The AI race is usually described as a contest in models, chips and capital. That is incomplete. A model becomes economically useful only after a physical chain has worked: chips are fabricated, racks assembled, sites permitted, grid connections secured, electricity delivered continuously, cooling installed and the system financed before the revenue is certain.
My central view is that the next stage of AI competition will be shaped by deployable energy sovereignty: the ability to turn resources, equipment, infrastructure, finance and legal access into dependable productive capacity at the required place and time. Modern mercantilism is increasingly aimed at controlling that full stack.
The 2026 challenge requires at least ten objectively resolvable binary forecasts covering AI, modern mercantilism or their intersection. [1] The probabilities below are judgmental estimates based on information available on 1 August 2026. They range from 34% to 90%, and Forecast 4 is deliberately framed as a below-50% call.
Part 1 — Your Forecasts
-
76% — US data-centre electricity share. There is a 76% chance that data centres will account for at least 8.0% of total US electricity consumption in calendar 2028, resolved by the first DOE/LBNL national estimate published by 31 December 2029 using a methodology comparable to the 2024 report, with the LBNL technical report controlling if a DOE summary reports a different value. [2]
-
79% — Global data-centre electricity demand. There is a 79% chance that global data-centre electricity consumption will be at least 800 TWh in calendar 2030, resolved by the first IEA Key Questions on Energy and AI, Energy and AI, or direct successor publication released by 31 December 2031, with the official IEA data product controlling if the narrative and dataset differ. [3]
-
71% — Interconnection delay. There is a 71% chance that the median time from interconnection request to commercial operation for US generation projects completed in calendar 2028 will exceed four years, resolved by the first LBNL Queued Up or direct successor report published by 31 December 2030 that reports request-to-operation durations for all covered regions with both dates available. [4]
-
34% — PJM capacity-price relief. There is a 34% chance that the unconstrained RTO clearing price in PJM’s Base Residual Auction for the 2031/2032 delivery year will be below $250 per MW-day of UCAP, resolved by PJM’s official results; if the auction is renamed or delayed, the first official successor auction covering that delivery year controls, while cancellation without a comparable replacement price resolves no. [5]
-
63% — European industrial electricity cost. There is a 63% chance that the EU-27 average electricity price excluding taxes for non-household consumers in Eurostat band 500–1,999 MWh will be at least €0.130 per kWh in the second half of 2028, resolved using the first published value in Eurostat series
nrg_pc_205for that period. [6] -
68% — Critical-mineral refining concentration. There is a 68% chance that the unweighted arithmetic mean of the largest refining country’s global share for copper, lithium, nickel, cobalt and graphite will remain at or above 65% in 2030, resolved by the first IEA critical-minerals publication published by 31 December 2031 that reports comparable 2030 refining shares for all five minerals. [7]
-
82% — Chinese mineral export controls. There is an 82% chance that on 31 December 2028 China will still require an export licence under operative national rules for at least three of the following five separately counted materials—gallium, germanium, natural graphite, antimony and tungsten—where a material counts as controlled if at least one item category expressly identified under that material in the operative unified dual-use list requires an export licence on the resolution date. [8]
-
90% — Advanced-chip controls on China. There is a 90% chance that on 31 December 2028 exports from the United States to China of the advanced-computing commodities specified under EAR §742.6(a)(6)(iii) as of 1 August 2026, including Nvidia H200- and AMD MI325X-class items, will remain subject to a licence requirement under the Export Administration Regulations, even if the provision is later amended or renumbered. [9]
-
67% — Trade-policy intervention. There is a 67% chance that the arithmetic mean of the monthly WTO–IMF Trade Policy Activity Index will exceed its 2025 annual average in both 2027 and 2028, resolved using the first index release published after 31 March 2029 that reports all three calendar years on a common methodology; if no comparable backcast or common-methodology series is available by 31 December 2029, the forecast will be recorded as unresolved. [10]
-
64% — Compute-for-investment diplomacy. There is a 64% chance that by 31 December 2029 the United States will sign and publicly release at least one additional government-to-government framework, beyond the UAE arrangement, that both expands access to named advanced-computing items and states a quantified US investment, localisation, security or infrastructure commitment. [11]
-
82% — Nuclear restart. There is an 82% chance that by 31 December 2030 at least one US commercial reactor previously recorded as permanently shut down or in decommissioning will return to NRC operating status and report positive net generation in EIA Form 923 for two consecutive calendar months, with at least one of those months implying a capacity factor of 25% or more using EIA nameplate capacity. [12]
-
69% — Strategic manufacturing construction. There is a 69% chance that the 2028 annual unadjusted total for US private manufacturing construction in the Census category “computer, electronic, and electrical” will be at least $50.0 billion, resolved using the annual value in the July 2029 or first later comprehensive revision of the Value of Construction Put in Place series. [13]
Resolution convention. Where a named successor publication is not released by the stated deadline, the forecast will be resolved using the named organisation’s closest directly comparable official dataset available by that deadline. If no comparable official observation exists, the forecast will be recorded as unresolved rather than inferred from a secondary estimate. Forecast-specific rules—including Forecast 4’s auction-cancellation rule and Forecast 9’s common-methodology requirement—take precedence over this convention.
Part 2 — Framework & Holistic Synthesis
The constraint beneath the AI race
My thesis is that the marginal constraint on AI will migrate from model capability towards delivered power and strategic inputs. The biggest risk is that efficiency gains and a faster-than-expected infrastructure response outrun demand, turning today’s scarcity into an overbuild rather than a durable regime.
Bridgewater’s country comparison identifies cheap energy, self-sufficiency and deployment speed as strategic constraints across the United States, China and Europe. [14] My extension is to separate that broad argument into distinct tests of demand, delivery, market pricing and state control, each with a threshold, probability and resolver.
The important variable is not energy in the abstract. It is dependable power delivered to a particular site before the commercial window closes. The same distinction applies to minerals and semiconductors: ore is not a processed magnet, a chip design is not a packaged accelerator, and a project in an interconnection queue is not operating supply.
Forecasts 1–4 test that chain from demand to market price. DOE/LBNL estimates that US data centres used 4.4% of US electricity in 2023 and could reach 6.7–12% by 2028. [2] The IEA’s April 2026 central projection rises from 485 TWh in 2025 to 950 TWh in 2030. [3] Those projections can fail through efficiency, cancellations or weak returns, but the physical adjustment already reaches utilities, turbines, transformers, transmission, permitting and capital markets.
Forecast 3 is the delivery bottleneck. At the end of 2025, roughly 8,200 US projects representing 1,312 GW of generation and 749 GW of storage were seeking interconnection, while the median request-to-operation duration for projects completed in 2025 exceeded five years. [4] Forecast 4 asks whether several more years of new entry, demand response, transmission, self-supply and market reform can finally produce price relief. Its 34% probability is the book’s explicit negative call: a sub-$250 result would show the scarcity trade being competed or regulated away.
Modern mercantilism is moving from tariffs to capacity
Traditional mercantilism protected trade and accumulated financial claims. The modern version also targets productive capacity through subsidies, export licences, procurement, investment screening, stockpiles and bilateral access agreements.
Forecasts 6 and 7 distinguish diversification from control. The IEA reports that critical-mineral refining remains highly concentrated and that the number of mineral tariff codes subject to Chinese export controls has tripled since 2023. [7] Forecast 6 allows meaningful diversification while still testing whether processing remains structurally concentrated. Forecast 7 tests whether licensing remains a durable instrument across named materials rather than relying on a vague claim that controls will “increase.”
Compute is entering the same system. BIS moved H200-, MI325X- and similar exports to China to case-by-case review in January 2026 but retained the licence requirement. [9] That is why Forecast 8 sits at 90%: the uncertainty lies in technical thresholds and exceptions, not in whether advanced compute remains strategically controlled.
Forecast 10 tests whether compute access becomes a repeatable diplomatic bargain. The UAE framework links access to advanced computing items with matching US investment and anti-diversion commitments. [11] A second signed framework would show that this is more than a special case.
Forecast 9 is the common-factor regime test. The WTO–IMF index reached a record in early 2026, with January–May activity nearly twice the 2024 level and about a quarter above the 2025 average. [10] Requiring both 2027 and 2028 to exceed the 2025 average is more demanding than simply forecasting that intervention remains above the much lower 2024 base.
Regional and market implications
The United States has domestic energy, capital markets and leading AI firms, but its weakness is deployment. China has greater hydrocarbon-import exposure but exceptional manufacturing, grid-building and mineral-processing capacity. Europe retains capable firms and large markets but faces high delivered electricity costs, fragmented policy and greater import dependence. [6] [14]
That produces a selective opportunity set rather than a generic long-AI trade. In the United States, the strongest economics should accrue to scarce enablers of delivery—firm generation, grid equipment, turbines, nuclear fuel and qualified processing capacity—rather than every company carrying an AI label. Europe offers selective infrastructure winners against a weaker backdrop for energy-intensive production. China has the strongest deployment machine but greater policy and foreign-access risk.
A correct physical thesis is not automatically a profitable trade. The expression fails when scarcity is already fully priced, new capacity arrives before earnings materialise, or regulated returns prevent suppliers from retaining the economic rent.
AI capex can lift measured growth without producing a proportionate employment boom because data centres are equipment-heavy. Bridgewater estimates that AI capex could add about 140 basis points to US growth in 2026 and 150 basis points in 2027 while supporting relatively little employment for the amount invested. [15] Forecast 12 tests the physical investment footprint of that argument rather than restating the GDP estimate.
The single most important falsifier is sustained improvement in delivery. If interconnection times fall below three years and PJM prices normalise while data-centre load continues to rise, the supply system is adapting faster than the framework assumes. Lower demand, faster mineral diversification, weaker controls and European power prices returning towards the pre-2021 regime would reinforce that conclusion.
Part 3 — Analytical Appendix
Energy, delivery and market pricing
Forecast 1 uses an 8% threshold against DOE/LBNL’s 6.7–12% range for 2028. [2] It does not require the aggressive end of the build-out, but it can still miss if electricity supply grows faster than expected, utilisation disappoints or efficiency improvements dominate induced demand.
Source: DOE/LBNL [2].
Forecast 2 uses 800 TWh against the IEA’s updated 950 TWh central projection. [3] The gap allows for project failure and forecast error while preserving the claim that data-centre demand becomes material at the global-system level.
Source: IEA [3].
Forecasts 3 and 4 share the same driver but are framed in opposite directions: persistent delay supports Forecast 3 and reduces the probability of price relief in Forecast 4. LBNL’s queue contains proposed projects, not a supply forecast; only 13% of capacity requesting interconnection from 2000–2020 had reached operation by the end of 2025. [4]
Source: LBNL, Queued Up [4].
PJM’s 2028/29 auction cleared 138,318 MW at the $325/MW-day cap and remained 6,831 MW short of its reliability requirement. [5] That is evidence of scarcity, but not a pure market signal because the demand curve, accreditation rules and price collar are administered. Forecast 4 therefore resolves on the official 2031/32 price but should be interpreted alongside reliability procurement, new entry and any market redesign. Cancellation without a comparable replacement price resolves no because the forecast is specifically about an observable PJM clearing price, not an inferred scarcity condition.
Forecast 5 allows further European price normalisation while testing whether the industrial cost regime remains elevated. Eurostat’s EU-27 ex-tax price for band 500–1,999 MWh fell from €0.1986/kWh in 2022-H2 to €0.1511/kWh in 2025-H2. [6]
Source: Eurostat [6].
Forecast 11 requires both NRC operating status and sustained EIA generation rather than a token commissioning month. Two consecutive positive months, including one at an implied capacity factor of at least 25%, provide a cleaner test of an operating commercial restart. Palisades has returned to an operational licensing basis, but execution risks remain in commissioning, component condition, financing and litigation. [12]
Strategic inputs and policy
Forecast 6 fixes the basket at copper, lithium, nickel, cobalt and graphite and uses an unweighted average of the leading refining country’s share. The 65% threshold allows meaningful diversification while still testing whether processing remains structurally concentrated. [7]
Forecast 7 counts gallium and germanium separately and asks only whether an export licence remains legally required. Gallium, germanium, graphite and antimony were consolidated into China’s unified dual-use export-control list from 1 December 2024, while the February 2025 tungsten measure updated that list. A material counts if at least one expressly listed item category under that material remains licence-controlled; the forecast does not require every chemical form, product or technology associated with the material to be controlled. The operative unified list or its direct successor controls resolution. [8]
Forecast 8 freezes the reference class at the commodities covered by EAR §742.6(a)(6)(iii) on 1 August 2026. That prevents later renumbering from changing the substance of the event. The January 2026 rule moved qualifying H200- and MI325X-class exports to case-by-case review but retained the licence requirement. [9]
Forecast 9 is deliberately harder than the previous version: both 2027 and 2028 must exceed the 2025 annual average. The common-vintage rule prevents a methodology change from creating a false comparison. If the publisher cannot provide the three years on a comparable basis by the deadline, the forecast remains unresolved rather than combining incompatible vintages. It remains a common-factor forecast rather than the book’s most differentiated item. [10]
Forecast 10 requires a signed and public framework that contains both sides of the bargain. A commercial investment announcement without expanded government-authorised compute access does not count, and an export-control change without a quantified reciprocal commitment does not count. [11]
Capital formation, price interpretation and updates
Forecast 12 uses the Census annual unadjusted total for one fixed private-manufacturing category and one fixed revision vintage. The $50 billion threshold is approximately six times the 2019 annual total of $8.308 billion, but remains well below the $105.914 billion recorded in 2023. It therefore tests whether the strategic-manufacturing build-out persists after the peak investment wave rather than merely whether nearly all of it disappears. [13]
The 2020 negative-WTI episode remains a useful control case, but not a forecast driver. The broad equity crash followed shutdowns, collapsing expected cash flows and a dash for liquidity; the May WTI contract’s negative settlement reflected slow supply adjustment, scarce storage and physical delivery at Cushing. [16] The lesson is that prices are most useful when paired with quantities, legal rules and settlement mechanics.
The forecasts are correlated. Forecasts 1 and 2 share demand; 3 and 4 share delivery; 6 and 7 share mineral concentration; 8 and 10 share US control over compute; and 9 and 12 share the industrial-policy regime. They should not be read as twelve independent bets.
If Forecast 1 resolves yes while Forecast 3 resolves no and Forecast 4 resolves yes, demand grew but the grid and capacity market adapted faster than expected. If Forecast 9 resolves yes while Forecast 12 resolves no, policy activity increased without producing durable domestic investment. If Forecasts 6 and 7 both resolve no, diversification moved faster than the framework assumed.
| Indicator and trigger | Source | Forecasts | Likely update |
|---|---|---|---|
| US data-centre demand estimate falls below 7% for 2028 | DOE/LBNL [2] | 1, 2, 4, 12 | Reduce 1, 2 and 12; raise 4 |
| Median interconnection time falls below three years for two consecutive completion cohorts | LBNL [4] | 3, 4 | Reduce 3; raise 4 |
| PJM clears below $175/MW-day while meeting its reliability requirement without a rule-driven cap change | PJM [5] | 4 | Raise sharply or resolve yes |
| Leading-refiner average falls below 60% before 2029 | IEA [7] | 6, 7 | Reduce |
| BIS removes the licence requirement under EAR §742.6(a)(6)(iii) for China-bound advanced-computing commodities | BIS [9] | 8, 10 | Reduce sharply |
| A common-vintage index release places either 2027 or 2028 at or below the 2025 annual average | WTO–IMF [10] | 9 | Reduce or resolve no |
| EU ex-tax non-household prices fall below €0.11/kWh for two consecutive halves | Eurostat [6] | 5 | Reduce |
| A previously shut reactor reaches NRC operating status and satisfies the two-month EIA generation test | NRC/EIA [12] | 11 | Resolve yes |
References
[1] Bridgewater Associates and Global Citizen, “Forecasting the Future 2026,” 2026. [Online]. Available: Challenge page. [Accessed: Aug. 1, 2026].
[2] U.S. Department of Energy and Lawrence Berkeley National Laboratory, “2024 United States Data Center Energy Usage Report,” 2024. [Online]. Available: DOE release and report access. [Accessed: Aug. 1, 2026].
[3] International Energy Agency, “Key Questions on Energy and AI,” Apr. 16, 2026. [Online]. Available: Executive summary. [Accessed: Aug. 1, 2026].
[4] J. Rand et al., “Queued Up: 2026 Edition—Characteristics of Power Plants Seeking Transmission Interconnection as of the End of 2025,” Lawrence Berkeley National Laboratory, 2026. [Online]. Available: Queued Up. [Accessed: Aug. 1, 2026].
[5] PJM Interconnection, “PJM Capacity Auction Procures 138,318 MW of Generation Resources,” Jul. 14, 2026. [Online]. Available: Official auction release. [Accessed: Aug. 1, 2026].
[6] Eurostat, “Electricity Price Statistics,” dataset nrg_pc_205, 2026. [Online]. Available: EU industrial-consumer data. [Accessed: Aug. 1, 2026].
[7] International Energy Agency, “Global Critical Minerals Outlook 2026,” 2026. [Online]. Available: Executive summary. [Accessed: Aug. 1, 2026].
[8] Ministry of Commerce of the People’s Republic of China and partner agencies, export-control records for the five named materials: Unified Dual-Use Items Export Control List, Announcement No. 51 of 2024; Gallium and germanium, Announcement No. 23 of 2023; Graphite, Announcement No. 39 of 2023; Antimony, Announcement No. 33 of 2024; and Tungsten, Announcement No. 10 of 2025. [Accessed: Aug. 1, 2026].
[9] U.S. Bureau of Industry and Security, “Revision to License Review Policy for Advanced Computing Commodities,” 91 Fed. Reg. 1684, Jan. 15, 2026. [Online]. Available: Federal Register final rule. [Accessed: Aug. 1, 2026].
[10] World Trade Organization and International Monetary Fund, “Updated WTO–IMF TPA Index Shows Continued Rise in Global Trade Policy Activity,” Jul. 23, 2026. [Online]. Available: WTO update. [Accessed: Aug. 1, 2026].
[11] U.S. Bureau of Industry and Security, “Department of Commerce Eases Export Controls for UAE,” Jul. 10, 2026. [Online]. Available: BIS release. [Accessed: Aug. 1, 2026].
[12] U.S. Nuclear Regulatory Commission, “Palisades Nuclear Plant—Restart Inspection Report,” Feb. 2, 2026. [Online]. Available: NRC inspection report. [Accessed: Aug. 1, 2026].
[13] U.S. Census Bureau, “Annual Value of Private Construction Put in Place, 2012–2023,” Apr. 1, 2024. [Online]. Available: Annual private-construction table and current annual series and historical data. [Accessed: Aug. 1, 2026].
[14] K. Karniol-Tambour, I. Singer, D. Hochman and J. Ng, “Power Politics: Energy Self-Sufficiency in a Modern Mercantilist World,” Bridgewater Associates, Oct. 16, 2025. [Online]. Available: Research note. [Accessed: Aug. 1, 2026].
[15] G. Jensen, D. DeBois and A. Zimbler, “The Macro Implications of the AI Capex Boom,” Bridgewater Associates, Jan. 7, 2026. [Online]. Available: Research note. [Accessed: Aug. 1, 2026].
[16] U.S. Energy Information Administration, “Low Liquidity and Limited Available Storage Pushed WTI Crude Oil Futures Prices Below Zero,” Apr. 2020. [Online]. Available: EIA analysis. [Accessed: Aug. 1, 2026].