Artificial intelligence (AI) is best understood not as a stand-alone technology but as a global ecosystem built on interdependent layers. Control rests in the physical infrastructure where critical inputs such as computing hardware are scarce, and sourced from a tightly concentrated supply chain. This concentration shapes the capabilities of AI models while its effects on downstream applications are more diffuse as diversity expands further along the chain. This analysis focuses on frontier (including generative), large-scale AI models whose training and deployment depend on global computing, cloud and data infrastructures, where fragilities and hidden costs are most pronounced. With AI now at the centre of geopolitical competition and industrial policy – from the US–China contest over advanced semiconductor chips to the UK’s Bletchley Declaration – the key question is how governments and industry can build resilience across an increasingly fragile AI ecosystem. 

At the foundations of the AI ecosystem are the design of graphics processing units (GPUs) and the fabrication of advanced semiconductors, which are concentrated amongst only a few companies: US-based NVIDIA and Broadcom, Taiwan-based TSMC and the Dutch multinational corporation ASML. This concentration creates chokepoints vulnerable to export controls, tariffs or technical delays. The ‘cloud’, comprising vast interconnected data centres, is also central to the AI ecosystem. Three US-based hyperscalers – Amazon, Microsoft and Google – together account for more than 60% of cloud infrastructure worldwide, creating both vendor lock-in and geopolitical exposure. The severity of these supply-chain vulnerabilities varies: the US dominates cloud services but relies on global supply chains for semiconductors and the critical materials needed to produce them; Europe depends heavily on foreign providers across both layers; and China, while constrained by export controls on advanced semiconductors, retains a strategic advantage in upstream inputs such as rare-earth minerals and component manufacturing. That said, these players depend on the same complex, globally distributed supply chains that underpin AI development and deployment.

Data access is becoming increasingly uneven. The amount of data produced worldwide continues to grow rapidly, but access to the high-quality, large-scale datasets needed to train advanced AI models is narrowing. Large, vertically integrated firms such as Amazon and Google generate, collect and control vast proprietary datasets through their search, retail and cloud platforms, giving them a structural advantage over smaller players. Smaller developers often cannot match this access, and the ongoing tightening of data-protection and sovereignty rules further restricts what can be collected, shared or exported. Copyright disputes such as Anthropic’s US$1.5 billion settlement with authors, privacy restrictions such as Italy’s temporary ban on ChatGPT, and national data-sovereignty regimes in China (Personal Information Protection Law–PIPL; Data Security Law–DSL; Cybersecurity Law–CSL), the EU (General Data Protection Regulation–GDPR; AI Act) and India (Digital Personal Data Protection Act–DPDP) illustrate how the space for freely gathering and using data is shrinking. 

Model development is also highly concentrated. A handful of frontier AI labs such as OpenAI, Anthropic, Google DeepMind and China’s DeepSeek dominate the development of the most capable generative models. While these firms currently lead in scale and resources, the question of who leads truly cutting-edge research remains open, as new entrants and national research programmes compete to close the gap.

Most organisations will not build their own models but will use AI through commercial applications built by these frontier firms. However, these downstream systems depend on the same hardware, cloud infrastructure and data pipelines, so when problems arise at the top of the stack, their effects are inherited by every user. Costs and friction are already arising as energy demand strains grids, water use sparks local opposition, carbon emissions draw regulatory scrutiny, and talent shortages slow deployment. The lesson is clear: frontier AI is not plug-and-play; it depends on an ecosystem that must be managed with the same resilience, efficiency and governance as any other strategic infrastructure.

The fragility of the AI ecosystem is most visible when supply-chain chokepoints impact on environmental or operational constraints. Hardware shortages have affected projects worldwide: in May 2024, a South Korean supplier to NVIDIA, SK Hynix, announced that almost all the high-bandwidth memory chips it could produce in 2024 and 2025 had already been prepurchased. In February 2025, OpenAI delayed the rollout of GPT-4.5 to ChatGPT Plus users due to GPU scarcity. 

The strain on cloud infrastructure is another pressure point. In Ireland, where data centres consumed over 20% of national electricity in 2024, community opposition has intensified, forcing the government to weigh the economic gains of being Europe’s leading hub against environmental limits. Similar pushback, over water use, has derailed Google projects in Uruguay and Chile, while in Malaysia nearly one-third of applications to establish data centres in the state of Johor have been rejected for failing to meet sustainability standards. To manage these pressures, firms such as Microsoft are now building or financing their own energy-generation capacity, including on-site gas plants and agreements to secure zero-emissions green-hydrogen power for one of their European data centres. While this may ease grid strain in the short term, it raises broader questions about who controls critical energy infrastructure, how emissions from private generation are accounted for, and whether national regulators can keep pace with the scale and speed of AI-driven demand.

Another problem is data access. The EU’s AI Act now requires model providers to disclose sufficiently detailed summaries of their data, while regulators in Singapore, India and China are, for varying reasons, tightening rules on personal-data use and cross-border data transfers. These measures primarily affect access to large-scale datasets for model training, where restrictions on source material and data movement can slow or fragment development. Local firms that use customer-feedback loops to fine-tune or adapt existing models within national boundaries may be less affected, but they still face higher compliance costs and reduced flexibility in sourcing or sharing data internationally. The overall effects are rising costs, operational fragmentation and uneven access to AI capabilities across regions.

Security risks compound these vulnerabilities. Recent research has shown that even trusted enterprise systems can be exposed to prompt-injection weaknesses (i.e., where hidden instructions mislead the model), with demonstrations revealing how Slack AI and OpenAI connectors could be manipulated to leak private data. While there have been no confirmed large-scale prompt-injection attacks, cyber criminals are increasingly using uncensored or jailbroken (i.e., modified to bypass built-in safety and content restrictions) large language models, such as WormGPT and Mixtral, to develop and distribute malicious code. These trends highlight how vulnerabilities across the ecosystem – from insecure integrations to misuse of open models – can expose firms to legal, operational and reputational risks, while placing consumers at risk of privacy breaches or manipulated outputs. Without proactive security design, continuous testing and robust governance across the infrastructure, data and application layers, there is a risk that AI ambitions will be slowed or fragmented by cascading vulnerabilities. Emerging efforts such as the Bletchley Declaration and the UK’s Safe AI initiative mark early attempts to build shared standards for model safety, security and transparency, but they are works in progress and largely voluntary. Embedding these principles across the global AI ecosystem will be essential in order to move from mere aspiration to enforcement.