely. But industrial abundance gave us penicillin and also the McDouble. Both came from the same cost collapse. Neither was predictable from the other. The nation of geniuses will cure diseases and it will also be deployed on tasks that today seem trivial or absurd, because in a world of AI-generated abundance, contested ground is where relative advantage lives, and relative advantage is always worth paying for.

The Centaur fails not because it’s wrong about capability but because it assumes a fixed game with a verifiable outcome. The copper mine wasn’t that: every time you make the verifiable work cheaper, more contested work appears. Keynes assumed static desires. We are assuming a fixed quantity of contested work. The actual picture: a nation of geniuses generating new contested games faster than any single game gets resolved. Not a threat to human economic activity. The definition of it.

Last week Anthropic released Claude for Excel. And while it got less attention than Cowork and robot social networks I think it’s the clearest signal yet that AI capex is undersized, not oversized, and that the skeptics are measuring the wrong market.

Claude in Excel doesn’t work the way you might expect. It doesn’t click through menus or enter formulas. It takes your natural-language request, writes Python to solve it, and presents the results in Excel. The spreadsheet is the interface, not the engine.

This matters because the technique, turning a question into code into an answer, isn’t specific to spreadsheets. It works for any task that can be solved programmatically.

We’ve already seen what happens when this technique reaches a population of users. In the two years since AI code generation went mass market, Google and Microsoft went from no AI-generated code to 30% AI-generated. Frontier labs are apparently above 80%. That adoption curve happened with technical users; software developers who could already code and who could evaluate and debug the output.

Excel extends the same technique to a much larger population: people doing analytical work who aren’t developers. The users are less technical on average, which might slow adoption. But the task is also simpler on average, the interface more forgiving, and the output is a finished artifact (a model, a chart, a report) rather than code that needs further integration.

Anthropic has already reported the efficiency case. The Norwegian Sovereign Wealth Fund says it’s saving thousands of person-hours. Nate Jones is building complex financial models (multiple scenarios, sensitivity analysis, and integrated statements) in minutes rather than hours.

But there’s a second effect. When the barrier to analysis drops far enough, people don’t just do existing work faster. They do work that was never worth doing before.

Last week I wanted to backcast 30 companies through 1,300 deals across 20 funds A year ago, I wouldn’t have attempted it. Not because it was impossible, but because the effort-to-insight ratio was too uncertain. The question was interesting; the work to answer it wasn’t justified. With Claude in Excel, I had it in an afternoon. That is demand that would not have appeared in any prior budget. Multiply it across every analyst who now thinks “I wonder if...” and follows through because following through is suddenly cheap.

Efficiency and expansion both drive compute demand, and both mean the spend has ROI. The efficiency case means existing workflows move to accelerated infrastructure and become more productive. The expansion case means new workflows appear that produce value that wouldn’t have existed at all. Either way, worth paying for more GPUs.

There is a precedent for workloads migrating to accelerated compute at this scale. A major tailwind for Nvidia over the past decade was social media, search, and e-commerce companies moving to recommender systems, which is why Meta is thought to be a top-5 Nvidia customer despite having no datacenter business. The workload migrated to a form that demanded GPUs, and the companies doing that workload at scale became massive consumers of accelerated compute.

That was consumer internet. Now we’re watching enterprise software start the same transition. Excel, the world’s most-used analytical software, begins to require accelerated compute. Once LLMs are embedded in the workflow, usage compounds. People find adjacent problems. The technique spreads to reporting, presentations, data pipelines, to any task where the bottleneck was translating intent into execution.

History suggests that expansion effects from fundamental cost reductions tend to be underestimated. Before the Bessemer process, steel was expensive and rare, used only where nothing else would do. Bessemer made it cheap and abundant. What followed was not just cheaper existing uses, it was steel-framed buildings, railroads at scale, eventually automobiles. Entire categories of demand that no one sizing the pre-Bessemer steel market would have predicted.

The capex skeptics have a coherent case. They look at the hundreds of billions being poured into AI infrastructure and ask: where’s the revenue to justify this? They see foundation model companies with impressive growth but not yet impressive profits. They see enterprises experimenting but not yet deploying at scale. They conclude we’re in an overbuild.

What they’re missing is that the addressable market isn’t “AI software” as currently defined. It’s every task where the bottleneck is going from question to answer programmatically, a horizontal layer across the knowledge economy, not a vertical.

Clearly overbuilds happen and timing is hard. The history of technology investment is littered with correct theses applied at the wrong moment. But two precedents point in the same direction. What happened to coding is happening to knowledge work: the same technique, reaching a much larger population. What happened to consumer workloads is happening to enterprise workloads: the same migration to accelerated compute. Claude in Excel is where these trends become visible. The demand is coming, and it will be worth paying for.

Strange corporate behaviour is usually a warning sign. When companies tolerate years of poor returns, blur boundaries with competitors, or subordinate near-term optimisation to opaque long-term goals, the standard conclusion is mismanagement or institutional drift. Yet at the technological frontier normal behaviour often fails, and success may instead require institutions that look wrong by conventional standards. Extreme ultraviolet lithography (EUV) is a clean example. No one wanted to build it. It was expensive, fragile, and delayed for decades. And yet, when the semiconductor industry hit a hard wall, EUV became unavoidable, and the only organisations capable of delivering it were those willing to behave in ways the market consistently disliked. The story of EUV is not just about how advanced chips are made. It is a case study in how unusual institutional behaviour can signal work happening at the edge of the possible, and why long-term investors should learn to recognise it.

The wall

For much of its history, semiconductor progress rested on a kind of disciplined optimism. Optical lithography improved, then improved again, and was then stretched slightly further than its designers had intended. The industry learned to live close to the Rayleigh limit, extracting more performance from the same wavelength by increasing numerical aperture, refining masks, and leaning heavily on computational tricks. Deep ultraviolet lithography at 193 nanometres, especially once immersion techniques were introduced, became one of the great triumphs of incremental engineering. It worked for far longer than almost anyone expected.

By the time leading manufacturers reached the 7-nanometre generation, however, even if the physics itself had not changed, the economics clearly had. Multi-patterning, once a clever workaround, became an organising principle of entire fabs. Critical layers that would have required a single exposure a decade earlier now needed two, three, sometimes four passes, each with its own mask, alignment risk, and yield penalty. Mask counts ballooned, overlay errors accumulated, cycle times stretched, and costs began to rise faster than transistor density fell.

This was not a temporary inefficiency that scale could smooth away. It was a structural wall. At around 7 nanometres, it was still possible, just, to proceed without a step change. At 5 nanometres, it ceased to make economic sense. Continuing down the same path would have required such elaborate patterning schemes that the cost per transistor would have risen rather than fallen. The industry had spent decades turning Moore’s Law into a financial proposition as much as a technical one, and at 7 nanometres that alignment finally broke.

What followed is often retold as a triumph of engineering. It is more accurately described as a process of elimination. Over the years, many alternatives to conventional optical lithography had been explored. Electron-beam systems promised exquisite resolution but proved hopelessly slow at scale. Nanoimprint techniques could replicate fine features but struggled with defectivity and throughput. X-ray lithography had an impressive theoretical pedigree and an equally impressive history of failure. Directed self-assembly generated excitement in laboratories and frustration in fabs. Each approach worked somewhere, in some context, for some period. None could carry the full weight of high-volume, leading-edge manufacturing.

EUV – the least bad option

By the mid-2010s, extreme ultraviolet lithography remained. It was neither elegant nor widely loved, but everything else had fallen away. EUV’s 13.5-nanometre wavelength was short enough to restore single-exposure patterning on the most critical layers, yet it was also deeply unforgiving. At that wavelength, light is absorbed by both glass and air, rendering conventional lenses useless. Everything has to happen in vacuum, using mirrors that reflect rather than refract, each polished to tolerances that border on the absurd.

For years, EUV was described as perpetually five years away, delayed not because of managerial dithering but because it demanded solutions at the edge of what was physically controllable. Generating sufficient source power required firing a high-energy laser at microscopic droplets of molten tin tens of thousands of times per second. Capturing and directing that light required mirrors composed of dozens of alternating atomic layers, each deposited with extreme precision. Keeping those mirrors clean in a vacuum filled with tin debris and resist outgassing required an entirely new discipline of contamination control. None of these problems yielded to shortcuts.

Strange firms at the centre

What is striking, looking back, is not how hard EUV was, but how few organisations remained willing to absorb that difficulty over time. By the late 2000s, most potential rivals had stepped away. The field narrowed to a single system integrator, ASML, and a single supplier capable of producing the necessary optics, Carl Zeiss SMT.

For many years, ASML did not look like a company optimising for shareholder returns. It spent heavily on EUV research long before there was a product to sell, tolerated repeated delays and public scepticism, and acquired suppliers not to expand margins but to stabilise learning curves. Its machines grew larger, more complex, and more expensive, culminating in systems with tens of thousands of components, each pushing the limits of precision engineering.

Zeiss was an even stranger partner. Its semiconductor optics business sits within a broader organisation owned by a foundation rather than public shareholders, a structure that places hard constraints on profit distribution and soft constraints on time. Zeiss invests in apprenticeship, in craft, and in optical perfection in a way that resists easy scaling. It is comfortable saying no to work that would dilute capability, even when that work would be profitable in the short term. For EUV, this mattered. The mirrors at the heart of an EUV scanner must be polished to surface errors measured in fractions of a nanometre, a process that cannot be rushed and for which there is no alternative supplier waiting in reserve.

Together, ASML and Zeiss behaved less like a conventional vendor and more like custodians of an industrial capability. That distinction helps explain a second feature of the EUV story that still looks odd: the way customers responded.

Unmarket structure

In 2012, with EUV still struggling to reach production readiness, three of the world’s largest chipmakers stepped in to fund its completion. Intel, TSMC, and Samsung Electronics invested billions of euros in ASML, taking minority stakes and earmarking capital specifically for EUV development. These firms were, and remain, fierce competitors. Under normal circumstances, one would expect them to free-ride on each other’s investments or wait for a supplier to bear the risk alone. They did neither.

This arrangement departed sharply from how markets usually operate: competitors collaborated; customers underwrote supplier R&D; ownership structures blurred; and engineers were embedded across firm boundaries. Yet it was rational. Without EUV, advanced logic scaling would stall. Without shared funding, EUV might fail. Each firm could see the same constraint. Cooperation was not a gesture of goodwill but a response to necessity.

The effect was to turn a supply chain into something closer to an institution. And when EUV finally entered volume production at 7 nanometres and beyond, those who had funded it found themselves inside a small circle. Others were excluded by default, leaving advanced semiconductor fabrication in the hands of 3 firms.

The EUV guild

This is where the language of guilds becomes useful, provided it is used carefully. A guild is not a club formed for comfort so much as a mechanism for preserving scarce skill under conditions where dilution would be fatal. Guilds are exclusionary by design. They prioritise mastery over scale and often look inefficient until they become indispensable.

The EUV ecosystem had distinctly guild-like features. The number of participants was small, the knowledge involved was largely tacit, and learning curves spanned decades rather than quarters. Those working within the system were often motivated by a desire to make the machine work at all, not merely to make it profitable. At the same time, the outcome of that effort conferred enormous economic advantage. Once EUV succeeded, it became a gatekeeper technology: access to it determined who could manufacture at the leading edge and who could not.

For investors, this combination is uncomfortable. Guilds repel capital early. They require patience and resist clean narratives. Their economics are lumpy and delayed, and they can look, for long periods, like exercises in stubbornness rather than foresight. Yet when they succeed, they often dominate precisely because the path they followed was unattractive to most others.

What all this may mean

It is tempting to conclude that EUV is a one-off: semiconductors are unusually constrained; the physics is unforgiving; the capital intensity is extreme; and there may be no other technology that demands quite this mix of precision, scale, and endurance. That may be true. It is also possible that EUV is simply a particularly clear instance of a broader pattern that appears whenever development pushes far enough beyond the median.

At such extremes, normal optimisation breaks down. Learning curves lengthen. Feedback becomes sparse. Institutions that insist on near-term validation fall away. Those that remain often look strange. They tolerate behaviours that markets dislike, blur boundaries that governance textbooks prefer to keep neat, and invest in capabilities whose value cannot be fully articulated in advance.

This is not a theory of everything so much as a warning flag. When highly sophisticated organisations behave in ways that appear inefficient, unfashionable, or outright irrational, and when those behaviours persist coherently over long periods, it may be worth paying attention. Sometimes the market is correctly identifying waste. Sometimes it is mispricing the cost of learning at the frontier. The EUV story suggests that the difference matters, and that it is not always obvious in real time.

Competition in physical industries is changing. The winner-takes-most dynamics long associated with digital markets are beginning to operate in parts of the physical economy, as learning becomes faster, cumulative, and embedded directly in machines themselves. Where this happens, competitive advantage compounds rather than stabilises, and market structure shifts accordingly. This matters because the physical economy (transport, manufacturing, energy, agriculture, logistics, aerospace) is vastly larger than the digital sectors that produced the last generation of dominant firms. Where this logic takes hold, it reshapes industrial concentration and the distribution of returns across the physical economy.

Digital concentration was structural, not accidental

Over the past two decades, global markets have been shaped by a small group of technology companies that dominated their respective domains, capturing extraordinary economic value while operating in parts of the economy that were, in aggregate, relatively narrow. Search, online advertising, cloud hosted software, and social platforms do not account for a massive share of global output. Yet firms that achieved durable dominance within them became the most valuable companies in history.

This is because the outcome was driven by structure rather than size. Digital markets rewarded scale in ways that traditional industries did not. Near-zero marginal costs meant that growth did not bring proportional expense. Network effects meant that each additional user increased the product’s value to others. Most importantly, learning occurred continuously and at speed. Interactions generated data, refining algorithms in real time. Improvements were deployed instantly across the entire user base. Scale did not merely increase revenue; it improved the product itself.

Once this dynamic took hold, competition behaved differently. In markets governed by diminishing returns, scale eventually stabilises rivalry as costs rise and coordination becomes harder. In digital markets, those brakes were weak or absent. Leading firms became more efficient as they grew, not less. Rivals were not simply smaller versions of the leader; they were structurally disadvantaged, operating with inferior data, weaker learning loops, and slower improvement cycles. Early advantages hardened into durable positions.

Valuations reflect this. Firms that demonstrated increasing returns are rewarded not simply for growth, but for the durability of that growth, reflecting the expectation that competitive pressure would diminish over time rather than intensify. A relatively narrow slice of the economy was therefore sufficient to support companies of unprecedented value, provided dominance within it was secure.

Why this logic once stopped at the digital boundary

For a long time, this logic was as largely limited to digital goods and services. Physical industries were governed by constraints that limited the speed and scope of learning: capital intensity rose with scale; supply chains introduced coordination friction; engineering cycles were slow, tied to certification, tooling, and production schedules; and learning occurred through redesign and process optimisation rather than through continuous feedback from use. Even where experience reduced costs, learning curves flattened as systems matured.

Competition in physical industries therefore behaved differently. Large manufacturers tended toward oligopoly rather than dominance. Market share shifted slowly and technological leadership rotated over time. Data did not materially alter this pattern: sensors and telemetry were widespread by the late twentieth century, but the data they generated was narrow, local, and difficult to reuse. It improved individual machines or production lines, not entire fleets. Feedback loops existed, but they were slow and fragmented. Learning remained subordinate to engineering.

That distinction reflected the technological realities of the period.

What changed: learning became embedded in machines

That condition is now breaking down. In a growing number of physical industries, learning has become embedded directly in products themselves. Machines now generate operational data continuously as a core feature of use. That data feeds machine-learning systems capable of extracting patterns and improving performance at scale. Crucially, those improvements can be deployed across entire fleets in days or weeks rather than years. Learning is no longer local or episodic; it now operates at fleet scale.

Artificial intelligence accelerates this shift. Machine-learning systems can process operational data at a volume and velocity that was previously unattainable. Improvements no longer depend on occasional redesigns or isolated field trials; they emerge from use itself. As AI becomes embedded in physical systems, most visibly in robotics, autonomy, and complex machinery, the feedback loops that once characterised digital platforms begin to appear in the physical world as well.

The consequence is not simply better machines. It is a change in the logic of competition.
The significance is that the winner-takes-most dynamics that reshaped digital markets are beginning to operate in parts of the physical economy, changing the organising logic of industrial competition.

What makes this shift particularly consequential is that its earliest examples emerged under far more limited forms of machine intelligence than those now being deployed at scale. Systems such as SpaceX’s launch platform or John Deere’s precision agriculture stack achieved learning-driven advantages before the advent of large-scale generative models and embodied AI, suggesting that the underlying dynamic is robust and that its effects will deepen as machine learning becomes more capable, more autonomous, and more deeply embedded in physical systems.

What learning-driven competition looks like in practice

The clearest illustration of this shift is found in aerospace launch. SpaceX’s Falcon 9 system has accumulated hundreds of landings with reliability approaching 99 per cent. The significance of this achievement lies not in reusability alone, but in cadence. SpaceX operates on a tempo that allows learning to proceed faster than traditional engineering cycles. Each launch generates high-resolution operational data across thousands of sensors. Each landing feeds models that inform subsequent flights. Reused boosters carry forward the accumulated experience of the fleet.

Traditional aerospace programmes operate on multi-year development and certification cycles. SpaceX compresses comparable learning into months. The gap between SpaceX and its competitors widens even as rivals invest heavily. The advantage lies less in execution or risk tolerance than in learning speed.

The same logic appears in more constrained form elsewhere. John Deere provides a useful illustration of how learning-driven competition can emerge in a capital-intensive, asset-heavy industry without the speed or flexibility of software or aerospace.

Over two decades, Deere has embedded sensing, connectivity, and machine learning into its equipment, allowing operational data from the field to inform performance across deployed fleets. Precision agriculture systems generate continuous information on soil conditions, crop health, and application accuracy. That data feeds models that improve outcomes over time, with improvements reflected in deployed equipment as well as in design, manufacturing, and service.

The pace of this learning is necessarily slower. Agriculture is seasonal, environments vary, and deployment cycles are measured in years rather than weeks. Yet the underlying mechanism is the same: performance improves through use rather than redesign alone; learning accumulates across machines and geographies; and competitive advantage derives less from marginal cost reduction than from the steady compounding of operational insight.

Deere’s experience shows that learning-driven dynamics are not confined to founder-led technology companies or unconstrained digital markets. They can emerge in established industrial sectors where incentives align and feedback loops are allowed to operate, even if the resulting outcomes fall short of outright winner-takes-most dominance.

When learning outruns engineering

These examples share a defining feature. Learning proceeds faster than competitors can respond so improvement cycles outrun engineering cycles. Once that threshold is crossed, competition stops stabilising and begins to tip. Leaders pull ahead and followers struggle to close the gap even when capital is available. Market structure begins to resemble software rather than manufacturing.

This does not happen everywhere. It requires specific conditions: tasks must be sufficiently repeatable for learning to generalise, outcomes must be measurable, data must accumulate at scale, improvements must be deployable rapidly, and regulation must permit iteration. Where these conditions are absent, physical industries continue to behave as they always have.

For example: construction remains resistant because tasks vary too widely; healthcare devices are probably too constrained by regulatory cadence; and (most) consumer appliances lack the economics to support sophisticated learning systems. Many industrial domains will remain governed by diminishing returns and stable competition.

The idea of reindustrialisation deserves to be taken literally. The US $7 trillion of forecast AI-infrastructure spending is only the visible front of a much larger mobilization, perhaps US $30–50 trillion of investment that the Western world will need to undertake over the coming decades to rebuild the physical systems on which its prosperity rests.

This is not a cyclical investment boom but a structural project that governments now require to happen. It will be financed, guaranteed and accelerated by states that can no longer rely on extended supply chains or external manufacturing bases. Economically, it combines the ambition of the computer revolution with the mobilisation of the post-war rebuild.

With our friends at Brightwave we have put together a report detailing the key investment themes and opportunities from this wiring. This is available on request and below are some of the immediate opportunities in rare earths, battery grade graphite, and advanced packaging/HBM.

From global logic to structural bind

The modern global economy rests on five interlocking systems: energy, logistics, manufacturing, intellectual property and finance. For much of the industrial age the West controlled all five with the oil shocks of the 1970s showed what losing even one pillar could do. But beginning in the 1980s the logic of globalization took hold:: by retaining command of finance and of intellectual property (particularly in compute) the West could afford to let the capital-intensive, low-margin work of extraction, logistics and fabrication migrate abroad.

That bargain produced three decades of cheap goods, low inflation and strong asset returns. It also hollowed out the industrial core that had supported those same systems. Production, skills and middle-class stability were traded for efficiency and balance-sheet scale.

At home, this separation of thinking and doing eroded capability. The West retained world-class talent in design and research but lost the ability to execute at scale. Abroad, China, first through cheap labour and then through cheap capital created by domestic financial repression, methodically captured the lower three pillars. It became the workshop of the world, then the assembler of its machines, and finally the refiner of its inputs.

Deng Xiaoping’s ambition that China should become ‘to rare earths what Saudi Arabia is to oil’ has been realised: Beijing now controls close to ninety per cent of global magnet production and almost all heavy-rare-earth processing. The pattern repeats across logistics and refining: 98% of the world’s ship-to-shore cranes are Chinese-made; China produces no crude oil but operates refining capacity on par with the United States, etc.

And China did not stop there. It refused to accept a world in which it performed the labour of production while the West collected the rents of invention. It built an integrated high-technology economy in which the five pillars are fused again. In artificial intelligence, robotics, electric mobility and power electronics it now competes at the frontier; all with state coordination that the liberal economies can no longer match.

Debt as substitute for growth

During the post-industrial decades, as productivity growth slowed, governments filled the gap with debt. The system could sustain higher leverage because the West still owned the institutions in which the world trusted. But debt is a substitute for growth, not a replacement. In the long run a post-industrial economy simply grows more slowly than an industrial one.

The consequences are now structural. Western prosperity depends not only on being rich in absolute terms but on remaining more advanced in relative terms. The control of the upper layers of finance and intellectual property attracts global capital and supports sovereign credit. The legitimacy of that credit, and therefore of Western fiscal power, rests on continued technological and institutional primacy.

The West must restore productivity or watch its financial architecture weaken. It must defend both absolute and relative wealth. Losing the former invites decline; losing the latter undermines the very logic of its economic order.

AI as the hinge

Artificial intelligence is the hinge on which the next industrial era turns. It is both the cause and the remedy of the West’s industrial drift: a technology that exposes the limits of today’s physical capacity even as it offers a way to rebuild it.

Training and deploying modern AI systems demand vast new infrastructure; data centres, power generation, advanced chips, and materials that do not yet exist at the required scale. At the same time, AI promises a path out of the long stagnation in productivity by turning energy into cognition; by making brainpower, like horsepower before it, a function of watts.

And because embodied AI such as autonomous machines, robotics, precision manufacturing learns fastest when design, hardware, and data are tightly coupled, it rewards vertical integration. The companies that build and own the full loop of making and learning will advance the frontier.

AI, in that sense, is not only the symbol of a new industrial age but its engine.

The over-determined response

The Western response is being driven simultaneously by security, economics and politics. Governments must rebuild productive capacity to reduce strategic dependence and to restore the fiscal base of their economies. Entrepreneurs want to rebuild it because vertical integration again confers competitive advantage. Markets are rewarding it because tangible investment, for the first time in decades, is yielding outsized returns.

This alignment resembles the post-war mobilisation more than any recent policy cycle. The objectives are similar: to restore capacity, contain a systemic rival and reaffirm the credibility of liberal capitalism. Then, as now, the state is using every instrument at its disposal from direct subsidy to concessional finance, tariffs, offtake and regulatory preference.

The political pendulum has swung toward industrial policy across the alliance. In Washington, even after the scaling-back of the Inflation Reduction Act, core production credits and loan-guarantee authority remain intact. The CHIPS and Science Act continues to fund semiconductor and packaging capacity, and the Department of Energy’s loan programmes now anchor much of the critical-minerals build-out. Europe has replicated the model through its Chips Act and Critical Raw Materials Act; Japan, Korea and Australia have aligned capital deployment with allied security goals.

Industrial policy is the organising principle of the decade.

Proof points in the build-out

The scale of the coming reinvestment is vast, but the logic is visible already in three early theatres: rare-earth magnets; battery-grade graphite; and advanced semiconductor packaging, each revealing how government support, scarcity and strategic urgency combine to create investable rebuilds.

Rare-earths and NdFeB magnets

China’s dominance in magnet production and heavy-rare-earth processing has forced Western governments into a degree of market participation unseen in peacetime. In the United States, the Department of Defense’s partnership with MP Materials includes a guaranteed offtake for ten years and price floors for neodymium-praseodymium oxide. Similar backing supports Lynas Rare Earths’ new Texas facility and Energy Fuels’ separation line in Utah. These projects are not speculative: they are instruments of industrial strategy designed to anchor allied supply chains.

Battery graphite anodes

When Beijing imposed export controls on purified graphite in 2023, it weaponised a monopoly. The Western response has been immediate. The Department of Energy has underwritten new capacity from companies such as NOVONIX and Syrah Resources, pairing direct grants with production tax credits that cover up to forty per cent of capital cost. For automakers, the qualification window for new anode suppliers is two years, creating natural moats for first movers. Here too the state has become the bridge across the “valley of death” that once killed Western projects.

Advanced packaging and high-bandwidth memory

Not all chokepoints are geopolitical. The AI boom has exposed a different kind of vulnerability: the ability to assemble and package at scale. Advanced substrate capacity is concentrated in a handful of Japanese firms (in particular Ibiden and Shinko) and in Korean HBM manufacturers such as SK Hynix. Western governments are now treating this as infrastructure. The CHIPS Act’s recent disbursements include grants for Amkor’s Arizona packaging facility with Intel and for Micron’s new HBM fabs in New York and Idaho. These are not defensive subsidies but attempts to rebuild capability in the middle of the supply chain, the connective tissue of the AI economy.

Each of these sectors reveals the same pattern: acute dependence on a single supplier base, long qualification cycles that reward early entrants, and public policy willing to guarantee demand to secure strategic independence. They are the first nodes of what will become a much broader industrial renewal.

The new industrial order

The shape of the next stage of globalisation is becoming clear. It will not be a retreat from trade but a re-anchoring of production within trusted networks. The West is not abandoning openness; it is redefining it around reliability. Finance and technology will remain global, but supply chains will become regional and allied. Energy, logistics and manufacturing will again be treated as instruments of sovereignty.

This is not a temporary reversion to state planning; it is a structural correction to an era that over-financialised efficiency and under-invested in capability. For the first time since the 1940s, the leading economies are investing to secure the foundations of their relative power rather than assuming them. The scale rivals post-war reconstruction; the driver is existential rather than discretionary.

Artificial intelligence has made that necessity visible. It binds the abstract to the material, forcing a reintegration of thought, capital and production. The outcome will be a new industrial order, one in which Western economies rebuild not for nostalgia but for survival, and in doing so re-establish the coherence that made their earlier success possible.

The question for capital is not whether this will happen, but how quickly, and who will build it.

The world’s trade moves through a few narrow corridors: oil through pipelines and straits; containers through ports and shipping lanes; data through cables and satellites; and capital through payment networks and clearing houses.

For most of the last generation these channels were part of the unipolar landscape, quietly profitable and largely unexamined. But for the century before scale itself was strategic: railways, refineries, and shipyards were built as instruments of power as much as of profit. Only in the long calm of globalization did scale become a matter of efficiency rather than intent. That distinction is ended and that ending has brought huge opportunity.

Two companies illustrate the change: SpaceX, which rebuilt the economics of access to orbit, and COSCO, which turned China’s shipping fleet into a global logistics network. They operate in different systems and under different political conditions, but both mark the return of strategic scale: the fusion of visionary entrepreneurs, technology, capital, and state purpose that once defined industrial power.

This fusion is reopening entire classes of industry once thought mature. Energy, logistics, compute, communications, and finance each contain room for their next generation of SpaceXs and COSCOs: industrial platforms built to compound at venture speed.

The system of global trade

Global trade depends on five connected systems: energy, logistics, compute, communications, and finance. When they change, the pattern of the world economy changes with them.

Despite the ballyhooed death of globalization, global trade itself is of course not disappearing. Production will shift, capacity will move closer to home, and supply chains will shorten, but the movement of goods, energy, and information across borders remains the central fact of the world economy. These five systems are what make that movement possible.

Over the past three decades, the drive for efficiency produced extraordinary scale. Standardisation and cost discipline concentrated each system into a few global players. That structure worked while the environment stayed predictable; it was an efficient solution to its moment. As technology and geopolitics reset the conditions of competition, above all the rise of China and the end of the unipolar moment, those same structures have become the starting point for renewal.

Efficiency and concentration

By the early 2010s, each of the five systems had reached a form of structural maturity. Costs were low, margins thin, and the logic of scale left little room for entrants. The details differed across sectors.

Launch was technologically stagnant. Costs remained high, cadence low, and hardware bespoke. Procurement was almost entirely governmental. The industry was strategic but closed.

Shipping was intensely competitive but structurally consolidated. A handful of alliances coordinated over ninety per cent of container capacity; the barriers to entry were capital, coordination, and time. Margins fluctuated, but the industrial architecture itself was fixed.

Both conditions, stagnation in launch and oligopolistic efficiency in shipping, deterred newcomers. Both were industries whose underlying structures had stopped changing, even as competition within them remained intense. Scale and capital requirements protected the incumbents that efficiency had created. Those were the environments into which COSCO and SpaceX inserted themselves.

COSCO: the logic of the state-backed network

In 2016, Beijing merged two large state-owned carriers, China Ocean Shipping and China Shipping Group, into a single enterprise, China COSCO Shipping. The merger was one of several “national champion” consolidations of the period and came with balance-sheet support on a historic scale. Between 2010 and 2018 COSCO received roughly $132 billion in policy-bank lending and direct subsidies. For comparison, that was more than the combined market capitalisation of Maersk and Hapag-Lloyd at the time.

The capital bought reach. COSCO expanded its fleet to more than five hundred vessels, acquired or built terminals at thirty-eight ports, and took stakes in facilities from Piraeus and Zeebrugge to Abu Dhabi and Chancay. It integrated port operations, shipping finance, and logistics data into one system. The firm’s operating logic changed from cyclical profit to strategic continuity: it could maintain service on unprofitable routes, invest through downturns, and absorb short-term losses in pursuit of long-term position. It became both a commercial carrier and an instrument of state capacity.

Underneath that consolidation lay an ambitious program of digital modernization. COSCO integrated its terminals through real-time logistics platforms, deployed sensors across ports, and became an early adopter of blockchain for cargo documentation and settlement. The result was not technological frontier in the SpaceX sense but systemic intelligence: a network that could see and manage global trade as a single, data-driven organism.

COSCO’s transformation also has a human author. Xu Lirong, a career mariner who began as a deck officer, spent decades moving between the bridge and the boardroom. When Beijing decided to merge its carriers, Xu saw an opportunity not merely to rationalise assets but to create an enterprise capable of competing with Western giants on scale. He orchestrated the merger of fleets, terminals, and finance into one network and persuaded policymakers to fund it at levels private markets could never match. His genius lay in understanding what the state wanted: security of supply, projection of capability; and what the industry needed: capital, consolidation, and time. He aligned those interests tightly enough that COSCO could act commercially while serving national strategy.

The structure that followed is distinctive. COSCO controls ships, terminals, and finance. It can direct freight through its own network, use port stakes to influence routing, and coordinate with China Merchants and other state actors to protect supply chains. It also gathers data. The same logistics systems that track cargo provide intelligence on trade flows and demand. COSCO’s ports handle around one in ten containers moved globally; the company now sits astride the key routes of physical trade.

To Western eyes, COSCO’s scale looks defensive, and in its connections to the PLA worryingly offensive. But most of all it is a demonstration of what the combination of patient capital and technological integration can achieve in an industry others consider mature. COSCO used the balance sheet of the Chinese state to buy permanence in a volatile business and technology to bind it together. The returns are political before they are financial, but the network itself is durable and immensely valuable.

SpaceX: the logic of technological compounding

SpaceX started with no such protection. When Elon Musk founded the company in 2002, launch was dominated by a few aerospace primes operating on cost-plus government contracts. Each vehicle was bespoke, prices were measured in tens of thousands of dollars per kilogram, and cadence was measured in months.

SpaceX attacked with physics. Reuse was the core idea; make the first stage a flying asset, not a consumable. It built vertically integrated manufacturing in California and Texas, wrote its own flight software, and tested in a loop tight enough that each launch fed directly back into design. Costs fell by an order of magnitude, reliability improved, and cadence rose.

The financial architecture followed the engineering. NASA’s commercial cargo and crew programmes provided milestone payments tied to performance, not subsidy. They gave SpaceX predictable revenue without compromising control. The company used those contracts to fund its own development, later adding commercial launches and, crucially, Starlink, the low-Earth-orbit constellation providing global broadband. Starlink is both a product and a captive customer, filling the launch manifest while generating recurring cash flow.

The result is a flywheel. More launches improve technology and lower cost; lower cost attracts more launches; Starlink fills any slack and adds margin. Government missions anchor the floor; private and international customers expand the ceiling. SpaceX turned a bureaucratic procurement market into a self-sustaining industrial machine.

That compounding is what distinguishes SpaceX from earlier state-aligned giants. The same physical assets, launchpads, factories, constellations, double as data systems that learn with every use. Each iteration lowers marginal cost and raises informational yield. In effect, SpaceX married the capital intensity of an industrial program with the margins more like a digital platform.

Financing ambition too large for the market

Ambition at this scale requires long-duration capital, tolerance for technical risk, and customers willing to buy what does not yet exist. Private markets rarely provide all three at once. They can be created only when the state steps in through credit, procurement, or regulation.

China provided COSCO with credit: long-dated, low-cost funding that matched the cycles of shipbuilding and trade. The purpose was national resilience. Commercial returns were acceptable but secondary. The state absorbed volatility to maintain strategic capacity.

The United States provided SpaceX with demand: milestone contracts that paid for capability, not promises. The purpose was performance. NASA and the Department of Defense bought services, not equity. That arrangement forced delivery and kept innovation on the critical path.

Both companies exploited what the market would not fund. Both used state partnership to cross the gap between ambition and solvency. But COSCO converted that support into stability; SpaceX converted it into acceleration.

TSMC: the logic of globalized mastery

At first glance, TSMC could seem part of the same story as COSCO and SpaceX: a visionary engineer-entrepreneur fusing technical mastery, innovative business models, and state support to dominate a global chokepoint. Morris Chang built the company with a clarity of purpose that rivals Xu Lirong or Elon Musk, using engineering excellence and national backing to command a system the world cannot do without.

Yet TSMC, and the ecosystem that surrounds it (ASML in lithography, ZEISS in optics) are products of a different world. They are the purest expression of globalization’s logic: cooperation through specialization, neutrality as strength, and reinvestment as defense. TSMC’s discipline in execution, refusal to chase margins, and relentless capital expenditure created a monopoly, reinforced by a network of equally specialized partners. Together they formed the industrial spine of the global era: privately run, globally sourced, and geopolitically neutral.

That neutrality is now impossible. The United States wants their capacity and know-how onshore; China wants access to their output; both seek to ensure the other does not control them. Their challenge is no longer market competition, they mastered that through cost, scale, and execution, but strategic competition, where geography and politics define risk. The global dispersion that once secured efficiency has become exposure.

TSMC shows how the logic of globalization has reached its limit. It can look like a peer to COSCO or SpaceX but it belongs to another order. Its success depended on a world where production and capital could flow freely; COSCO and SpaceX are products of its aftermath, firms built for a world where power, not efficiency, again sets the terms of scale.

The opportunity for investors

The five great systems, energy, logistics, compute, communications, and finance, are all dominated by incumbents built for the conditions of the last thirty years. They are large, stable, and reluctant to change. Their stability once made them safe. It now makes them vulnerable.

Technology and geopolitics are shifting the economics of these systems. The build-out of compute capacity, the need for secure and redundant communications, the restructuring of supply chains, and the search for greater resilience in finance and logistics all demand new capital and new approaches. Governments have moved from being referees to being participants: investors, customers, and regulators at once. That combination is prying open sectors that had looked closed.

These are enormous, strategically important industries where the combination of technology and state partnership now makes real disruption possible. The same features that once made them forbidding mow create the opportunity. The task is to find entrepreneurs able to use both tools: technology that changes the economics and state alignment that lowers the cost of capital.

These are demanding markets: slow, capital-intensive, and heavily regulated. But they can support very large companies. They already contain businesses worth tens or hundreds of billions of dollars, and there is room for more. They are not the only places where compounding will occur, but they are among the few where technology, capital, and policy now intersect to make it possible. The same combination that produced COSCO and SpaceX will define the next generation of strategic scale.

It is natural to reach for the dotcom bubble when trying to make sense of today’s artificial intelligence boom. The parallels are tempting: capital surging, demand curves stretched to infinity, valuations that seem unmoored from reality. But dotcom is the wrong story. That episode ended in collapse, erasing both capital and much of the technology. AI is not dotcom. It is shale.

The shale revolution did not fail. It succeeded so completely that it reshaped the global energy order. New drilling and fracking techniques turned the United States into the world’s top producer of oil and gas, reduced dependence on the Middle East, and re-anchored American power. Yet most investors in shale were ruined. Productivity gains drew in too much capital. Debt-fuelled independents flooded the market. OPEC’s attempt to crush shale only accelerated the cycle. The paradox was that shale’s abundance triumphed geopolitically but starved capital of returns.

AI is tracing the same arc of abundance. Every new GPU cluster, every megawatt energised, every advance in quantisation or inference efficiency makes the system more productive. But productivity invites more capital, which risks overshooting demand. The lesson from shale is not that success evaporates. It is that success does not guarantee profits for all. There will be returns in AI; but only for those who sit inside the envelope now being constructed around it.

China as OPEC, Scale as Strategy

Shale’s decisive moment came in 2014, when Saudi Arabia flooded the market to bankrupt U.S. producers. Shale survived because its productivity curve was steeper than OPEC expected. In AI, China will play that role.

Beijing cannot yet match the exquisite assets of the industry: frontier chips and leading models. Instead, it follows a familiar playbook: drop to the lowest defensible point in the value chain, mobilise scale, and climb upward. It has done this before. In oil refining, China became dominant despite lacking crude of its own. In rare earths, it cornered the midstream. In solar, batteries, and now electric vehicles, it moved from defensive scale to outright leadership.

For AI that means subsidised power, cheap state-directed finance, Ascend-class accelerators, and permissive open-source models. The aim is both defensive and offensive: ensure sovereignty at home; and set the global clearing price from below abroad. A flood of cheap inference into the Global South could resemble OPEC’s oil glut, depressing returns for Western independents, even if the frontier remains American.

The Western Response: The Technodollar Envelope

Unlike oil in the petrodollar era, where U.S. and OPEC interests were often aligned, AI sits inside a rivalry. It cannot be left to market forces because it has become a direct arena of competition. A new architecture of power is emerging: the technodollar. Where the petrodollar rested on oil priced in dollars and U.S. security guarantees, the technodollar rests on controlled access to the AI stack: chips, cloud regions, frontier models, and the applications that run on them.

The technodollar is an envelope, and three mechanisms define its shape:

· Certification and compliance. Governments are creating trusted lanes. Only certified providers will be allowed to handle public-sector and critical workloads. This favours hyperscalers and large neutral colos that can absorb the compliance burden and locks out leaner independents.

· Procurement. Multi-year contracts and capacity payments will turn episodic training spikes into baseload demand. Governments will not leave sensitive compute to market cycles. They will guarantee utilisation for those inside the envelope, smoothing returns and underwriting investment.

· Power policy. A data centre is a machine for turning megawatts into tokens. Chips are useless without electricity. The true bottleneck is not GPUs but energised, contracted power. Interconnect queues and transmission bottlenecks mean that the scarce input is firm supply that can be delivered quickly and guaranteed for decades. Gas is fast but volatile. Low-carbon baseload (nuclear uprates, hydro, geothermal, gas with CCS) wins because it provides predictability, unlocks subsidies, and aligns with allied procurement rules. Power producers and grid enablers are thus central to the technodollar architecture.

This envelope is not purely American. It must hold together a coalition. Washington cares most about security and scale. Europe and Canada insist on carbon intensity and climate alignment. Japan and Korea emphasise supply chain resilience and domestic champions. The technodollar’s rules of entry are shaped by all these priorities. Security, procurement, and carbon standards are the glue that binds the Western bloc and distinguishes it from the Chinese alternative.

For other states, the choice is starker. Some will try to hedge, seeking access to both U.S. and Chinese systems. But these swing sovereigns risk losing privileged entry to the technodollar stack if they straddle both sides for too long.

Together, these mechanisms create the envelope. They ensure that even in a glut, capacity aligned with the technodollar survives, consolidates, and earns. Which means that returns are concentrated rather than erased.

Who Wins Inside the Envelope

• Hyperscalers and leading edge model providers — the supermajors. Microsoft, Amazon, Google, likely OpenAI. They will absorb years of thin returns, then consolidate weaker capacity. Distribution moats and political capture make them indispensable.

• Power producers and grid enablers. Power is the new right-of-way. Utilities and IPPs that can deliver firm megawatts at speed, and the transmission developers and EPCs who energise them, will command strategic rents.

• Specialist data-centre landlords — the midstream. Multi-tenant, power-adjacent campuses with re-tenanting optionality and policy lanes will consolidate into critical infrastructure, much like pipelines in energy.

• Efficiency innovators — the shale services. The firms that bend cost curves: compilers, serving frameworks, inference-first silicon, immersion cooling. They decide how brutal the overshoot becomes.

• Early-cycle suppliers — the picks and shovels. HBM memory, advanced packaging, high-speed optics, transformers, switchgear. Acute bottlenecks today, strong returns for the next year or two, normalisation later.

Everyone else, the leveraged independents, the single-tenant shells, the swing sovereigns who try to sit across both systems, will face the paradox of abundance without the shelter of the envelope.

The Shape of the Cycle

Phase One: bottleneck boom, as scarce kit captures outsized rents.
Phase Two: efficiency gains and Chinese competition drive down effective costs, stressing independents.
Phase Three: Western policy enforces the technodollar; procurement backstops demand; consolidation begins.
Phase Four: a regulated oligopoly of hyperscalers, power platforms, and certified colos, operating as critical infrastructure.

Investing in Abundance

The paradox of abundance is that it delivers both geopolitical success and financial disappointment. Shale demonstrated that abundance could transform America’s place in the world, even as it left investors disappointed. AI will do the same but with an important difference. The technodollar ensures that abundance does not wipe out returns altogether. It channels them toward those players inside the envelope.

For long-horizon allocators that means: anchor in firm power and interconnects; back the campuses with policy lanes; accumulate exposure to bottlenecks while they last; and fund the efficiency layer that keeps the West competitive even in a price war; and hold the hyperscalers/model leaders because they will consolidate stranded assets and emerge as the regulated utilities of AI.

Artificial intelligence will succeed spectacularly, just as shale did. But this time, there will be returns, not for everyone, but for those who position themselves inside the technodollar envelope, where policy, power, and productivity intersect.

Every venture cycle invents a theory of abundance.

In the 2010s, it was that every vertical could produce a platform. A clean interface and enough capital could turn almost any service into a network-effect machine. That theory gave us food delivery apps, mobility startups, and a wave of VC-subsidised convenience.

In today’s AI cycle, the illusion has changed form. This time, it lives at the application layer: in the wrappers, copilots, and workflow tools that grow quickly but stand on borrowed ground.

Most of the real winners will come from the hard and expensive work or changing user behavior and building real data moats. But the mistake is not overpaying for the best, it’s assuming there are more of them than there really are.

The Shape of Scarcity

It is a strange feature of financial history that markets, which are said to be composed of intelligent participants, routinely forget the same things. Chief among them: the unevenness of return.

We have had the data for a long time. In public equities, just four percent of stocks have accounted for nearly all net wealth creation over the last century. The median stock does not outperform short-duration bonds. The average only rises because the best are so good that they compensate for everything else.

This a structural fact rather than a statistical curiosity. A fact that holds even more strongly in venture. Private markets carry fewer signals, offer no exit path for most participants, and endure long stretches without price discovery. The outcome is predictable: one or two companies determine the fate of an entire portfolio; the rest fade, quietly or otherwise.

And yet, in every cycle, we behave as if greatness is not just possible but abundant. Capital flows not only into the company that is working, but into anything that looks sufficiently like it. The illusion takes hold that there are many more winners to be found; that the original was not exceptional, but representative.

This is the defining feature of a venture bubble. Not that valuations exceed reason, but that conviction exceeds scarcity. Not too much paid for the best, but too many treated as if they might be the best.

The mistake is not in the pricing, it is in the count. And it is happening as much in the current AI wave as in any cycle to date.

Last Time: The Illusion Was Breadth

The nountech boom of the 2010s was built on a powerful intuition: that mobile interfaces would reorder large parts of the economy. If the smartphone could intermediate human attention and action, then surely any service delivered offline could be reimagined as software.

This gave rise to a simple theory: any category, transportation, food, cleaning, housing, insurance, could produce its own Uber. The only requirements were a sleek interface, a well-subsidised experience, and enough time to capture demand. Everything from laundromats to car leases briefly appeared to be a platform in waiting.

At a surface level, the model worked: users flowed in; adoption was rapid; and in many cases capital kept the flywheel spinning until growth could be read as validation. But most of these companies were not changing consumer behavior. They were smoothing it. They relied on habits that already existed. The subsidy did not create a new market, it simply paid for an old one to appear modern.

Beneath it all was the real driver: a capital stack pushed out of shape. Interest rates sat below their natural level. Allocators with fixed liabilities could not meet return targets in traditional places. The result was not exuberance but compulsion.

Too much capital chased too many copies of a winner that was never really reproducible. Uber may well have been an outlier, a company that escaped by force of leadership and an almost reckless willingness to deploy capital before incumbents could react. It was not a template. It was, in a way, a fluke.

But it was mistaken for a pattern. And so the illusion took hold: that breadth, the extension of a model across many verticals, would yield a generation of giants. It did not.

This Time: The Illusion Is Surface

The AI boom is different. The technology is stronger and the gains, in some domains, are compounding. It is more plausible, at least on the surface, that we are living through a foundational shift.

But the illusion has returned. It has simply changed form. Where nountech mistook vertical proliferation for inevitability, AI risks mistaking surface traction for durable value.

Every week brings news of another AI-native company with explosive growth: vertical copilots, workflow wrappers, decision assistants. They grow quickly because they do not ask users to change much. The interface improves. The friction lessens. But the underlying behavior is familiar. In some cases, the user is simply being given autocomplete for an existing task.

The metrics are impressive. Revenue arrives early. But the advantage is borrowed. These companies build atop foundation models they do not control. They rely on cost structures that are artificially low. They exist in ecosystems where the model providers are not dormant infrastructure, but active competitors.

The Windsurf acquisition highlights this. The model layer will not remain passive. It will not watch as applications siphon margin and loyalty. It will acquire, integrate, and internalise the best ideas. Not out of spite, but because it can, and because, increasingly, it must.

What this means for the current generation of application-layer AI startups is uncomfortable. Many will not fail outright. They will simply be absorbed. They will be reinterpreted as product exploration exercises, not companies in their own right.

Just as nountech startups mistook app delivery for defensibility, many AI companies today are mistaking distribution for moat. The abundance is again an illusion. The surface looks wide. But the layer is thin.

Where the Real Opportunity Lies

If we are honest about what the power law demands, then we must also be honest about how little of today’s startup landscape will matter. Not because it is unserious, or poorly run, or even overvalued. But because it is not sufficiently different.

To create a company that escapes the gravitational pull of the model stack will require more than a clean interface or a faster sales cycle. It will require either a meaningful change in user behavior, or the creation of differentiated data that feeds back into product performance.

The former is hard. Users do not change easily. But those that do, those who form new habits, accept new workflows, or learn to trust machine judgment in previously human domains, can create lock-in that outlasts the initial feature advantage.

The latter may be harder still. Proprietary data, especially in enterprise or scientific contexts, is rarely just "found." It is cultivated. Reinforcement learning offers one path. Closed-loop systems, where usage refines prediction, offer another. But these are not shortcuts. They are commitments.

In both cases, what matters is the feedback. Not just from users, but from the system itself. A company that gets better as it grows, not only in revenue, but in precision, efficiency, and decision quality, is one that can survive the consolidation to come.

That is where the opportunity lies. Not in abundance, but in rarity. And not in what the market is excited about now, but in what it will still need when the illusion clears.

The temptation, especially if capital loosens again, will be to forget this. To believe once more that escape velocity is common and that scarcity can be scaled.

But it never has been, and it won’t be this time either.

For a century and a half, industrialization powered economic expansion by transforming energy into physical labor. Steam, electricity, and combustion scaled muscle and motion far beyond what humans or animals could supply. The correlation between energy consumption and GDP was nearly one-to-one; more watts meant more output.

Then, around 1970, the relationship broke. Growth in the West slowed, productivity gains narrowed, and per capita energy use flatlined. We entered what I hope to have been an interregnum: a fifty-year pause in the great arc of industrial progress. In its place came services, software, and finance. Important industries, but never delivering the kind of growth that industrialization had once made routine.

The Decoupling of Energy and Growth

Take the United States. Today, the average American consumes about 290 million BTUs of primary energy per year, roughly equivalent to a barrel of oil a week, or about 97 human-days of physical labor per day. But about 60% goes to heating, transport, and residential electricity. Only a fraction, 10 to 20% at most, supports knowledge work, even though such work constitutes nearly two-thirds of value-added GDP.

At first glance, this seems like a triumph of efficiency. We use little energy to produce the majority of value. But knowledge work, as currently organized, is not a growth engine. Its productivity gains have been incremental, not transformational. It does not scale like factories scaled; it does not turn power into output.

That shift, from adding workers to making workers marginally more productive, is the essence of the post-1970 growth slowdown. Industrialization brought whole new layers of labor into being. The post-industrial economy has merely squeezed more out of what we already had.

Why Industrialization Stalled and why we should care

Why did industrialization stall? Standard accounts point to global labor competition, oil shocks, environmental regulation, and the rise of digital services. These factors all matter. But there may be a deeper structural dynamic at work. As physical systems scale, accelerating them demands disproportionately greater innovation. Like aerodynamic drag, progress in the world of atoms becomes harder the faster we go. By the 1970s, we may have exhausted the available leverage without a new engine.

Should we care? Growth, after all, is a contested virtue. But there are three strong reasons to take it seriously. Philosophically, living systems export entropy. Decline often masquerades as efficiency. Historically, growth correlates with everything we associate with progress: lifespan, literacy, abundance. Practically, our institutions were built for a growth regime. Pensions, entitlements, and debt presuppose expansion. They may not survive its absence.

A New Engine: AI as Cognitive Labor

AI promises to reverse the slowdown in growth. It is not merely a tool for augmenting knowledge work, but a means of performing it. This places it closer to the general-purpose technologies of the early industrial era, like steam and electricity, which transformed the productive base of the economy by turning energy into labor. In AI we find the modern equivalent: a machine for converting energy into cognition.

The naive image is of AI agents replacing doctors or lawyers. But that is just the surface. In the first industrial revolution we started by imitating horses and ended up reshaping chemistry. From Stephenson's Rocket to the Haber-Bosch process, the arc bent toward the unimaginable. Fertilizer as a direct descendant of the steam engine would have seemed absurd in 1830, but the high pressures of the boiler birthed the reactor.

There is also an evolutionary parallel. Living things did not evolve to run fast, lift heavy, or think clearly; they evolved to survive and reproduce in a cold and indifferent world. That means making tradeoffs. Biological systems are full of compromises and evolution favors good-enough strategies, not optimal performance. Machines are not so constrained. It is no surprise we reached artificial superspeed with the locomotive: the iron horse did not need to rest, eat, or survive predators.

Intelligence follows the same logic. As David Deutsch and others have argued, intelligence is not the only thing our brains do. In fact, it is only a subset of our neural function, and not necessarily the dominant one. The cognitive illusions explored by Kahneman and Tversky are not evidence of failure but of adaptation: humans are not perfectly rational because perfect rationality is not an optimal evolutionary strategy. Heuristics, bias, emotion, these are tools for surviving, not solving.

So we should expect to reach artificial superintelligence quickly, as we did with locomotion, but by trading off, not copying. The Bitter Lesson teaches us that generality arises not from modeling ourselves but from letting machines discover their own solutions. We see this already in reinforcement learning: the most powerful systems do not mimic humans; they outperform them by doing what works, not what feels familiar. The path to AI performing knowledge work is not automation of the familiar, but the invention of the unfamiliar.

Why Software Came First

Today, the leading edge of this transition is software. Software development is one of the earliest domains to see mass adoption of AI tools. At Microsoft and several advanced labs, more than half of new code is now generated by AI systems. This is often explained by technical factors: code is a highly structured language, rich in pattern, and its "in silico" nature allows for instant testing and iteration.

These are valid points. But the deeper explanation is historical. Just as early steam pioneers emerged from mining communities, where steam was already in use and where moving heavy loads was a constant problem, so too have today's AI pioneers emerged from environments steeped in software. They are technically sophisticated and produce vast quantities of code. They turn to AI for help not because code is uniquely tractable for machines, but because it is a product they know intimately and can immediately use.

This suggests a likely pattern for what comes next. Fields with similarly technical cultures and clear feedback loops, such as the "money sciences" of drug discovery and materials design, are already moving quickly. The experimental method gives them a natural framework for validating AI-generated outputs. And in conversations with leaders across professional services, especially in law and finance, a striking theme emerges: many believe the work of junior staff, the basic synthesis, drafting, and pattern-matching, is ripe for AI.

From Software to Specialization

Today, we deploy AI like software: automating tasks, parsing text, writing code. But these are merely the early rails. As systems grow more agentic, more modular, more composable, they will begin to do for knowledge what the assembly line did for goods: not just accelerate production, but enable new categories altogether.

Specialization will follow. Just as steam engines begat locomotives, pumps, and generators, so too will transformer models fork into bespoke synthetic researchers, planners, negotiators, and designers. This specialization will inevitably blend into the industrial economy, reinvigorating it and perhaps overcoming the innovation drag that stalled growth in the 1970s.

And as the cost of thought falls, the returns to energy rise.

Power as the Strategic Input

This makes power the crucial input once again, specifically electrical power. Every AI agent is a conversion function: energy in, cognition out. As that conversion improves, the importance of generation, transmission, and storage will also grow.

In that light, investing in power infrastructure is no longer just about keeping the lights on. It is the upstream leverage point for economic strategy, the missing link between compute, cognition, and growth. The great interruption of 1970–2020 may soon appear as a brief pause, a time when we learned to conserve rather than to create. But energy is once again becoming labor. And the growth engine is starting to turn.

In a previous piece, I laid out why vertical AI is the most promising path to value creation in the post-foundation model era. Foundation models have leveled the cost of raw intelligence, but intelligence alone does not build companies. In vertical AI the most durable moats will emerge from experience: from agents embedded in real workflows, capturing data that no one else sees, and learning from it faster than anyone else.

This is where the Bitter Lesson meets the Era of Experience. Sutton taught us that generality emerges from learning, not from hand-coded knowledge. But to learn agents need data. And to learn fast they need feedback. In vertical AI, the companies that win will be those that convert customer interaction into experience traces, cycle them through simulation, and return to production with sharper policies in less time.

This essay tries to be more tactical:

It maps the two dominant go-to-market paths (tech-out and customer-in) and the way they converge; and

It introduces a new metric stack for tracking whether a company is really in the Era of Experience or just relabeling automation with AI.

I doubt these metrics will be ‘right’; they are more about trying to move away from using traditional Saas engagement metrics or foundational model performance metrics in a vertical AI space that rewards velocity above all.

Intelligence Is Table Stakes. Motion Is the Moat

The rise of foundation models has standardized raw intelligence as a platform capability. Call an API, get a baseline level of performance: search, summarize, interpret, and predict. Every product can now embed "smartness" as a default. But this flood of intelligence has not flattened the field. It has intensified the race for motion.

What will separate vertical AI winners is not how smart their model is at rest, but how fast it sharpens with use. The next great moats will form around experience loops: companies that accumulate proprietary traces, train on them in high-fidelity environments, and improve their models in ways others cannot match.

Experience Traces and the Feedback Flywheel

An experience trace is the raw evidence of interaction: what the agent saw, what it did, and how the world responded. These traces become the input to simulation environments, controlled spaces where models are retrained, tested, and refined.

The creates a loop of: customer interaction → trace → simulation → updated policy → redeployment. Each cycle should compound. Over time, the system becomes faster, more confident, and more autonomous.

This is what takes vertical AI beyond RPA-on-steroids. The difference is not task execution so much as adaptive performance. Companies that simulate their own agents become not just software vendors but learning organisms.

The Two Archetypes: Tech-Out vs Customer-In

Most startups follow one of two foundational go-to-market paths. These archetypes, Technology-Out and Customer-In, shape everything from product architecture to capital intensity to the slope of the experience curve. For vertical AI:

A. Technology-Out: Building from First Principles

Technology-Out startups begin by building general-purpose agents, simulators, or architectures without committing immediately to a specific customer problem. Their founders are often researchers or platform engineers, convinced that a core capability (self-improving agents, world-modeling infrastructure, or multi-modal execution) will eventually unlock large markets. Examples include firms like P-1, which is betting on a simulation-native architecture for agent learning, Mechanize, which is building agentic devtools from the ground up, or Lila’s scientific super intelligence.

This approach is expensive and often slow. Without customer constraints the product may drift or chase benchmarks that fail to translate to commercial traction. Yet when it works the upside is enormous. These companies are positioned to enter multiple verticals, reuse the same substrate across contexts, and build compounding moats through infrastructure reuse and data network effects.

The danger is clear: build a brilliant engine that never leaves the lab, but the potential upside enormous.

B. Customer-In: Learning Through Embedded Workflows

Customer-In startups begin by embedding themselves in a specific vertical or workflow, such as automating research for strategy teams, drafting contracts for lawyers, or summarizing clinical encounters for doctors. These companies get to real users quickly, gather high-fidelity experience traces from day one, and begin the improvement loop through real-world exposure.

This approach is leaner, faster, and usually generates revenue sooner. But the risk is that the model becomes too entangled with a narrow use case. Optimization occurs within the walls of one job-to-be-done, and extensions into new workflows require painful architectural changes or retraining.

In short: you get signal fast, but you risk becoming a point solution.

C. The Strategic Tradeoff—and the Path to Convergence

The essential tradeoff is between efficiency and extensibility. Customer-In is efficient, but may entrench you. Tech-Out is extensible, but may strand you. Most companies will eventually converge: either the Tech-Out firm acquires a flagship vertical, or the Customer-In firm starts investing in horizontal substrate once its wedge is secure.

The dream is of course Customer-In verticals with substrate potential: solving a clear immediate problem that gets you to revenue quickly, but that also give you access to data that is broad and interesting enough to build generally applicable simulations and get to data network effects. That is one of the reasons that healthcare and pharma use cases seem on their face to attractive for vertical AI plays, and why compliance use cases may also be high potential, not so much for the size of the compliance market itself but as a way to get to a train on business critical data, and data which by definition needs to exist in similar form across multiple companies in an industry.

The Metric Stack: Recognizing the Era of Experience

Traditional SaaS metrics such as ARR, active users, and retention fail to capture the dynamics of learning-based systems. In the Era of Experience, at least at these early stages, what matters is less the level of performance than the rate of improvement. The companies to watch are not those with the largest customer base, but those where every interaction feeds a compounding loop.

To track this, we need a new stack of slope metrics:

Trace Accumulation

Are you generating proprietary traces faster than your competitors?

Are those traces growing in diversity and depth, not just volume?

Automation Slope

What share of a given workflow is handled autonomously and how is that share growing?

How long does it take to reach a threshold of autonomy with each new customer?

Learning Velocity

How quickly do new traces lead to observable improvements in behavior?

Are policy updates accelerating over time across tasks, cohorts, or edge cases?

Simulation capacity multiplies all three. The richer your synthetic environment, the faster you can improve without waiting for the world to generate the data you need These are the signs that a company is not just automating, but learning and so operating a flywheel.

Strategic Prompts for Builders and Backers

For founders: What experience traces are you capturing that others cannot? Can you simulate the world your agents act in? Are you building a point solution, or a general substrate disguised as one?

For investors: Is the company’s learning velocity accelerating? Are traces being recycled into better policies? Is there early evidence of generalization beyond the initial wedge?

For incumbents: Can you retrofit your org to capture experience and simulate improvement? Or will vertical-native upstarts learn faster, automate more, and gradually hollow out your core?

Conclusion: From Foundation to Flywheel

The age of models is becoming the age of motion. Foundation models gave every company access to intelligence on demand. But only some will learn how to sharpen it through interaction.

The winners will not be those who deploy the largest model. They will be those who move fastest through experience, build better agents through simulation, and compound learning into dominance.

In the Era of Experience, who learns fastest wins.

We are living through a narrative breakdown. The neoliberal order that defined the post–Cold War world, open markets, global capital, liberal immigration, and U.S.-led multilateralism, has lost its grip. Three ruptures mark the moment:

· Tech has replaced finance, insurance, and real estate (FIRE) as the primary source of economic power.

· The U.S. is actively dismantling the international rules-based system it once enforced.

· The core tenets of neoliberalism, especially on immigration, climate, and globalization are being questioned across the political spectrum.

This is more than a geopolitical pivot. It’s a collapse of the model that explained the world and in its place, new narratives are struggling to emerge. And if we want to see where those narratives are forming we cannot look to the old institutions. Instead we must look to Substack, Twitter, and the markets where knowledge now leaks and coagulates in real time.

When Models Become More Useful as They Become Wrong

We often say that all models are wrong, but some are useful. Yet some models, especially governing narratives like neoliberalism, become more useful as they become more wrong, because usefulness isn’t just about accuracy, it is about coordination.

When a model becomes embedded in institutions it gains inertia. It provides a shared language, a mental map, and a rationale for policy and power. Its defenders believe in it and are often professionally, politically, and financially invested in maintaining it.

Model decay follows two common paths:

· Facts change, but the model doesn’t. For example, the rules-based international order made sense when the U.S. was an energy-independent, industrial exporter and global military hegemon. By the 1970s, America was energy-dependent, facing industrial rivals, and turning inward, yet the Cold War frameworks endured.

· The model is useful to elites but it drifts from the experience of the majority. Globalization, for instance, was long framed as an unambiguous good, even as it hollowed out Western industrial employment. The model persisted not because it described the world but because it justified it.

Today we’re past the petrodollar paradigm. The U.S. is energy-independent again. The industrial base remains hollowed, but the new driver is compute. Tech rather than oil, steel, or even high finance, anchors power. The old models no longer apply.

Why Traditional Institutions Can’t Keep Up

In past cycles we might have turned to mainstream academia, legacy media, or Wall Street research to help us make sense of change. But these institutions are themselves embedded in the collapsing narrative. They are, in effect, part of the problem.

Worse, their business models no longer support real-time, high-resolution sense-making:

· Academia struggles with incentive misalignment and is epistemically rigid.

· Mainstream media has lost authority and runs on outdated ad-driven dynamics.

· Bank research is now largely client-gated and less public-facing, and the rise of alts means much cutting edge work is deliberately kept secret.

They are incapable of updating and economically unable to adapt.

The Rise of Narrative Leak Platforms

In this vacuum, new platforms have emerged to fill the interpretive gap. Chief among them: Twitter, Substack, and Polymarket. Each offers a different piece of the new narrative infrastructure:

1. Knowledge Leaks - When the dominant model fails to track reality, experts outside the institutions begin leaking real insight. These are people with deep, domain-specific experience, engineers, geologists, commodity traders, who now have channels to reach a wide audience.

Example: Doomberg, the anonymous energy writer, has upended assumptions about U.S. energy needs and grid demand. Insights that once lived in paywalled consultancies or trade publications are now public.

2. Narrative Entrepreneurs - In periods of breakdown, new narratives emerge from the fringe. These entrepreneurs often speak in speculative or extreme terms, but they serve a purpose: to question what had been unsayable.

Curtis Yarvin, for instance, offers a radical critique of liberal democracy that resonates in tech circles, not because it’s widely accepted, but because he has the credentials and daring to say what others won’t. From the left MMT advocates or rationalists have played a similar role in challenging economic orthodoxy.

3. Market Signals of Narrative Collapse - Polymarket is the most interesting development here. It represents Hayekian knowledge: decentralized consensus compressed into price. When Polymarket is “right” and traditional forecasts are “wrong,” it doesn’t just suggest informational failure. It signals narrative failure: a large group of people not only disagrees with the consensus, but believes the divergence is actionable.

These platforms have to some degree aggregated, and importantly provided and economic model for, what was previously dispersed among blogs and message-boards. Podcasts, too, play a significant role, though they do not yet have a similar aggregating platform, and Twitter and Substack play an important role in surfacing new and niche pods.

Much tends to be made of closed groups on telegram, Discord, whatsapp, etc. This is mostly because of the glamour of secrecy rather than importance. The principle that ‘if it is on TV, it is a TV show’ holds here: in the attention economy an idea has value once it starts to gain attention, so it is of limited value to try and look upstream of public platforms.

Post-Truth Is a Symptom, Not a Diagnosis

It is common to wring hands over our "post-truth" era. But truth is not the core issue, consensus is. We are not post-truth; we are post-model. The narratives that once structured our understanding have decayed. No shared map means no shared reality.

When sense-making collapses, moral language rushes in to fill the void. It becomes easier to condemn than to interpret. But beneath the outrage is disorientation: the old frameworks don’t work, and no replacement has achieved legitimacy.

Take Jamie Dimon's recent remarks: he linked immigration to stagnant wages in the bottom 20%. Ten years ago there would have been a chorus of op-eds explaining why Dimon was wrong, five years ago he would have been cancelled. Today the statements attract attention but little more. That shift tells us something fundamental has changed.

Narrative Regime Change and the Investment Horizon

There is no consensus replacement yet. The tech/right narrative appears ascendant, but is far from dominant. It may yet be overtaken by a progressive alternative, perhaps something like the “abundance agenda.” But the establishment is unlikely to produce it. The break down is too profound and the almost all establishment institutions too tied to the old model. What we are watching is narrative regime change. The traditional sense-making class, legacy media, academia, even mainstream think tanks, has been sidelined. The new narratives are being forged in real time, on messy, entrepreneurial platforms.

This matters for investors. Getting the petrodollar story right gave you a 50-year investment tailwind. I doubt the next model will last as long but it will be just as powerful while it holds.

My bet is on the technodollar: a regime where compute, not capital or commodities, becomes the organizing principle of global power. But what matters now is knowing where to watch: the leaks, the prophets, the price signals. In a collapsing model landscape, these are our compasses.

What Makes Vertical AI Compound

We are entering a phase where every enterprise wants to put AI to work—not in general, but in context. Not to summarize or autocomplete, but to reason through regulatory change, assess operational risk, guide investment planning, etc. That is the promise of vertical AI: systems built not simply with domain data, but for domain reasoning.

The typical vertical AI startup begins with a familiar idea: gather proprietary data, fine-tune a model, and let scale drive improvement. More customers bring more data. More data strengthens the model. A stronger model produces a more capable product. The loop should compound.

In practice, the system often flattens early. The model ceases to improve in meaningful ways. Outputs begin to converge, shaped less by learning than by templates that already reflect its limits. Each new deployment adds surface area, but not depth. The system expands its reach without increasing its capacity for reasoning. Over time, what was intended to be a learning engine settles into a static interface. The product stops improving, customers notice, and eventually churn.

This is because vertical AI systems do not improve because they collect data. They improve because they are designed to learn from use. That requires feedback, structure, and trust. Not as separate components, but as interdependent systems that shape how the product evolves.

The strongest products in this space will be those whose decisions are trusted to carry weight, and whose performance improves with every interaction. That improvement comes through two loops: a learning loop, where agents refine their reasoning through structured feedback, and a trust loop, where usefulness earns permission to take on increasingly important tasks. One loop compounds technical capability. The other compounds institutional acceptance.

Not every problem requires this. Domains such as logistics routing, risk scoring, and dynamic pricing reward data access and statistical learning. Others may be won through distribution or compliance. But in domains where ambiguity matters, where judgment is at the center of the task, these loops will define whether the system compounds or simply accretes.

Much of the below is based on trying to understand the implications of the bitter lesson and the era of experience. And so are part of my continuing series ‘this investor generated superior risk adjusted returns with one simple trick’, looking to find that one paper that answers it all.

The Learning Loop

In the most serious domains such as law, medicine, or finance there is a baseline level of data required for a model to function. One needs access to precedent, records, and the structural logic of the field. Below that threshold, the system cannot perform. But beyond it, simply adding more data does not reliably lead to improved reasoning.

For a model to improve, it needs more than exposure. It needs a structured environment in which its actions carry consequence, where it can take steps, observe results, and adjust accordingly. This is what we mean by a world model.

This is not a schema, or a database, or a graph. It is a structured representation of the domain; its entities, relationships, causal structures, and potential failure modes. It is the terrain through which an agent learns to navigate. It resembles what Palantir describes as an ontology but applied beyond a customer’s data, ideally to a whole industry or market.

The goal is not to reproduce the expert’s workflow. It is to solve the underlying problem. The agent does not learn by imitation. It learns by doing. It explores the world model, tests its reasoning, encounters friction, and refines its policy accordingly.

That is what creates the learning loop:

A better world model allows for smarter actions
Smarter actions generate clearer feedback
Feedback reshapes the model
The system improves

What compounds is not the amount of data. It is the structure that allows the model to learn from use.

The Trust Loop

Even when a system is capable of learning, it rarely begins with permission to act. In high-stakes domains decision-making is gated and trust must be earned.

That process does not begin with autonomy. It begins with constrained responsibility; tasks that require reasoning, sit close to judgment, but remain insulated from operational risk. This is what makes deep research such a common starting point. Almost every major lab has led with some version of it, not because research is easy or marginal, but because it is a broadly applicable tool that allows the system to demonstrate value without requiring trust in execution. It allows the model to prove its ability to reason long before it is asked to decide.

The same pattern holds across verticals. The challenge is not to automate the core decision on day one. The challenge is to identify a narrow, high-context task, close to the center of the judgment loop but shielded from consequence, and use that as a foothold. If the system performs well, it earns the right to move upward: from synthesis to recommendation, from drafting to action.

This is how the trust loop begins to form. Performance earns confidence. Confidence opens the door to higher-value tasks. Those tasks, in turn, generate better feedback, which sharpens the model further. Over time, what begins as a research assistant becomes a decision-making partner.

Why SaaS Metrics Do Not Apply

Some problems will continue to be solved by traditional ML workflows, especially those that live inside closed loops with abundant data and clear outcomes. These domains reward prediction and pattern recognition. They do not require structured reasoning, nor do they depend on trust.

But the most high value domains operate differently: they are ambiguous; they rely on institutional context; and they demand interpretation, not just recall.

SaaS products were built for another kind of problem. Their defensibility rests not on how much better they get with use, but on how deeply they are embedded in process. They succeed by capturing workflows, structuring data, and becoming operationally central. Integration becomes inertia. And inertia becomes retention.

Vertical AI systems are different. They do not aim to become indispensable by owning the workflow. They aim to become indispensable by exercising good judgment. Over time, the system is trusted with more decisions, first at the edge, then at the core. And as that trust accumulates, the product becomes harder to remove. Not because it holds the data. Because it holds more of the customer’s reasoning.

What matters most is delegation. The measure of success is not how much of the system is used, but how much of the outcome it shapes.

This shift is visible in the architecture:

[ Data Layer ]

- Centralized infrastructure (Snowflake, Databricks)

↓

[ World Model/ontology ]

- Structured domain representation built to support learning

↓

[ Agent Layer ]

- Adaptive systems trained through interaction, not imitation

↓

[ Application Interface ]

- Where judgment is surfaced and trust is earned

The strongest vertical AI systems will not be the ones that manage the most workflows. They will be the ones whose outputs carry the most weight because they have earned the right to reason on the customer’s behalf.

A Playbook

If learning and trust are the conditions under which vertical AI compounds, then building in this space requires a different orientation, one that treats product, data, and customer relationships as part of the same architecture.

1. Identify modelable problems - Focus on areas with structure, variation, and consequence. Look for reasoning loops that sit upstream of important decisions, even if they do not look like traditional workflows.

2. Build a world model - Start from what is minimally necessary to structure the domain. Prioritize clarity over completeness. The model should support exploration, not just indexing.

3. Deploy at the boundary of action - Enter through reasoning tasks that matter but do not carry risk. Let the model demonstrate capability in low-friction contexts.

4. Climb the decision stack - Use performance to earn permission. Use permission to unlock feedback. Let each level improve the system’s ability to reason through the next.

5. Track what matters - Do not measure integration. Measure influence:

(a) How much judgment has been delegated

(b) How visible improvement is to the user

(c) How central the system has become to decision-making

What matters is not how often the system is used. It is how much of the outcome the system shapes.

The Two Loops

Vertical AI will compound because it learns. That learning depends on two interlocking systems.

The first is the learning loop. It relies on structure. It turns action into feedback, and feedback into progress. The second is the trust loop. It relies on performance. It turns usefulness into permission, and permission into responsibility.

Together, they determine whether the system remains a tool or becomes a decision-maker.

The strongest vertical AI companies will not be those with the most integrations
They will be those whose outputs carry weight and whose performance improves with every use.

Europe’s relative tech underperformance is often explained in terms of culture, regulation, or capital. These are symptoms, not causes. Regulation is uneven but not uniquely burdensome. Capital is abundant, perhaps disproportionately so considering the size of Europe’s native tech sector. And in my experience there is very little difference in the work ethic or risk appetite of the best European and U.S. startups.

What’s missing is not effort, talent, or money. It is the scaled platforms that make an ecosystem compound over time.

This missing industrial layer now matters more than ever. The rise of AI is not just a new product category. It’s an architectural shift in the global economy. And it intersects directly with Europe’s traditional strengths: automotive, pharma, chemicals, and supply chain and logistics. The age of the technodollar is not just about tech dominance, it is about industrial transformation. Europe needs an industrial policy built around energy, fostering AI powered industrial platforms, and attracting (or retaining) the people needed to run them.

The Missing Middle: Platforms that Compound

Innovation economies rest on three pillars. Universities advance the frontier of knowledge. Startups absorb new tools and mutate them into fresh business models. But between them lies a critical middle layer: scaled firms that turn insight into infrastructure.

This layer does two things. First, it performs the heavy lifting of systematization. As W. Brian Arthur noted, new technologies do not change economies through invention alone. They must be embedded: converted into platforms, products, and operational standards that people can build on. Second, scaled firms become schools of scale for executives, advisors, and investors, who can take these lessons into the next generation of startups.

Firms like Amazon, Google, and Nvidia don’t just dominate markets, they become infrastructure for new innovation. Their platforms offer APIs, capital, and talent networks that lower the cost of building adjacent ventures.

Europe excels at the bookends. Its universities remain world-leading. It has no shortage of ambitious founders and has produced high-quality firms such as ASML, SAP, Spotify, and Adyen. But these firms do not function as platforms in the North American sense. They are tightly bounded, optimized for excellence within vertical domains, and generally lack ecosystem-wide spillover. They generate value but not gravity.

The result is a structural absence at the heart of Europe’s innovation economy. Without scaled tech incumbents, there is no machinery to absorb frontier science, no internal labor market to circulate talent, no recursive loop through which knowledge compounds. Europe becomes a source of ideas and a site of consumption but not the ground on which the next industrial regime is assembled.

Deepmind fits awkwardly into this mix as it is now a key part of Google’s scaled R&D efforts but began as a startup not a lab. But it also shows where Europe’s issues are and aren’t: Deepmind had no problem raising money from global investors; but when it needed to get hold of truly scaled compute and capital West Coast US tech giants were the only answer. It is unlikely that Deepmind could have achieved its various breakthroughs as a standalone entity, and unthinkable as part of Siemans or SAP.

Why the Gap Exists: The Industrial Policy Mistake

Marios Draghi’s report on European competitiveness highlights tech as a key difference between European and US productivity. But Draghi’s framing, comparing Europe to the U.S., misses the more instructive contrast: China. The divergence between the two is a live A/B test in industrial strategy.

China responded to the emergence of the digital economy with strategic intent. It pursued a recognizably Listian approach to industrial policy. The Great Firewall was also a digital tariff wall, insulating local platforms as they scaled. At the same time, export-oriented subsidies flowed to firms like Alibaba, Tencent, and Huawei, enabling them to build capacity and eventually compete abroad.

Europe, in contrast, maintained an open market posture but failed to build internal defences. Industrial policy was not simply neglected, it was actively discredited, viewed as incompatible with the principles of an efficient, liberalized economy. Meanwhile, competition policy, guided by a narrow focus on consumer pricing and choice rather than national or bloc development, became a barrier to the strategic consolidation or nurturing of domestic champions. In practice this meant that while Europe welcomed global tech entrants it too often failed to cultivate its own.

Pride played a role too. Europe, especially in the peak neoliberal moment between German reunification and the 2008 financial crash, saw itself at the very least as a peer to the U.S. rather than an emerging economy in need of protection. Indeed, for much of the time it appeared as if we had achieved some kind of alchemy: Germany, France, and the UK in particular converging on U.S. GDP per capita while maintaining a more humane form of capitalism.

All this, and the global underestimation of the digital economy in the wake of the dotcom crash, led to an orientation toward legacy strength rather than developing and protecting new domains.

The Technodollar Moment

The technodollar marks a shift in the basis of U.S. global leverage: from finance and force to compute and code. It buys influence over industrial standards, talent flows, and supply chain dependencies. Compute is the new chokepoint.

For Europe, the implications are twofold. On the supply side, the United States and China are now vertically integrating AI from end to end, owning the chips, the models, the infrastructure, and the distribution layers. Europe has no sovereign stake in any of these domains. It is dependent on foreign supply chains for the very tools that will determine the pace and shape of industrial renewal.

On the demand side, European industry is facing direct competition from AI-enabled challengers. This is most pronounced in automotive but similar dynamics are unfolding in chemicals, pharma, logistics, and beyond. AI is not simply augmenting these fields, it is redefining them, and it is doing so largely outside of European control.

Even Europe’s intangible champions, luxury goods and finance, are at risk. Prestige economies don’t float on sentiment alone. Without the ballast of industrial leadership, brand and talent eventually drift elsewhere.

The quandary of the German car industry, caught between Tesla and BYD, or BASF, forced to significantly relocate to China in search of affordable energy, may be coming to all of Europe’s industrial champions.

Let Go of the AI Powerhouse Fantasy

The impulse to compete head-on is understandable but self-defeating. Europe will not win a race in general-purpose AI. The stack is too deep, the capital too entrenched, and the geopolitics too locked in.

Yet many European leaders still speak in slogans: “AI sovereignty,” “AI powerhouse,” “the next Silicon Valley.” These ambitions are not just unrealistic, they misdirect attention and resources.

Europe must reframe the goal. Not AI at the frontier, but applied AI in sectors where Europe already leads. The next industrial champions won’t be those who train models, but those who wield them to reshape existing value chains and create new ones.

A Strategic Response

To build industrial tech strength, Europe needs a coordinated strategy across three axes: energy, platforms, and people.

Energy is foundational. AI consumes immense computational power, and the infrastructures it enables. robotics, semiconductor fabs, automated supply chains, etc, are power-hungry. Clearly climate ambitions must be balanced against industrial requirements. But power supply is actually a bigger question: it will not be possible to compete in an electrified economic world when close to two-thirds of our power comes from imported fossil fuels.

Second, Europe must re-embrace the logic of infant industry protection, though not in the traditional sense. It should target sectors where it retains latent strength: pharmaceuticals, materials science, chemicals, auto, and supply chains. Within these, the priority must be to develop new platforms that apply AI to real industrial problems: molecule simulation, logistics optimization, circular manufacturing, etc. The tools of economic statecraft such as targeted subsidies, strategic procurement, mission-oriented R&D must be directed toward challenger firms. The next industrial champion may already exist, buried inside a legacy corporate, or latent in a university lab. But it will need help to emerge.

Finally, Europe must reckon with its human capital dilemma. Too many of its best operators build elsewhere. It is not enough to tweak tax codes or offer golden visas. What’s needed is a mission-driven talent strategy. One that actively recruits experienced builders and embeds them in strategically important sectors. European governments should be on the hunt for Morris Changs and willing to support them to build TSMCs from automotives to chemicals. Taiwan built an industrial giant by backing one visionary with the right tools and mandate. Europe must do the same and at greater scale.

Conclusion: The Clock Is Ticking

Europe missed the digital wave. It cannot afford to miss the AI-industrial one. The technodollar shift is a window, and, if it acts now, Europe can anchor itself in the next phase of global industrial history. Not as a peer to Silicon Valley, but as a sovereign producer in applied AI.

In 1945, as Europe began to rebuild and planners debated the shape of the postwar economy, Friedrich Hayek published The Use of Knowledge in Society. Hayek claimed that no central planner, indeed no one at all, could ever know enough to direct an economy. Knowledge was distributed, embedded in local experience, intuition, and shifting circumstance. Prices were not just signals of value but compressed representations of scattered knowledge.

What made Hayek’s article powerful was not just its defense of markets but its recognition of uncertainty, of how little even the smartest can truly see. But it also left a question: if prices reflect distributed knowledge, can we build systems that interpret that knowledge more deeply?

This question is not just philosophical. In the early 2010s, two of the world’s most sophisticated real asset investors, Blackstone and Brookfield, made large, opposing bets on the future of American malls. Brookfield believed malls would retain their value as community anchors. Blackstone, reading the same market signals, reached a different conclusion: warehouse demand, e-commerce acceleration, and last-mile logistics were pointing to a structural shift. The difference was about interpretation, not information. Blackstone had a sharper internal model of how the world was changing.

Below I go from Hayek’s insight into knowledge and prices, to how complexity economics helps us move beyond linear prediction toward causal reasoning in dynamic systems, and on to how artificial intelligence will lower the cost of running simulations, testing assumptions, and refining institutional judgment. The next frontier is productized world models: systems that help firms, investors, and public institutions understand why things are changing and act accordingly at scale.

The edge will no longer be in access or speed. It will be in the quality of the model.

The Hayekian Constraint

Hayek’s theory of prices is a theory of knowledge and its limits. He challenged the central conceit of socialist planning: that an economy could be managed like a machine if only the right levers were known. He claimed that in a modern economy, knowledge is dispersed, local, tacit, often unspoken. No single actor can grasp it all.

What makes coordination possible is the price system. Prices compress scattered signals into a single figure. They don’t explain what’s happening but help people act as if they understand enough to decide. When the price of tin rises, a buyer doesn’t need to know whether a mine collapsed in South America or demand spiked in China. The price alone is enough to shift behavior. Prices enable cooperation without comprehension.

The Limits of Price Alone

But comprehension, of course, matters. Knowing why a price is moving can be essential and while prices are visible meaning is often not.

This distinction matters even more in environments where regimes shift, where old signals mislead, and where small differences in interpretation lead to large outcomes. The most successful operators are often not those who react fastest but who interpret most effectively.

Interpretation is not art or science. It is a cultivated capacity. Within an institution it emerges from experience, narrative, analytics, and feedback. It’s built through repetition, pattern recognition, and shared sensemaking. Over time it hardens into an internal model; an evolving framework for understanding what matters, how change propagates, and how to act under uncertainty.

These models rarely have formal names. They live in memos, plans, and conversations. But they shape how the institution thinks.

From Judgment to Structure

Insights from these informal models are often presented in shorthand ‘We knew e-commerce was accelerating: we saw it in warehouse leasing’. But behind such remarks lie deeper, slower-moving judgments. These judgments form the scaffolding of what might be called a world model: a shared internal logic about how events connect, how systems behave, and where leverage lies.

The difference between Blackstone and Brookfield during the shift in retail real estate is illustrative. Both firms had deep access to information. Both had teams of experienced professionals. Both understood the headline trend: e-commerce was growing. But only Blackstone reorientated its portfolio aggressively toward warehouses, logistics, and last-mile distribution.

Blackstone’s strength in that moment wasn’t informational advantage so much as interpretive edge. Their internal understanding of how retail demand, infrastructure, and digital commerce were interacting, and how that interaction would reshape asset valuations, gave them conviction to act before the consensus had caught up.

These kinds of judgments are rarely the product of a single insight. More often, they emerge from an evolving internal model, one that may not be formalized, but which shapes how signals are perceived, debated, and operationalized. Over time, the best institutions find ways to structure these judgments: to embed them in workflows, to communicate them across teams, and to update them as conditions change.

Complexity Economics and Institutional Cognition

For much of the 20th century, mainstream economic theory treated the economy as a system tending toward equilibrium: predictable, legible, and governed by rational agents. The models that emerged from this view often assumed away the very uncertainty and feedback that real institutions contend with daily.

But over the past two decades a different intellectual lineage has gained ground. Complexity economics begins with a different premise, that the economy is a complex adaptive system: one in which agents learn, adapt, and interact in ways that produce emergent, often unpredictable outcomes. In this view, behavior is not governed by equilibrium but shaped by history, structure, and feedback.

Markets are more like ecosystems than machines. Investment environments evolve rather than returning to baseline. In such settings the goal is not perfect prediction but adaptive orientation. Understanding how change might unfold, what tipping points may be near, and how local decisions interact with global effects.

The tools of complexity, agent-based modeling, system dynamics, scenario analysis, offer a way to explore these possibilities. For institutions that have already internalized the need to “see around corners,” these tools offer formal scaffolding for what was once intuition alone.

Modelling the World at Scale

What Blackstone and old-school macro speculators embodied in company culture and philosophy some of the largest tech companies have embodied in systems.

Amazon offers perhaps the clearest example. Its forecasting, fulfillment, and delivery systems are guided by models that continuously absorb new data and re-estimate demand. These are used to make real decisions in real time and at their best allow the company to operate as if it understands what is coming next.

Tesla works similarly. Its supply chain planning, manufacturing scheduling, and autonomy development are driven by internal expectations about regulation, market demand, hardware capabilities, and software learning curves. These expectations are not fixed and their model is never static.

What distinguishes these firms is not perfection (Amazon’s modelling famously overestimated the stickiness of pandemic demand and Blackstone has made plenty of bad investments). It is the presence of an institutional architecture that allows judgment to be expressed, tested, and revised continuously. These architectures are explicit representations of belief; structured enough to be queried, simulated, and debugged.

What AI Makes Possible

AI doesn’t eliminate uncertainty but it changes how institutions engage with it.

Where researchers once spent weeks building a scenario, AI now enables thousands of variations in minutes. Institutions can test assumptions, explore nonlinear effects, and reason across scenarios at scale.

AI lowers the cost of simulation. From power grids to capital markets, AI-powered models explore what happens under changing conditions—not to forecast exactly, but to generate insight: What matters most? What regime are we approaching? Where might nonlinearities appear? Good decisions are not based on isolated facts. They depend on context, how facts are connected and interpreted. AI helps construct that scaffolding: the relationships and causal structures that give signals meaning.

This is the foundation of a new kind of institutional capacity: systems that ingest data, compare evolving signals against internal models, update assumptions, and suggest adaptive responses. Not hard-coded rules, but flexible heuristics. A model always learning, always partial, always in dialogue with the world.

The goal is a Soros in a box—not a system that knows the future, but one that reasons about it, updates reflexively, and anticipates shifts by analogy and simulation.

Palantir and the Service Layer

If the earlier examples show how institutions can build their own models of the world, Palantir offers the scaffolding for others to do so.

Palantir sells a structure, a way of organizing data, mapping relationships, and reasoning across complexity. Its Foundry and Gotham platforms impose ontologies on institutional data, linking operations, assets, risks, and signals into a shared, navigable frame. They make it possible to simulate scenarios, test interventions, and identify second-order effects. They turn disconnected information into a usable model.

For Palantir’s customers this service above all about coherence. Enabling institutions to act in ways that are internally consistent, causally informed, and accountable over time: a world model as a service.

From Service to Product

Most institutional models today are still hand-built. They emerge from custom deployments, intensive consulting, and close coordination between software providers and domain experts. But that model doesn’t scale easily and it doesn’t generalize.

The next frontier is productized institutional modeling platforms that offer structured ways to interpret the world, out of the box, within defined domains. These systems don’t try to model everything. They focus on sectors where the stakes are high, the feedback is rich, and the causality can be usefully structured.

Some efforts begin technology-out: Google’s work on climate forecasting, supply chain risk, and weather modeling reflects this approach: large models refined until they can be deployed across markets. These are deep infrastructure bets.

Others go customer-in. Altana, for example, is building a living map of global trade and enforcement, helping firms and governments model the flow of goods, identify weak links, and navigate compliance. Crux is developing tools to simulate infrastructure risk and investment scenarios in energy, while Brightwave is helping institutions bring their hard won tacit knowledge to new deals in a systematic way.

This work is not just for ‘AI native’ businesses but for businesses with a deep understanding of, and dataset on, complex industries. Crunchbase, for example, is increasingly becoming a model of private tech markets, capable of making predictions down to the individual company level.

These are more than dashboards. They are systems that contextualize signals, test decisions, and suggest actions. And increasingly they will act on those decisions — triggering workflows, allocating resources, even making strategic adjustments without waiting for human input.

Investing in Interpretation

If prices are signals, and the best institutions are those that know how to interpret them, then the next frontier for investors, founders, and policymakers alike lies in building and backing systems that improve that interpretation.

This means moving beyond data and even beyond AI in the narrow sense to institutional models that are causally informed, dynamically updated, and operationally embedded. Systems that don’t just answer questions but help formulate better ones.

For investors, this opens up several clear paths:

(a) Vertical AI platforms that encode domain-specific logic: logistics, energy, trade, infrastructure, governance.

(b) Simulation and scenario engines that let institutions reason through complexity rather than optimize against oversimplified assumptions.

(c) Agentic systems that manage portfolios of strategies or policies and adjust in real time based on evolving context.

(d) Interpretation infrastructure for the state: civic software that gives governments the capacity to model their environments and respond reflexively in trade, enforcement, climate, or crisis response.

We are entering the age of institutional cognition. The institutions that learn how to build, refine, and act on models of the world will be the ones that shape it.

The petrodollar system is ending. For decades, oil priced in dollars and backed by U.S. security guarantees allowed America to sustain persistent trade deficits, dominate global capital markets, and maintain alignment across a network of energy exporters and surplus holders. But that architecture—reliant on military power projection, financial recycling, and industrial offshoring—has reached its limits.

The United States is beginning to organize a new system in its place. Influence through controlled access to the ‘stack’ that underpins artificial intelligence: compute, cloud infrastructure, models, and the applications that run on them. This system, still informal and in motion, is not built on security guarantees or reserve flows. It is being built on selective provisioning and state-backed industrial coordination.

It was already visible in deals like the 2024 U.S.-UAE chip clearance, where access to NVIDIA hardware was granted only after Abu Dhabi’s G42 cut ties with Chinese vendors, but became most apparent in May’s sweeping infrastructure and AI deals in Saudi and the Gulf. It is reinforced by the U.S. dominance in export-controlled chips, the gating of frontier models, and cloud region segmentation.

This emerging system is the technodollar. It is not yet a regime, but it is a strategy for broad economic and geopolitical alignment in the age of AI. A strategy with significant investment implications, with winners being the national champions of the various layers of the stack, protected, trusted and integrated into US policy.

1. The End of the Petrodollar Order

The petrodollar system was a response to the breakdown of Bretton Woods. After the gold window closed in 1971, the United States needed a new mechanism to stabilize the dollar and a way to fund deficits. That emerged through a set of bilateral understandings: oil would be priced in dollars, and the surplus revenue would be recycled into U.S. Treasuries. In exchange, the U.S. would provide military protection and access to its markets.

This arrangement proved remarkably durable. It allowed the U.S. to finance deficits, absorb global production, and underwrite the international order. But by the 2010s, the foundations had begun to crack.

First, the fiscal side deteriorated. Low interest rates and repeated stimulus—most notably after the COVID-19 shutdowns—exposed the limits of deficit finance, with national interest payments surpassing defence spending in 2024.

Second, strategic returns diminished. The U.S. military remains globally deployed, but its ability to compel behavior has declined. The withdrawal from Afghanistan, the limits of deterrence in Ukraine, and China’s rise across technology and trade all make the post-Cold War ‘unipolar’ security model look less credible.

Third, and most importantly, the technological basis of power has begun to shift. Influence is less about trade volume or military bases and more about control over the stack: chips, energy, data centers, and the models built on top.

American firms—particularly NVIDIA, Microsoft, Amazon, and OpenAI—control these inputs. And US policy is being organized around them. We are seeing not a new doctrine, but a series of deals, controls, and alignments that resemble a system.

2. The Technodollar as Strategic Pivot

The US government’s initial foray into geopolitical AI policy was in line with unipolar muscularity: in 2022 export limits on advanced chips explicitly intended to hold back Chinese AI research. By 2023, thought the United States had begun to shift its approach, offering selective access, provisioning AI capabilities under licensing terms and alignment conditions. ‘Provisioning’ is the theme of the technodollar: selectively granting access to critical capabilities like compute, models, or cloud infrastructure, under conditions of alignment, rather than through open markets.

The clearest early case was the UAE. Out in the cold in 2022 by 2024it could access NVIDIA’s most advanced chips but only after G42 (Abu Dhabi’s national champion) agreed to sever its partnerships with Chinese firms, including Huawei. The same year Microsoft committed to deploying OpenAI’s infrastructure through Azure in the Gulf, effectively bringing inference under U.S. compliance. In both cases, the principle was the same: access to the U.S. stack required strategic alignment.

Where the petrodollar system reinforced itself through recycling surplus, allowing producers to recycle dollars into U.S. capital markets, the technodollar reinforces itself through integration. Once a country builds its AI infrastructure on U.S. terms, the cost of exit will be high. Inference pipelines, cloud dependencies, and compliance layers are not easily replaced. Interoperability becomes lock-in.

3. How the Stack Holds Together

The technodollar system functions through a layered stack, each level governed by access, licensing, or service. Where the U.S. controls a layer directly, it sets the terms. Where it does not, it aligns adjacent layers to structure interdependence.

(i) Energy The US is the world’s largest LNG exporter and these exports are shaping the physical geography of AI infrastructure. In countries like Vietnam and Poland, long-term LNG deals are being bundled with investment in data centers. These arrangements make energy a platform—not just for electricity, but for alignment. In energy rich regions like the Gulf the US is looking to partner in physical infrastructure and of course control the digital infrastructure above.

(ii) Cloud InfrastructureCloud platforms like Azure and AWS function as jurisdictional control systems. Sovereign cloud zones embed policy into data residency, latency, and access. To build on these platforms is to inherit their governance.

(iii) Compute NVIDIA’s hardware—especially the H100 and forthcoming B100 chips—is a tightly gated resource. Access is regulated by the U.S. Commerce Department. Exports to China have been blocked, while access for allies is conditional.

(iv) Model providers operate under implicit governance. Access is mediated through APIs, compliant cloud platforms, and usage controls. The result is not just a product ecosystem but a compliance regime.

(v) Applications The application layer is dominated by American firms, themselves deeply tied to the US stack, and no doubt eyeing these recent Gulf deals and weighing the prospects of becoming national champions themselves.

The narrowness of the stack is strategic. It reflects a logic from earlier industrial transitions: protect the emerging layer, consolidate early, embed into policy. The priority is provisioning architecture rather than a competitive marketplace.

4. Capital and Alignment

The technodollar is not evolving toward open competition. It is moving toward state-backed consolidation, enforced through interoperability with the government allocator, gatekeeper, and enforcer.

As such we should see the emergent national champions not as tech incumbents of highly priced startups but as strategic infrastructure: trusted, embedded, and protected. And while there will still be returns to marginal software most value will accrue to those able to get to the heart of this convergence of energy, compute, models, and applications.

The most helpful historical guide will be past episodes of ‘infant industry’ support, whether in industrial Germany and North Asia, or the early days of US steel. The pattern is government promotion through ‘export discipline’ (those who can sell abroad get access to government support in various ways) leading to consolidation into a small number of very large national champions

The main challenge to this architecture is open source. Foundation models with permissive licenses may offer flexibility. But they still run on chips, clouds, and compliance. Even China’s DeepSeek, a sovereign open-source initiative, reflects this logic: DeepSeek is (now) a government backed open source project as an attempt to disrupt the stack, but the model layer is probably not sufficiently foundational to really do so.

Stack Overview

All these representative firms could make a case to be national champions, and the most likely contenders are those included in May’s Saudi and Gulf deals.

I suspect that over the next year we will be able to add a ‘robotics’ layer with strategic players in automated/ AI driven science, drones, and industrial manufacturing. At the moment these are areas without clear leaders (like automated/AI driven science) or (like drone and industrial manufacture) where the leaders are Chinese. But the prospect of these systems being built as or around ‘embodied agents’ will shake up these markets and their strategic importance will lend them to infant industry support national champions.

5. Conclusion

Grand pronouncements about changes to the global order always leave you hostage to fortune and subject to (very fair) accusations that ‘oh look, this week a VC has decided to be an expert on global trade deals’. It Is of course very possible that May’s Middle Eastern deals had more to do with secret deals over Ukraine or Iran, or some other thing, than a strategic pivot around AI. But they do seem to form part of a pattern in which the AI stack is becoming a tool for US strategy rather than just another stick to add to ever vaster and ever more used collection of sticks.

And the world’s most important industries and commodities have always been tightly managed. Tech, in its networked software incarnation of the late 1980s through 2020, may prove to be an exception. More a reflection of underestimation of the technology’s power and the financial privilege of the unipolar moment than a pattern for the future.

If we return to form it will mean a return to integration and to national champions, with investment themes (broadly – own the stack) flowing from that.