Mariana Mazzucato: Who benefits from AI abundance?
Public investment and shared resources underpin the technology. The economist argues that how we account for those contributions should change who shares in the returns.
We are being promised a world without limits. Unlimited compute. Unlimited data. Unlimited intelligence, cheap enough to pour into every corner of the economy. When the Chinese startup DeepSeek briefly wiped nearly $600 billion off Nvidia’s market value in January 2025, the reassurance arrived within hours, in the form of an idea from a Victorian economist: the Jevons paradox. As Microsoft’s chief executive Satya Nadella explained on X, as artificial intelligence becomes more efficient and accessible, we will simply use far more of it.1
He may be right about demand. But notice what the promise of limitlessness does to the conversation. It turns every question about artificial intelligence into a question of how much and how fast, and quietly retires the question I think matters most: AI for what?2
I ask it because abundance, in an economy structured like ours, has a habit of producing scarcity. Not scarcity as a natural limit, but scarcity as a residue—the livelihoods displaced, the minerals extracted, the water drawn down, the creative work ingested without consent or compensation. They are what a particular way of creating value leaves behind, and our economics has taught us to file them under “externalities” or “market failures” rather than count them as costs.
The most influential account of scarcity today is the one Ezra Klein and Derek Thompson advance in Abundance: that plenty—of housing, energy, infrastructure, and now of intelligence—is within reach, and that what holds it back is an excess of rules, procedures and conditions. It is a serious thesis, and I share its ambition and its impatience with institutions that block rather than build. But applied to AI, my theory is precisely the opposite. What is holding back shared abundance is not too many conditions on capital. It is too few.
Conversations at the frontier, in your inbox
SubscribeWe have a financialized, extractive capitalism with default settings that allow artificial intelligence to serve the interests of capital over labor, and private interests over public ones. Public money flows to the AI industry without sufficient conditions. The scarcities this issue is devoted to are not the product of government standing in the way; they are the product of extraction going unconditioned.
Value creation and value extraction
Value can be created through collective investment in new capabilities, products and services that expand what an economy can do. Or it can be extracted, by moving existing value around, capturing a position or charging for passage. Both appear as profit. Both are counted in GDP. Neither our national accounts nor our dominant theory of value distinguishes between them, because modern economics largely gave up asking what value is, settling instead for the assumption that value is whatever fetches a price.3
Everything without a price becomes invisible. The foundations of AI—decades of high-risk public investment in the internet, GPS, touchscreens, voice recognition and the early research behind machine learning—were financed by taxpayers long before any company could see a return, and they carry no price on any company’s balance sheet.4 Private firms have since invested enormously and made genuine breakthroughs of their own. But they built on a platform whose earliest and riskiest stages the public paid for, and whose returns the public do not share.
The commons on which today’s models are trained tells a similar story. Some of it is in the public domain, and some is now licensed: a growing number of publishers, news organizations and platforms have struck deals with AI companies5. But those deals came late, often after litigation, and reach mainly the rights holders large enough to negotiate. Much of the rest was used without permission. In 2025, Anthropic agreed to pay $1.5 billion—about $3,000 a book—to settle claims over nearly half a million books downloaded from pirate libraries, after a federal judge ruled that acquiring them that way was not fair use, even as he found that training on a lawfully purchased book was.6 The pattern is clear enough: the commons was treated as free until someone with the means to do so said otherwise.7
Count the collective investment at zero, the commons at zero, the costs at nothing, and the abundance looks miraculous. It is not miraculous.
Count the collective investment at zero, the commons at zero, the costs at nothing, and the abundance looks miraculous. It is not miraculous. A great deal of it is a transfer: from the public that financed the earliest, riskiest research and from the creators whose work was ingested, to firms our accounting flatters as a technology’s sole authors. Even the way we talk about these companies conceals this. Calling these firms “tech giants” implies they invested in the technologies from which they profit, when it was largely taxpayers who funded the underlying breakthroughs.8
Consider how the returns are structured. Private investors take real risks, and many of their bets fail. So do the state’s; the difference lies in who shares in the rewards. In 2009 and 2010, the U.S. Department of Energy backed loans of $535 million to the solar company Solyndra and $465 million to Tesla. Solyndra went bankrupt, Tesla repaid its loan early before the government’s right to share in its rising stock could take effect, leaving taxpayers with modest interest payments9. No venture investor would build a portfolio that way. Even where the state did attach conditions—as with Tesla’s warrants—they were structured too weakly, or cashed out too early, to capture meaningful upside.
Tax arrangements then leave much of the value lightly taxed: in 2024, the European Union’s top court upheld a ruling that Apple owed Ireland €13 billion in back taxes, and one estimate puts the gap between headline tax rates and what six of the largest U.S. tech companies actually paid over the past decade at $278 billion.10 Labor is squeezed too. A 2024 Senate HELP Committee investigation led by then-Chairman Bernie Sanders found that Amazon’s warehouses recorded more than 30% more injuries than the industry average in 2023. Further down the AI supply chain, workers in Kenya labeling violent and abusive content to make ChatGPT safer were paid less than $2 an hour11. The profits that result are not simply the reward for a good product. In my reading, they are also the compound return on a public investment that was never priced, taxes that were not fully paid and work that was not fully compensated.
What extraction leaves behind
Consider what the promise of unlimited compute requires, and who supplies it.
Data center electricity consumption is on course to roughly double between 2025 and 2030, to around 950 terawatt-hours—roughly what Japan consumes in a year.12 Data centers are joining grid-connection queues that were already congested, and where they win capacity, others wait. In 2022, parts of west London’s grid reached capacity after a wave of data center construction, and new housing in Ealing, Hillingdon and Hounslow was temporarily halted. A London Assembly inquiry has since named data center growth as a key factor in the capital’s housing delays. In Ireland, data centers now use 23% of all metered electricity, more than all urban households combined, up from 5% in 201513.
A midsize data center can consume roughly a million liters of water a day, much of it in basins already under water stress.14 The hardware starts from the ground, in minerals such as copper, lithium, cobalt and rare earths; demand for these is projected to increase over the coming decades.15 Having served as a special representative to South Africa’s G20 Presidency, where our task force developed the G20’s first Critical Minerals Framework, I watched supply chains reproduce the oldest pattern in development economics: raw material flows out of the Global South, and the value accrues elsewhere.16 I have put the resource footprint bluntly before: AI is guzzling resources at planet-eating rates.17 In drought-stricken Santiago, Google planned a data center that would have drawn up to 7.6 million liters of drinking water a day, until residents challenged it, and a Chilean environmental court partially reversed its permit in 202418.
Then there is labor. Whether AI augments workers or simply replaces them is not a property of the technology. It is a choice made by firms responding to incentives. A study of 238 regions in 21 European countries found that regions with more AI-related patenting saw labor’s share of income decline between 2000 and 2017, particularly in industrial areas19. We can interpret this association as a sign that AI is a capital-biased technology, with gains flowing to owners rather than workers.
These scarcities share a cause: an economy financialized to reward extraction. S&P 500 companies have spent a record $1.02 trillion buying back their own shares, capping more than $7 trillion of repurchases over the past decade.20 Not every buyback is wasteful, but at this scale the opportunity cost is hard to ignore; every dollar returned to shareholders is a dollar not invested in workers, research or productive capacity. William Lazonick found that between 2003 and 2012, 449 companies in the S&P 500 devoted 54% of their earnings to buybacks and a further 37% to dividends. A system that treats stock buybacks as a legitimate use of corporate surplus, while treating wages, training and long-term research as costs to be minimized will systematically generate the outcomes we are now calling scarcity.
Follow the money
None of this was unforeseeable. With Tim O’Reilly, Ilan Strauss and Rufus Rock, I have argued that one class of technological risks is knowable in advance: the risks arising from a misalignment between a company’s incentive to monetize its model and society’s interest in how that model is deployed.21 If you want to know where AI is heading, follow the money.
The money tells a story we have seen before. In 2019, I warned about what I called digital feudalism: platforms that began by creating genuine value then used their positions as gatekeepers to extract it—a process Cory Doctorow calls enshittification.22
We call this algorithmic rents, and specifically, algorithmic attention rents—the use of AI and machine learning not to create genuine value but to concentrate market power and extract wealth from users and from the smaller players who depend on these systems.23 In our study of more than 2,000 Amazon search queries in 2023, the most-clicked sponsored products were on average 17% more expensive than the top organic result, and ranked 33% lower by Amazon’s own quality-optimizing algorithm.24 The gap between what the algorithm knows is best for you and what it is paid to show you is a rent—extracted at scale, automatically, by code. Rents like these also help fund salaries no university can match: roughly 70% of new AI PhDs in North America now go into industry, up from about 40% in 201125.
Generative AI shows signs of the same pattern: models trained partly on content used without permission, value increasingly centralized within AI services, and less of it flowing back to the creators and developers on which they depend. 5 Beneath the applications sits infrastructure that is anything but open: Amazon, Microsoft and Google together account for 63% of global enterprise spending on cloud infrastructure.26
When DeepSeek arrived, I asked whether more efficient approaches might break the stranglehold the major cloud companies have established through their control of computing resources—or whether other forms of market concentration would simply emerge in their place.2 Model performance has been commoditized. But capital has moved to the layers that cannot be: the five largest American hyperscalers are on course to spend around $750 billion on capital expenditure in 2026, increasingly financed by debt.27 DeepSeek punctured something else, too: the belief, treated inside the industry as a law of nature, that models must grow ever larger and are therefore ever hungrier for energy, water and minerals. Karen Hao has documented how the logic of “race” rewards ever larger models over more efficient ones.28 Meanwhile, Sebastian Mallaby’s account of the race to superintelligence shows how widely that assumption—that more scale, rather than more efficiency, is the path to progress—came to be shared at the frontier.29 Scale is a design choice, not a law of physics. And design choices can be governed.
Abundance upstairs is precisely what makes the tollbooth downstairs worth owning.
That capital intensity builds chokepoints. The cheaper and more abundant intelligence becomes at the application layer, the more valuable ownership of what remains scarce beneath it becomes: compute, advanced chips, grid connections, distribution. Abundance upstairs is precisely what makes the tollbooth downstairs worth owning. That is not a market failure. It is a business model, and a rational one given the incentives we have set.
Fixing markets versus shaping them
Here, I want to be careful, because the obvious inference—that what AI needs is more regulation—is not quite my argument, either. There is a false divide between innovation policy and regulation, and both camps in this debate are trapped inside it2.
The framework beneath that divide is market fixing: markets allocate resources efficiently, government intervenes only to correct failures and regulation is therefore always a brake. This is why governments end up perpetually behind technology, regulating after the fact and redistributing after the damage.30
Badly designed regulation is not neutral; it is a subsidy to incumbents. The abundance school’s diagnosis is that plenty is blocked by an excess of conditions. Mine is that in artificial intelligence, the causality runs the other way. The conditions that matter—on how public money is used, on the direction of innovation, on how the commons is compensated, on how gatekeepers treat those who depend on them—are largely absent, and their absence is exactly what lets extraction crowd out creation. An unconditioned subsidy becomes a buyback. An unconditioned platform becomes a tollbooth. In AI, the gatekeepers are not planning committees. They are rent-extracting chokepoints, and deregulation does not remove those gates. It hands over the keys.
None of which means conditions cannot be designed badly. Klein’s notion of the “everything bagel”—a project loaded with so many competing requirements that it never gets built—is a real failure mode. But that is a design failure, not an argument against conditions. With Andrea Laplane, I set out the case for socializing the risks and rewards of public investment.31 With Dani Rodrik, a taxonomy of how conditionalities work in practice: well-designed conditions do not slow growth, they direct it.32 Germany’s KfW energy efficiency programs returned roughly four euros to the state for every euro invested, while generating 64,000 full-time jobs and cutting 700,000 tons of CO2 a year.
The alternative to bad regulation is not no regulation; it is market shaping. Not picking winners, but creating the conditions and incentives that steer markets toward outcomes we want as a society.
Well-designed rules have repeatedly unleashed innovation. In 1956, an antitrust consent decree required Bell Labs to license its patent portfolio royalty-free. The result was not stagnation, but an explosion of follow-on innovation outside the Bell system—the seeding of the modern electronics industry.33 What is arguably the single greatest act of technology-sharing in industrial history was not the removal of a rule. It was a rule.
Without conditions on food and medicine, we would be eating and swallowing things that kill us. Drug regulation can be micromanaging and obstructive, or it can push an industry to innovate toward products that are safe and work. The failure mode is bad design, not regulation as such.
The choice was never innovation versus regulation. It is market fixing versus market shaping.
The choice was never innovation versus regulation. It is market fixing versus market shaping.
Steering AI toward public value
AI is not a sector. It is a general-purpose technology that is already reshaping every sector of the economy,34 and like every transformative technology before it, it can create tremendous value or cause serious harm. That is precisely why its direction cannot be an afterthought. Three shifts in policy follow.
Set the direction: Missions, not sectors
Treating AI as a sector produces sector policy—subsidies, national champions and a race to be an “AI superpower.” That is a question about rate. Missions are questions about direction. A mission-oriented approach starts with problems worth solving, such as decarbonizing the grid, extending healthcare to those without it, and improving how we grow food, and invites solutions from across the economy.35 Gabriela Ramos and I have argued that AI can enhance our lives in exactly these ways, but that without effective governance, it will create new inequalities and amplify existing ones.36
The deeper problem today is that AI has no mission at all. It is AI for its own sake, while climate has slipped into the background and governments have settled into reactive mode. Governing AI for good cannot be bolted on once the models are trained and the data centers are constructed. It has to be built from the beginning: the common good is not a redistributive afterthought, it is a design principle.37
Invest with conditions: Share risks and rewards
Governments everywhere are pouring public money into AI. When the U.K. published its AI Opportunities Action Plan, Tommaso Valletti and I argued that it missed the mark, casting the state as customer and cheerleader.38 An entrepreneurial state does it differently: it builds pre-distributive structures that share the risks and rewards of innovation fairly from the start, rather than socializing the risks, privatizing the rewards and hoping to redistribute afterward.
Conditions on access, pricing, reinvestment and openness are not red tape. They are what any serious investor asks for: terms. Public money that comes with reinvestment requirements rather than freeing up private cash for buybacks is the difference between subsidizing capability and subsidizing share price.
Conditions should extend to the commons. AI is trained on the accumulated creative and intellectual work of humanity, much of it taken without payment. With Fausto Gernone, I have proposed a levy on the revenues of large AI companies, channeled into funds that sustain cultural production, journalism and creative labor.7 It is the ordinary principle that a firm pays for its inputs, extended to the one input this industry has largely obtained for free. An economy that consumes its commons without replenishing it liquidates an inheritance and books the proceeds as growth.
Rebuild public capacity and public alternatives
It is impossible to regulate a system you do not understand—so what happens when all the knowledge is concentrated in five private companies?
Layered on top is what Rosie Collington and I have called the “Big Con”: decades of outsourcing the government’s own thinking to consultants, hollowing out the state’s capacity until it must ask the industry it regulates how to regulate it. It’s the danger NASA’s head of procurement dubbed “brochuremanship” in the 1960s.39 You cannot attach intelligent conditions to an investment you cannot evaluate.
Where a layer of the stack is a natural site of rent extraction, the answer is a public alternative: what Francesca Bria calls the “EuroStack”: digital infrastructure governed as public goods rather than monopolistic enterprises.40 The question for Europe and the U.K. is not whether they can become AI superpowers, nor for the U.S. whether it can remain one, but whether they can help build an AI ecosystem that serves the common good.
The choosing
Scarcity, in an economy like ours, is rarely a fact of nature. It is chosen, though further upstream than we usually look: in the economic framework through which we see technology at all.
We chose an economics that cannot distinguish value creation from value extraction, then wondered why so much innovation optimizes for the latter. We chose to let public money flow without conditions, to let public capacity drain into the firms the public must govern, and to file every resulting scarcity—of shared prosperity, of water, of grid capacity, of trust—in someone else’s account.
The abundance we already enjoy—the internet in your pocket—was not deregulated into existence. It was shaped into existence.
The test of abundance was never how much, or how fast. It is whether what it consumes is counted, whether what it displaces is compensated, and whether what it creates is shared. Which returns us to the question we began with: AI for what? And who gets what and why?
Notes
1 Nvidia’s loss was reported at $589–593 billion depending on the basis used; see Reuters and CNBC coverage, 27 January 2025. Nadella, S. [@satyanadella], post on X, January 27, 2025. https://x.com/satyanadella/status/1883753899255046301
2 Mazzucato, M. (2025, February 3). “AI for What? Public value creation versus extractive rents.” Mission Economics. Substack. https://marianamazzucato.substack.com/p/ai-for-what
3 Mazzucato, M. (2018). The Value of Everything: Making and Taking in the Global Economy. Allen Lane.
4 Mazzucato, M. (2013). The Entrepreneurial State: Debunking Public vs. Private Sector Myths. Anthem Press.
5 For a running tally of publisher and platform licensing deals, see Media and the Machine, “AI content licensing deals,” June 2026. https://mediaandthemachine.substack.com/p/ai-content-licensing-deals-june-2026
6 Bartz v. Anthropic (N.D. Cal.). The June 2025 summary judgment ruling held that training on lawfully acquired books was fair use but that downloading pirated copies was not; the $1.5 billion settlement covers roughly 482,000 works at about $3,000 each. NPR, “Anthropic to pay authors $1.5 billion in settlement over chatbot training material,” September 5, 2025. https://www.npr.org/2025/09/05/g-s1-87367/anthropic-authors-settlement-pirated-chatbot-training-material
7 Mazzucato, M., & Gernone, F. (2025, July 25). “AI Should Help Fund Creative Labor.” Project Syndicate. https://www.project-syndicate.org/onpoint/how-ai-profits-can-help-fund-cultural-production-by-mariana-mazzucato-and-fausto-gernone-2025-07
8 Mazzucato, M. (2019, October 2). “Preventing Digital Feudalism.” Project Syndicate. https://www.project-syndicate.org/commentary/platform-economy-digital-feudalism-by-mariana-mazzucato-2019-10
9 Mazzucato, M. (2013). The Entrepreneurial State. Anthem Press, on Tesla and Solyndra. On the early repayment and the warrants that never took effect: Green Car Reports, “Tesla repays $465-million DoE loan,” 23 May 2013. https://www.greencarreports.com/news/1084340_tesla-repays-465-million-doe-loan-chrysler-bites-back-at-claim
10 Court of Justice of the European Union, Case C-465/20 P, Commission v Ireland and Others, judgment of 10 September 2024. Fair Tax Foundation (2025), The Silicon Six and their enduring global tax gap: an effective corporate income tax rate of 18.8% against headline rates of 27–30% for Alphabet, Amazon, Apple, Meta, Microsoft and Netflix over the decade to 2024. https://fairtaxmark.net/silicon-six-global-tax-gap/
11 Perrigo, B. (2023, January 18). “Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic.” TIME. https://time.com/6247678/openai-chatgpt-kenya-workers/
12 International Energy Agency (2026). Key Questions on Energy and AI. IEA—data center consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030. https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf
13 London Assembly Planning and Regeneration Committee (2025), Gridlocked: how planning can ease London’s electricity constraints. https://www.london.gov.uk/who-we-are/what-london-assembly-does/london-assembly-press-releases/data-centres-delay-housing-delivery. Central Statistics Office (Ireland), Data Centres Metered Electricity Consumption 2025, July 2026. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/
14 On mid-sized and hyperscale facility water consumption and withdrawals in water-stressed basins, see the Environmental and Energy Study Institute, https://www.eesi.org/articles/view/data-centers-and-water-consumption and Ceres, Drained by Data (2025). https://www.ceres.org/resources/reports/drained-by-data-the-cumulative-impact-of-data-centers-on-regional-water-stress
15 International Energy Agency (2025). Global Critical Minerals Outlook 2025. https://www.iea.org/reports/global-critical-minerals-outlook-2025/overview-of-outlook-for-key-minerals
16 G20 (2025). G20 Critical Minerals Framework, G20 South African Presidency. https://www.g20.org.za/wp-content/uploads/2025/04/G20-TF1-FINAL-Deliverable-3-G20-Critical-Minerals-Framework.pdf
17 Mazzucato, M. (2024, May 30). “The ugly truth behind ChatGPT: AI is guzzling resources at planet-eating rates.” The Guardian. https://www.theguardian.com/commentisfree/article/2024/may/30/ugly-truth-ai-chatgpt-guzzling-resources-environment
18 Data Center Dynamics, “Chile partially reverses Google data center permit over water use concerns,” February 2024. https://www.datacenterdynamics.com/en/news/chile-partially-reverses-google-data-center-permit-over-water-use-concerns/. Karen Hao also reports on the Cerrillos case in Empire of AI.
19 Trades Union Congress (2026). Artificial Intelligence, Business and the Future of the Workforce. TUC, citing CEPR (March 2026) on AI-related patenting and declining labor income share in European regions. https://www.tuc.org.uk/research-analysis/reports/artificial-intelligence-business-and-future-workforce
20 S&P Dow Jones Indices, S&P 500 buybacks, 12 months to September 2025: a record $1.020 trillion. https://www.spglobal.com/spdji/en/corporate-news/article/sp-500-q3-2025-buybacks-post-modest-62-gain-to-249-0-billion-after-declining-20-1-amidst-uncertainty-in-q2/
21 O’Reilly, T., Strauss, I., Mazzucato, M., & Rock, R. (2024, April 11). “To understand the risks posed by AI, follow the money.” The Conversation. https://theconversation.com/to-understand-the-risks-posed-by-ai-follow-the-money-225872
22 Doctorow, C., on “enshittification”; discussed in Mazzucato (2025), “AI for What?”
23 UCL IIPP / Omidyar Network, Algorithmic Attention Rents research program. https://www.ucl.ac.uk/bartlett/public-purpose/policy/digital-technology-and-artificial-intelligence/digital-economy-and-algorithmic-rents
24 Rock, R., Strauss, I., O’Reilly, T., & Mazzucato, M. (2024). Behind the clicks: Can Amazon allocate user attention as it pleases? Information Economics and Policy, 69, 101115.
25 Amazon.com, Inc., 2024 Form 10-K (advertising services revenue of $56.2 billion). Stanford Institute for Human-Centered AI, AI Index Report 2024 (share of new North American AI PhDs entering industry: 70.7% in 2022, against 40.9% in 2011). https://hai.stanford.edu/ai-index/2024-ai-index-report/education
26 Synergy Research Group (2026), cloud infrastructure market share, Q1 2026: AWS 28%, Microsoft Azure 21%, Google Cloud 14%. https://www.srgresearch.com/articles/cloud-market-annual-revenue-run-rate-topped-half-a-trillion-dollars-in-q1-as-growth-surge-continues
27 CreditSights hyperscaler capex estimates, 2025–26; on debt financing and cash-flow absorption see FactSet, 2026. https://know.creditsights.com/insights/tech-raising-hyperscaler-capex-2026-estimates/
28 Hao, K. (2025). Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI. Penguin Press.
29 Mallaby, S. (2025). The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence. Penguin Press.
30 Mazzucato, M., & Gernone, F. (2024, March 11). “Governments Must Shape AI’s Future.” Project Syndicate. https://www.project-syndicate.org/magazine/governments-must-shape-ai-business-not-only-regulate-redistribute-after-the-fact-by-mariana-mazzucato-and-fausto-gernone-2024-03
31 Laplane, A., & Mazzucato, M. (2020). Socializing the risks and rewards of public investments: Economic, policy, and legal issues. Research Policy, 49, 100008.
32 Mazzucato, M., & Rodrik, D. (2026). Industrial policy with conditionalities: a taxonomy and sample cases. Industrial and Corporate Change, dtaf063.
33 Watzinger, M., Fackler, T. A., Nagler, M., & Schnitzer, M. (2020). How antitrust enforcement can spur innovation: Bell Labs and the 1956 consent decree. American Economic Journal: Economic Policy, 12(4), 328-359.
34 Perez, C. (2003). Technological revolutions and financial capital. Edward Elgar Publishing.
35 Mazzucato, M. (2021). Mission Economy: A Moonshot Guide to Changing Capitalism. Allen Lane.
36 Ramos, G. & Mazzucato, M. (2022, December 26). “AI in the Common Interest.” Project Syndicate. https://www.project-syndicate.org/commentary/ethical-ai-requires-state-regulatory-frameworks-capacity-building-by-gabriela-ramos-and-mariana-mazzucato-2022-12
37 Mazzucato, M. (2026). The Common Good Economy: A new compass. London, UK: Penguin Allen Lane. ISBN: 9780241722244
38 Mazzucato, M., & Valletti, T. (2025, February 11). “Governing AI for the Public Interest.” Project Syndicate. https://www.project-syndicate.org/commentary/governing-ai-in-the-public-interest-starmer-plan-misses-the-mark-by-mariana-mazzucato-and-tommaso-valletti-2025-02
39 Mazzucato, M., & Collington, R. (2023). The Big Con. Allen Lane; the “brochuremanship” warning is discussed in Mazzucato (2025), “AI for What?”
40 Bria, F., “EuroStack: Building European Digital Independence,” AI Now Institute. https://ainowinstitute.org/publications/x-european-digital-independence-building-the-eurostack; discussed in Mazzucato (2025), “AI for What?”