A Blueprint for Multinational Advanced AI Development

A Blueprint for Multinational Advanced AI Development by Adrien Abecassis, Jonathan Barry, Ima Bello, Yoshua Bengio, Antonin Bergeaud, Yann Bonnet, Philipp Hacker, Ben Harack, Sophia Hatz, Joachim Henkel, Holger H. Hoos, Kit Kitamura, Ranjit Lall, Yann Lechelle, Constance de Leusse, Charles Martinet, Nicolas Miailhe, Julia C. Morse, Maximilian Negele, Kyung Ryul Park, Miro Pluckebaum, Murielle Popa-Fabre, Benjamin Prud’homme, Yohann Ralle, Mark Robinson, Charbel-Raphael Segerie, José-Ignacio Torreblanca, Lucia Velasco, K. VijayRaghavan November 2025 This memo is authored by: Adrien Abecassis — Paris Peace Forum Jonathan Barry — Mila – Quebec Artificial Intelligence Institute Ima Bello — Future of Life Institute Yoshua Bengio — Mila – Quebec Artificial Intelligence Institute Antonin Bergeaud — HEC Paris Yann Bonnet — Paris Peace Forum Philipp Hacker — European University Viadrina Ben Harack — Oxford Martin AI Governance Initiative Sophia Hatz — Uppsala University Joachim Henkel — Technische Universität München (TUM) Holger H. Hoos — RWTH Aachen University, Leiden University Kit Kitamura — Independent expert Ranjit Lall — University of Oxford Yann Lechelle — Probabl Constance de Leusse — AI & Society Institute (ENS - PSL) Charles Martinet* — Oxford Martin AI Governance Initiative, Centre pour la Sécurité de l’IA (CeSIA) Nicolas Miailhe — AI Safety Connect, PRISM Eval Julia C. Morse — Oxford Martin AI Governance Initiative, University of California, Santa Barbara Maximilian Negele — Oxford Martin AI Governance Initiative Kyung Ryul Park — Korea Advanced Institute of Science and Technology Miro Pluckebaum — Oxford Martin AI Governance Initiative Murielle Popa-Fabre — AI & Society Institute (ENS - PSL) Benjamin Prud’homme — Mila – Quebec Artificial Intelligence Institute Yohann Ralle — The Future Society Mark Robinson — Oxford Martin AI Governance Initiative Charbel-Raphael Segerie — Centre pour la Sécurité de l’IA (CeSIA) José-Ignacio Torreblanca — European Council on Foreign Relations Lucia Velasco — Oxford Martin AI Governance Initiative K. VijayRaghavan — Former Principal Scientific Advisor to the Government of India *Lead drafter The authors are grateful to Toby Ord, Milo Rignell, Philip Fox, Lily Stelling, and Rafael Andersson Lipcsey for their comments and suggestions. A Blueprint for Multinational Advanced AI Development Thesis: an international advanced AI research & development partnership of AI bridge powers1(1) can feasibly produce frontier AI models; and (2) is essential for safeguarding the sovereignty, democratic values, economic competitiveness and growth, technical innovation, and national security of these bridge power states. Executive Summary The global race to develop advanced AI has entered a new phase marked by staggering investments, rapid technical breakthroughs, and intensifying geopolitical competition. The United States now controls approximately 75% of global AI compute capacity, China 15%, and the EU 5%.2This concentration of compute, alongside concentrations of AI development talent, data, and AI model ownership suggests that mid-sized economies likely face insurmountable barriers to independent frontier AI development. At the same time, economic, cultural, and security infrastructures are coming to rely ever more on frontier models. States that are unable to develop their own frontier models or access the computing hardware required to train them will have to choose between dependency and weakness: •Dependency: if states adopt U.S. or Chinese AI systems, these frontier AI states3can then exploit their privileged position in ways that harm dependent states, for example through data theft, service restrictions, selectively withholding frontier capabilities, embedding values in foundation models, and unfavorable terms of trade. •Weakness: if, on the other hand, states limit their adoption of frontier systems to avoid dependency, frontier AI states may achieve breakthrough capabilities—in economic productivity, in scientific discovery, in military operations—that create widening gaps in economic and military capabilities. Yet, mid-sized economies are also AI bridge powers, possessing substantial AI development capabilities and resources that, if combined, would allow them to challenge the status quo. By working together and strategically choosing their AI development approaches, AI bridge powers can develop competitive frontier models: •First, pooled computing infrastructure can support frontier-scale development. Coordinated deployment of existing, planned, and within-reach European and other bridge power AI compute capacity is likely to provide sufficient computational resources to produce frontier AI models in the next few years, although significantly more investments are probably required to keep up with the moving frontier. •Second, a significant portion of top AI talent has ties to AI bridge power countries. 87 of the 100 most-cited AI researchers originate from or currently work in countries outside the United States and China. Bridge powers could “call home” leading researchers if they had an inspiring vision backed by sufficient resources and an ethical development path. •Third, while most of the data used to train frontier models is already public, bridge powers could pool domain-specific data and resources for data cleaning and expert labeling efforts. •Fourth, bridge powers should make strategic, frontier development bets, leveraging shared digital infrastructures (e.g. pooled pre-training) and R&D efforts to focus on promising areas that do not rely on matching scale elsewhere, in order to reach and then track or even surpass the AI frontier. •Fifth, building reliable AI represents an unmet market need where bridge powers have structural advantages. High-value industries require control over AI tools and confidence in their reliability before deploying them at scale. Bridge powers can act as trusted brokers by leveraging strong 3 A Blueprint for Multinational Advanced AI Development data protection regimes, robust rule of law, and responsive governance to speed up sustainable adoption. A multinational partnership could enable members to preserve sovereignty, have more weight in shaping global AI governance, and lead through ethical stewardship. Some precedents of similar multilateral projects exist through CERN or Airbus4, and the capabilities exist through collective action. The question is then whether bridge powers will act decisively before dependencies deepen and the bipolar structure consolidates. 4 A Blueprint for Multinational Advanced AI Development Contents A. The Strategic Downside of Bipolar Frontier AI 6 B. A Multinational Frontier AI Partnership 7 C. Feasibility and Timeliness of a Multinational Partnership 10 D. Reversing Strategic Vulnerabilities: Benefits for Member Countries 11 5 A Blueprint for Multinational Advanced AI Development Introduction The strategic implications of concentrated AI investment are now undeniable. With U.S. companies expected to spend over $300 billion and China nearly $100 billion on AI infrastructure in 2025, the computational divide has become a chasm: American control of approximately 75% of global AI compute capacity, combined with talent, data, and AI model ownership concentrations, creates barriers that individual mid-sized economies cannot overcome through national efforts alone. The question is no longer whether bridge powers face disadvantages, but whether they will act collectively before those disadvantages harden into structural constraints. G7 minus US 79,000 China 231,000 USA 1,079,000 Figure 1: Global Distribution of AI Compute in October 2025 (H100 equivalents)5 However, AI bridge powers have substantial AI development capabilities of their own. By partnering strategically and coordinating their investments, compute, talent, data, and governance, they can participate in the frontier of AI development and safeguard their sovereignty and values. A. The Strategic Downside of Bipolar Frontier AI AI is poised to become the defining asset of the 21st century. Frontier AI systems already demonstrate rapidly improving abstract thinking and reasoning skills, and match or exceed human experts across wide-ranging domains. They are increasingly enmeshed with economic, cultural, scientific, and security processes, directly and through myriad applications,6but this may just be the tip of the iceberg if AI advances continue according to current trends. Many experts believe that superhuman general AI will be achieved within the next 5–10 years. While such claims are highly uncertain, the pace of development over the past two years suggests that human-level AI is plausible in the near future. AI advances are likely to undermine the sovereignty of non-frontier states. If states cannot independently develop, train, or modify frontier AI systems, they may find themselves faced with a choice between two different forms of vulnerability: either buy AI systems from frontier states and become structurally dependent on them (dependency), or be left behind by frontier states altogether (weakness).7 In a dependency scenario, non-frontier states are dependent on frontier states’ AI systems. This allows frontier AI states to exploit the power differential between them and the states that depend upon them. For example, frontier states could copy sensitive data exchanged with their AI models and use it for economic or political advantage.8They could also selectively modify AI access, threatening service degradation or even a full cut-off;9and their values and design choices will be embedded in foundation models, impacting all downstream applications.10 Dependency may also make terms of trade more unfavorable to non-frontier states. Efforts to limit adoption in order to avoid dependency may lead instead to weakness: a world where frontier states use AI to achieve breakthroughs in economic productivity, scientific discovery, and military operations, allowing them to build out asymmetric capabilities (such as AI-enabled cyber operations that make conventional defenses obsolete11) or automate core economic and security 6 A Blueprint for Multinational Advanced AI Development functions. Such gaps could widen dramatically as AI models are used to further accelerate their own development. Bridge powers thus face a strategic choice between fundamentally different approaches to accessing frontier AI. Each strategy implies having access to sufficient computing power12, and involves distinct trade-offs between capability, sovereignty, and cost: Table 1: Strategic Alternatives for Bridge Power Access to Frontier AI13 Strategy Frontier-competitive? Sovereignty protected? Financially viable? Import closed models -Yes, but several months behind; access can be restricted qVulnerable to service denial, licensing restrictions ¥Low upfront cost; high ongoing dependency Adopt open models qBehind frontier; released open models lag more than 6 months behind closed models, which could grow due to national security restrictions -Reduces but doesn’t eliminate foreign dependencies ¥Low direct cost National champions -Behind frontier; fragmented efforts -Partial sovereignty; may require significant foreign ownership qExpensive per country; duplicates infrastructure Multinational partnership ¥Achievable through pooled resources and strategic development choices ¥Reinforced sovereignty through collective governance, guaranteed AI access, improved domestic AI ecosystems ¥Shared costs; economies of scale As illustrated by Table 1’s first three strategies, importing, adopting, or domestically developing frontier AI each involve unacceptable trade-offs: a multinational partnership is the only viable path to achieving frontier competitiveness while preserving sovereignty at manageable cost. Although they differ in many ways, AI bridge powers have common interests in safeguarding sovereignty and protecting their values and way of life. These common interests imply that there is potential for international cooperation on frontier AI development. B. A Multinational Frontier AI Partnership A multinational frontier AI partnership offers a viable and scalable strategic option. A joint AI bridge power partnership has a much greater chance of reaching the technological frontier than individual bridge powers or national champions operating alone. Advantages include pooled computing infrastructure, talent, and data. Advantage 1: Pooled Computing Infrastructure The logic of compute pooling is rooted in the fundamental economics of AI development. AI inference, or use, involves costs that are decentralized and scale proportionally with the number of users in each country. In contrast, AI training and development costs are unrelated to the number of users in a country. Those costs require a massive, concentrated expenditure of compute resources. This economic reality means that by pooling resources to cover the high, fixed costs of training a frontier model, a group of bridge powers can achieve a level of capability and competitiveness that no single member could reach independently. While individual bridge powers cannot match the scale of initiatives like OpenAI’s Stargate project (>$100B/year announced), groups of bridge powers collectively possess significant datacenter capacity.14 Frontier AI development costs may exceed several billions per model by 2028 (see Fig7 A Blueprint for Multinational Advanced AI Development ure 2), and the required infrastructure investments will require tens of billions. Given this, the costs of maintaining a frontier AI program under existing paradigms using only domestic compute—building not just one model, but maintaining frontier competitiveness through continuous R&D15—are arguably of an order that no single bridge power can sustain on its own, short of shifting into a war-time type of economy with politically untenable costs.16 0 5B$ 10B$ 15B$ 20B$ 25B$ 0.5B$ 1.1B$ 2.8B$ 6.6B$ Projected Cost of Leading Training Run 3B$ 5.7B$ 10.8B$ 20.6B$ Projected Cost of Leading Compute Cluster 2025 2026 2027 2028 Figure 2: Projected Hardware and Training Costs for State-of-the-Art AI Systems (USD, Billions)17 Sources: Pilz et al., 2025;Cottier et al., 2024;Epoch AI, 2025. The infrastructure foundation is already taking shape: in the EU, AI Factories and public supercomputers like Jupiter (Germany, 24,000 GPUs, operational) and Alice Recoque (France, exascale, 2026) are deploying near-term capacity, while in the mid-term, five Gigafactories will deliver 100,000+ specialized AI chips each by 2027.18 Coordinated through a multinational partnership, these assets—valued at €20+ billion in EU commitments alone—could support frontier development at scale. In the short term, it is neither feasible nor necessary for bridge powers to match the scale of U.S. or Chinese investment. Indeed, frontier competitiveness derives not just from the sheer amount of resources, but how they are deployed, as demonstrated by DeepSeek and Mistral, which, through architectural innovations and strategic focus, achieved performance below but comparable to frontier models while spending less. By coordinating investments and adopting novel approaches, a multinational partnership can achieve outcomes at or near the frontier without matching the spending in leading countries.19 Furthermore, a targeted focus can alleviate resource requirements:20 rather than attempting to eliminate all dependencies across the AI stack simultaneously, the partnership could concentrate on developing frontier models with general capabilities in reasoning, planning, trustworthiness, and multimodal understanding. Downstream applications can leverage these, and they would be valuable cards to bridge powers in future negotiations around global AI geopolitics. The multinational partnership should aim to minimize its use of compute located in non-member states and compute owned by foreign entities. Although it could begin by renting compute, this reinforces and creates substantial vulnerabilities over time.21 Bridge power investments in AI infrastructure must therefore continue and increase significantly. It is important to highlight the key role within the AI stack of frontier AI algorithms for capability advances: if capabilities approach or surpass human-level as current trends suggest, frontier AIs themselves will enable rapid innovation and improvements to the other components of the stack. Advantage 2: Pooled AI Talent AI bridge powers have good prospects of recruiting and retaining significant AI talent through preexisting ties to AI researchers. Evidence suggests that a substantial pool of elite AI talent is needed to reach the frontier. Despite an abundance of compute and data resources, companies like 8 A Blueprint for Multinational Advanced AI Development Meta appear to have struggled to reach the frontier in part because they lack a critical mass of leading researchers. At the same time, of the 100 most cited AI researchers in the world, 87 come originally from countries other than the U.S. and China, or are currently working in them.22 Thus, a group of bridge powers might “call home” a large proportion or even a majority of leading researchers, drawn by substantial salaries and the chance to work on civilization-defining questions within collective governance structures. The brain drain is reversible when infrastructure meets values and a galvanizing, visible, and feasible project. Advantage 3: Pooled Data Data pooling offers a significant advantage to partnership members. While most of the data needed for frontier AI R&D is publicly available on the internet, the most expensive and hard-to-get data, which comes from expert labeling and annotation, is usually privately generated and owned. Bridge powers can thus pool the cost of data cleaning and labeling efforts, which companies in frontier AI states fund independently at massive scale. Furthermore, some companies may be willing to share their data with the partnership through preferential access or licensing arrangements. Combined with the proprietary datasets of member nations, and in some cases those shared by their domestic industries, these pooled data resources would provide a scale and differentiation advantage. To attract a diverse and durable pool of AI researchers, a partnership would require an inspiring vision, credible and sufficient capacity and resources, and exciting, leading-edge tasks. Bridge powers could jointly offer this combination by pooling their talent and computing infrastructure. They could also offer access to substantial portions of the world’s AI data. Together, these three resources—compute, talent, and data—comprise the core inputs for frontier development. Adopting or replicating existing AI models are no silver bullets: One might think that bridge powers need not develop frontier models themselves, but could instead adopt open-weight releases (like Llama) or rapidly replicate frontier advances from behind (“fast-following”). For many commercial applications, this may prove adequate, since some open models now nearly match closed frontier models in certain capabilities, with only a 3-month lag. Yet, this strategy faces fundamental limitations: 1. Licensing and access restrictions impose limits on use. Foreign closed-source models offer no guaranteed access, with providers able to restrict, degrade, or terminate service.23 But even “open source” models carry licensing restrictions: Meta’s Llama prohibits most military applications, while others may impose commercial-use limitations. Defense planning and critical infrastructure should not rely on systems where legal access may be revoked or denied in the first place. 2. The most advanced closed-source models have reportedly already exceeded national security risk thresholds that trigger strong security measures to protect their weights.24 When open-weight models approach that point, there will likely be strong pressure from national security agencies in frontier states to prevent their release to mitigate the risks that these models be weaponized in dangerous ways. This would allow the gap with leading closed-source models to grow. Historical cases of other dual-use technologies (e.g., nuclear technology,cryptography) show how logics of national security and industrial advantage justified strict controls on international transfer. But AI differs in scope: in addition to security, frontier models may underpin a substantial portion of economic activity, aggravating the consequences of dependence or exclusion. Countries controlling frontier AI will have limited structural incentives to diffuse these capabilities globally. 9 A Blueprint for Multinational Advanced AI Development Bibliography •Adamson, G., & Allen, G. C. (2025). Opportunities to Strengthen U.S. Biosecurity from AI-Enabled Bioterrorism: What Policymakers Should Know.https://www.csis.org/analysis/opportunities-stren gthen-us-biosecurity-ai-enabled-bioterrorism-what-policymakers-should •Anthropic. (2024). Powering the next generation of AI development with AWS.https://www.anthro pic.com/news/anthropic-amazon-trainium •Anthropic. (2025a). Activating AI Safety Level 3 protections.https://www.anthropic.com/news/act ivating-asl3-protections •Anthropic. (2025). Disrupting the first reported AI-orchestrated cyber espionage campaign.https: //assets.anthropic.com/m/ec212e6566a0d47/original/Disrupting-the-first-reported-AI-orchest rated-cyber-espionage-campaign.pdf •ARIA. (n.d.). Safeguarded AI.https://aria.org.uk/opportunity-spaces/mathematics-for-safe-ai/sa feguarded-ai •Belfield, H. (2025). Domestic frontier AI regulation, an IAEA for AI, an NPT for AI, and a US-led Allied Public-Private Partnership for AI: Four institutions for governing and developing frontier AI (No. arXiv:2507.06379; Version 1). arXiv. https://doi.org/10.48550/arXiv.2507.06379 •Bengio, Y. (2025). Introducing LawZero. Yoshua Bengio.https://yoshuabengio.org/2025/06/03/ introducing-lawzero/ •Chavez, P. (2024). Sovereign AI in a Hybrid World: National Strategies and Policy Responses. Lawfare.https://www.lawfaremedia.org/article/sovereign-ai-in-a-hybrid-world--national-strat egies-and-policy-responses •Clymer, J., Duan, I., Cundy, C., Duan, Y., Heide, F., Lu, C., Mindermann, S., McGurk, C., Pan, X., Siddiqui, S., Wang, J., Yang, M., & Zhan, X. (2025). Bare Minimum Mitigations for Autonomous AI Development (No. arXiv:2504.15416). arXiv. https://doi.org/10.48550/arXiv.2504.15416 •Congressional Research Service. (2001). Encryption Export Controls.https://web.archive.org/we b/20190228095041/http://www.au.af.mil/au/awc/awcgate/crs/rl30273.pdf •Cottier, B. (2024). How far behind are open models? Epoch AI. https://epoch.ai/blog/open-model s-report •Cottier, B., Rahman, R., Fattorini, L., Maslej, N., Besiroglu, T., & Owen, D. (2025). The rising costs of training frontier AI models (No. arXiv:2405.21015). arXiv. https://doi.org/10.48550/arXiv.2405. 21015 •Donnelly, J. M. (2022). Cost to rebuild US semiconductor manufacturing will keep growing.https: //techxplore.com/news/2022-02-rebuild-semiconductor.html •Emberson, L., & Somala, V. (2025). Build times for gigawatt-scale data centers can be 2 years or less. Epoch AI.https://epochai.substack.com/p/build-times-for-gigawatt-scale-data •Epoch AI. (2024). Machine Learning Trends. Archive.Ph. https://archive.ph/Nzogc •EuroHPC JU. (2024). New Call to Procure the European Exascale Supercomputer, Alice Recoque.https: //www.eurohpc-ju.europa.eu/new-call-procure-european-exascale-supercomputer-alice-recoq ue-2024-09-09_en •European Commission. (2025). AI Factories.https://digital-strategy.ec.europa.eu/en/policies/a i-factories •Farrell, H., & Newman, A. L. (2019). Weaponized Interdependence: How Global Economic Networks Shape State Coercion. International Security,44(1), 42–79. https://doi.org/10.1162/isec_a _00351 •Goodin, D. (2025). New hack uses prompt injection to corrupt Gemini’s long-term memory. Ars Technica. https://arstechnica.com/security/2025/02/new-hack-uses-prompt-injection-to-corru pt-geminis-long-term-memory/ •Gooding, M. (2025). Cities across Europe line up to host EU’s AI “gigafactories.” https://www.datace nterdynamics.com/en/news/cities-across-europe-line-up-to-host-eus-ai-gigafactories/ •Google Scholar. (n.d.). Researchers tagged “Machine Learning.” https://scholar.google.com/citatio ns?view_op=search_authors&hl=en&mauthors=label:machine_learning&before_author=vsT-_ 16 A Blueprint for Multinational Advanced AI Development _H1DgAJ&astart=0 •IEEE Spectrum. (2025). Europe’s First Exascale Supercomputer JUPITER Powers Science.https: //spectrum.ieee.org/jupiter-exascale-supercomputer-europe •International AI Safety Report. (2025). First Key Update: Capabilities and Risk Implications.https:// internationalaisafetyreport.org/publication/first-key-update-capabilities-and-risk-implications •Kleinman, Z. (2025). What caused the AWS outage—And why did it make the internet fall apart? https://www.bbc.com/news/articles/cev1en9077ro •Knowledge Centre Data & Society. (2025). Invest AI initiative. EN. https://data-en-maatschappij. ai/en/publications/europese-commissie-invest-ai •Lambert, N. (2023). DeepSeek V3 and the cost of frontier AI models.https://www.interconnects.ai /p/deepseek-v3-and-the-actual-cost-of •Lee, S. (2025). Securing the Backbone of Artificial Intelligence: Protecting Data Centers. New America. http://newamerica.org/future-security/reports/securing-the-backbone-of-ai/ •Maletta, G. (2025). Non-proliferation, Nuclear Technology and Peaceful Uses: Examining the Role and Impact of Export Controls. •McGuire, S., & Islam, N. (2015). Indigenous technological capabilities, emerging market firms and the aerospace industry. Technology Analysis & Strategic Management,27(7), 739–758. https: //doi.org/10.1080/09537325.2014.959482 •Moretti, E. (2021). The Effect of High-Tech Clusters on the Productivity of Top Inventors. American Economic Review,111(10), 3328–3375. https://doi.org/10.1257/aer.20191277 •OpenAI. (2025b). Announcing The Stargate Project.https://openai.com/index/announcing-the-s targate-project/ •Pancevski, B. (2020). Officials Say Huawei Can Covertly Access Telecom Networks. Wall Street Journal.https://www.wsj.com/articles/u-s-officials-say-huawei-can-covertly-access-telecom-n etworks-11581452256 •Pang, J., & Pomfret, J. (2024). Chinese researchers develop AI model for military use on back of Meta’s Llama. Reuters. https://www.cnbc.com/2024/11/01/chinese-researchers-build-ai-model -for-military-use-on-back-of-metas-llama.html •Pilz, K. (2025). Data on GPU clusters. Epoch AI. https://epoch.ai/data/gpu-clusters •Pilz, K. F. (2025). The US hosts the majority of GPU cluster performance, followed by China. Epoch AI. https://epoch.ai/data-insights/ai-supercomputers-performance-share-by-country •Pilz, K. F., Sanders, J., Rahman, R., & Heim, L. (2025). Trends in AI Supercomputers (No. arXiv:2504.16026). arXiv. https://doi.org/10.48550/arXiv.2504.16026 •Rosenberg, S. (2025). Mark Zuckerberg fuels AI’s talent lottery with high pay offers. Axios. https: //archive.ph/2FYgO •Sett, G. (2024). How AI Can Automate AI Research and Development.https://www.rand.org/pubs/ commentary/2024/10/how-ai-can-automate-ai-research-and-development.html •Shen, X. (2025). China’s AI capital spending set to reach up to US$98 billion in 2025: BofA. South China Morning Post. https://www.scmp.com/tech/tech-war/article/3315805/chinas-ai-capital-s pending-set-reach-us98-billion-2025-amid-rivalry-us •Smith, T. (2023). ‘The economics of trading equity for compute are not great’—Mistral releases its first model. Sifted. https://sifted.eu/articles/mistral-releases-first-ai-model/ •Stanford HAI. (2025). The 2025 AI Index Report.https://hai.stanford.edu/ai-index/2025-ai-index -report •Stevens, R., Matsuoka, S., Valero, M., Catlett, C., Foster, I., & Gagliardi, F. (2025). The Trillion Parameter Consortium (TPC). •Subin, S. (2025). Tech megacaps plan to spend more than $300 billion in 2025 as AI race intensifies. CNBC. https://www.cnbc.com/2025/02/08/tech-megacaps-to-spend-more-than-300-billion-i n-2025-to-win-in-ai.html •Suominen, A., Kauppinen, H., & Hyytinen, K. (2021). ‘Gold’, ‘Ribbon’ or ‘Puzzle’: What motivates researchers to work in Research and Technology Organizations. Technological Forecasting and 17 A Blueprint for Multinational Advanced AI Development Social Change,170, 120882. https://doi.org/10.1016/j.techfore.2021.120882 •Tan, J., Jackson, B., Berjon, R., & Coyle, D. (n.d.). Airbus for AI: A global strategy for public value creation. Bennett School for Public Policy, University of Cambridge.https://publicai.co/airbus-for -ai.pdf •Temple-Raston, D. (2021). A “Worst Nightmare” Cyberattack: The Untold Story Of The SolarWinds Hack. NPR.https://www.npr.org/2021/04/16/985439655/a-worst-nightmare-cyberattack-the-u ntold-story-of-the-solarwinds-hack •The Ezra Klein Show. (2024). What if Dario Amodei Is Right About A.I.? The New York Times. https://www.nytimes.com/2024/04/12/opinion/ezra-klein-podcast-dario-amodei.html •Tiku, N. (2024). Top AI researchers say OpenAI, Meta and more hinder independent evaluations. The Washington Post.https://www.washingtonpost.com/technology/2024/03/05/ai-research-l etter-openai-meta-midjourney/ •Todd, B. (2025). When do experts expect AGI to arrive? 80,000 Hours.https://80000hours.org/2 025/03/when-do-experts-expect-agi-to-arrive/ •University of Bristol. (2025). UK-France Summit: University of Bristol to lead a supercomputing partnership with France.https://www.bristol.ac.uk/news/2025/july/france-ai-partnership.html •Verification for International AI Governance. (n.d.). Oxford Martin AIGI.https://aigi.ox.ac.uk/publ ications/verification-for-international-ai-governance/ •Vipra, J., & Korinek, A. (2023). Market Concentration Implications of Foundation Models (No. arXiv:2311.01550). arXiv. https://doi.org/10.48550/arXiv.2311.01550 •Wang, Z., Shi, T., He, J., Cai, M., Zhang, J., & Song, D. (2025). CyberGym: Evaluating AI Agents’ Real-World Cybersecurity Capabilities at Scale.https://rdi.berkeley.edu/blog/cybergym/ 18 A Blueprint for Multinational Advanced AI Development 19