Wolfgang Wopperer Notes Essays

The Political Economy of AI, Part 2: Fissures and Footholds

・27 min read

The Political Economy of AI, Part 2: Fissures and Footholds

We’re in a bubble produced by the AI capex system – a capital cycle in which present value is extracted, future liabilities accumulated, and narratives deployed to keep the latter invisible to the actors making decisions about the former.

But the bubble is not the system’s only product. The other is the rubble that will be left behind when the bubble bursts: the gigawatts, the data centres, the fabrication capacity, a whole industry infrastructure built ahead of the demand that could pay for it. The mechanism that bills the future also builds the assets it will inherit.

This process is part of a recurring pattern that Carlota Perez has mapped across two centuries of technological revolutions.The framework is implicit in much of the debate around the state and future of AI, particularly the question of an AI bubble. Thompson (2025) is one of the few who name it explicitly. In the late installation phase of a techno-economic paradigm, financial capital floods in, builds infrastructure ahead of profitability, and generates a frenzy. The bubble that ends the phase is not a deviation but the mechanism by which financial capital force-funds the infrastructure the deployment phase will later use cheaply. The burst is the turning point that disciplines capital and recouples it to production.Perez (2002) is the canonical statement; Perez (2024) updates the argument for AI specifically. I disagree with her reading of AI as part of the still-unfolding ICT surge, though; using W. Brian Arthur’s framing (Arthur 2009), I believe AI captures a new class of phenomena (patterns learnt from data, not rules specified by a programmer), and therefore counts as its own body of technology with its own surge. Her analysis has to make weird moves (introducing a “second installation phase”) to avoid this conclusion.

Which deployment follows is not settled in advance. The aftermath, Perez writes,

can enable a ‘golden age’ or only a modified, but still unstable version of the ‘gilded age’. It can establish institutions for increasing social cohesiveness, improving income distribution and general well-being or it can try to reinstate the ‘selfish prosperity’ of the frenzy phase.Perez (2002), 53.

The productivity narrative is critical to the economics of AI: the bubble bursts when belief in it breaks. But belief does more than underwrite valuations – it secures consent, the sense across the workforce and the public that the build-out is inevitable and broadly a good thing. Narrative and consent fail together, and a burst on this scale is not only a market event. It is a hegemonic crisis:On the concept and for an analysis of the broligarchic hegemony that has emerged since 2016, see Movement Ecology Collective (2025). the order that produced AI keeps its grip on power but loses its hold on consent.

What happens after the burst therefore hinges on two questions: who inherits the asset the crash devalues, and whether the order that built it survives the loss of consent. The people left holding the liability are the constituency that can claim the asset; the fissures are where the coalition of those holding them has begun to break; the footholds are what we can place in the cracks to step on, the pieces of the asset we can already hold.

Whoever has the money

Will the firms leading after the burst be the ones leading now? If the dotcom crash is any model: some of them. Amazon and Google came through; Pets.com and eToys did not. What distinguished the survivors was real cash flow independent of the bubble’s circular financing, ownership of a critical layer (infrastructure, distribution, a user base) rather than a story, and the capital to buy distressed assets on the way down. The crash concentrated the industry more than resetting it.

But which of them will survive and lead barely matters. Whoever emerges does so with capital, and capital in this paradigm is broligarchyCarole Cadwalladr coined this term for the emergent global oligarchy of tech billionaires (Cadwalladr 2025). capital – it carries the values of the current hegemonic coalition, of tech billionaires, authoritarians and toxic masculinity whatever the logo on the building. The deployment model they will implement is the gilded-age branch of Perez’s fork: the asset re-enclosed, the productivity gain captured as rent.Enclosure turns a commons – a bundle of use-rights held by many – into an exclusive, rent-bearing title, and the platform runs that operation on the AI stack. The re-commoning proposed below would run it in reverse. See Enclosure converts a commons into a compounding claim. And the same devaluation that could put the infrastructure within the public’s reach puts it within easier reach of whoever already owns the rest. The window in which the rubble is up for grabs is the window of the hegemonic crisis, and it closes as the survivors consolidate.

This broligarchy’s deployment model is the platform. As Brett Christophers, Guy Standing, Yanis Varoufakis and others have pointed out, the economy’s centre of gravity has shifted, from making things to owning the gates between people and what they need: housing and credit, but increasingly also software that is rented, not owned, listening histories locked into streaming platforms, tractors whose buyers are forbidden to repair them.Christophers (2020) and Standing (2016) call this rentier capitalism, Varoufakis (2023) calls it technofeudalism; the differences in periodisation don’t matter here. The tractor example refers to John Deere. AI productivity poured into a vessel built for capture will follow the course Cory Doctorow has called enshittification: the platform first serves its users, then its business customers, then only its shareholders. A value gain captured inside such a vessel does not become prosperity. It becomes rent.The enshittification concept is developed across Doctorow’s Pluralistic blog. Enshittification is enabled by switching costs, network effects, and the legal environment that protects them.

And the gain is wrung from inputs the build-out never paid for. The models are pre-trained on the knowledge commons – text, code and art scraped without permission, now the subject of dozens of copyright suits and a $1.5bn settlement after Anthropic was found to have pirated millions of books.Bartz v. Anthropic (N.D. Cal., 2025): Anthropic downloaded some seven million books from shadow libraries; the $1.5bn settlement covers the roughly 500,000 works in the certified class (Allyn 2025). More than fifty AI-copyright suits are pending in US federal courts (Copyright Alliance 2025). It’s important to note, though, that the suits are actually enclosure against enclosure: the commons of the mind had already been enclosed once, in the decades of copyright-term extension, software patents and database rights James Boyle named a ‘second enclosure movement’ (Boyle 2003). The labs’ scraping is the next turn of the same operation, pitting the first enclosure’s rights-holders against the second’s. They run on an energy footprint set to double the electricity demand of data-centres to roughly Japan’s national consumption by 2030. They need hundreds of billions of litres of water, much of it drawn from already water-stressed ground.The IEA (2025) projects data-centre electricity demand more than doubling to around 945 TWh by 2030 – “slightly more than the entire electricity consumption of Japan today”. The water figure follows Li et al. (2023); on the concentration of data centres in water-stressed areas, see S&P Global (2025). And they are made usable by microworkers, disproportionately in the Global South, who label the training data and filter violence and abuse out of model outputs for minimal wages, often under clauses forbidding them to even describe their work at home.Gray & Suri (2019), Crawford (2021) A gilded-age deployment keeps every one of these bills open – its prosperity for some is built on deepened extraction from others.

The socio-economic shape of the deployment period is determined by a redistribution choice, and this choice is a site of struggle. In the past, the gain reached workers only through their own structural power and through external shocks that forced wealth redistribution.See Capital concentrates wealth upward by default and Major power shifts require both organised leverage and external shock. The burst will be a shock – thanks to its size and severity it will discipline capital and reopen some political room sealed by neoliberalism.Discussions about wealth taxation have already shifted gear. Gabriel Zucman has put it back on the G20 agenda: a coordinated minimum tax requiring the world’s roughly 3,000 billionaires to pay at least 2% of their wealth a year, where many currently pay an effective 0–0.5%. (Zucman 2024) The question is whether workers are organised enough to claim it.

Where the cracks start to show

Far downstream from microworkers, the group currently most affected by AI are the people who write the platforms’ software. They feel two pressures of the build-out: one is layoffs and wage compression as employers reach for AI; the other is the work that remains – more features shipped in less time, more slop, less craft. This creates mobilisation potential that can simply begin with “can you still do the work you wanted to do?” In early May 2026 UK staff at Google DeepMind voiced dissent over Google’s Pentagon contract and the IDF’s use of DeepMind models by voting to unionise almost unanimously – the first union at a frontier lab, and a reminder that some means of production sit on the desks of people who can refuse to operate them.Nolan (2026). 98% of the votes were pro union.

Withdrawal of consent already reaches past the workforce. In May 2026 more than sixty MAGA-aligned figures, Steve Bannon among them, signed an open letter demanding mandatory testing and government approval of frontier systems before deployment.Gold & Nichols (2026). Other named signatories include Amy Kremer and Brendan Steinhauser. The letter’s central frame: “America did not become the greatest nation in the world by allowing unelected elites to experiment on the public without safeguards.” It was one expression of a broader turn that Axios calls an “AI hate wave”: over 70% of Americans say AI is advancing too quickly, only 18% of under-thirties feel hopeful about it, and Morgan Stanley now treats public pushback as a binding constraint on the data-centre build-out.Mills (2026), aggregating Gallup, Economist/YouGov (9–11 May 2026) and Stanford global-opinion data. The political space for action on AI capex is wider than the Silicon Valley consensus assumes.

Communities asked to host the build-out’s externalities are refusing in growing numbers. Maine’s legislature passed the first state-level moratorium on data centres above 20 MW (vetoed, but a clear signal), and the Data Center Watch tracker counts more than 140 local groups that have blocked or delayed over $64bn of projects in about a year, with cancellations rising from six in 2024 to twenty-five in 2025.Data Center Watch (2026); Tilton (2026) on Maine’s veto. The concrete concerns are water, electricity, and the health of the people nearby. The consumer advocate Erin Brockovich now runs a crowdsourced map that tracks the build-out town by town – and the opposition with it, turning scattered local fights into a single visible pattern.Binder (2026). Brockovich made her name in the 1990s exposing Pacific Gas & Electric’s contamination of the groundwater at Hinkley, California – a case that ended in a $333M settlement, then the largest direct-action settlement in US history. Her AI Data Center Reporting site lets residents log operational, under-construction and proposed facilities; submitters name water, electricity and community health as their chief concerns. It also tracks consequences: more than fifteen local moratoria or pauses, and four local councillors removed from office after a data-centre vote.

The build-out is also inherently more fragile than its headline numbers suggest. Of the roughly 16 GW of US capacity slated for 2026 only about 5 GW is actually under construction, and Sightline Climate expects 30–50% of the pipeline to slip; the 2027 picture is worse.Wang (2026). The gap between announced and actually-building capacity is the leading indicator most worth watching. Fragility on this scale makes the hegemonic coalition vulnerable to organised pressure.

The supply chain is tightening at the same time. Surging demand has roughly doubled the price of high-bandwidth memory (HBM) and DRAM since late 2025, and suppliers warn the shortage could run past 2027.Uko (2026) on the Samsung/SK Hynix warning that the shortage could run past 2027; The Guardian (2026) on the knock-on rise in consumer hardware prices. The Iran war has compounded it: the closure of the Strait of Hormuz disrupted up to a third of global helium supply, non-substitutable in EUV lithography and HBM stacking.Iyengar & Lu (2026); Lichtenberg (2026). Liquid helium evaporates in ~45 days, so stranded inventory is lost, not delayed – structurally different from an ordinary supply shock.

But even under the most optimistic assumptions, Joachim Klement calculates for the FT, “the implied return on investment is highly negative for all of them except Amazon” – “if the hyperscalers continue on the current trajectory, the AI boom will become a story of one of the largest destructions of shareholder value in history.”Klement (2026). He contends that this could be avoided either by the hyperscalers inreasing their revenue by 100–300% (unrealistic) or the planned investments never materialising (a bubble burster).

On their own these are grievances; together they are a beginning withdrawal of consent across every constituency the build-out depends on – workers, voters, host communities, even financiers reading the slippage. The crisis is hegemonic, not merely financial. And each crack can be a foothold – once there is somewhere worth climbing toward.

An alternative vision

In 1968 Douglas Engelbart demonstrated the mouse, hypertext and real-time collaborative editing in a single 90-minute session, all of it in service of one goal: augmenting human intellect, and especially the collective intellect of groups working together.Engelbart (1962, 1968). The 1968 demo is on YouTube; the collective-augmentation framing is most explicit in Augmenting Human Intellect. Almost everything that defined how we have used computers in the last four decades was previewed in this “mother of all demos”. In the 1970s and 80s Alan Kay built languages people could think in (Smalltalk) and conceptualised the computer for the rest of us (the Dynabook, foreshadowing tablet and smartphone); Bret Victor builds notations and rooms-as-interfaces that change what is thinkable. The computer, in Steve Jobs’s famous dictum, as a “bicycle for the mind” – this is a very different vision for what technology can enable.Jobs used the image from around 1980; the filmed version is Memory & Imagination (Krainin & Lawrence 1990): the computer as “the equivalent of a bicycle for our minds”.

A decade focused on replacing human labour through generative AI lost this strand from view. But it will outlive the burst exactly because it never rested on the productivity narrative inflating the bubble: its ambition was never more, but better – not maximise output per worker, but the capacity to solve hard problems together. An echo of this vision survives in DeepMind’s protein-folding research and in Altman’s claim that AI could “cure cancer”DeepMind’s AlphaFold (Jumper et al. 2021) predicted protein structures at scale and shared the 2024 Nobel Prize in Chemistry. On disease, Altman (2025) writes that “maybe with 10 gigawatts of compute, AI can figure out how to cure cancer”. – but it has become part of the justifying hegemonic narrative instead of shaping agendas and investment decisions.

Maggie Appleton names what this narrative misses in practical terms: once agents make implementation cheap, “the hard question is no longer how to build it. It’s should we build it” – “agreeing on what to build is the new bottleneck.”Appleton (2026) What constrains a group is not individual throughput but coordination: alignment on intent, judgement and shared context. Current tooling makes that harder, funnelling a flood of agentic output into workflows built for a slower pace and stripping out the moments where teams gathered context and corrected course. An alternative use of AI needs to reverse this and focus on coordination – the augmentation of collective intelligence that Engelbart, Kay, Victor and others have been working towards.

The same lineage also fed cyberlibertarian individualism and slid smoothly into platform feudalism, though – the augmentation commons enclosed in its turn. To resist that enclosure it has to hold on to two things. It has to be anti-extractionist: knowledge retention not extraction, idiosyncratic not generic tools, running on hardware the user controls. And it has to serve collective intelligence, not just maximise individual capability and output. The productivity boost these tools can give is real (I rely on it daily in my own work), but whether it accrues to the user and their community or to a platform’s flywheel is decided by those two conditions, not by the technology per se.

That this is my experience and my worry marks my position on a chain of extraction: the desk where an agent can be switched off sits far downstream of the labour and the land the models are built on. To take the foothold within my reach for the foothold would rebuild the colonial pattern one rung down: agency for those already closest to it, grievance for everyone further out.

A tactical baseline

The artisan doubly displaced by replacement and deformation sits close to me on the chain. For them, the vision offers a third direction: keep the craft, the architectural thinking and the judgement – and hand the writing-out to a tool that cannot enshittify because no one owns its channel to you. You use technology to improve coordination, not output, and ship better rather than more.

Local AI is approaching a point where it can be the substrate for this alternative vision. Open-weight models are now good enough to build on. LLMs that run on a developer’s own machine sit within a few points of last year’s proprietary leaders, and nine of the thirteen current frontier models are open-weight, almost all of them from Chinese labs that, ironically, US export controls pushed toward efficiency.Artificial Analysis’s leaderboard shows open-weight models dominating the Pareto frontier for Intelligence-vs-Price at half to one-sixth the cost of proprietary equivalents. The open-weight frontier as a whole has closed from ~13 points behind the proprietary leaders a year ago to 3–6 points today. (Artificial Analysis 2026) Close-to-frontier capability no longer requires entering the trillion-dollar capex race to obtain it.

Platform owners are slowly waking up to this: Thanks to their unified-memory architecture, Apple’s Mac mini and Mac Studio models are so popular for running local AI agents that their basic configurations are sold out – Apple is expecting “that the Mac mini and Mac Studio may take several months to reach supply/demand balance”.MacRumors (2026), quoting outgoing CEO Tim Cook on the Q2/26 earnings call. Nvidia, teaming up with Microsoft, just released the RTX Spark, a hardware platform and software stack specifically focused on enabling local AI; they call it “a new age of PC”.Huang (2026). Microsoft aims at users of agentic AI even more directly by advertising their new flagship laptop based on this platform as “the cure for token anxiety”. (Microsoft 2026)

Of course a model running on hardware you own cannot parallelise work the way the cloud can: it spins up roughly one agent at a time, not a hundred. Inside a token-maximising paradigm, against OpenClaw’s hundred parallel agents and its $1.3M month, that looks disqualifying. In a paradigm that focuses on maintenance and coordination, one that trades token volume for judgement and alignment, it is close to a perfect fit. That paradigm does not want a hundred parallel agents; it wants one good tool under the user’s control. The constraint and the paradigm select for each other.

The tools needs to be actually good enough, though. Open-weights models might be close to the frontier in general intelligence, and in agentic coding might be just 6–12 months behind. But the gains that frontier models made in this time are enormous – in the most reliable available benchmark, Opus doubled its success rate just between versions 4.6 and 4.8.DeepSWE (2026). DeepSWE is a more realistic and reliable coding benchmark than the widespread SWE Bench Verified, in which open-weights model tend to perform unrealistically well. To make their models such good coding agents, the frontier labs have invested heavily into post-training (e.g. refinforcement learning from human feedback or with verified results), for which they need high-quality training data – real development projects, real conversations between users and agents. Which is what subsidised developers deliver to Anthropic, OpenAI and Cursor – but not to open-weights labs. Catching up could be a collective effort, with crowdsourced data and distributed finetuning. On Huggingface, a community is emerging around exactly these techniques.A platform for sharing and using open wight models and data sets. The vibe is somewhere between early GitHub and bio-hacking.

These shifts have wider political-economic consequences. At scale a credible local alternative converts captive demand for platform subscriptions into mobile demand looking for something better – including what the platforms structurally cannot offer: democratic governance, commons-based pricing, no enshittification. A constituency that can leave is a constituency that can be organised around something better.

Developers are starting to organise seriously around open and local AI. In the words of organiser Manolo Remiddi, they refuse to “surrender their creative and strategic agency to centralised monopolies”.Remiddi (2026). The Augmentatism movement calls for co-owned, locally-run alternatives to centralised models, built by self-described “AI Artisans”. Its proposed governance layer is blockchain-based, unlike the public-commons model developed below; the shared ground is the diagnosis and the refusal it organises. But they are also clear: a local open-weight model on its own is “a non-negotiable tactical baseline”, not a strategic solution.

If your sovereignty relies entirely on a detached, open-source artifact without a self-sustaining social infrastructure around it, you are merely hiding in a clearing that can be bought and cleared by corporate landlords tomorrow. You cannot trust an architecture that you do not actively co-own.Ibid.

Local AI is a bootstrap, not an endpoint. It removes one dependency on platform owners and changes what an individual or a group can do, but the weights are still produced by someone else, on hardware made by someone else, trained on data taken from everyone, and local inference is less extractive than cloud inference, not non-extractive – the microworker labour and the memory-and-helium supply chain are upstream of the weights wherever they run.

This dependency is critical. A foothold that frees the developer while still resting on the labeller’s piecework and the miner’s helium is not an exit from extraction but a more comfortable seat within it. The unit that has to be freed is the chain, not the node – which means the developer’s foothold is incomplete until the people upstream hold one too.

Local AI is also not ownership or democratic control of the means of production, and it inherits the limitations of the current paradigm: an architecture that needs enormous data and compute to produce intelligence is a poor basis for autonomous technology.

The next paradigm

Some of the architects of the current AI paradigm have already begun pointing past it, towards leaner approaches. OpenAI co-founder Ilya Sutskever declared the scaling era over in late 2025; Yann LeCun left Meta in March 2026 to bet that large language models will prove a dead end on the road to AGI. Gary Marcus has made the diminishing-returns case for years and now finds a wider audience.Sutskever (2025), LeCun (2026), Marcus (2025–2026)

Their argument is that a large language model is a probability distribution over its training set, and however well it fits that distribution it cannot generalise beyond it – which is what intelligence in any non-trivial sense requires. A system that can truly generalise has been the long-standing target of AGI research, displaced by the LLM detour. After the bubble we can return to it.

Such a system need not be big – in LLMs intelligence just happened to be a function of their size. The smaller labs Sutskever and LeCun have founded work on ideas that don’t rely on scale, at the few-hundred-million and billion-dollar scale – tiny compared to the frontier labs’ funding.

And the constraints the LLM paradigm refuses – energy budgets, supply-chain limits, the fact that built infrastructure outlives a hardware generation – will be the starting conditions for whatever follows: they favour smaller architectures that run on commodity hardware, on the residual capacity of the post-bubble data-centre stock, and on the consumer devices people already own.

Re-commoning the asset

Pre-burst, a commons stack cannot out-train the frontier labs; the capital, talent and supply-chain access required are an order of magnitude beyond what public or commons effort can muster on any relevant timescale. A project that begins with “we need our own frontier model” is not a project but a wishlist, and the wishlist justifies indefinite delay. The frontier labs go unchallenged because they are the only ones with the resources for pre-training.

In a post-burst situation, this asymmetry will have disappeared. The multi-billion-dollar cost of training existing frontier models has already been absorbed during the bubble, on someone else’s balance sheet, and the newly disciplined capital will not invest in training new ones. What remains – fine-tuning, alignment, evaluation, deployment, governance of what’s already there – is not as capital-intensive as pre-training, can more effectively re-use outdated GPUs, and could more realistically be funded collectively. And the work that layer consists of – labelling, fine-tuning, evaluation, moderation – is largely the work microworkers already do. The layer a commons can realistically afford to own is the one their labour constitutes, which makes them not suppliers to the stack but a constituency that can hold it.

Any such approach has to be based on reclaiming ownership of the stack rather than renting space on it. As Trebor Scholz and Mario Esposito argue, a democratic AI “cannot simply rent space on the extraction stack”: workers, communities, cooperatives and public institutions have to own the infrastructure layer by layer, “from the earth to the cloud”, treating the technology as collective rather than artificial intelligence.Scholz & Esposito (2026)

Read as a description rather than a slogan, “from the earth to the cloud” names who has to hold the stack: the community whose land, power and water a data centre occupies and uses; the labeller and moderator whose work turns a base model into a usable one; the developer at the inference endpoint. None of them has to own the stack outright – that would take us back to the wishlist that licenses indefinite delay. What re-commoning needs is the form English tenants once held between owning and renting: copyhold, a secured, heritable, custom-bounded use-right – real security without freehold.Copyhold was land held at the lord’s will but bound by the custom of the manor and recorded on the manorial court roll – heritable and transferable, security without freehold. Its weak point was the fine the lord could levy when the holding changed hands: where that charge escaped custom and tracked the market, it became the lever that dispossessed the tenant. Hence the design rule – bind the cost of entry and transfer to collective rule, or the commons re-encloses through it. See Enclosure converts a commons into a compounding claim.

Each layer holds its own bundle of rights: the host community a claim on the compute its land carries and a say in what runs on it; the labeller a standing share in the models their work makes usable, rather than a one-off wage; the developer a guaranteed right to run a model, adapt it, and leave with their data. Whether this is a genuine alternative or only a softer enclosure is decided by whether these rights reach the first two, not merely the last. A stack that frees the Global North desk while still resting on Southern piecework and an extractive supply chain has just rebuilt the old stack in commons colours.

The three holdings pull access to compute, labour and the knowledge commons back out of the market and under collective rule. In Karl Polanyi’s terms, this is a re-embedding: the disembedding from communities the build-out performed, run in reverse. The withdrawal of consent already under way is the countermovement that makes it possible.Polanyi (1944) calls land, labour and money ‘fictitious commodities’ – things not made for sale that the market nonetheless prices as if they were. Compute, energy and water, human labour, and the knowledge commons are this build-out’s fictitious commodities; their forced subjection to the market is the disembedding, and the protective reaction it provokes the countermovement.

Such a stack – an AI Eurostack, to borrow Cory Doctorow’s term for non-US technology infrastructureThe term is borrowed from Cory Doctorow who uses it for his proposal of a “non-USA internet” (Doctorow 2026). – would have to work on four timescales at once:

  1. Now: run Chinese open-weight base models, treating them as upstream rather than a sovereignty problem to defer. The open-weight frontier already supplies a frontier-grade base, and the weights run outside Chinese jurisdiction.The residual concern is training-data provenance and the possibility of embedded behaviour, both auditable in a way a closed proprietary model is not. The exposure is real but manageable, and asymmetric against the alternative of total dependence on US labs.
  2. Near-term: build sovereign post-training on top of them – fine-tuning, alignment, evaluation, deployment, governance – where sovereignty is reachable at commons-feasible cost and governance can be designed in rather than retrofitted.
  3. Medium-term: fund cheap divergent research for the next paradigm, on the scale of the Sutskever and LeCun bets, with minimum intervention and maximum attention to how innovation actually emerges. Three to five paradigm-divergent efforts could put a commons-governed alternative in position to lead the deployment phase.
  4. Long-term: develop communal ownership and democratic governance of the next paradigm.

The Public-Common Partnership, developed by Kai Heron, Keir Milburn and Bertie Russell, is a template for communal ownership and democratic control of productive assets. It combines a worker-managed Joint Enterprise, a public body supplying legitimacy and resources, and a territorially-rooted Common Association, with surplus communally allocated and asset locks preventing privatisation.Heron, Milburn & Russell (2025) Those asset locks are the custom that keeps the holding from being enclosed again, the modern form of the manorial rule that once bound a lord’s discretion.A public register of who holds which use-right plays the part of the copyhold court roll, which favours institutional governance over the blockchain layer some propose. Shared compute also needs stinting – the open-field cap on each commoner’s use of the common pasture – so that a governed commons is not depleted the way an open-access one would be (Ostrom 1990). Its emphasis on contested reproduction – two economic orders coexisting and struggling – makes it a good fit for the moment: a Eurostack does not wait for the burst before it exists, it grows inside the conditions the burst will open.

None of this fits the institutions that would have to fund it. EU programmes reward incremental, legible work; national programmes still back private firms; an understanding of how innovation emerges from complex systems is largely absent. The first capital may have to come from wealth holders willing to fund infrastructure they will not own – what is built needs to be open, commons-based and asset-locked from the first day, its governance designed to resist the enclosure that usually arrives the moment it works. Re-enclosure comes at the point of transfer, not in the day-to-day charge: a fee set there by the market rather than by custom can quietly buy the commons back.This draws on conversations with people at the interface of public innovation policy and industrial research, who stress how badly the current institutional setup is suited to projects of this ambition. That is a constraint, not an excuse – and exactly why the political project has to be larger than a funding programme.

The bet underneath this proposal is usefully asymmetric. If the bubble bursts and the frontier flattens, the commons project is positioned to define what comes next; if it does not burst on the predicted timeline, the post-training infrastructure still serves an actual user base and the divergent-research bets are cheap enough to carry.

The shape of a coalition

Several interests could align around such a project.

The frontier labs themselves are not the homogeneous actors they want to appear as. The Musk v. OpenAI lawsuit has brought several areas of conflict and contestation to light.Musk v. Altman (N.D. Cal.): the litigation disclosed emails and filings on OpenAI’s founding intentions and its for-profit conversion before a jury found for OpenAI in May 2026 (Washington Post 2026). While Google employees unionise for political and ethical reasons, Anthropic refused to let the US military use its models for autonomous weapons and domestic surveillance, and had its contract terminated and itself branded a “supply-chain risk” for it.The US Department of War terminated a roughly $200M contract and designated Anthropic a “Supply-Chain Risk to National Security” (27 February 2026) after its usage policy barred fully autonomous weapons and mass domestic surveillance; the designation was then contested in court (Hendrix 2026). OpenAI published a white paper proposing redistributive policies its own lobbying then worked to defeat.OpenAI’s April 2026 white paper Industrial Policy for the Intelligence Age proposes a 32-hour week, capital-gains tax increases and a public wealth fund granting citizens an equity stake in AI labs; over the same period OpenAI-linked money funded a super PAC opposing a congressional candidate who had proposed comparable AI taxation (Sun 2026). Project Tapestry, launched in Paris in May 2026 with Yann LeCun as chief science advisor, aims to build an open platform for “globally federated development of frontier AI models” – and has IBM and Meta as co-founders.AI Alliance (2026). The Alliance is a 200+ member coalition whose founding partners include IBM and Meta; the May 2026 Paris founding workshop convened ML researchers, systems engineers, compute providers, governments and universities. These fissures are small, but we might fit in some footholds nonetheless.

Tech workers would gain from a Eurostack the autonomy the AI capex system denies them, the standing threat of enshittification or being priced out defused. Even the firms locked in the capital race might find aligned interests once the narrative driving it is contested and a different deployment model exists. And the geopolitical case is easy to make: less dependence on US infrastructure increasingly turned against its allies, and no blind trust in Chinese open weights for everything, leaves our own – governed not as state or corporate assets but as a democratic commons.

Tech-worker organising is a viable starting point – not because the harm is sharpest there but because the leverage and the technical understanding are. The harm is greatest upstream, among the labellers and the host communities; what is concentrated on the developer’s side is the capacity to act, and the fact that some means of production sit on these workers’ desks. As it will become more apparent that sustainable productivity is still dependent on human contribution (alignment, architecture, acceptance), the leverage of the humans in these positions will grow with the agentic power at their disposal. Their activation might run from “my job feels shit” through “this is structural” and “we have to act together” to “I’m part of a movement”.I won’t claim to know the labeller’s and the resident’s paths to the same movement – they are theirs to name. Unions like UTAW in the UK can build on this. The leverage is weaker than industrial leverage once was, but it compounds when stacked with other pathways.

Those pathways reach past tech labour, as does the potential in collective agency. The moderators who founded the African Content Moderators Union in Nairobi frame their work not as grievance but as standing: “We know our worth. Social media does not exist without us,” organiser Kauna Ibrahim Malgwi put it at its founding.Kauna Ibrahim Malgwi at the May 2023 founding of the African Content Moderators Union (Labour Solidarity 2023). A Nigerian psychology graduate, she moderated content for Facebook through the outsourcer Sama and was named to the TIME100 AI list in 2024. In South Memphis, where xAI has built a data centre and its gas turbines in a historically Black neighbourhood, KeShaun Pearson of Memphis Community Against Pollution makes the same claim: “it’s not just that the people who are proximate to the pain have the stories or have the suffering, we also have the solutions.”Pearson, in Küpper (2026). His group, Memphis Community Against Pollution, formed around the build-out of xAI’s “Colossus” in Boxtown, a South Memphis neighbourhood founded by formerly enslaved people.

Public sentiment has already turned, with the “AI hate wave” as evidence. The task is now to convert diffuse concern into organised pressure. The fragility of the build-out makes that pressure count: minimal demands that would be mere aspiration in a confident market become winnable in a stalling one – interoperability, disclosure of upstream labour conditions, the opening of proprietary stacks that sit on public infrastructure, disclosure of the productivity and failure-rate evidence enterprises are sitting on.

Beyond that, datacentre activism could act as a trigger for the bubble to burst and therefore much wider and more rapid change, as David Karpf suggests:

The local fights create friction for industrial deployment. The friction slows down the cadence. Slowing down the cadence makes it impossible to sustain the bubble. And after the bubble pops, we gain much more agency to shape the trajectory of the digital future.Karpf (2026)

None of this is a coalition yet. It exists in pieces – the communities blocking the data centres, the unionising engineers, the artisans co-owning their tools, the researchers building federated alternatives – held apart by the sense that the build-out is inevitable. The hegemonic crisis is what dissolves that sense, and the burst is what brings it on – the external shock that, historically, has been the precondition for any gain reaching the people who made it.

Before the rubble is cleared and the survivors consolidate, the people who paid for it can still claim it. We need to get ready for that moment.

Disclosure

I used Claude Opus 4.6, 4.7 and 4.8 at various stages of writing this essay.

References