From Storage to Engine
For three decades, the data center was a passive structure. It stored bits, served pages, and ran transactional workloads at predictable, incremental growth rates. Artificial intelligence has dissolved that pattern. Training a frontier model is not storage; it is sustained, concentrated electrical consumption at industrial scale. Inference at hyperscale is not a transient query load; it is a continuous burn. The shift from storage to compute engine has redefined the data center as a power-consuming, water-consuming, land-consuming installation whose siting decisions now register on national grid planning documents and foreign ministry briefings. The geographic concentration of that compute — and the legal jurisdiction over the entities that operate it — is the dominant strategic fact of the AI build-out.
The IEA Numbers: 415 TWh to 945 TWh
The International Energy Agency’s Electricity 2025 report establishes the baseline. Global data center electricity consumption reached approximately 415 TWh in 2024 (Fact, High). The IEA’s central projection places that figure at approximately 945 TWh by 2030 — a 2.3× increase over six years (Fact, High). For comparative scale, 945 TWh is roughly equivalent to Japan’s total national electricity consumption.
The driver is unambiguous: GPU-dense AI workloads. A single GPT-4-class training run consumes several gigawatt-hours; inference at production scale is continuous and additive, not episodic. US data centers alone consumed approximately 200 TWh in 2023 and are projected to reach 260–290 TWh by 2026 under IEA estimate ranges (Fact, High). As a share of global electricity, data centers represented roughly 1–1.5% in 2024 and are projected to reach 2–3% by 2030 under the central scenario.
The 2–3% figure understates the strategic significance. The consumption is not distributed; it is geographically concentrated in a small number of grid nodes that were not engineered for industrial loads of this density. The question is no longer whether the global grid can supply the electrons. It is whether specific substations in Loudoun County, Virginia, or specific transmission corridors in Singapore can deliver them without years of construction lead time.
Northern Virginia: Correcting the Myth, Maintaining the Significance
A persistent claim states that “70% of the world’s internet traffic” passes through Northern Virginia. TeleGeography, the primary commercial source for global traffic measurement, has explicitly stated this figure is not supported by their data (Fact, High). The number circulates because it is rhetorically convenient, not because it is measured. It should not be reproduced as fact.
The corrected figure is more interesting. Northern Virginia — the Ashburn metro cluster, marketed as “Data Center Alley” — represents approximately 13% of global operational data center capacity by megawatts (TeleGeography/CBRE 2024 data) (Fact, High). That makes it the single largest data center market in the world by operational capacity. Ashburn hosts one of the most peered internet exchange points in the United States — the Equinix Ashburn / DC-IX complex — and concentrates campuses operated by Amazon Web Services, Microsoft Azure, Google Cloud, Equinix, Digital Realty, and QTS.
The origins are layered. The original ARPANET nodes at Virginia universities, the Herndon/Reston corridor as an early ISP cluster in the 1990s, Virginia’s low corporate tax regime, cheap land prior to 2010, proximity to federal government contracting in Washington DC, and the consolidation of MAE-East into what became the Equinix Ashburn exchange — each layer accreted onto the next (Fact, High). The cluster is now self-reinforcing: new tenants locate in Ashburn because peering partners are in Ashburn.
The bottleneck has become structural. Dominion Energy, the regional utility, and the Northern Virginia grid more broadly are now capacity-constrained. Grid interconnection queues for new data center projects have exceeded five years in some cases, per Dominion’s 2024 Integrated Resource Plan filing (Fact, High). Construction is no longer rate-limited by capital or by permitting alone; it is rate-limited by transmission and generation capacity. Land has been bought; megawatts have not arrived.
The Global Geography of Compute
Beyond Northern Virginia, the concentration is similarly stark. The top five global data center markets by operational capacity (CBRE/JLL 2024) are Northern Virginia (~13%), Silicon Valley / Bay Area (~7%), Singapore (~5%), London (~5%), and Amsterdam (~4%) (Fact, High). Together these five markets account for roughly 34% of global operational data center capacity — two US hubs, one UK hub, one EU hub, one Southeast Asian hub. The remaining 66% is distributed across dozens of smaller markets.
China is a separate ecosystem. Alibaba Cloud, Tencent Cloud, Huawei Cloud, and ByteDance operate at hyperscale but serve primarily domestic users and adjacent Asian markets, constrained by Chinese data localization law. Beijing’s inland data center strategy designates Guizhou province — specifically the Gui’An New Area — as a primary national zone, exploiting lower land and energy costs, hydropower proximity, and physical distance from coastal attack vectors (Fact, High). This is state-directed compute geography, an explicit instrument of Digital Sovereignty and of the Digital Silk Road export strategy.
India is the fastest-growing emerging market, with the Mumbai / Navi Mumbai cluster anchoring domestic and intra-Asian capacity. The Digital Personal Data Protection Act 2023 introduces a localization layer that, when fully enforced, will channel data flows into Indian-jurisdictioned facilities (Assessment, Medium).
The pattern is consistent: AI compute concentrates in a small number of grid-rich, fiber-rich, and jurisdictionally favorable nodes. The connective tissue between these nodes is the global submarine cable system covered earlier in this series (Fiber Optic Transmission). The chokepoint logic from that system applies here in compressed form: a handful of physical sites carry a disproportionate share of the workload.
The Hyperscaler Capex Wave
The capital deployment is unprecedented. Microsoft announced an $80 billion capex plan for fiscal year 2025 dedicated primarily to global data center buildout (Fact, High, January 2025 announcement). Amazon’s implied 2025 capex for AWS and adjacent data center infrastructure sits in the $75–80 billion range. Google’s parent Alphabet guided to approximately $75 billion in 2025 capex, the majority on data centers, per the Q4 2024 earnings call (Fact, High). Meta guided to $60–65 billion in 2025 capex with heavy AI infrastructure weighting.
Combined, the four largest US hyperscalers will deploy approximately $300–320 billion in infrastructure capex in 2025 alone (Fact, High). This exceeds the annual defense budgets of every nation other than the United States and China. It is concentrated in a predictable geography: Virginia, Texas, Iowa, and Ohio in the US; Ireland, Sweden, Finland, and the Netherlands in Europe; Singapore, Malaysia, Japan, and India in Asia-Pacific.
Three structural implications follow. First, the capex wave is a barrier to entry that locks in incumbency. No new market entrant can match a $300 billion annual deployment; the AI compute layer is consolidating into the same four-to-five operators that already dominate cloud. Second, the geographic distribution is itself an alliance map — the build-out follows the contours of US treaty relationships and Five Eyes jurisdictional comfort. Third, the capex is forward-loaded against revenue: the operators are betting that AI demand will materialize at scale sufficient to amortize the spend, and that bet is now too large to unwind without macroeconomic consequences (Assessment, Medium).
Water, Energy, and Land
Data centers consume water primarily for cooling. Estimates range from approximately 1.0 to 5.0 liters per kWh depending on cooling architecture — air-cooled designs at the low end, evaporative cooling at the high end. GPU-dense AI clusters tend toward the higher water intensity because of thermal density.
Disclosed figures provide scale. Google reported approximately 21 billion liters of fresh water consumed globally in 2022 (Google Environmental Report 2023), with the figure growing. Microsoft disclosed 6.4 billion liters globally in fiscal year 2023 (Fact, High). Aggregate hyperscaler water consumption is not centrally reported, but extrapolation from these disclosures places it in the low hundreds of billions of liters annually and rising.
Water stress has become a site-selection constraint, not merely an environmental footnote. The Phoenix, Arizona data center cluster faces measurable water scarcity, and several proposed expansions have been contested by state and municipal authorities (Fact, High). Conversely, Nordic markets — Finland, Sweden, Iceland — offer cold-climate air cooling with minimal water draw, which is one reason the geography of new AI build-out increasingly tilts north.
Energy and water now function as hard physical limits, not optimization variables. The transmission queue in Northern Virginia and the water-rights contestation in Arizona are early signals of a broader pattern: physical infrastructure constraints are catching up to compute demand. Energy sufficiency is becoming a precondition for compute sovereignty.
CLOUD Act: When Server Location Stops Mattering
The legal layer is decisive and underappreciated. The Clarifying Lawful Overseas Use of Data Act (CLOUD Act), enacted in March 2018, authorizes US law enforcement to compel US-based cloud providers to produce data held outside US borders, subject to executive agreements with foreign governments (Fact, High). The legal mechanism does not depend on the physical location of the server. It depends on the corporate domicile of the provider.
The implication is structural. A European company storing data in a Frankfurt or Dublin AWS region is, as a legal matter, still subject to US government access requests routed through Amazon’s US corporate entity. Physical sovereignty of the server room does not confer legal sovereignty over the data on the drives. The same logic applies to Microsoft Azure regions in the EU and to Google Cloud regions worldwide.
Bilateral CLOUD Act executive agreements have been concluded with the United Kingdom (2022) and Australia (2023); EU-level negotiations remain ongoing as of 2025 (Fact, High). The agreements streamline cross-jurisdictional requests but do not narrow the underlying extraterritorial reach.
European responses are explicit. France’s “Cloud de confiance” framework and the EU Cybersecurity Certification Scheme for Cloud Services (EUCS) attempt to define a tier of cloud provision walled off from CLOUD Act exposure, typically by requiring European corporate control. The Gaia-X initiative is the federated infrastructure layer underneath that legal framework. (Assessment, Medium): CLOUD Act has deterred a measurable subset of European public-sector and defense customers from US cloud adoption and is driving demand for sovereign cloud alternatives, though the gap in capability between US hyperscalers and European sovereign offerings remains wide.
This is the jurisdictional intelligence layer. Hyperscaler geography becomes an intelligence question, not merely a commercial one. Which company holds which workload, under which corporate domicile, determines which government has lawful access. That is the analytical core of Digital Sovereignty in the AI era.
Military Compute: JWCC and the Classified Layer
The US Department of Defense cloud trajectory illustrates the institutional answer. The JEDI single-vendor contract awarded to Microsoft Azure in 2019 was cancelled in 2021 amid litigation and was replaced by the Joint Warfighting Cloud Capability (JWCC), awarded in December 2022 to four vendors: Microsoft, Amazon, Google, and Oracle, for a ceiling of up to $9 billion over ten years (Fact, High).
The four awardees are all US-domiciled, FedRAMP High-certified, and cleared at DoD Impact Levels IL5/IL6 for sensitive and classified workloads. The structural choice is multi-vendor for resilience and competitive pricing, but the jurisdictional surface is uniform — all four are US entities subject to US law. Palantir’s Gotham and AIP platforms run on this DoD-cleared cloud substrate for intelligence analysis and AI-assisted targeting workflows.
The Chinese equivalent is operated through state-owned cloud providers — China Electronics Technology Group (CETC), Huawei Government Cloud, and adjacent entities — physically separated from the commercial internet under classified designation (Fact, High). The bifurcation is complete: US military AI runs on US hyperscaler substrate under US legal jurisdiction; Chinese military AI runs on Chinese state-controlled substrate under Chinese legal jurisdiction. There is no neutral third option at scale for serious military AI workloads.
Nordic Resilience and the Sovereign Cloud Alternative
The Nordic markets offer the closest approximation to a non-US, non-Chinese compute geography with serious capacity. Iceland operates on 100% renewable electricity — geothermal and hydroelectric — with cold-climate air cooling that reduces water consumption to near zero, full GDPR compliance, political stability, and physical isolation as an island with no land borders (Fact, High). Finland, Sweden, and Norway combine political stability, NATO membership following the 2022–2023 accessions, renewable-heavy grids, and dense submarine cable connectivity to continental Europe via the Baltic Sea and North Sea routes (covered in the earlier installments of this series and in The Internet’s Plan B — LEO Satellites, Terrestrial Alternatives, and the Resilience Paradox).
The constraint is latency. Distance from Western European and East Asian population centers raises round-trip times beyond what is tolerable for real-time inference serving urban end-users. Nordic compute is therefore optimal for training runs, archival storage, backup workloads, and batch inference — workloads where milliseconds of latency are immaterial. It is not a substitute for edge inference in London, Frankfurt, or Tokyo.
The sovereign cloud alternative more broadly faces a recurring tension: capability versus jurisdiction. US hyperscalers offer the most advanced AI tooling, the largest model catalogs, and the deepest ecosystem of pre-built services. Sovereign alternatives offer legal insulation but at higher per-unit cost, with smaller ecosystems and a slower frontier-model release cadence (Assessment, Medium). The choice is not symmetric.
Strategic Implications
AI compute concentration is a dual chokepoint. Geography — Northern Virginia, Singapore, London, Frankfurt, Tokyo as critical hubs — concentrates the workload in a small number of physical sites. Supply chain — TSMC fabrication → NVIDIA accelerators → hyperscaler integration → AI service delivery — concentrates the upstream dependencies in a parallel set of single points of failure. Disruption at either layer propagates to the other.
Water and energy are now binding physical constraints. The Northern Virginia grid queue and the Phoenix water contestation are early manifestations. The 945 TWh figure by 2030 is not a forecast that the grid will absorb passively; it is a forecast that energy sufficiency will become an explicit instrument of compute sovereignty. States that command surplus baseload generation — Nordic hydro, Gulf-state gas, French nuclear, Chinese inland hydro — acquire a structural advantage independent of their position in the chip supply chain.
The CLOUD Act subordinates data sovereignty to US legal jurisdiction regardless of physical server location. For non-US states and their public-sector and defense customers, the choice is binary in structure: accept US cloud dependency with implicit CLOUD Act exposure, or accept sovereign cloud with higher cost, narrower capability, and a slower AI frontier. The middle position — European-located US cloud regions — does not resolve the legal exposure; it only obscures it.
China’s Guizhou strategy demonstrates the state-directed counter-model: centralize compute in inland provinces, insulate from foreign access vectors, power with sub-cost hydroelectric resources, and designate the installations as critical infrastructure subject to military protection. The model is replicable in form but not in scale; few states command the combination of land, energy, and capital to operate it.
The 945 TWh figure is, at root, a geopolitical figure. Whoever controls the generating capacity, the cooling water, the land for new campuses, and the legal jurisdiction over the operating entity controls where AI runs. The build-out of the next half-decade will redraw the map of compute as decisively as the build-out of submarine cables redrew the map of bandwidth in the prior decade. The two maps overlap, but they are not identical, and the divergences are where the next contests will be fought.
For the broader argument linking this article to the rest of the series, see SYNTHESIS. For the physical substrate of internet transport that underlies all compute geography, see How the Internet Physically Works.
Sources
Primary — High Confidence:
- International Energy Agency (IEA), Electricity 2025 report — global data center consumption 415 TWh (2024); projection 945 TWh (2030).
- TeleGeography — global IP traffic and data center capacity datasets; explicit clarification that the “70% of internet traffic through Northern Virginia” figure is not supported by their data.
- CBRE Global Data Center Trends 2024; JLL Data Center Outlook 2024 — top-five market share by operational capacity.
- Dominion Energy 2024 Integrated Resource Plan filing — Northern Virginia interconnection queue durations.
- Microsoft Q2 FY25 earnings disclosure (January 2025) — $80 billion FY25 capex.
- Alphabet Q4 2024 earnings call — ~$75 billion 2025 capex guidance.
- Amazon and Meta 2024 annual reports — capex guidance ranges.
- Google Environmental Report 2023 — 21 billion liters fresh water (2022).
- Microsoft Environmental Sustainability Report FY23 — 6.4 billion liters water.
- US Department of Defense — JWCC contract announcement (December 2022), four-vendor award.
- CLOUD Act, Public Law 115-141, March 2018; UK-US CLOUD Act Agreement (2022); Australia-US CLOUD Act Agreement (2023).
Secondary — Medium Confidence:
- European Commission communications on EUCS and Gaia-X.
- French government materials on “Cloud de confiance” framework.
- IEA US data center range estimates (200 TWh 2023 → 260–290 TWh 2026).
- Public reporting on China Guizhou data center designation and Chinese state cloud architecture (CETC, Huawei Government Cloud).
Gaps:
- No publicly verified figure for aggregate hyperscaler water consumption.
- Limited disclosure on Chinese hyperscaler capex on a comparable basis to US hyperscalers.
- No public dataset measuring CLOUD Act request volume by jurisdiction, requested provider, or outcome — a structural transparency gap.
- No consolidated public measurement of grid interconnection queue lengths across US data center markets beyond Dominion’s filings.