AI Series: The Downside
“Every form of refuge has its price.”
— The Eagles, “The Last Resort”, 1976The productivity dividend from AI is real. So is the cost. Technologies do not arrive as pure gifts — they arrive with disruptions, dependencies, and power asymmetries that are rarely visible in the initial deployment phase. The honest accounting of AI requires examining both sides of the ledger with equal rigour.
Labour Market Disruption: The Jobs Question
The labour market impact of AI is the most politically sensitive and analytically contested dimension of the technology. The historical debate between technological optimists (who argue that automation always creates more jobs than it destroys, as each prior wave has done) and structural pessimists (who argue that cognitive AI is fundamentally different from physical or narrow automation) has not been resolved. What can be said with confidence is that the magnitude and speed of current disruption exceeds prior technology transitions by a significant margin.
The most structurally damaging aspect of AI disruption may not be aggregate job displacement but the hollowing out of entry-level roles in white-collar professions. Junior analyst, paralegal, entry-level coder, and customer service positions are the roles where professionals acquire the foundational skills that enable career progression. If AI absorbs those tasks, the question is not merely “how many jobs are lost?” but “where do professionals build the judgment that comes from doing those tasks?” This pipeline disruption has no clean historical precedent.
Legal research and document review — tasks that junior associates bill at high hourly rates but can now be performed by AI systems in minutes — represent a material portion of law firm revenue at the entry level. Thomson Reuters’ CoCounsel and Harvey.ai are already deployed in major law firms. The structural question is whether the billable hour model survives intact when the tasks underlying it can be automated. Early evidence suggests it is being restructured rather than immediately collapsed: firms are pricing AI-assisted work differently while retaining human oversight requirements, and using AI productivity gains to expand volume rather than purely to reduce headcount. But the trend direction is not in doubt.
Investment banking and equity research face a similar dynamic. AI systems can now produce a credible first-draft earnings preview, sector comparison, or discounted cash flow model in minutes — tasks that previously occupied significant junior analyst bandwidth. The productivity gain is real for firms deploying these tools. The career impact on the analyst development pipeline is an unresolved question with long-term implications for the quality of senior research talent a decade from now.
Software engineering is the most instructive case study because it is the domain where AI capability is most advanced, deployment is most widespread, and hiring data is most transparent. GitHub Copilot, Cursor, Claude Code, and similar tools have materially changed the productivity profile of software development. In 2024–25, multiple major technology companies — Google, Meta, Amazon, Salesforce — reported that AI tools were handling 20–30% of new code generation. Amazon CEO Andy Jassy stated in 2025 that AI tools had already reduced the need for new software engineer hires by a material amount.
This shows up in hiring data: software engineering job postings in the United States fell approximately 30% from their 2022 peak through 2025, even as the broader economy remained near full employment. Some portion of that reflects post-COVID normalisation, but the AI impact is visible in the data. Entry-level software engineering roles — where bootcamp graduates and CS graduates historically built their careers — have been disproportionately affected. The roles that remain are more senior, more architectural, more requiring of judgment that AI tools cannot yet replicate.
Energy, Water & Resource Pressure: The Infrastructure Cost of Intelligence
The energy implications of AI are large enough to materially affect utility sector investment theses, grid planning horizons, carbon trajectory models, and — increasingly — the geopolitics of resource access. They are also among the most underreported dimensions of the AI buildout in financial media, which tends to cover the technology through the lens of the companies building it rather than the infrastructure it consumes. This section examines the electricity demand profile, water stress, semiconductor supply chain resource requirements, and the emerging dynamic by which AI’s physical infrastructure needs are beginning to reshape both corporate strategy and state-level resource policy.
The Electricity Demand Shock
U.S. data centers consumed approximately 4% of total national electricity in 2024. The IEA projects this will more than double by 2030. A single ChatGPT query consumes approximately 10 times the electricity of a standard Google search. Training a large frontier model at GPT-4 scale requires electricity equivalent to the lifetime consumption of roughly 100 U.S. households. These are not abstract statistics — they aggregate into grid planning problems that utility commissions, transmission operators, and capacity market designers are now grappling with in real time.
The geographic concentration compounds the challenge. Data center demand is not distributed evenly across the grid. Northern Virginia (Loudoun County alone hosts more than 25% of the world’s data center capacity) accounts for a disproportionate share of PJM load growth. The PJM capacity auction for 2025–26 saw data centers contribute an estimated $9.3 billion price increase, translating to residential electricity bill increases of $16–18 per month in Ohio and western Maryland. Pew Research cites a Carnegie Mellon University study estimating that data centers and cryptocurrency mining could increase average U.S. electricity bills by 8% by 2030, with local increases exceeding 25% in the highest-demand markets.[1] This is a regressive transfer: household ratepayers in AI-proximate states bear infrastructure costs for technology whose economic benefits accrue globally.
The grid investment response is already underway but faces a multi-year lag. PJM has approved the largest transmission expansion in its history — approximately $50 billion in new transmission investment — with AI demand as the primary driver. Similar programmes are active in ERCOT (Texas), MISO, and SPP. For utility analysts, the investment thesis is straightforward: rate base expansion at regulated utilities serving AI-dense territories is a multi-decade growth story. The risk is demand-side: if AI efficiency improvements or a hyperscaler capex rationalisation reduces data center build rates, utilities face stranded capital in transmission and generation assets that ratepayers ultimately bear.
The most durable utility investment thesis within AI is not generic data center exposure but specifically nuclear-adjacent utilities and those with heavy industrial load in AI-dense territories. Virginia, Texas, Georgia, and Arizona are the primary geographies. Nuclear-powered utilities (Constellation Energy, Dominion Energy with North Anna and Surry) are uniquely positioned: 24/7 dispatchable, carbon-free, and capable of direct PPA structures with hyperscalers that bypass traditional utility rate structures. Microsoft’s 20-year Three Mile Island PPA is the template. Utilities with large renewable portfolios but insufficient dispatchable backup are less attractive — data centers require reliability that intermittent generation alone cannot guarantee.
The Carbon Lock-In Problem
The energy source question is where AI’s environmental paradox sharpens into a genuine policy crisis. AI is being deployed, with genuine effect, to accelerate clean energy materials science, improve grid management efficiency, and sharpen renewable generation forecasts. Simultaneously, the electricity demand it creates is being met, in significant part, by gas-fired generation — because clean alternatives cannot be deployed fast enough to match the pace of data center growth.
Greenpeace East Asia reported a 4.5× year-on-year increase in AI chip manufacturing emissions between 2024 and 2025 — driven by the energy-intensive nature of advanced semiconductor fab processes at TSMC, Samsung, and Intel foundries.[4] This is upstream of the operational electricity story and largely invisible in corporate carbon accounting frameworks that focus on Scope 2 emissions (purchased electricity) rather than Scope 3 (supply chain). The full lifecycle carbon footprint of AI infrastructure — chip fabrication, data center construction, cooling systems, operational power — is substantially higher than the operational electricity figures suggest.
The carbon lock-in dynamic operates through committed infrastructure. When a hyperscaler signs a 15–20 year data center operating agreement, the power demand profile that agreement creates is effectively locked in for that period. If the generating capacity contracted to serve it is gas-fired (because new nuclear and utility-scale solar with storage cannot be commissioned fast enough), that gas demand is similarly locked in. At the aggregate scale of the 2024–2026 hyperscaler capex programme — $400+ billion per year — the locked-in fossil demand being created is a material input to climate transition models that is not adequately reflected in most net-zero scenario analyses.
Water Stress: The Less-Visible Constraint
Data centers cool their servers primarily through evaporative water cooling systems. The water demand is large and geographically concentrated. A University of Houston study estimates that data centers in Texas alone will use 49 billion gallons of water in 2025, rising to 399 billion gallons by 2030 — equivalent to drawing down Lake Mead by over 16 feet annually.[2] ChatGPT reportedly uses approximately 500 millilitres of water per conversation for cooling — a figure that, at 100 million daily active users conducting multiple sessions, aggregates to volumes that are material at the watershed level.
The conflict with agricultural and human consumption is already live in water-stressed geographies. In central Arizona, data centers approved under pre-AI-boom planning assumptions are now competing for water allocations with agricultural users in a state that already overdraws the Colorado River. In Goodyear, Arizona, a proposed data center campus was rejected by the city council specifically because its projected water consumption exceeded available supply — the first prominent example of water, rather than power or planning permission, becoming the binding constraint on AI infrastructure expansion.
The international dimension is starker. In Chile — a leading data center hub for South America — AI infrastructure projects are being built in the Atacama Desert region, where water scarcity is an extreme constraint and where the same lithium extraction operations that supply battery supply chains for AI hardware also consume enormous volumes of water. In India, data centers proposed in water-stressed states such as Maharashtra and Rajasthan face local opposition from agricultural communities that mirrors the solar farm conflicts of the previous decade.
Critical Minerals: The Upstream Resource Dependency
The semiconductor supply chain that powers AI is itself deeply resource-intensive and geographically concentrated. The minerals required for advanced chips, cooling systems, server hardware, and the clean energy infrastructure needed to power data centers sustainably represent a set of supply chain dependencies that are underappreciated relative to the attention given to the compute layer itself.
| Resource | AI Application | Geographic Concentration | Supply Risk |
|---|---|---|---|
| Gallium | Compound semiconductors; GaN chips used in power electronics for data center efficiency | China produces ~80% of global supply; imposed export controls Oct 2023 | CRITICAL |
| Germanium | Semiconductor substrates; optical fibre for data center interconnects | China produces ~60% of global refined supply; included in Oct 2023 export restrictions | CRITICAL |
| Rare Earth Elements | Permanent magnets in cooling fans, electric motors; neodymium in server hardware | China controls ~85% of global rare earth processing capacity | HIGH |
| Cobalt | Lithium-ion batteries for UPS systems; electric vehicles for logistics | DRC produces ~70% of mined cobalt; Chinese firms control major refining | HIGH |
| Lithium | Battery storage for grid balancing supporting AI power demand; EV logistics fleets | Chile, Australia, Argentina (production); China (~65% of processing capacity) | MEDIUM-HIGH |
| Silicon (High-Purity) | Semiconductor wafers; the foundational substrate for all AI chips | Production concentrated in Germany, US, Japan; polysilicon in China | MEDIUM |
| Copper | Data center wiring, power distribution, GPU interconnects | Chile, Peru, Congo; more diversified than others | MEDIUM (demand driven) |
| Water | Data center cooling (primary); semiconductor fab processing water | Locally concentrated in AI data center geographies | HIGH (geographically) |
China’s October 2023 export restrictions on gallium and germanium were not an isolated trade measure — they were a deliberate demonstration of leverage over the semiconductor supply chain. In a geopolitical context where the United States has used chip export controls to constrain Chinese AI capability, China has demonstrated an ability to apply reciprocal pressure through mineral supply chain chokepoints. The strategic calculus is explicit: the same resource dependencies that the AI buildout is creating are becoming instruments of geopolitical coercion. This dynamic connects directly to the Fracture scenario in Part IV and to the broader pattern of resource-competition-driven foreign policy examined there.
The critical minerals supply chain is the most structurally underpriced dimension of the AI investment landscape. Markets are pricing AI through semiconductor and hyperscaler multiples but have not fully priced the upstream resource constraints that a sustained AI buildout implies. Gallium and germanium are small markets — China’s ability to restrict them creates supply shocks that dwarf the commodity prices themselves in their impact on semiconductor economics. The investment case for non-Chinese gallium and germanium producers, rare earth processors outside China’s supply chain, and lithium producers with North American or Australian exposure is partly an AI infrastructure play and partly a geopolitical hedge — and it is significantly less crowded than the obvious semiconductor trade.
The Energy Transition Paradox in Full
The paradox at the heart of the AI energy story is not merely that a technology promising to accelerate clean energy is consuming dirty energy. It is that the investment capital and engineering talent currently being deployed into AI infrastructure — data centers, chips, cooling systems, power distribution — is capital and talent that might otherwise be deployed into grid decarbonisation, battery storage, and clean generation. The opportunity cost is real even if it is invisible in the financial accounts of the companies making these investments.
The IEA’s April 2025 Energy and AI report projects that data center electricity demand will equal the entire current electricity consumption of Japan by 2030 — a remarkable figure that has received far less analytical attention than it deserves. The path to meeting that demand cleanly requires a nuclear renaissance, an acceleration of utility-scale storage, and transmission grid investment at a pace that no major electricity market has achieved in the modern era. The probability that all three materialise simultaneously, at sufficient scale, to prevent the AI energy buildout from adding net carbon load is, in Fenrir’s assessment, low. The Depletion scenario in Part IV is partly a consequence of this dynamic playing out over the next decade.
Capital Concentration: Who Captures the AI Dividend
The distributional question in AI is not merely about jobs. It is about who owns the systems, who captures the productivity surplus they generate, and whether that surplus flows broadly through wages and lower prices or narrows into the accounts of capital owners and equity holders at a handful of dominant firms.
The concentration data is stark. Crunchbase data for 2025 shows that OpenAI and Anthropic alone captured 14% of all global venture investment — two companies out of hundreds of thousands globally.[3] The five hyperscalers (Microsoft, Amazon, Alphabet, Meta, Oracle) account for the overwhelming majority of AI infrastructure capex. The pattern is self-reinforcing: scale in data and compute creates capability advantages that attract more customers, generating revenue that funds further scale investment. The barriers to entry at the frontier model level have grown, not shrunk, as the field has matured.
| Concentration Dimension | Current State | Direction of Travel | Risk Assessment |
|---|---|---|---|
| Foundation Model Developers | ~5–8 credible frontier labs globally | Consolidating; compute cost barriers rising | HIGH |
| Compute Infrastructure | Nvidia ~80%+ GPU market share; 3 hyperscaler clouds | Increasing; custom silicon creates lock-in | HIGH |
| Training Data | Internet-scraped corpora; proprietary user data | Incumbents accumulating proprietary datasets | MEDIUM-HIGH |
| Application Layer | More fragmented; hundreds of startups | Consolidation underway; winners emerging by vertical | MEDIUM |
| Geographic Distribution | US + China dominate; EU/UK/India lagging | Bifurcation risk; sovereign AI emerging | HIGH |
| Talent Pool | Highly concentrated in handful of metro areas | Remote work helps but network effects persist | MEDIUM |
The Productivity Surplus Distribution Problem
The critical macroeconomic question is not whether AI generates a productivity surplus — it does and will at increasing scale — but whether that surplus flows broadly through the economy via lower prices, higher wages, and improved public services, or whether it accumulates as supernormal returns for the companies owning the AI systems. Prior technology waves have generally distributed gains relatively broadly over time, via competition eroding rents. The concern with AI is that network effects, data advantages, and compute scale create moats that competition cannot erode — resulting in persistent monopoly or oligopoly rents that concentrate capital rather than diffusing it.
Misuse, Misinformation & the Regulatory Vacuum
The dual-use problem — that the same capabilities that make AI valuable for legitimate purposes also enable harm at scale — is not a theoretical concern. It is already manifesting across a range of domains where the regulatory framework is inadequate.
The asymmetry between the cost of generating synthetic media and the cost of detecting and debunking it is the core structural problem. A convincing deepfake video can be produced with consumer hardware in minutes; fact-checking it requires forensic expertise, provenance analysis, and platform-level infrastructure that does not exist at scale. The 2024 election cycle saw significant synthetic media interference in multiple democracies — most visibly in Bangladesh, Slovakia, and the United States — demonstrating that the threat is not theoretical.
The disinformation-at-scale problem compounds this. AI tools can generate thousands of unique, topically relevant, grammatically correct pieces of content in an hour — saturating information environments in ways that human content moderation cannot match. The production economics of disinformation have been permanently altered. A state actor, political campaign, or criminal organisation with access to commercial AI tools now has capabilities that five years ago required significant specialist resources.
AI systems trained on historical data inherit and can amplify the biases embedded in that data. The evidence base is established. Facial recognition systems have shown materially higher error rates for darker-skinned individuals in multiple peer-reviewed studies (Buolamwini & Gebru, 2018; NIST, 2019). Recidivism prediction algorithms used in U.S. criminal courts (notably COMPAS) have shown racially disparate outcomes in independent analysis. Credit scoring models trained on zip code and behavioural data can effectively proxy for race in jurisdictions where residential segregation is a historical legacy.
The problem is not merely one of bad intentions by AI developers. It is structural: training data reflects historical human decisions that were often discriminatory, and AI systems that learn to predict those decisions will reproduce their patterns unless explicitly corrected. The correction — removing protected characteristics and their proxies from training data or model inputs — is technically and commercially complex, and currently governed inconsistently across jurisdictions.
Lethal autonomous weapons systems (LAWS) — systems that can select and engage targets without human authorisation — are moving from research programmes to operational deployment faster than international law is developing frameworks to govern them. Drone swarms, autonomous maritime vehicles, and AI-assisted targeting systems are already being used or tested by major military powers. The United States, Russia, China, Israel, and South Korea are all pursuing capabilities in this space.
The accountability gap is the central problem. When an autonomous system makes a lethal targeting decision that results in civilian casualties or a war crime, existing frameworks of international humanitarian law — designed around human decision-makers — do not assign accountability clearly. Campaign to Stop Killer Robots and a growing number of legal scholars argue that weapons systems incapable of human oversight cannot comply with the laws of armed conflict. Negotiations at the UN Convention on Certain Conventional Weapons have stalled, with major military powers resistant to binding constraints.
AI has lowered the skill threshold for cyberattacks in ways that security professionals regard as a step-change in the threat environment. Phishing emails generated by LLMs are grammatically fluent, contextually aware, and nearly indistinguishable from legitimate communications — eliminating the tell-tale signs (poor grammar, generic salutations) that email filters and human recipients relied on. Automated vulnerability scanning and exploit generation tools are accelerating the window between vulnerability disclosure and weaponisation.
The concentration risk in critical infrastructure is underappreciated. As utilities, water systems, financial market infrastructure, and healthcare systems deploy AI tools — often from the same handful of providers — correlated failures become possible at scale. A vulnerability in a widely deployed AI system creates a single attack surface that, if exploited, could affect multiple sectors simultaneously. This is not a hypothetical concern: the 2021 Colonial Pipeline ransomware attack demonstrated how a single technology dependency can cascade into systemic infrastructure disruption.
The Regulatory Patchwork
| Jurisdiction | Framework | Approach | Coverage Gaps |
|---|---|---|---|
| EU | EU AI Act (in force 2024) | Risk-tiered; prohibits certain uses; high-risk systems require conformity assessment | Foundation model obligations limited; enforcement capacity uncertain |
| United States | Executive Order (Oct 2023, partially rescinded 2025); sector-specific guidance | Light-touch, innovation-first; no comprehensive federal AI law | No federal liability framework; no mandatory disclosure; LAWS ungoverned |
| China | Algorithmic Recommendation Regulation; Generative AI Measures (2023) | Sectoral; requires content alignment with “socialist values”; data localisation | Primarily controls political/social use; limited on economic AI harms |
| United Kingdom | Pro-innovation principles; AI Safety Institute | Non-statutory guidance; sectoral regulators apply existing law | No binding rules; reliance on voluntary commitments from frontier labs |
| India | Digital India Act (draft); AI Policy 2023 | Light-touch; focus on opportunity; advisory guidelines only | Comprehensive regulation not yet enacted; significant governance gap |
The Creative Economy: What AI Destroys Before It Builds
The creative economy is experiencing AI disruption in real time, and it is worth distinguishing between the short-run effects (which are destructive for specific categories of workers) and the longer-run picture (which may see AI expand the creative economy while redistributing value within it).
The near-term harms are concentrated and measurable. Stock photography platforms have reported revenue declines of 30–50% in certain categories since the introduction of high-quality text-to-image models. Survey data from the Graphic Artists Guild (2024) found that 87% of illustrators reported AI competition affecting their income. Entry-level copywriting rates have compressed significantly in markets where AI-generated content is accepted by clients. The workers most affected are those doing standardised, repeatable creative tasks at volume — a category that turns out to be quite large within the creative economy.
The legal framework is unresolved at a structural level. The New York Times v. OpenAI litigation, Getty Images lawsuits, and numerous class actions by artists contesting the use of their work in training datasets have not yet produced settled law. The EU AI Act requires disclosure of copyrighted training data, but retroactive remedy for works already incorporated into existing models is legally and practically complex. Until training data compensation frameworks are established, the economic incentive structure systematically favours AI developers at the expense of the human creators whose work trained the systems.
The Cybersecurity Crisis: AI as Weapon, AI as Target
Cybersecurity is the downside risk that every major institution acknowledges and almost none has adequately addressed. AI has not invented new categories of cyberattack — it has fundamentally changed the cost, scale, and sophistication of existing ones, while simultaneously creating new attack surfaces through the AI systems themselves. The threat environment in 2025–26 is not an extension of the 2020 threat environment; it is qualitatively different.
Three changes in the threat environment are structural. First, AI has eliminated the skill floor for sophisticated attacks. Hyper-personalised phishing — emails that reference a target’s recent activity, mirror their writing style, and correctly name their colleagues — was previously the work of skilled social engineers operating at low volume. LLMs now generate thousands of such messages per hour, each contextually accurate, each indistinguishable from legitimate communication. The State of AI Cybersecurity 2026 report identifies hyper-personalised phishing as the top concern (50%), followed by automated vulnerability scanning and exploit chaining (45%) and adaptive malware and deepfake voice fraud (40% each).
Second, AI has compressed the full attack lifecycle. IBM’s X-Force Threat Intelligence Index 2026 found that over five years, major supply chain and third-party breaches quadrupled — reflecting a shift toward targeting interconnected systems and trusted integrations rather than a single organisation’s perimeter. CrowdStrike documented that over 90 organisations had legitimate AI tools exploited to generate malicious commands and steal data, with ChatGPT mentioned in criminal forums 550% more than any other model. Agentic AI — autonomous agents with access to internal data and APIs — is what IBM calls the most helpful insider threat: broad access is required for the agent to function, and that access becomes the attack vector.
Third, nation-state escalation is unambiguous. In January 2026, Salt Typhoon, the China-linked threat actor, was confirmed to have achieved persistent access to US congressional communications. This is not classic espionage; it is systematic infrastructure penetration enabling both intelligence collection and, in an escalation scenario, operational sabotage. The first half of 2025 saw a 65% year-over-year increase in ransomware incidents affecting government bodies — 208 confirmed attacks. Healthcare remains the highest-cost target at $9.77 million per breach, with documented cases of patient harm from ransomware-induced operational disruption.
The systemic risk dimension is the one markets are least pricing. AI-managed power grids, water systems, and financial market infrastructure present concentrated attack surfaces whose disruption would cascade across sectors. The 2021 Colonial Pipeline incident — a single ransomware attack on scheduling software that caused fuel shortages across the US Southeast — is a preview of what concentrated AI infrastructure dependency enables. AI security spending is projected to grow from $25.35 billion in 2024 to $93.75 billion by 2030 at a 24.4% CAGR, but most of that spending will chase yesterday’s threat rather than tomorrow’s attack vector.
Cybersecurity is the clearest anti-fragile AI investment: demand grows under every scenario, driven by the same AI deployment that expands the attack surface. The investment differentiation is between AI-native security platforms — using AI to detect AI-enabled threats across identity, cloud, endpoint, and network in a single integrated architecture — and AI-washed incumbents. Platform consolidation is accelerating: 93% of security professionals now prefer integrated platforms over point products. The platform winners have durable competitive advantages that individual-product vendors cannot match.
The Mineral Race: Resource Competition, Coercion, and the Risk of Conflict
The AI buildout is, at its physical base, a competition for specific rocks. Advanced AI chips require gallium, germanium, and rare earth elements. The clean energy infrastructure to power them requires lithium, cobalt, and copper. The geographical distribution of these materials — and critically, the processing capacity to turn raw ore into usable inputs — is deeply concentrated and, in critical cases, controlled by states that are strategic competitors of the leading AI powers. This is not background context. It is becoming a primary driver of state foreign policy behaviour and a structural risk to the AI investment thesis.
The US State Department stated explicitly at the February 2026 Critical Minerals Ministerial: “Today, this market is highly concentrated, leaving it a tool of political coercion and supply chain disruption, putting our core interests at risk.” The same day, the US signed bilateral critical minerals frameworks with multiple partner nations and launched the Forum on Resource Geostrategic Engagement. The US National Security Strategy, released November 2025, highlighted securing critical mineral supply chains as a matter of national security. In January 2026, President Trump signed an executive order on adjusting imports of processed critical minerals to reduce foreign source reliance.
Critical minerals have become the 21st-century “oil” — lithium, cobalt, rare earths, and semiconductor inputs are now explicit levers of state power, as demonstrated by China’s 2024–2025 export controls and the scramble by the US and allies to secure alternative supplies. Geology drives geopolitics: these minerals are geologically fixed and highly concentrated. By mid-2025, with the two superpowers locked in a trade war, Beijing demonstrated its willingness to weaponise its control of heavy rare earths through a series of export restrictions, though China eventually agreed to suspend implementation until late 2026.
When major producers have announced export controls on gallium, germanium, and graphite, it caused immediate price spikes and panic among manufacturers worldwide — not random market fluctuations, but calculated geopolitical moves creating a dangerous feedback loop of price volatility, disrupted investment, and supply chains used as diplomatic weapons.
The pattern of US foreign policy engagement with resource geographies in 2024–25 reveals a strategic posture that has moved beyond commercial competition into coercion: Venezuela’s coltan and oil, Greenland’s rare earth deposits and Arctic position, DRC cobalt, Ukrainian titanium and lithium deposits in Russian-occupied territories. In each case, the US has employed sanctions, political pressure, military posturing, or direct deal-making. Trump’s stated interest in Greenland — including suggesting military options against a NATO ally’s territory — represents a public articulation of coercive resource logic that would have been inconceivable in the post-Cold War era.
The convergence of these dynamics — AI capability competition, mineral supply concentration, weakening international institutions, and the accelerating pace of technological change — is assembling the structural conditions for resource conflict. This is not a prediction of imminent great power war. It is an observation that the post-WW2 UN-led order, which constrained great power resource competition within diplomatic channels, is visibly under strain in ways that direct economic and security incentives are driving rather than moderating. The AI mineral race is one of several forces simultaneously degrading the institutional friction that historically prevented resource competition from escalating into open conflict.
Markets are pricing AI through semiconductor multiples and hyperscaler revenue growth — neither captures the supply chain disruption risk from a sustained Chinese rare earth embargo, a DRC political collapse, or an Arctic territorial dispute that escalates. The investment response is not to avoid AI exposure but to hold it alongside explicit resource security hedges: non-Chinese rare earth producers, critical mineral ETFs (REMX, LIT), North American lithium developers, and defence infrastructure companies with resource security programme exposure. These positions are cheap insurance against the tail risks most absent from consensus AI positioning.
Bottom Line
The costs of AI are real, concentrated, and structurally underpriced. Labour market disruption in white-collar roles is already visible. The energy and water demands of AI infrastructure create measurable negative externalities. The cybersecurity threat environment has been qualitatively transformed — attack volumes, sophistication, and state-actor capability are all accelerating simultaneously. And the mineral supply chains underpinning the entire buildout are becoming instruments of geopolitical coercion that risk escalating into resource conflicts more serious than markets are pricing.
These are not arguments against AI deployment. They are arguments for clear-eyed accounting of who bears the costs of a transition whose benefits are real but unevenly distributed — and for regulatory, security, and diplomatic frameworks that address these structural failures. Part IV examines where these trajectories lead if extended to their logical conclusions.
Sources & Citations
- Pew Research Center. (Oct 2025). What We Know About Energy Use at U.S. Data Centers Amid the AI Boom. Citing IEA Energy & AI Report (Apr 2025) and Carnegie Mellon University study.
- Lincoln Institute of Land Policy. (Feb 2026). Data Drain: The Land and Water Impacts of the AI Boom. Citing HARC / University of Houston study.
- Crunchbase. (Dec 2025). 6 Charts That Show The Big AI Funding Trends of 2025.
- J.P. Morgan Private Bank. (Apr 2026). Job Destroyer? Here’s What You Need to Know About AI and Labor Markets.
- Goldman Sachs Research. (2023). The Potentially Large Effects of Artificial Intelligence on Economic Growth.
- IMF. (Jan 2024). Georgieva, K. et al. AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity.
- Greenpeace East Asia. (Oct 2025). AI Supply Chain Decarbonisation Report. Via Greenpeace International.
- Buolamwini, J. & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. FAT* Conference Proceedings.
- NIST. (2019). Face Recognition Vendor Test Part 3: Demographic Effects.
- NPR / Erik Brynjolfsson. (Aug 2025). AI Could Widen the Wealth Gap and Wipe Out Entry-Level Jobs.
Leave a Reply