‘Industrial AI’ is the watchword of the world’s most uninspiring AI strategies. Every industrial country that gets into AI policy sooner or later stumbles across the same idea: its legacy industries allegedly form inescapable bottlenecks and provide uniquely valuable data. They conclude that protecting the same incumbents as always is the best AI policy, whether that’s South Korean chaebol or German automakers. That’s strangely convenient.
But it’s not entirely wrong: industrial assets really are central to AI strategy. If you think the Americans will keep running away with AI, you need to make sure they’ll continue to need you. On the search for complementary assets that middle powers could hold and America could need, manufacturing is at the top of the list: to turn ‘a country of geniuses in a datacenter’ into real-world value, the Americans will need manufacturing capacity. Many of the long-promised applications of AI require physically building things: cancer cures, hypersonic missiles, humanoid robots.
Allies who can provide manufacturing capacity can capture a share of AI-driven growth and gain leverage over America’s AI industry in return. Those could be the foundations of a strategically stable post-AI Western alliance.
Getting to this good version of industrial AI means avoiding two failure modes. The first is arrogance: countries assume they can go it alone without American AI models to supercharge their industries, then fall behind and get outcompeted. The other is naïveté: they invite the Americans in, only for them to take what they need, reshore the industries, and ultimately discard allies they no longer need. Getting this right requires a balance: integrate industries with American AI enough to benefit from the boost to R&D; protect them enough that they don’t get hollowed out and shipped to the Sun Belt.
Access First
The first step in any reasonable industrial AI strategy remains securing frontier access. The question of how to do that has kept me on the road over the last few months, and from publishing here as frequently as I would’ve liked. I feel we’ve made a lot of progress, and when I talk to political decisionmakers now, they treat access to frontier systems as a genuine geopolitical priority. Two recent publications set out some of that work:
With a group of co-authors: ‘A Transformative AI Strategy for Europe’, to which I contributed chapters on compute-for-access, semiconductor supply chain leverage, and frontier model moonshots.
With Sam Winter-Levy: ‘AI Breakout Capacity’, arguing that middle powers should neither build frontier AI today nor give up completely, but retain the fallback capacity to sprint to the frontier later on.
But even the best case for frontier access alone is a prudent but unsatisfying pitch. Let’s assume we have a frontier model up and running, but so do many of our rivals and adversaries—now what?
Critics are right: hosting foreign data centres is not a geopolitical strategy; building and sustaining a publicly funded AI lab isn’t a sustainable economic model. Frontier access is not the endgame, and it’s not even the midgame. It’s a phase we need to make it through to start playing to our strengths.
The midgame is industrial AI.
This concept understandably has a poor reputation among people who take transformative AI seriously. It’s usually read as a coping strategy: ‘You Americans can build your fanciful superintelligences. We have proprietary data and will use it to build efficient systems with real-world value.’ That version of the idea is haunting European discussions, but it’s also taking root in other highly industrialised middle powers like South Korea and Japan. The prospect of small, specialised industrial models and of treasure troves of data makes business elites and policymakers think they can make do without American tech.
That version of industrial AI is less a strategy than a bet on the Americans being catastrophically wrong. It flies in the face of what we’ve learned in the last few years about the performance of general-purpose models, the returns to scale in data and compute, and the beginnings of recursive self-improvement. General-purpose models are outcompeting narrow models on enabling robotics; millennium maths problems are falling to large swarms of even larger agents; and the AI developers are betting big on the usefulness of their models for life sciences, chip design, and many other core industrial domains.
This is why everyone in frontier AI groans when they read the phrase ‘industrial AI’. It’s why they—too quickly—recoil at Mario Draghi’s recent suggestion that Europe could win on AI by using its data. It’s also why they hold their heads in their hands watching their industrialised home countries squander such a good hand just because they don’t seem to get it. To those at the frontier, every day brings new evidence that the old version of industrial AI won’t work.
But there’s still a good version of industrial AI. It builds on deep mutual integration: European manufacturing running in a tight feedback loop with American frontier AI models. Europe provides the scarce, hard-to-build industrial capacity that enables America to turn AI tokens into real-world value. America provides the frontier AI that will be necessary to rapidly improve and automate research, development, and innovation, preventing European industry from falling behind its overseas rivals. They make the tokens, we make the robots and the cancer cures.
Incentive Aligned
I’m optimistic about this idea, despite how far it is from the current state of relations between the US and its allies, because it makes too much sense to ignore. I’ll offer you three ways of coming to this conclusion:
Economically, it’s comparative advantage. American allies are good at manufacturing by virtue of specialised workers and existing assets, but not at building frontier AI models. That’s because building today’s frontier AI models requires an extraordinary concentration of talent and capital. Silicon Valley can do it; China barely manages to keep up. Middle powers’ policymakers know this, but hope they can sidestep the scaling race by building smaller, specialised models. They’d prefer for the cliché to hold true. Much like the Ford Expedition or the Buc-ee’s truck stop, general-purpose AI models are just American excess.
But that view seems mistaken: large, expensive general-purpose models increasingly look like the best route to specialised capabilities well beyond language or software engineering. Recent maths breakthroughs draw on general-purpose models, as do some of the best-performing robotics models. Anthropic is betting on this approach generalising to life sciences; OpenAI is applying it to chip design.
The combination of highly capable general-purpose models with highly specialised post-training pipelines seems to allow for breakthroughs in any domain for which these pipelines are built. And the efficiency curves are brutal: today’s general-purpose breakthroughs become commercially available, highly efficient systems within months. Small, specialised approaches seem ineffective on cost and on quality.
That does not mean specialised data is useless. Specialised data is, in fact, extremely useful for getting the most out of frontier models. That’s why frontier labs are happy to pay exorbitant prices for datasets, reinforcement-learning environments, and the start-ups that build them. But it’s the frontier labs that can use them well, because only they have base models at sufficient scale to turn the data into value. You need both: models only the Americans can build, and the data that is still scarce across the entire alliance.
Conversely, the Americans are much better at AI than at manufacturing. Much has been said about the reasons why American attempts at reindustrialisation are struggling: American capital markets are drawn to the rapid return cycles offered by software investments, the existing manufacturing base has atrophied, and the next generation of automated manufacturing is not ready to leapfrog it.
More speculatively, I suspect a crowding-out effect reminiscent of Dutch disease is affecting American industrial investment. The marginal use for a dollar of venture funding is a start-up that rides the AI wave. The marginal use for a gigawatt of power is a data centre that hosts a frontier training run. And the marginal use of skilled workers, construction capacity and sites—resources that could go into reindustrialisation—is to build data centres instead. Because AI is eating the American economy, it’s also eating the American ability to reindustrialise beyond the core supply chain of chips, models and data centres.
Allies, who don’t need to worry about building superintelligence any time soon, remain less afflicted by this disease. That, combined with the long lead times of standing up AI manufacturing, is why I believe the current allied strength in manufacturing can remain a stable source of comparative advantage: as long as the marginal American dollar flows into advancing the AI frontier, allies are free to carve out and defend their complementary manufacturing niche.
If we need industrial capacity and AI development across an alliance of countries, it makes a lot of sense for the Americans to build the AI and the rest of us to do the industrialisation.
Strategically, it’s allied scale for AI. America produces barely 15% of global manufacturing output to China’s growing 32%. America and its allies account for 43%. That mismatch has prompted US analysts Rush Doshi and Kurt Campbell to suggest an approach of ‘allied scale’: only with allied manufacturing capacity could America hope to match China in economic competition or a prolonged military conflict. But not all allies are necessarily interested in aligning their manufacturing output with US strategic ambitions in the Pacific. And even if they were, allied manufacturing bases—especially in Europe—are falling behind already, and will atrophy even further as China continues to automate.
Deeper integration between American frontier AI and allied manufacturing solves both of these problems. Allies will be incentivised to align their output with American interests because their American partners will get a say in what they build. And they will be more competitive with their Chinese rivals because, with American help, they too will have access to deep capital markets and cutting-edge AI models that they’d otherwise lack.
Politically, it’s playing to entrenched constituencies. It’s exceedingly difficult to convince allied legacy economies to pivot to AI, whatever that means. It’s comparatively easy to tell all the established industry players that they’ll remain the cornerstone of national economies. Everything will change, but everything will stay the same—that’s why even the bad version of the industrial AI story is so exceedingly comfortable.
Of course, it won’t be that easy: industrial integration will also bring disruption. Many jobs would disappear, tasks would change, and AI would still deprive many people of their livelihoods. This will be a disruptive time either way. But an industrial strategy allows for much greater continuity between the pre- and post-AI economy: economic power concentrated in similar regions and sectors and with similar firms, fundamentally similar requirements on the distribution of power and resources. That is an easier transition to manage than a more decisive pivot that requires abandoning all rural economies in favour of a tech industry rooted in capitals. Just look at the UK, where the last 20 years have concentrated economic power in London—that has been very hard on the British political economy. If we can help it, we should make sure AI does not coincide with transitions of that sort.
All that matters because any ambitious AI transformation will be hard to sell to the economic incumbents. If things move too slowly, middle powers will fall behind; but if they move too fast, their electorates may reject the pivot. It’ll be much easier to strike a sustainable and equitable bargain if you can plausibly wrap the transformation in assurances that much will stay the same. Industrial AI lets you tell an AI story that neither unions nor legacy industries hate—that’s worth a lot.
Reindustrialise, Without the Sweaters
Putting this into practice is difficult.
The first step is to ensure access as a geopolitical precondition. Spread general understanding that frontier access is important, create security conditions to enable it, create compute-for-access deals to lock it in. So much, so familiar. But another important corollary of expanding frontier access is commodification.
Middle-power policymakers talk a lot about commodifying the frontier. The idea is that fierce competition between frontier labs building similar products compresses their margins so much that the surplus value of AI can be captured elsewhere. Middle powers would prefer that world, because more value is captured on the layers of the stack they control, and less power accrues to American developers. That discussion is often just aspirational: ‘Wouldn’t it be nice if the frontier were commodified?’ Or it is hopeful: ‘We choose to believe the frontier will commodify’.
But middle powers can increase the odds of commodification by ensuring access to competing frontier models. One way for frontier developers to protect their revenues is to make their products hard to substitute for, so they do not have to compete on price. But they’re not managing to differentiate themselves through model capabilities, because all frontier labs pursue recursive self-improvement and are therefore building models geared towards that goal.
Another way to find a defensible niche is through exclusive agreements: only providing their models to a select group of cleared buyers, or only using them internally in vertically integrated business units. Middle powers, with some shrewd negotiation, can crack open that approach: ensure that international firms and governments have access to different frontier models and can switch between them at market-determined costs instead of being locked in by geopolitics. That would compress developers’ margins and allow more value to be captured elsewhere. Ensuring frontier access in that broader sense shifts the frontier AI market as a whole toward value capture at the adoption level and makes industrial AI much more attractive.
Building the Ship
But once we’re in the game, we still have to win. We don’t yet know what industrial AI will look like, so much of what follows is guesswork. In that spirit, here’s my guess:
Successful industrial integration means joint ventures and exclusive deals between middle powers’ industrial firms and American AI developers. Industrial firms offer long-term offtake guarantees and exclusive rights to the use of their data; American developers offer to use that data to develop and deploy exclusive systems fine-tuned to boost industrial output. Different developers compete for long-term deals with different industrial firms in an open market.
The basic arrangement looks like this: on one side, a factory makes a physical product. It’s staffed by skilled workers and industrial robots, it’s highly modular and adaptable, and it’s equipped with sensor equipment to allow intimate digital visibility into its production process. On the other side, we have a datacenter running a proprietary, specifically post-trained version of an American frontier model. It’s staffed by forward-deployed company engineers that deeply integrate the model into the workings of the factory. When the model detects inefficiencies, engineers and factory staff work to fix them; when the model develops hypotheses for new materials, processes or products, the factory can adjust to experiment and test them. The more sophisticated the manufacturing plant, the quicker these feedback loops can be—call it RSI-ready manufacturing.
The process generates specialised data that can be used to improve the proprietary model. The proprietary model itself never leaves this local data centre; it’s not fed into general training on the developer side. The intended bargain is that the factory needs the model to remain competitive, while the developer needs the factory to earn a return on the specific model it has trained and cannot use elsewhere—interdependence.
Industrial AI strategy is successful if it helps create these setups. It has failed if these setups fail to manifest. There are two concrete interventions to make this happen: encourage the deals while legacy industry still drags its feet; and stabilise the deals so they don’t turn extractive.
Encouraging the deals is dicey industrial policy. If this argument is correct, you’d expect the deals to happen anyway: both parties have a commercial interest, and government should stay well away. But I still suggest government action for three reasons.
First, I expect unfair competition: America is subsidising its reindustrialisation efforts, and China’s unfair trade practices in support of its industrial base are well-known. If we are to create a new manufacturing paradigm in the allied world, we need to play by the same standards. That includes tax advantages for favoured integration arrangements, trade barriers for competing products and favourable ally-US trade conditions for the products of joint manufacturing ventures. It should be cheap to do this and easy to sell its products to everyone involved.
Second, legacy industry in many allied countries is being slow on the uptake. I’ve had more of these conversations than I can count, and I’ve talked to even more senior officials who in turn take their marching orders from legacy industry. The impression I take from these conversations is clear: many legacy industrial firms simply don’t believe any of this is happening. That’s why they’re not committing to long-term purchases of computing capacity, or engaging in joint ventures, or seeking at least financial exposure to AI-driven growth. They simply don’t get it, just as they didn’t get what was happening with electric vehicles before. Governments should guide their hand.
Third, middle powers are sometimes actively sabotaging good industrial AI through arcane data regulation. Restrictions and uncertainty around industrial data make cross-border data-sharing very difficult; and in an attempt to prevent their industries from selling out to China or America, middle powers are frequently tempted to tighten restrictions rather than relax them. That instinct responds to a real coordination problem with the wrong measure; it risks undermining the competitiveness it is meant to protect. At the risk of being glib: we shouldn’t make it illegal to do this.
Stabilising these arrangements comes down to data governance. There’s a baseline stability to what I propose because lead times are long: it’ll always be much harder for allies to stand up their own model development rather than finding compromise, just as it’ll always be harder for the US to quickly build out physical manufacturing capacity at home than to draw on allied capacity. But deals of that shape can still be unstable in two directions. First, American AI developers could siphon off all the trade secrets until they’re able to recreate that manufacturing capacity at home. Second, allied industry could skim off all the surplus created by AI until industrial collaboration becomes such a low-margin activity that the AI labs stop engaging in it.
Ideally, allied firms would want to own the entire layer on which their data is used—one proprietary harness, into which they can plug frontier or sub-frontier models at will, nothing ever flows back. Ideally, AI labs would want to integrate all the data they can get—feed it back into their general training and use each collaboration to make the foundational model better and better. Neither is tenable, because each concentrates too much economic power in one party’s hands.
The stable equilibrium is mutually exclusive integration: American AI developers get to use the data for proprietary post-training and integration into industrial processes, but only in the context of this one specific deployment. The data flows to the local version of the model owned by the frontier developer, not back to San Francisco. The allied firms will hate this because it raises switching costs; the frontier labs will hate it because it slows down the flywheel effect. But if we’re right about these trends, this is still enormously profitable and mutually beneficial. It’s an equilibrium worth keeping.
While We’re Sailing
People who know more about manufacturing than I do will tell you the above is half-baked. We don’t yet know which models are most useful for industrial applications, and I’m asking firms to enter joint ventures without a clear business case. They’re right, but I still don’t see another way: our situation is so dire that the time calls for audacious bets.
The first commercially promising results will emerge only after substantial integration: in El Segundo and Shenzhen, if not in Stuttgart and Ulsan. But if allies wait until they happen elsewhere first, they’ll be too late to seek out good industrial AI on their own terms. They’ll already be outcompeted by more AI-forward competitors, their assets will have lost value and their revenues will be collapsing. From that position of future weakness, they will not be able to make a deal that preserves their bargaining power, and they’ll have squandered what could have been a decisive asset.
If we wait too long for certainty, we may lose the chance to make a good deal. I know that middle powers don’t like the current deal very much. But they can either come to the table and make it better now, or be left praying that it doesn’t get altered even further.