The Physical Layer of AI: Copper, Energy, and Critical Mineral Bottlenecks — Cross-Provider Synthesis

Executive Summary

• Copper is the single most constrained commodity for AI scaling, with a projected structural supply deficit of 10 million metric tons by 2040 (25% below demand), driven simultaneously by AI data centers (up to 50,000 tonnes per hyperscale facility), grid expansion, EVs, and renewables — while mine development timelines average 16–29 years and ore grades have fallen 40% since 1990. Price forecasts range from $12,500–$15,000/tonne by 2026–2035, up from ~$9,500–$11,700 today.

• Electricity demand for AI data centers will roughly double by 2030, from ~415–460 TWh globally in 2024 to 945–1,000 TWh (IEA base case), potentially reaching 1,300–2,000 TWh by 2035 in high-growth scenarios. The U.S. alone faces an 11+ GW capacity shortfall by 2025, with grid interconnection queues averaging 4 years — making transmission infrastructure, not generation, the binding near-term constraint.

• China's processing monopoly over critical minerals constitutes an existential geopolitical risk: China controls 89–91% of rare earth refining, 75% of cobalt refining, 75% of lithium chemical processing, and >90% of graphite processing. The 2025 escalation of export controls on heavy rare earths, gallium, and germanium — with extraterritorial jurisdiction clauses — has already caused factory shutdowns in the U.S. and Europe, demonstrating that this is an active weapon, not a latent risk.

• Water is an underappreciated hard constraint: Large AI data centers consume up to 5 million gallons per day, and ~two-thirds of U.S. data centers built since 2022 are in high water-stress regions. U.S. direct data center water consumption is projected to quadruple from 17 billion gallons (2023) to 68 billion gallons by 2028, with indirect consumption (from power generation) representing an additional 12x multiplier.

• The 2029–2032 window is the critical inflection point: Copper deficits become structural (2–4.5 Mt annual gap), rare earth functional scarcity peaks, regional electricity capacity gaps become acute, and water constraints force facility relocation decisions. Decisions made in 2025–2027 — on power contracts, mineral supply agreements, facility siting, and cooling technology — will determine competitive positions through 2035 and beyond.

Cross-Provider Consensus

1. Copper Supply-Demand Gap Is Structural and Severe

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok, Gemini-Lite (all six) Confidence: HIGH

All six providers independently confirmed that copper faces a structural supply deficit driven by the convergence of AI data center demand, grid expansion, EVs, and renewables. Specific figures vary (see Contradictions section), but the directional consensus is unambiguous: existing mine pipelines cover only ~70% of projected 2035 demand (IEA), with a potential 10 Mt annual deficit by 2040 (S&P Global). All providers noted declining ore grades, 15–29 year mine development timelines, and the inelasticity of AI demand to copper price increases.

2. AI Electricity Demand Will Roughly Double Globally by 2030

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok, Gemini-Lite (all six) Confidence: HIGH

Every provider cited the IEA base case of ~945 TWh global data center electricity demand by 2030 (from ~415–460 TWh in 2024), representing approximately 3% of global electricity. All providers noted that AI workloads are the primary growth driver, with U.S. data centers potentially reaching 8–14% of national electricity demand by 2030. The consensus is that generation capacity is nominally buildable, but transmission infrastructure and grid interconnection queues are the binding near-term constraint.

3. China's Processing Monopoly Creates Systemic Geopolitical Risk

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok, Gemini-Lite (all six) Confidence: HIGH

All providers independently confirmed China's dominant position in critical mineral refining: 85–91% of rare earth processing, 75% of cobalt and lithium refining, >90% of graphite processing. All noted that China has demonstrated willingness to weaponize this position through export controls (2023 gallium/germanium, 2025 heavy rare earths, 2025 battery supply chain controls). All providers flagged this as the highest geopolitical risk to AI supply chains.

4. Water Consumption Is a Localized Hard Constraint

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok (five of six) Confidence: HIGH

Five providers confirmed that large AI data centers consume up to 5 million gallons/day, that ~two-thirds of new U.S. data centers are in high water-stress regions, and that U.S. data center water consumption will roughly double to quadruple by 2028. All noted the indirect water footprint from power generation (often 12x direct consumption) as a critical but underreported factor.

5. Lithium Faces a 38–40% Supply Gap by 2035

Providers: Perplexity, OpenAI, Anthropic, Grok (four of six) Confidence: HIGH

Four providers confirmed IEA projections of a ~40% lithium supply shortfall by 2035 under stated policies scenarios, with demand growing 4–5x from 2021 levels by 2030. All noted Australia (40%), Chile (25–30%), and China (processing, 75%) as the dominant players, and flagged the 2022–2024 price crash as having suppressed investment in new supply, creating conditions for future shortages.

6. Cobalt's DRC Concentration Is a Geopolitical Flashpoint

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok (five of six) Confidence: HIGH

Five providers confirmed DRC controls ~68–76% of global cobalt mining, with China controlling ~75% of refining. The DRC's February 2025 export ban (and subsequent quota system permitting only ~18,125 tonnes for the remainder of 2025 vs. ~96,600 tonnes annually for 2026–2027) was confirmed by Anthropic and Gemini as a live supply shock, with prices rallying ~170% from January 2025 lows.

7. Hyperscale AI Facilities Require Dramatically More Copper Per MW Than Conventional Data Centers

Providers: Perplexity, OpenAI, Anthropic, Gemini, Grok (five of six) Confidence: HIGH

Five providers confirmed the ~27–33 tonnes of copper per MW figure for AI data centers, versus 5,000–15,000 total tonnes for conventional facilities vs. up to 50,000 tonnes for hyperscale AI facilities. This 3–10x intensity differential is a key driver of the demand surge.

8. Grid Transmission Infrastructure Is the Binding Near-Term Constraint, Not Generation

Providers: Perplexity, OpenAI, Gemini, Grok (four of six) Confidence: HIGH

Four providers independently identified transmission bottlenecks, interconnection queue delays (averaging 4 years in North America), and substation upgrade timelines as the primary near-term constraint on data center expansion — more acute than generation capacity itself. Perplexity noted transmission expansion lags generation by 2–5 years; OpenAI cited the 4-year average connection wait; Gemini cited the "cost allocation problem" for transmission as a regulatory barrier.

9. Nuclear Power Is Emerging as the Preferred Baseload Solution for AI

Providers: Anthropic, Gemini, Grok (three of six, with supporting data from OpenAI) Confidence: MEDIUM

Three providers documented the wave of corporate nuclear commitments: Microsoft's Three Mile Island restart (837 MW, 20-year PPA), Meta's Clinton Clean Energy Center PPA (1.1 GW), Google's Kairos Power SMR agreement (500 MW), and Amazon's X-energy investment ($500M, 5 GW by 2039). Anthropic noted SMR commercial deployment is unlikely before 2030–2031 at scale, creating a near-term gap filled by natural gas.

Unique Insights by Provider

Perplexity

• Detailed scenario analysis with three quantified futures: Perplexity uniquely provided a structured three-scenario framework (Supply Responds / Geopolitical Friction / Supply Crisis) with specific financial outcomes: AI capex increases of 8–12%, 18–28%, and 40–60% respectively, and deployment delays of 0, 6–18, and 24–36 months. This is the only provider to quantify the capex impact of supply scenarios, making it directly actionable for investment planning.
• Aggregate capex requirement of $2.4–$3.6 trillion (2026–2035): Perplexity is the only provider to synthesize total infrastructure investment requirements across mining ($85–130B), energy ($2.0–3.1T), grid transmission ($220–280B), and water ($8–29B) into a single aggregate figure. This contextualizes the scale of the challenge relative to global GDP (~0.3–0.5% annually).
• Dysprosium-specific deficit quantification: Perplexity uniquely identified a specific 400–2,300 Mt dysprosium deficit by 2032 (vs. production capacity of 3,800–4,500 Mt), with price projections of $180–250/kg (vs. $100–150 baseline). This granularity on heavy REE sub-categories is absent from other providers.
• Monitoring dashboard with alert thresholds: Perplexity uniquely provided specific price alert thresholds (copper >$13,000/Mt, lithium >$18,000/Mt LCE, cobalt >$25/lb, dysprosium >$200/kg) and production monitoring targets, enabling real-time supply chain risk management.

OpenAI

• Ireland as a canary-in-the-coalmine case study: OpenAI uniquely highlighted Ireland's data center electricity consumption reaching 21% of national supply, with grid operators halting new connections around Dublin until 2028 and projections of one-third of national power by the end of the decade. This is the most concrete real-world example of a developed nation hitting a hard grid capacity wall due to AI infrastructure.
• Amazon's direct copper procurement deal: OpenAI uniquely documented Amazon's two-year agreement to purchase the first newly mined U.S. copper in decades from an Arizona mine specifically for AWS data centers — a concrete example of hyperscaler backward integration into commodity supply chains that signals a strategic shift in procurement models.
• Price inelasticity of AI copper demand: OpenAI uniquely articulated the mechanism by which AI demand is price-insensitive to copper — copper represents <0.5% of data center project costs, so even a doubling of copper prices adds only a rounding error to total capex. This explains why AI demand will not self-correct through price signals, making the supply gap more severe than in price-elastic sectors.

Anthropic

• DRC cobalt export ban as a live 2025 supply shock: Anthropic provided the most detailed and current account of the DRC's February 2025 cobalt export ban, including the specific quota figures (18,125 tonnes for remainder of 2025, 96,600 tonnes annually for 2026–2027 — less than half of previous export levels), the 72% drop in Chinese cobalt imports in July 2025, and the ~170% price rally. This transforms cobalt from a theoretical risk to a documented supply crisis.
• China's October 2025 extraterritorial jurisdiction clause: Anthropic uniquely documented China's October 2025 announcement of extraterritorial jurisdiction over any product containing ≥0.1% Chinese-sourced rare earths by value — the most aggressive mineral export control measure yet, effectively weaponizing the entire global rare earth supply chain. The subsequent U.S. 130% tariff response and November 2025 one-year suspension are also uniquely documented.
• DoD's $400M MP Materials partnership with price floor: Anthropic uniquely documented the U.S. Department of Defense's finalized $400M partnership with MP Materials establishing a price floor for neodymium and praseodymium, and Apple's $500M commitment to source magnets manufactured in Texas — concrete examples of government-industry coordination to rebuild domestic rare earth supply chains.
• U.S. copper mine permitting timeline of 29 years: Anthropic uniquely cited the specific figure that U.S. copper mine permitting averages 29 years (second longest globally), compared to under 5 years in many other jurisdictions, and noted the U.S. has only three primary copper smelters (two operating), none modern by international standards. This makes the U.S. particularly vulnerable despite having 48 Mt of identified copper resources.

Gemini

• U.S.-Brazil Rare Earth Partnership (January 2026): Gemini uniquely documented the January 2026 U.S.-Brazil "Rare Earths Partnership," including $600M+ in DFC/EXIM Bank financing commitments, Brazil's 21 million tonne rare earth reserves (second largest globally), and the EU-Mercosur agreement for joint investment in Brazilian lithium and rare earths. Critically, Gemini also noted the 15–20 year timeline for Brazil to develop independent refining capabilities — tempering optimism about near-term diversification.
• Extraterritorial rare earth controls timeline and diplomatic resolution: Gemini provided the most complete chronological account of China's escalating export control regime (2023 gallium/germanium → April 2025 heavy REEs → October 2025 extraterritorial rule → November 2025 one-year suspension), including the specific diplomatic mechanism of the November 9, 2025 MOFCOM announcement. This timeline is essential for understanding the current state of play.
• Hyperscale campus scale: 50,000-acre, 5 GW proposals: Gemini uniquely cited early-stage proposals for 50,000-acre AI campuses consuming up to 5 GW continuously — exceeding the capacity of the largest single nuclear or gas plants in the U.S. — as a concrete illustration of the unprecedented localized power density challenge facing grid operators.
• Humanoid robotics as an additional copper demand vector: Gemini uniquely noted that humanoid robots could add 1.6 Mt of annual copper demand at scale, and that AI itself could help optimize mining and exploration — two forward-looking demand and mitigation vectors absent from other providers.

Grok

• Defense sector as a competing copper demand vector: Grok uniquely quantified defense sector copper demand at ~1 Mt by 2040, adding a non-AI, non-EV demand vector that compounds the supply gap. This is particularly relevant given simultaneous defense spending increases across NATO and Indo-Pacific nations.
• BNEF 106 GW U.S. data center power demand by 2035: Grok cited BloombergNEF's projection of 106 GW of U.S. data center power demand by 2035 (up from ~25 GW operating in 2024) — a 4x increase that is more aggressive than most other provider figures and useful for stress-testing infrastructure planning assumptions.
• AI's contribution to specific mineral demand percentages: Grok uniquely cited IEA data showing AI/data centers could add ~2% to global copper demand, ~3% to REEs, and up to 11% to gallium demand by 2030 — providing proportional context for AI's role within the broader critical minerals demand picture.
• Circular economy potential: Grok uniquely noted that a successful scale-up of recycling could lower new mining requirements by 5–30% by 2040, providing a quantified range for the recycling offset that other providers discussed qualitatively.

Gemini-Lite

• Gallium and germanium as "Extreme" constraint category: Gemini-Lite uniquely categorized gallium (>95% China production) and germanium (~60% China refining) as "Extreme" constraint severity — a higher rating than any other provider assigned to these materials — and explicitly linked them to AI chip manufacturing rather than treating them as peripheral. This framing elevates these materials to strategic priority status.
• Concise constraint severity matrix: While less detailed than other providers, Gemini-Lite's clean tabular format (Infrastructure/Power/Hardware/Operations × Primary Constraint/Strategic Outlook) provides a presentation-ready summary framework that synthesizes the key findings efficiently.

Contradictions and Disagreements

Contradiction 1: Copper Demand Projections Vary by 3–4x

The Disagreement: Provider estimates for annual data center copper demand by 2030 vary enormously:

• Macquarie: 330,000–420,000 tonnes/year (data centers only)
• Sprott/CRU: ~1.1 million tonnes/year (including associated grid infrastructure)
• S&P Global: 1.1 Mt in 2025 growing to 2.5 Mt by 2040 (all data center-related)
• Wood Mackenzie: ~700 kt cumulative in-facility to 2030, plus up to 5 Mt in T&D

Why It Matters: The difference between 330,000 and 1.1 million tonnes/year is the difference between a manageable demand increment and a market-disrupting surge. The discrepancy appears to stem from different scope definitions — whether "data center copper" includes only in-facility wiring or also the transmission and distribution infrastructure required to deliver power to facilities.

Resolution Attempt: The most defensible approach is to use the narrower in-facility figure (~330,000–700,000 tonnes/year) for data center operators' procurement planning, and the broader grid-inclusive figure (~1.1 Mt/year) for national infrastructure planning. Neither figure is wrong; they measure different things.

Flag for Investigation: Analysts should request explicit scope definitions from any copper demand forecast before using it in planning models.

Contradiction 2: Global Copper Supply Peak — 2030 vs. Ongoing Decline

The Disagreement:

• S&P Global (cited by Anthropic and Grok): Global copper production will peak in 2030 at 33 Mt, then decline to 22 Mt by 2040
• Perplexity's supply projections: Moderate scenario reaches 24–26 Mt by 2035, optimistic reaches 27–29 Mt — implying continued growth, not decline

Why It Matters: If S&P Global is correct that supply peaks and then declines, the supply gap by 2040 is catastrophic (10 Mt). If Perplexity's moderate scenario is correct, the gap is severe but manageable with investment. The difference determines whether the copper constraint is solvable through capital deployment or represents a hard physical limit.

Flag for Investigation: The S&P Global "Copper in the Age of AI" (January 2026) report appears to be the most recent primary source. Perplexity's figures may reflect an earlier analytical vintage. Readers should obtain the S&P Global report directly to verify the post-2030 production trajectory.

Contradiction 3: Lithium Market Timing — Surplus Now vs. Deficit When?

The Disagreement:

• Perplexity: Lithium in surplus through 2028 (+0.5–0.6 Mt LCE surplus), then deficit emerging by 2030 (-0.35 Mt), worsening to -0.93 Mt by 2035
• OpenAI: IEA projects lithium could enter deficit by 2030 without new mines; existing supply meets only ~35% of projected 2035 demand
• Anthropic: Annual demand could reach 2.5–3.3 MMt LCE by 2030 (from 1.2 MMt in 2024); supply gap of 38–40% by 2035
• Gemini: Lithium demand quadruples from 2021 levels; supply gap of 300,000–768,000 tonnes LCE by 2030

Why It Matters: The near-term surplus (2024–2028) has already caused an 80%+ price crash from 2022 peaks, suppressing investment in new supply. If the deficit emerges by 2030 as most providers suggest, the investment suppression of 2023–2025 will have created the conditions for a severe shortage — a classic commodity cycle trap.

Flag for Investigation: The 2022–2024 lithium price crash and its impact on project cancellations/deferrals is the critical variable. A systematic review of cancelled lithium projects since 2023 would sharpen the deficit timing estimate.

Contradiction 4: Cobalt — Structural Surplus vs. Active Supply Crisis

The Disagreement:

• Perplexity: Cobalt shows supply surplus throughout 2026–2035 (+85–95 kt annual surplus), with constraint being geopolitical risk rather than absolute tonnage
• Anthropic/Gemini: The DRC's 2025 export ban has already created an active supply crisis, with Chinese imports down 72% year-on-year in July 2025 and prices up 170% from January 2025 lows; the quota system permits less than half of previous export levels

Why It Matters: Perplexity's analysis appears to reflect pre-2025 data or a scenario where the DRC ban is not modeled. Anthropic's account of the live 2025 supply shock is more current and suggests the "nominal surplus" framing is dangerously misleading — the market is experiencing acute shortage conditions right now, not in 2032.

Flag for Investigation: This is the clearest case where data vintage matters. Perplexity's cobalt analysis should be treated as superseded by Anthropic's 2025 data. The key question is whether the DRC quota system (96,600 tonnes/year for 2026–2027) represents a permanent structural reduction or a temporary negotiating position.

Contradiction 5: Electricity Capacity — Global Surplus vs. Regional Crisis

The Disagreement:

• Perplexity: Global electricity capacity is "nominally sufficient" by 2035 (10,385 GW projected vs. 9,300–10,150 GW needed), with the constraint being regional mismatch and renewable intermittency
• OpenAI/Gemini: The U.S. faces an 11+ GW shortfall by 2025 (38 GW demand vs. 26.7 GW supply), with a ~10 GW gap persisting through 2028; Ireland has already hit a hard wall; Southeast Asia faces a 40–60 GW gap by 2032

Why It Matters: The global aggregate framing obscures the severity of regional crises. A data center operator in Northern Virginia or Singapore cannot use surplus capacity in rural China. The regional framing is operationally correct; the global framing is analytically misleading for infrastructure planning purposes.

Resolution: Both are correct at different levels of analysis. Global capacity is adequate in aggregate; regional capacity is critically constrained in the specific geographies where AI infrastructure is being built. Planning should use regional, not global, capacity figures.

Contradiction 6: Rare Earth Supply — Aggregate Adequacy vs. Functional Scarcity

The Disagreement:

• Perplexity: Total REE production (800,000–950,000 Mt oxide equivalent) appears adequate vs. demand (78,000–108,000 Mt), but heavy REEs (dysprosium, terbium) face specific deficits of 400–2,300 Mt by 2032
• Anthropic/Gemini: China controls 89–94% of rare earth refining and magnet manufacturing, making the constraint primarily geopolitical rather than geological — adequate reserves exist globally but cannot be processed outside China at scale

Why It Matters: The distinction between "geological scarcity" and "processing monopoly" is critical for policy responses. If the constraint is geological, the solution is exploration and mining investment. If the constraint is processing monopoly, the solution is refining capacity investment outside China — a fundamentally different (and harder) problem.

Resolution: Both framings are correct and complementary. Heavy REEs face genuine geological scarcity (dysprosium deficit). All REEs face processing monopoly risk. The combination makes rare earths the most complex constraint to address.

Detailed Synthesis

The Physical Substrate of the AI Revolution

The AI industry has spent the past decade treating physical infrastructure as a solved problem — a commodity input to be procured at market rates. That assumption is now collapsing. The convergence of AI scaling, electrification, and geopolitical fragmentation has transformed copper, electricity, water, and critical minerals from background inputs into strategic assets that will determine competitive outcomes through 2035 and beyond [Perplexity, Anthropic, Gemini].

The core dynamic is straightforward but underappreciated: every major technology trend of the 2020s — AI, electric vehicles, renewable energy, grid modernization, and defense modernization — draws on the same finite pool of physical resources. These demand vectors are not sequential; they are simultaneous. And the supply side, constrained by decade-long mine development timelines, processing monopolies, and environmental permitting, cannot respond at the speed of software deployment [Grok, Anthropic].

Copper: The Irreplaceable Bottleneck

No commodity better illustrates this dynamic than copper. Copper is present in every layer of AI infrastructure: the 50,000 tonnes of wiring and busbars inside a hyperscale AI data center [OpenAI, Anthropic], the transformers and transmission lines delivering power to the facility (5,000–7,000 tonnes per GW of grid capacity) [Gemini], the cooling system heat exchangers, and the UPS battery systems providing backup power [Perplexity]. The industry average of 27–33 tonnes of copper per MW of AI data center capacity [Anthropic, Grok] means that a single 1 GW AI campus requires 27,000–33,000 tonnes of copper — more than the annual copper consumption of many mid-sized nations.

Global refined copper demand is projected to rise from approximately 28 million tonnes in 2025 to 42 million tonnes by 2040, a 50% increase [Anthropic, Grok]. AI infrastructure alone could add 2 million tonnes of annual demand by 2040 [Grok], with data centers consuming 330,000–1.1 million tonnes annually by 2030 depending on whether grid infrastructure is included in the scope [Perplexity, OpenAI, Grok]. The wide range reflects a genuine methodological disagreement about scope, not data quality — both figures are defensible for different planning purposes.

The supply side cannot respond at this pace. S&P Global projects that global copper production will peak in 2030 at 33 million metric tons and then decline to 22 million by 2040, creating a potential 10 million tonne annual deficit [Anthropic]. The IEA's more conservative analysis suggests existing mine pipelines cover only ~70% of projected 2035 demand [OpenAI, Perplexity]. Either way, the gap is structural, not cyclical. Average ore grades have fallen 40% since 1990 [Anthropic, Perplexity], meaning more rock must be processed per tonne of copper recovered. Mine development timelines average 16–17 years globally and 29 years in the United States [Anthropic, Grok] — the second-longest permitting timeline in the world. Even if capital were deployed today at unprecedented scale, new mines cannot enter production before the mid-2030s.

The price implications are significant but, critically, insufficient to self-correct the problem. Copper hit a record $11,705/tonne in December 2025 [Anthropic], with major bank forecasts ranging from $12,500 (JPMorgan) to $13,000 (UBS) by end-2026 and Goldman Sachs projecting $15,000/tonne by 2035 [Anthropic, Perplexity]. However, as OpenAI uniquely identified, copper represents less than 0.5% of data center project costs — meaning AI operators are effectively price-insensitive to copper. A doubling of copper prices adds only a rounding error to hyperscaler capex. This price inelasticity means the market's normal self-correcting mechanism (high prices suppressing demand) will not function for AI infrastructure, making the supply gap more severe than in price-elastic sectors [OpenAI].

The strategic response is already visible. Amazon has signed a two-year agreement to purchase the first newly mined U.S. copper in decades from an Arizona mine specifically for AWS data centers [OpenAI] — a concrete example of hyperscaler backward integration into commodity supply chains. Recycling offers a partial offset: current global copper recycling rates of ~50% could rise to 55–60% by 2035 [Perplexity], potentially covering 30–35% of the supply gap. But recycling cannot close a 10 Mt deficit on its own.

Electricity: The Transmission Bottleneck

The electricity challenge for AI is frequently framed as a generation problem, but the more acute near-term constraint is transmission and interconnection [Perplexity, OpenAI, Gemini]. Global data center electricity consumption is projected to roughly double from ~415–460 TWh in 2024 to ~945 TWh by 2030 in the IEA base case [Anthropic, Grok], potentially reaching 1,300–2,000 TWh by 2035 in high-growth scenarios [Perplexity, Gemini]. In the United States, data centers could account for 8–14% of national electricity demand by 2030, up from ~4–5% in 2023 [Anthropic, OpenAI].

The U.S. already faces an 11+ GW capacity shortfall by 2025 (38 GW demand vs. 26.7 GW supply), with a ~10 GW gap persisting through 2028 [OpenAI]. But the more fundamental constraint is that the average wait for a new data center grid connection in North America is approximately 4 years [OpenAI, Perplexity], driven by interconnection queue backlogs and substation upgrade timelines. Transmission expansion lags generation capacity by 2–5 years [Perplexity], meaning that even if generation capacity is built, it cannot reach data centers in time.

Ireland provides the starkest real-world example: by 2023, Irish data centers were consuming 21% of national electricity — more than all urban households combined — forcing grid operators to halt new connections around Dublin until 2028 [OpenAI]. Virginia's data centers already consume 26% of state electricity [Anthropic]. These are not projections; they are current operational realities.

The generation mix is shifting in ways that create new dependencies. Natural gas currently supplies over 40% of U.S. data center electricity [Anthropic], and orders for large gas turbines hit a 20-year high of 14 GW in 2024, accelerating to 18 GW in H1 2025 [Anthropic]. Coal retirement plans have been delayed or reversed for 37+ GW of capacity since 2023 [Anthropic]. The renewable energy buildout is real — solar and wind are projected to meet nearly half of additional data center demand through 2030 [Gemini] — but intermittency requires storage, and storage requires lithium and cobalt, creating circular dependencies across the mineral supply chain.

Nuclear power is emerging as the preferred long-term solution for AI baseload power, with Microsoft, Meta, Google, and Amazon all signing major nuclear agreements [Anthropic, Gemini]. However, SMR commercial deployment at scale is unlikely before 2030–2031 [Anthropic], leaving a 5–7 year gap filled by natural gas. The nuclear renaissance is real but delayed.

Water: The Invisible Constraint

Water is the most underappreciated physical constraint on AI scaling, primarily because its impact is localized rather than global [Perplexity, OpenAI, Anthropic]. Large AI data centers consume up to 5 million gallons per day — equivalent to the water needs of a town of 10,000–50,000 people [Anthropic, Gemini]. U.S. data center direct water consumption reached 17 billion gallons in 2023 and is projected to quadruple to 68 billion gallons by 2028 [Anthropic, OpenAI]. Texas alone could see data center water consumption reach 399 billion gallons by 2030 [Anthropic, Gemini].

The indirect water footprint — from power generation — is often 12x the direct consumption [Anthropic, OpenAI], meaning a data center running on fossil fuel power has a total water footprint far larger than its cooling systems suggest. Approximately two-thirds of U.S. data centers built since 2022 are in regions projected to face greater water scarcity by 2050 [Anthropic], creating a structural mismatch between facility location and water availability.

The constraint is already manifesting in permitting conflicts. A proposed data center in Newton County, Georgia requested more water per day than the entire county's existing usage [OpenAI]. The U.S. Southwest faces acute conflicts between data center water rights and agricultural water rights [Perplexity]. Immersion cooling can reduce water consumption by up to 91% [Anthropic], but adoption requires significant capital investment and operational changes. Microsoft has launched a zero-water cooling design [Anthropic], but mass adoption remains years away.

Critical Minerals: The Geopolitical Chokepoint

The critical mineral supply chain for AI infrastructure is characterized by a paradox: geological reserves are generally adequate globally, but processing capacity is catastrophically concentrated in China [Perplexity, OpenAI, Anthropic, Gemini]. China controls 89–91% of rare earth refining, 75% of cobalt and lithium chemical processing, >90% of graphite processing, >95% of gallium production, and ~60% of germanium refining [Anthropic, Gemini-Lite, OpenAI]. For a remarkable 19 out of 20 important strategic minerals, China is the leading refiner, with an average market share of 70% [Anthropic].

This is not a latent risk — it is an active geopolitical weapon. China's escalating export control regime has moved from theoretical leverage to operational disruption: 2023 gallium/germanium controls, April 2025 heavy rare earth controls (dysprosium, terbium, and five others), and the October 2025 extraterritorial jurisdiction clause covering any product containing ≥0.1% Chinese-sourced rare earths by value [Anthropic, Gemini]. The October 2025 measures triggered U.S. tariffs of 130% on Chinese goods and caused factory shutdowns across the U.S. and Europe [Anthropic]. A November 2025 one-year suspension provided temporary relief, but the underlying leverage remains intact [Gemini].

Cobalt presents the most acute near-term supply crisis. The DRC's February 2025 export ban — targeting cobalt hydroxide shipments — has already reduced Chinese cobalt imports by 72% year-on-year [Anthropic]. The subsequent quota system permits only ~18,125 tonnes for the remainder of 2025 and 96,600 tonnes annually for 2026–2027, less than half of previous export levels [Anthropic]. Cobalt prices have rallied ~170% from January 2025 lows [Anthropic]. Perplexity's analysis showing a cobalt surplus through 2035 appears to reflect pre-2025 data and should be treated as superseded by current market conditions.

Lithium faces a different timing profile: a current surplus (driven by the 2022–2024 price crash and overproduction) masking a structural deficit emerging by 2030 [Perplexity, OpenAI, Anthropic]. The price crash has suppressed investment in new supply, creating the classic commodity cycle trap: low prices today → underinvestment → shortage tomorrow. Annual lithium demand could reach 2.5–3.3 MMt LCE by 2030 (from 1.2 MMt in 2024), requiring supply to more than double [Anthropic]. The IEA projects a ~40% supply shortfall by 2035 under stated policies [Anthropic, Grok].

Western responses are accelerating but face long lead times. The January 2026 U.S.-Brazil Rare Earths Partnership, with $600M+ in DFC/EXIM Bank financing, represents a significant diplomatic achievement [Gemini]. The DoD's $400M partnership with MP Materials establishing a price floor for neodymium and praseodymium [Anthropic] and Apple's $500M commitment to Texas-manufactured magnets [Anthropic] demonstrate government-industry coordination. But Brazil's rare earth refining capacity is 15–20 years from maturity [Gemini], and the U.S. Mountain Pass mine expansion adds only ~180,000 Mt/year by 2035 [Perplexity] — insufficient for full independence from Chinese processing.

The Geopolitical Stakes

The physical layer of AI is now a theater of great power competition. China's control over mineral processing, combined with its domestic AI infrastructure buildout (capacity surplus of 400–800 GW by 2035 [Perplexity]) and state-directed resource allocation, gives it structural advantages that cannot be overcome through software innovation alone [Perplexity, Anthropic, Gemini]. The nations and companies that secure diversified, reliable access to copper, electricity, water, and critical minerals will determine the geography of AI leadership through 2035.

The investment required is staggering: Perplexity estimates $2.4–$3.6 trillion in total infrastructure investment (mining, energy, grid, water, facilities) over 2026–2035, averaging $250–390 billion annually — roughly 3–5x the entire global venture capital industry's annual deployment. This is not a market that can be addressed through startup investment alone; it requires industrial policy, government coordination, and multi-decade capital commitments.

The 2029–2032 window is the critical inflection point [Perplexity, Anthropic]. Copper supply-demand gaps become structural (2–4.5 Mt annual deficit), rare earth functional scarcity peaks, regional electricity capacity gaps become acute, and water constraints force facility relocation decisions. The decisions made in 2025–2027 — on power contracts, mineral supply agreements, facility siting, cooling technology, and backward integration into mining — will determine competitive positions through 2035 and beyond.