Softbank Ohio 10GW Data Center: A Practical Guide for AI Infrastructure Decision Makers

1. Problem definition

Who is this article for: CTO/technical leader considering building AI infrastructure, cloud cost optimization manager, AI startup infrastructure architect

What problem does it solve:

With the advent of 10GW mega sites, AI infrastructure options are rapidly expanding
Existing cloud (AWS/Azure/GCP) vs. co-location vs. dedicated infrastructure Decision-making criteria are different
Unclear how to evaluate power acquisition bottlenecks, cooling infrastructure costs, and lead time risks

Scope of application: Organizations operating or planning 100kW-10MW AI workloads, MLOps/platform teams requiring GPU cluster cost optimization

Scope of application: Small-scale inference-only workload (single server level), regulatory environment that cannot be deployed on-premise

2. Evidence and comparison: Softbank 10GW project fact check

2.1 Project Overview (announced on March 20, 2026)

Item	Number	Source
Location	Piketon, Ohio (formerly DOE Uranium Enrichment Facility)	DOE official announcement
Total power capacity	10GW (9.2GW natural gas + 800MW initial)	Tom's Hardware, StateNews.org
Investment size	$30-40B (data center) + $33B (power plant) ≈ $66B+	Multiple Sources
Construction/operation	Construction begins in 2026 → 800MW begins operation in 2028 → Total in the late 2030s	DOE
Employment	Construction for up to 35,000 people + Operation for 2,500 people	StateNews.org
Site	3,700 acres (with Cold War high-density transmission infrastructure)	DataCenterDynamics

Caution: Some media reported "$500 billion" or "750 trillion won", but this is a misinformation. The actual confirmed amount is around $66B+.

2.2 What 10GW means

Comparison context:

8 times the output of Hoover Dam
1/3 of Ohio's total power generation (~30GW) in 2024
60% of new U.S. data center power (16.7GW) in 2026

Power Density Trend:

2020: 10-14kW/rack
2026: 100-300kW/rack (based on AI GPU)
2027 projection: 176kW/ft² (Programs.com statistics)

2.3 Comparison of alternatives: 10GW megasite vs. existing options

Criteria	Public Cloud	Colocation	Mega site (10GW level)
GPU unit price	$2-4/GPU-hr (On Demand)	Exclusive contract required	Large discounts available
Securing power	Cloud company responsibility	1-3 year lead time	Including own power plant
Scalability	Immediately	Several months - 1 year	Years
Initial investment	None	Tens of millions of dollars	Hundreds of millions of dollars
Suitable workload	Burst/Experiment	Medium-scale sustained	Large-scale learning

Judgment criteria:

Annual GPU cost $10M+ → Start co-location/dedicated review
Annual GPU cost $100M+ → Consider mega site partnership or self-construction
Lead time for securing power 2 years+ → Bridge to the cloud first

3. How to do it step by step: AI Infrastructure Capacity Planning

Step 1: Calculate current workload power

#Example of TDP standards for each GPU (requires actual PUE to be applied)
H100_TDP=700  # W
NUM_GPUS=1000
PUE=1.3       #Cooling included

TOTAL_POWER_KW=$(echo "($H100_TDP * $NUM_GPUS * $PUE) / 1000" | bc)
echo "Power required: ${TOTAL_POWER_KW} kW"
#Based on 1000 H100 → approximately 910kW

Step 2: 3-year forecast based on growth rate

#Simple capacity planning script
current_gpu_count = 1000
annual_growth_rate = 1.5  #Assuming 50% YoY growth
years = 3

for year in range(1, years + 1):
    projected = current_gpu_count * (annual_growth_rate ** year)
    power_kw = (projected * 700 * 1.3) / 1000
    print(f"Year {year}: {projected:.0f} GPUs → {power_kw:.0f} kW ({power_kw/1000:.2f} MW)")

Step 3: Lead time reversal schedule

Milestone	Lead time	Completion date starting in 2026
Power contract negotiation	6-12 months	Q1 2027
Grid connection approved	12-24 months	Q1 2028
Facility construction	18-36 months	2029 Q1
GPU procurement/installation	6-12 months	Q3 2029

Key Insights: If decided now, operation possible in 2029. Cloud bridge is essential if needed by 2027.

Step 4: Cost Modeling

#Annual Cost Comparison (Simplified)
cloud_hourly = 2.5  # $/GPU-hr
colo_monthly = 500  #$/GPU-mo (power+space)
utilization = 0.7   #70% operation rate

gpu_count = 1000
hours_per_month = 730

cloud_monthly = cloud_hourly * hours_per_month * utilization * gpu_count
colo_monthly_total = colo_monthly * gpu_count

print(f"Cloud: ${cloud_monthly:,.0f}/month")
print(f"Colocation: ${colo_monthly_total:,.0f}/month")
print(f"Saving rate: {(1 - colo_monthly_total/cloud_monthly) * 100:.1f}%")
#Cloud: $1,277,500/month
#Colocation: $500,000/month
#Savings rate: 60.9%

4. Pitfalls: Hidden costs of GW-level infrastructure

Pitfall 1: Difference between power contract and actual available power

Issue: Contracted capacity 100 MW ≠ Immediately available 100 MW

Cause: Substation capacity, grid stability, peak load limit

Prevention: Check “firm power” vs “interruptible power” contract terms, negotiate curtailment terms during peak hours

Recovery: Secure own backup generator (diesel/natural gas), at least N+1 redundancy

Pitfall 2: Underestimating cooling infrastructure costs

Problem: Air cooling not possible at 100kW+ rack density

Numerical: Air cooling limit 50-100kW/rack, AI rack actual 100-300kW/rack, cooling costs add 20-30% of total energy

Prevention: Reflected in immersion or direct-to-chip design, with additional 30% cooling CAPEX reflected in the design stage

Restoration: Existing air cooling facility → Immersion cooling retrofit cost $100K+/rack

Pitfall 3: Water conflict risk

Problem: GW-class cooling tower = millions of gallons of water consumed per day

Case: Data center water use regulations tighten in Arizona, Georgia and Texas, with permits denied or delayed in some areas

Prevention: Prioritize closed-loop cooling system review, require technology to achieve 80%+ water recycling rate, prioritize areas with abundant water resources (Ohio, Northern Europe)

Repair: Add air-cooled hybrid system → Increase cost by 50%+

Pitfall 4: Grid Queuing Risk

Issue: East US grid (PJM) connection queue 3-5 years

Numbers: PJM data center additional load 7.9GW in 2026, capacity cost increase $270/MW-day (9.3 times increase)

Prevention: “Behind-the-meter” self-generation review (Softbank model), evaluation of alternative grid regions such as ERCOT (Texas)

Repair: On-site natural gas/SMR installation → Additional $1B+ investment

Pitfall 5: 2030 power shortage scenario

Problem: US projected power shortage of 49 GW in 2028

Probability: Only ~54% of announced 157GW projects are feasible (Janus Henderson analysis)

Prevention: Multi-regional distribution strategy, early securing of long-term power agreement (PPA), review of renewable energy + battery hybrid

Recovery: Project delay → reflected in cloud bridge cost plan

5. Action Checklist

Items to check before making infrastructure decision

☐ Current GPU annual cost Accurately calculated (separate on-demand + reservation)
☐ 3-year workload growth rate Evidence-based forecast completed
☐ Power requirement Calculated based on PUE 1.3
☐ Lead time Matching inversion schedule and business schedule
☐ Cooling method Check availability for 100kW+ rack density
☐ Water resource regulation Check the current status of the area
☐ Grid queue Check average connection time
☐ Backup power Establish N+1 redundancy plan
☐ Cost modelComparison of cloud vs colocation vs dedicated
☐ Partner evaluation Review of Softbank/Oracle/AWS long-term contract terms

Definition of Done

"Infrastructure decisions are complete": A three-year capacity plan has been developed, a power contract LOI (Letter of Intent) has been signed, and a cloud bridge strategy has been documented.

6. Reference

7. Author's perspective

Recommended: If your annual GPU spend is over $10M and you plan to run AI workloads for more than 3 years, you should start negotiating power contracts now. Even if it starts in 2026, actual operation will be in 2029. If you ignore this lead time, there will be no alternative when cloud costs surge during the power shortage phase in 2028.

Not recommended: If the workload is uncertain or the scale is less than MW, cloud flexibility is more advantageous than dedicated infrastructure. If you forcefully enter colocation without the financial capacity to cover the initial investment payback period of 5 years+, the burden of fixed costs will slow down the speed of innovation.

If other choices are better:

High workload volatility → Cloud On-Demand + Spot combination
Latency sensitive inference → Edge co-location distribution
Regulatory requirements (data residency) → Cluster dedicated to your region

Conclusion: Softbank 10GW project is changing the rules of the AI infrastructure game. A new formula has been established: “Whoever secures power controls AI.” Whether you take this change as an opportunity or remain dependent on the cloud depends on your decision now.