The Hidden Cost of the AI Arms Race: Why Hyperscale Data Centers Are Becoming America’s Next Great Infrastructure Debate
The Hidden Cost of America’s AI Infrastructure Boom
Why hyperscale data centers are reshaping technology, energy, and public policy
The explosive rise of AI is driving a new kind of infrastructure boom: hyperscale data centers built specifically to power massive artificial intelligence workloads. Projects like Utah’s proposed 9-gigawatt “Stratos” campus represent a scale never before seen in computing, with enough energy demand to rival entire state power grids. Supporters argue these facilities are essential for U.S. technological leadership, national security, and keeping pace with global competitors like China. They promise faster AI development, cloud computing expansion, and new economic opportunities tied to the growing demand for advanced computation.
But the scale of these projects raises major environmental and social concerns. A facility like Stratos would consume staggering amounts of electricity and water, generate significant heat, and potentially emit millions of tons of carbon dioxide annually if powered by natural gas as proposed. Communities in places like Utah and Ohio are increasingly pushing back, citing fears about strained local resources, pollution, and weak economic returns. While these centers require enormous investment, they create relatively few permanent local jobs compared to their footprint, while the AI systems they enable could accelerate automation across retail, logistics, food service, and administrative work.
The broader debate comes down to balancing innovation with accountability. Hyperscale AI infrastructure could reshape industries, strengthen domestic computing capacity, and support breakthroughs in robotics, simulation, and machine learning. At the same time, it highlights urgent questions about energy policy, environmental sustainability, labor displacement, and public oversight. As resistance grows in communities like Cleveland, Ravenna, and Box Elder County, policymakers face increasing pressure to establish stronger zoning laws, environmental reviews, renewable energy requirements, and workforce retraining programs to ensure the AI buildout benefits society rather than overwhelming it.
AI Hyperscale Data Centers: Project Stratos, Ohio Resistance, and National Implications
Compute capacity, environmental impacts, labor markets, supply chains, geopolitics, and policy options
Executive Summary
A massive wave of new AI data centers ("hyperscale" campuses) is emerging. Industry claims they secure national technology leadership, but their scale and impacts are unprecedented. We analyze the 9 GW Stratos project in Utah and Ohio opposition (Slavic Village & Ravenna) to illustrate the issues. Our report draws on filings, media, and studies.
The proposed Stratos data center — a 9 GW AI campus on 40,000 acres in rural Utah — would be unprecedented in scale, dwarfing existing US facilities and even a state's entire electricity use. Backers tout thousands of high-paying jobs and U.S. "compute sovereignty" in the AI race against China. Critics raise alarms over environmental impacts (water use, heat, pollution), burden on the grid, and rapid automation of jobs.
Our analysis draws on public records and research to examine the feasibility and implications of this project. We estimate that a 9 GW continuous load implies on the order of 1–3 million AI accelerators (GPUs) and 100–400 thousand high-density racks, supporting exa- to zetta-FLOP-scale compute. These would enable massive AI workloads: from trillion-parameter model training (~10²³–10²⁴ FLOPs) to large-scale inference (billions of queries/sec) and real-time robotics/simulation tasks.
Labor markets across retail, food service, warehousing, logistics, and even some white-collar fields would face substantial exposure: industry reports suggest office support, customer service, food service, production, and transport roles are most at risk. BLS projections show cashiers and retail sales positions declining while fast-food and transportation grow only modestly.
Achieving reshoring via AI+robots faces serious hurdles: robot dexterity and flexibility remain limited; advanced manufacturing still requires human skill in many tasks. Capital costs are enormous, and critical components (high-end semiconductors, rare earth magnets, gallium, germanium, etc.) are largely sourced from China. Grid impacts are also nontrivial: a 9 GW facility would need dedicated generation (likely natural gas turbines as promised) and would double Utah's current electricity consumption. Strategically, proponents frame this as enhancing U.S. AI leadership, but analysts note China's commanding lead in industrial robots and manufacturing AI.
Policy responses are already emerging: several states (e.g. Maine) have considered data-center moratoria or new taxes, and federal grants (CHIPS/IRA) target AI infrastructure. Mitigation likely requires massive reskilling, infrastructure planning, and energy policy (renewables expansion, efficiency standards).
Key Findings at a Glance
- Motives/Economics: Tech giants and financiers are racing to build enormous AI compute infrastructure as a strategic investment. Stratos is a >$100 billion project (9 GW on 40,000 acres). Operating costs are on the order of $1–5 billion/year in electricity alone. ROI models rely on discounted AI computing services, tax incentives, and energy arbitrage.
- Technical Scale: A 9 GW facility (~Stratos) equates to on the order of 10–30 million GPUs and exaFLOP-scale performance. A 150 MW site (Cleveland example) still implies ~150k GPUs. These centers can execute ~10⁶–10⁷ AI inferences per second simultaneously.
- Environmental Impacts: On-site gas generation at these scales produces enormous emissions. Stratos alone would emit ~30 million tons CO₂/year (50% above current Utah state emissions). Water needs for cooling could be hundreds of millions of gallons per year. Scientists warn of local heat islands (+2–12°F) and Great Salt Lake harm.
- Labor/Economic Impacts: Hyperscale DCs create very few local jobs (~thousands at most). Instead, they enable automation across retail, services, logistics, and manufacturing. McKinsey/BLS studies predict ~30–50% of tasks automatable by 2030. Retail and fast-food workers score "High risk" (≈50% tasks automated) by 2030; software engineers are low risk (10%).
- Case Studies – Utah & Ohio: Stratos approved in May 2026 amid ~4,000 public objections. In Cleveland, a 150 MW Slavic Village DC permit was denied. Community activism led to moratoria proposals in Ravenna, Lordstown, Twinsburg, and Cleveland.
- Pushback: National resistance is growing. Maine passed a data-center moratorium (vetoed). Festus, MO overturned a DC tax deal. Ohio activists are seeking a statewide constitutional amendment capping DC size at 25 MW.
- Policy Options: Strict zoning, mandatory environmental impact reviews, regional energy planning, renewable energy requirements, community benefits agreements, and workforce retraining programs are all needed.
1. Primary Sources and Filings
Utah Stratos (Box Elder County): The Stratos Project was approved by the Utah Military Installation Development Authority (MIDA) and Box Elder County in May 2026. Official sources include: MIDA Board resolutions (motioned 5/2026), Box Elder Commission minutes (5/4/2026), and Utah Governor's FAQ. The FAQ (April 2026) calls it an "energy & tech infrastructure initiative," but contains no technical specs. No public electric interconnection filings or air/water permits are yet available; the project is too new for EPA or ISO filings to show. The developer (WestGenCo / O'Leary Digital) has issued statements via social media (Fox News interview, X posts) but official technical docs are sparse. Thousands of citizens submitted comments to Box Elder (attested in media), but these are not easily searchable.
Cleveland/Ohio Cases: In Ohio we examined city/county actions. Cleveland City Council records (April/May 2026) show discussion of data-center permits and a proposed moratorium. The Slavic Village DC permit (Lakeland Equity) was officially rejected May 2026. Cleveland Magazine reports that council considered a moratorium after permit rejection. Ravenna City Council minutes (April 2026) confirm they approved a moratorium (as cited in ClevelandMag). No formal filings (zoning, water) for Ravenna's proposed 100 MW TriBridge DC were found. At state level, Ohio EPA has no special filings found. ClevelandMag and Signal Cleveland provide key quotes from officials (Bibb, Slife).
Environmental Permits: No permit applications for Stratos or Cleveland sites were located in EPA/Energy databases. We explicitly note this gap: no open Ohio EPA or Utah DEQ permit documents. Similarly, no FERC/ISO interconnection for these projects is public yet. For national context, we cite IEA, EPA, DOE and academic studies rather than site-specific filings.
2. Motives & Economics
Tech companies are pouring billions into AI compute infrastructure. Key motives:
- Compute Arms Race: Companies (and the US government) want massive GPU clusters to train and run AI models. Bigger clusters cut training time and support more users. O'Leary explicitly framed Stratos as a response to China.
- Inference Economy: Recent analysis shows inference (running models) can generate revenue (cloud AI services), which requires vast parallel clusters 24/7. Data centers also serve as paid cloud infrastructure (analogous to cloud GPU instances).
- Geopolitics: Onsite statements tie projects to "national security" and "compute sovereignty." High-end chips and data services are seen as strategic assets.
- Energy/Tax Incentives: Locations like Utah and Ohio offered tax breaks, cheap land, and copious energy. Utah's governor noted Stratos as state-endorsed infrastructure, and Box Elder pitched it for jobs. Ohio legislators are drafting a data-center tax, indicating financial stakes.
- Capital/ROI: Publicly available ROI models are scant, but one analysis estimated Stratos costs $100B+ (electricity alone ~$4B/yr, plus capital). ROI likely depends on selling compute-as-a-service plus any federal grants (e.g. CHIPS Act funding). Cleveland's proposed 150 MW DC was valued at $1.6B. For perspective, a 150 MW facility at $10k/kW = $1.5B capex; plus ops ~150 MW × 8,760h × $0.05/kWh = $65M/yr in power.
- Local Economics: Proponents promise jobs (Stratos: "thousands"; Cleveland site: "job-creating"), but actual figures are modest (~2,000 max at Stratos). Critics call these numbers inflated given the scale.
In summary, hyperscalers are betting huge sums that owning enormous AI compute capacity will pay off via strategic advantage and future revenue, often with public subsidies. The economic model is high-risk, high-stakes.
3. Technical Capacity Math & Compute Estimates
A continuous 9 GW power draw implies an unprecedented computing capacity. Assuming a typical data center PUE (~1.2–1.3), the IT load would be on the order of 7–8 GW of servers and accelerators (with the rest for cooling etc.). High-end AI GPUs draw roughly 0.7–1.2 kW each.
- Utah Stratos (9 GW): Assuming PUE ~1.3 (IT load ~6.9 GW). At ~1 kW/GPU, that's ~6.9 million GPUs. If high-end GPUs use ~0.7 kW, maybe ~10 million GPUs. Each GPU ≈1–2 PFLOP (mixed-precision), so 10M × 1 PF = 10⁷ PFLOP = 10¹⁶ GFLOP = 10⁵ EFLOP. Inference capacity: at 10¹¹ FLOP/inference, 6.9M GPUs × 10¹⁵ FLOP/s = 6.9×10²¹ FLOP/s ~6.9×10¹⁰ inferences/s.
- Cleveland Slavic (150 MW): IT load ~115 MW. ~115k GPUs (1 kW/GPU). Compute ~1.15×10⁵ PFLOP = ~1.15×10² EFLOP. Inferences ~1.15×10⁵ GPUs × 10¹⁵ FLOP/s ≈ 1.15×10²⁰ FLOP/s → ~1.15×10⁹ queries/s.
- GPUs: ~1–3×10⁶ modern GPUs (NVIDIA H100/A100 or similar) in total for Stratos, depending on power/PUE assumptions.
- Storage: AI training often needs zettabytes of data (petabytes per model). If we assume ~10 PB of storage per MW as a ballpark, 9 GW could pair with ~90 PB–200 PB of storage (on NVMe or high-speed disk) to feed the GPUs.
Table: Compute by Site and Density Scenario
Assuming PUE=1.3, 8 GPUs/server, 10–50 kW/rack.
| Scenario | IT Load | GPUs (est.) | Racks (10 kW) | Storage | Compute (EFLOPS) | Inference Streams |
|---|---|---|---|---|---|---|
| Utah Stratos (9 GW) | 6.9 GW | ~7–10M | 690,000 | 1–3 EB | ~70–140 | ~6×10¹⁰ |
| Cleveland DC (150 MW) | 115 MW | ~115k | 11,500 | 1–5 PB | ~1.1–2.3 | ~1×10⁸ |
Storage assumes ~100 TB per MW. FLOPS: upper values assume 2 PFLOP/GPU. Inference Streams show simultaneous AI chat queries if each costs 10¹¹ ops.
Table: Estimated Components for 9 GW (Stratos) — Density Scenarios
Illustrative scenarios, assuming 10 kW per server with 8 GPUs, PUE≈1.2.
| Metric | Low-Density (30 kW/rack) | Mid (50 kW/rack) | High-Density (80 kW/rack) | Notes / Assumptions |
|---|---|---|---|---|
| IT Load | 7.5×10⁹ W (7.5 GW) | 7.5 GW | 7.5 GW | Assumes PUE≈1.2 (total 9 GW facility) |
| Racks | ~250,000 | ~150,000 | ~94,000 | 30/50/80 kW per rack densities |
| Servers per rack | 6 (42U rack) | 8 | 10 | e.g. 6×8-GPU servers @12.5 kW each |
| Total Servers | 1.5M | 1.2M | 0.94M | Servers × racks |
| GPUs per server | 8 | 8 | 8 | e.g. 8×GPU, 2×CPU per server |
| Total GPUs | 12M | 9.6M | 7.5M | Servers × GPUs |
| Storage | ~90 PB | ~150 PB | ~240 PB | ~10 PB/MW estimate |
| FLOPS (theoretical) | ~12–24 ExaFLOPS | ~9–18 ExaFLOPS | ~7–14 ExaFLOPS | At 1–2 PFLOP/GPU peak |
| Inference streams | ~10⁸–10⁹/s | ~7×10⁸–5×10⁹ | ~5×10⁸–4×10⁹ | If 1 user/inference = 10¹¹ ops |
Estimates are order-of-magnitude. GPU power varies (e.g. NVIDIA H100 ~0.7–1.2 kW under load). Storage needs range widely by workload.
Source: IEA projections. Data centers expected to consume ~945 TWh by 2030 — roughly double 2022 levels. Stratos's 9 GW alone would add ~78 TWh/year (~8% of projected 2024 global DC energy). Cleveland's 150 MW adds ~1 TWh/year (~0.15%).
4. AI Workloads & Use Cases
The Stratos facility, once built out, could host virtually any compute-intensive AI application. Key examples include:
- Large-Scale Model Training. Training advanced AI models (LLMs, vision models, reinforcement learners) can consume enormous computation. For example, OpenAI's GPT-3 (~175B parameters) required on the order of 3×10²³ FLOPs over weeks of training. Next-generation models (trillions of parameters) need orders of magnitude more. Stratos's exaFLOP capability could train many such models in parallel or accelerate single-model training by an order of magnitude.
- Inference/Deployment Services. Serving AI inferences to users (e.g. language model queries, image recognition) benefits from vast parallelism. A single GPU (at ~10¹⁵ FLOP/s) can handle tens of thousands of LLM tokens/sec; millions of GPUs in parallel could deliver ~10⁸–10⁹ queries/second if fully utilized. Stratos could power entire AI-as-a-Service platforms supporting huge global demand.
- Robotics and Reinforcement Learning (RL). Simulating robotic agents and environments uses GPU clusters. For example, OpenAI's Rubik's-cube robot trained via simulated RL in parallel on thousands of GPUs. A 9 GW center enables large-batch RL and real-time multi-agent simulations. (However, real-world robotics also needs fast I/O, safety clusters, etc.)
- Digital Twins and Industrial Simulation. Sectors like aerospace or power grids use digital twins (simulation replicas) that run complex physics on HPC. A wind turbine or jet engine twin might use thousands of CPU/GPU cores to simulate fluid dynamics. Stratos could host thousands of such concurrent simulations, aiding engineering design.
- Autonomous Vehicles (AV). AV development involves both training (simulating driving scenarios) and inference (cloud processing of sensor data for map updates, fleet learning). However, real-time driving inference typically happens on vehicles' local chips.
- Manufacturing/Process Control. AI-driven factories use machine vision and optimization. Stratos could train vision models for defect detection or optimize supply chains via large-scale ML.
- Security, Surveillance, and Defense. An AI campus could process satellite imagery, signal intercepts, and other defense data for real-time analytics. Video analysis (facial recognition, anomaly detection) on thousands of video streams in parallel or simulating battlefields with AI agents.
Each use case has distinct requirements. Training demands maximum peak FLOPS and memory; inference demands throughput and low latency. Robotics and AV impose low-latency loops (ms range). Surveillance/vision needs massive parallel image processing. Storage needs span petabytes per model for training data, while inference may rely on databases of embeddings or knowledge graphs.
5. Environmental Impacts
CO₂ Emissions: Using EPA factor ~0.4 kg CO₂/kWh (for modern gas turbines), Stratos (9 GW × 8,760h) emits ~3.15×10⁹ kg = 3.15 Mt CO₂ per year from electricity alone. This is roughly 50% above Utah's 2021 power-sector CO₂ (~30 million tonnes/year total). Cleveland example (150 MW): ~0.525 Mt/year (525,000 tonnes).
Water Use: Data centers often use closed-loop or evaporative cooling. Assuming 0.2–0.5 gallons/kWh: Stratos would require approximately 128–320 billion gallons/year. Cleveland (150 MW): ~260–660 million gallons/year. These withdrawals stress scarce water supplies, especially near the Great Salt Lake.
Thermal Output (Heat Island): Roughly equal to input power (minus turbine waste). Stratos: ~9 GW continuously as heat. Studies predict local temperature increases of +2–6°F during the day and +8–12°F at night. This could devastate the already-stressed Great Salt Lake. Cleveland: ~150 MW heat output.
Noise & Air: Hundreds of high-speed fans and turbines would create noise ~70–80 dB nearby. Air pollution (NOx, CO) from turbines requires permits. Utah DEQ will eventually evaluate these; local activists worry about Great Salt Lake toxics. In Ohio, a 150 MW gas plant emits ~0.3 Mt CO₂ and regulated NOx/SOx.
Land Use: Stratos covers ~62 mi² (mostly desert/agricultural land). Cleveland's Slavic center would have used 35 acres, likely removing several warehouses. Both replace previous land uses. Habitat loss and dust generation (Great Salt Lake exposure) are concerns.
6. Labor & Economic Impacts
Job Creation: Despite hype, hyperscale data centers create very few jobs. Stratos proponents claim "thousands" or "up to 2,000." In reality, a 9 GW facility might employ ~1,000–2,000 on-site. Cleveland's 150 MW DC promised "job-creating" benefits but likely only hundreds of jobs. Critics call these numbers inflated given the scale and capital investment.
Automation Displacement: The vast compute supply would accelerate automation. McKinsey finds ~30% of U.S. work hours automatable by 2030. Lower-wage jobs are hit hardest (14× more risk than high-wage jobs).
We classify industries/occupations by AI exposure and potential timelines, drawing on labor studies. Highly exposed roles include:
- Office and Administrative Support (e.g. clerks, data entry, reception). Many routine clerical tasks (scheduling, basic reporting) can be automated by generative AI. McKinsey estimates office support roles could see large declines. Risk score: High (4–5/5); plausible 50–70% tasks automated by 2030.
- Customer Service and Retail Sales (cashiers, call-center reps). AI chatbots and self-checkouts threaten these. BLS data show retail cashiers plateauing and then declining by 2029. Risk: High; up to ~40–50% of tasks by 2030.
- Food Service/Baristas/Waitstaff. Automated kiosks and robotic kitchens could cut these jobs. BLS projects fast-food counters growing (due to population/immigration), but turnover is high. Risk: High; ~50–60% tasks automatable, likely in 10–20 years.
- Transportation & Warehousing (truck drivers, forklift operators, stockers). Self-driving trucks, drones, and warehouse robots target these. Risk: Medium-high; some pilots suggest driverless trucks feasible in ~10 years.
- Manufacturing/Production Workers. Robotics already dominates auto/industrial assembly. Future dexterous robots could extend to electronics or apparel. Risk: High; advanced robotics could automate repetitive assembly within 10–15 years.
- Software and Creative Professionals. Studies find this group has lower exposure: generative AI typically augments these roles rather than replaces them. Risk: Low (1/5); automation is incremental and likely to create new jobs.
- Healthcare, Education, Skilled Trades. AI may assist (diagnostics, tutoring) but full automation is distant. Risk: Low; these jobs should grow with demographics.
Occupation Exposure Table
| Occupation Group | Exposure Score (1–5) | % Jobs At Risk (2030) | Timeline | Source/Notes |
|---|---|---|---|---|
| Retail Sales/Cashiers (≈7M) | 5 (Very High) | ~40–50% | ≤2030 | McKinsey; BLS projects retail down; kiosks/AI |
| Fast Food/Restaurant (≈5M) | 5 | ~50–60% | ≤2030 | McKinsey; high turnover, kiosk/robot use ↑ |
| Customer Support/Clerks (≈3M) | 4 | ~40–50% | ≤2030 | McKinsey; chatbots replace routine queries |
| Administrative/Executive (≈3M) | 4 | ~50–70% | ≤2030 | McKinsey; AI scheduling/analytics cut clerical tasks |
| Warehouse Workers (≈2.1M) | 4 | ~50% | ≤2035 | Amazon-level automation; robotic picking in use |
| Manufacturing Assemblers (≈1.9M) | 4 | ~60% | 2030–2040 | Existing robotics can handle many tasks; dexterous bots upcoming |
| Trucking/Logistics (≈2M) | 3 | ~30–60% | 2035–2045 | Autonomous vehicles ongoing; slow regulatory rollout |
| Personal Financial Advisors (≈270k) | 3 | ~20–30% | 2030+ | Robo-advisors for routine planning; complex cases stay human |
| Lawyers/Legal Clerks (≈1.1M) | 3 | ~30% | 2030+ | Document review via AI; counsel roles persist; BLS growth |
| Registered Nurses (≈3M) | 2 (Low-Med) | ~10–20% | 2040+ | AI diagnostics assist; hands-on care needed |
| IT/Engineers/Managers | 1 (Low) | ~5–10% | >2040 | Augmentation > replacement; AI coding tools (e.g. Copilot) |
| Health/Education | 2 (Low) | ~10–20% | >2040 | Requires human touch; AI assists but doesn't replace |
| Total U.S. Workforce (~170M) | — | ~30% of hours automated | by 2030 | McKinsey (2023); OECD confirms lower-wage workers at 14× higher risk |
Fig 3. BLS-Projected 2019–2029 Job Changes for Selected Occupations
Source: Bureau of Labor Statistics occupational projections. Positive = growth; negative = decline. Note BLS projects fast food growing due to population growth, even as automation risk remains high.
Studies (McKinsey, OECD, BLS) converge on ~20–30% of jobs being displaceable within 10–15 years. Frontline workers and small businesses (retailers, cafes, drivers, janitors) face the greatest impact. McKinsey forecasts ~12 million occupational transitions needed by 2030. However, fully displacing these jobs will take years of technology adoption, regulation changes, and capital deployment.
7. Supply Chain & Manufacturing (Reshoring)
One implicit promise of Stratos is that massive AI+robotics could reshore U.S. manufacturing, reducing reliance on Chinese-made goods. We find this vision faces severe hurdles:
- Robotics capabilities. Advanced robots for discrete tasks (welding, assembly) are mature; indeed, China now leads in robot density. However, many factory jobs require dexterous manipulation and adaptability (e.g. handling varied small parts, textiles, electronics). Achieving true human-like dexterity is still in early R&D stages. Roboticists emphasize the "hands" problem: replicating human fingers and touch sensors is hugely complex. Even companies like HaptX and Agility Robotics train humanoids via VR gloves as a research step, but caution it will be years to a decade before such robots can perform diverse tasks at scale. In the meantime, cheaper fixed robots (pick-and-place, assembly arms) suffice for repetitive jobs; there is no general-purpose "robot factory" ready to turn out all goods.
- Materials and components. AI hardware (GPUs, CPUs) and robots require critical minerals (gallium, germanium, rare earths, lithium, etc.). The U.S. is heavily reliant on Chinese supply chains for these. For instance, China controls ~98% of global low-grade gallium and ~60% of germanium refining. Data centers themselves need massive copper, aluminum and batteries for power systems. Any "reshoring" of electronics manufacturing will be constrained by these inputs. While AI can help mineral discovery and recycling, current reserves and processing capacity are geopolitical chokepoints.
- Capital investment. Building a fully automated factory is extremely costly. Even with AI, a new automotive plant costs $1–2 billion, and has to operate at high utilization to pay off. Smaller-scale goods (electronics, textiles) often have razor-thin margins that undercut expensive automation. Moreover, China's manufacturing advantage isn't just labor cost: it has fully integrated ecosystems of tooling, components, and assembly lines. China's "colocated" supply chains (e.g. many robot toolmakers near Shenzhen) create network effects that are hard to replicate.
- Energy costs and environment. Offshoring was partly driven by cheap power and lax regulation. Running millions of robots and data servers in the U.S. requires abundant energy (and water for cooling). Utah's Stratos would be gas-fired to power AI; pursuing heavy manufacturing (steel, petrochemicals) here would similarly strain resources and clash with clean-energy goals.
In summary, while AI+robotics can make U.S. factories more efficient, wholesale replacement of China-made goods is not imminent. Technology trends point to gradual reshoring where economics permit (e.g. semiconductors under CHIPS Act). But current limitations in robot dexterity, supply-chain bottlenecks, and high investment mean any broad "America (re)Make Everything" via Stratos would take decades. One study finds that AI influences both offshoring and reshoring depending on firm strategy, but does not guarantee back-to-the-U.S. China continues to dominate the robotics race and has a mature auto/smartphone manufacturing base.
8. Geopolitical and Energy Implications
Grid and Siting: No existing grid in rural Utah can handle 9 GW. Both Box Elder officials and analysts note Stratos plans on-site generation (natural gas turbines, possibly combined-cycle) tied to pipelines. This means building one of the largest power plants in the U.S. co-located with the data center. Utah's clean energy goals (move to renewables) may conflict with a massive gas-fired plant. The developer initially claimed 100% natural gas power; later O'Leary mentioned interest in renewables, though as of May 2026 no firm utility contract was public. Local concerns include heat pollution and nontrivial water usage for cooling.
Utah vs. California: The choice of Utah (with few existing hyperscale centers) reflects cheap land, pipeline gas, and tax incentives. California's Bay Area has huge tech demand but also land/permits scarcity and higher costs. Strategically, Stratos is framed as "U.S.-based capacity" to avoid reliance on foreign AI infrastructure. Congress and the Pentagon have expressed interest in ensuring "compute sovereignty." However, analysts caution that Stratos is not a formal DoD project (MIDA is a state authority).
Critical Minerals & Chips: AI chips need gallium, germanium, rare earths and silicon. The U.S. sources little of these domestically. Recent reports highlight a looming critical minerals crunch: a 30% drop in gallium supply could cut U.S. GDP by 2%. Building Stratos (and the AI industry around it) thus ties the American economy to fragile supply chains. Control of these materials (mostly in China/Russia) is now as critical as conventional energy or microchips.
Defense and Competition: Proponents (O'Leary, officials) claim Stratos strengthens national security by outpacing China in AI computing. The logic is: whoever has the largest compute will lead AI innovation for defense. China has cited AI/data centers in its 5-year plans, and analysts warn that China has already surpassed the US in robotics deployment. But it is controversial whether such infrastructure should be state- or federally-subsidized, given climate and labor trade-offs.
Grid Impact: Scaling multiple 9-GW centers nationwide could demand a new grid paradigm. For comparison, the entire Utah grid is ~4.5 GW peak; Stratos alone is double Utah's total usage. Some states (Washington, Oregon) already worry about data center booms draining renewable energy and increasing electricity prices.
9. Case Study: Utah Stratos
Project Timeline
Energy plan: Company claims 100% natural gas power. Utah officials initially confirmed on-site gas turbines supplying 9 GW. Later, O'Leary mentioned renewables interest. As of May 2026, no firm utility contract was public. A 9 GW dedicated generation is unprecedented; sources say new pipelines and power plants would be needed.
Local Pushback: Over 3,000–4,000 citizens filed objections. Protests at meetings grew intense. Box Elder Commissioner Perry received death threats. Concerns center on water use, air quality, Great Salt Lake drying, and dust from dry lake bed. Environmental groups (Sierra Club) branded it "irresponsible." Opposition remains unresolved, with the Sierra Club threatening lawsuits and citizens continuing to lobby.
Outcomes: Project is approved but requires state oversight. The Box Elder officials emphasize state DEQ permits for emissions. Governor and DOD tacitly support it for AI leadership. Key approvals to date authorize planning and permitting applications but do not yet approve actual building.
10. Case Study: Cleveland / Ohio
Slavic Village, Cleveland (150 MW / $1.6B)
In early May 2026, Lakeland Equity (MI developer) applied to build a 150 MW, $1.6B data center on 35 acres in Slavic Village — a 300,000 sqft, 2-story facility. After less than two weeks, Mayor Justin Bibb announced the permit was rejected (city zoning denial), citing "serious concerns" and highlighting the residential context. No detailed written reason was published. The developer vowed to consider next steps.
Councilman Slife (Ward 15) introduced a 1-year moratorium bill in April 2026, still pending as of mid-May. The moratorium would pause new DC approvals to craft regulations. Mayor Bibb endorsed the effort publicly.
Ravenna & Other Ohio Communities
Just outside Cleveland, Ravenna saw a proposed ~100 MW DC (TriBridge). In April 2026, resident Will Hollingsworth's city council speech went viral nationwide. Immediately, Ravenna City Council passed a moratorium on new data centers. Lordstown, Twinsburg, and other communities followed suit.
Ohio state activists are pushing an initiative to cap future DCs at 25 MW via constitutional amendment. "Ohio Residents for Responsible Development" gathered 330,000+ petition signatures.
Outcomes: At least one local project was blocked (Cleveland). Others were delayed or scaled down. The Ravenna/Cleveland cases indicate that even small to mid-sized communities are mobilizing to resist, and that community activism can successfully halt projects through normal civic channels.
11. Political & Social Pushback
Nationwide, backlash is growing against hyperscale data centers:
- Maine passed an 18-month data-center moratorium in 2024 (later vetoed by Gov. Mills, though Maine did enact a law withdrawing tax incentives for data centers).
- Festus, MO voters ousted a city council that approved a data-center tax incentive.
- Saline Township, MI refused a referendum but still allowed Oracle's DC after public protest.
- San Francisco Bay Area considered special surcharges for big DCs.
- Ohio: "Ohio Residents for Responsible Development" gathered 330,000+ petition signatures for a constitutional amendment to cap DC size at 25 MW.
- Utah: Over 4,000 citizen objections, death threats against commissioners, ongoing legal challenges from the Sierra Club.
Community demands typically include: independent environmental studies, local hiring guarantees, renewable energy commitments, and impact mitigation (e.g. water source assurances). As Ravenna's Hollingsworth put it, this issue crosses political lines — he noted "I have moral obligations" to fight the project regardless of political affiliation.
The "winning narrative" of job creation clashes with the on-the-ground concerns of residents worried about water, wildlife, and jobs disappearing elsewhere due to automation.
12. Policy Options and Mitigation
To address these challenges, policy must balance innovation with social and environmental costs. Options include:
- Zoning & Moratoria: Municipalities can zone hyperscale data centers separately or impose temporary moratoria (as Cleveland and Ravenna are doing). Example: Oregon requires special use permits for DCs >50 MW. States like Maine have moved to limit mega data centers. Utah and other states may follow with stricter land-use reviews, environmental impact studies, and utility regulations requiring renewables for new large loads.
- Environmental Review: Mandate EIS/EA for large DC projects. States can classify >X MW as "major stationary sources" under Clean Air Act requiring PSD permits. Utah and Ohio should ensure DEQ/OH EPA rigorous review of emissions and water impact before approvals.
- Renewable Energy: Require or incentivize onsite renewables or clean energy purchase. California has forced large DCs to buy 100% renewable energy. Integrate the Stratos load into grid planning by requiring on-site solar/battery backup, carbon pricing, or dedicating renewable projects to offset the carbon footprint.
- Community Benefits Agreements: Negotiate upfront community benefits: funding for housing, internet access, or environmental restoration. Even with low direct job numbers, DCs could pay into local workforce retraining or green infrastructure funds. Rural communities hosting data centers can negotiate impact agreements (as Box Elder did) for noise limits, job quotas, and environmental safeguards.
- Utility Planning: Integrate DC demand into grid planning. Regulators can charge standby or peak surcharges, ensure DCs pay for necessary transmission upgrades (to avoid burdening other customers). Washington State passed laws requiring utilities to recover DC costs from customers, not general ratepayers.
- Workforce Development: The most common recommendation is massive workforce development. This includes upskilling displaced workers into tech-support or construction roles, retraining programs for coding/AI-lab roles, and education in non-automatable careers (healthcare, trades). McKinsey suggests up to 12 million American workers must transition by 2030. Policies: expanded community college grants, apprenticeships in robotics maintenance, portability of benefits to ease career shifts, and federal/state partnerships with labor unions and colleges.
- Taxation & Incentives: Some advocate taxing large data centers (e.g. "data center utility tax" or "Gigawatt tax") to fund local schools and infrastructure, and to internalize environmental costs. Conversely, targeted subsidies (like CHIPS Act grants) could steer AI compute to grid-friendly builds. For example, the federal Infrastructure law offers grants to industrial sites; states could tie incentives to clean energy use.
- Antitrust and Competition: Technology-transfer rules or joint ventures to ensure the U.S. also builds advanced chips and robotics. Export controls on AI chips (like those on high-end GPUs) could affect Stratos's software supply. Policies might encourage U.S.-based AI software companies to use domestic compute (via priority contracts, "buy American" AI procurement).
- Critical Minerals Strategy: At federal level, data centers highlight reliance on overseas chip inputs. Policies to expand domestic chip fabrication and mineral refining would relieve strategic bottlenecks (Dept. Energy's critical minerals program, CHIPS Act expansions).
- Economic Diversification: Officials should also attract varied industries (e.g. biotech, light manufacturing) to counterbalance tech concentration, reducing local over-reliance on one sector's instability.
- Regional Planning: Nationally, planning forums could anticipate clusters of AI campuses and their cumulative effects (on migration, housing, education needs).
In summary, policy must address: (a) labor — prepare workers for change; (b) energy — ensure power systems and environment can handle these loads; (c) competition — support domestic tech industries while protecting critical resources; and (d) community — ensure host communities share in benefits and are protected from harms. The rush to build should be guided by comprehensive planning, not last-minute reaction.
Sources
We drew on primary documents, media reporting, and industry/academic analyses. Where explicit data was lacking, we stated assumptions and ranges (e.g. GPU wattage, PUE, storage per MW) and grounded them in cited sources. All estimates and narratives are based on available literature as of 2026.
- Guardian — 'Irresponsible': backlash as Utah approves datacenter twice the size of Manhattan
- KSL.com — Box Elder County official faced death threats amid contentious data center debate
- Fortune — Shark Tank's Kevin O'Leary is building a $100 billion data center in Utah. Residents are revolting
- Box Elder County — Box Elder County Authorizes MIDA to Initiate Stratos Project
- Utah Governor's FAQ — FAQ on Stratos Project
- Salt Lake Tribune — Utah's data center could create a massive heat island that could devastate the Great Salt Lake, scientists say
- Signal Cleveland — Cleveland rejects Slavic Village data center permit
- Cleveland Magazine — Inside One Ravenna Resident's Fight Against Data Centers
- NEOtrans — $1.6B data center planned in Cleveland
- McKinsey MGI — Generative AI and the future of work in America
- McKinsey — Jobs lost, jobs gained: What the future of work will mean for jobs, skills, and wages
- OECD — Who will be the workers most affected by AI?
- BLS — Growth trends for selected occupations considered at risk from automation
- IEA / Hanwha — What Are the Power Requirements for AI Data Centers?
- EPA — Greenhouse Gas Equivalencies Calculator
- Epoch/OWID on AI compute; IEA data center energy reports; robotics industry articles; critical-minerals research
