The Window Is Closing

$650 billion is chasing AI compute this year.
Half of it can't find power.

50%

of US data centers canceled or delayed

12 GW

planned — only 4 GW being built

3-5 yrs

to build grid transmission

The bottleneck isn't money or demand. It's power delivery. Transmission costs $1-3M per MW and takes years. Meanwhile, flare gas burns at oil wells, solar farms pay negative prices to dump power, and remote hydro sits idle.

The companies who deploy compute at the energy source in the next 12-18 months will lock in the market before hyperscalers do it themselves.

1 / 7

The Product

A shipping container that turns power into AI revenue.

Three standard ISO containers. Built in 50 days. Installed in 48 hours. No construction permits.

Container	What's Inside	Capacity	Cost
Power (40ft)	800kW UPS, batteries, PDU, monitoring, fire, CCTV	Supports 3-4 racks	$496K
IT (20ft)	1 GPU rack, liquid cooling CDU (160kW)	1 rack each	$129K
Chiller (20ft)	Water chiller (500kW), cold storage, pumps	Supports 2-3 racks	$66K

Power and chiller are shared infrastructure — deploy once, then add IT containers as you scale. Built by KSTAR — China's #1 UPS manufacturer, publicly traded. Compare: traditional DC costs $500M+ and takes 2-3 years.

2 / 8

How It Scales

Each additional rack costs less. Power and chiller are already there.

Rack #	What You Add	Cost
1st	Power + Chiller + IT + GPU rack	$6.4M
2nd	Just IT container + rack	$5.85M
3rd	IT container + smaller rack	$4.0M

2nd rack saves $550K. 3rd saves $2.4M. You don't duplicate the $496K power container or $66K chiller — they have capacity for 3 racks.

Power Required at the Wall

1 rack (104 B200)	205 kW
2 racks (172 B200)	340 kW
3 racks (240 B200)	479 kW
AMD MI300X (104 GPU)	115 kW
Huawei Ascend (128 GPU)	72 kW

Includes GPUs + chiller + CDU pumps + UPS losses. PUE ~1.4x.

How Power Flows

Continuous feed required. Grid or generator power flows through the UPS to racks 24/7.

Batteries are backup only. Standby for grid drops — 15 min to 4 hrs bridge power. Not a battery system.

3 / 8

Choose Your Investment

Same $690K container. Three GPU tiers.

Huawei Ascend 910C

$0.6M

128 GPUs · 1,800 tok/s each

DeepSeek, Qwen models

Total site power: 72 kW

11 month payback · $0.65M/yr profit

Pilots & emerging markets

Best Value

AMD MI300X

$1.9M

104 GPUs · 3,500 tok/s each

All open-source models

Total site power: 115 kW

7.1 month payback · $3.2M/yr profit

80% of NVIDIA performance, 25% of the price

NVIDIA B200

$6.4M

104 GPUs · 10,000 tok/s each

All models, max throughput

Total site power: 205 kW

7.8 month payback · $9.9M/yr profit

Scale to 3 racks (240 GPUs): $14.3M → $22.4M/yr

GPUs are swappable. Start with AMD or Ascend. Upgrade to NVIDIA later. The container doesn't change.

Total site power includes GPU servers + cooling (CDU pumps, chiller compressor) + UPS conversion losses + monitoring. PUE ~1.4x.

4 / 8

The Revenue

94-96% gross margins. 10-50x more per kWh.

Self-Hosted AI Margins

Model	Market Price	Your Cost	Margin
Llama 4 Maverick 400B	$1.25/M tok	$0.05	96%
Qwen 3.5 72B	$0.91/M tok	$0.04	96%
DeepSeek V3.1	$0.32/M tok	$0.02	94%

Input vs output: Output tokens cost 3-8x more (each requires full model read from memory). Typical blend: ~$0.95/M. We model at $0.50 — conservative.

Revenue Per kWh

Grid sales

$0.05

Bitcoin mining

$0.08

AI inference

$0.50 — $3.00 (10-50x)

Power is only 2-12% of token revenue. Whether you pay $0.02 or $0.15/kWh barely moves the math. The token price is the business.

Even at expensive power ($0.15/kWh), the Standard config still nets $9.2M/year.

5 / 8

Why You Have to Move Now

This window exists for 12-18 months. Then it closes.

1. The US grid is failing

12 GW of data centers were supposed to open in 2026 — only 4 GW are being built. Big Tech is spending $650B+ but can't get power. Transformers have 3-year backlogs. Crusoe, Giga Energy already proved container-at-wellhead works with Bitcoin — we upgrade to AI at 10-50x the revenue.

2. Open-source AI just reached parity

Llama 4, DeepSeek R1, Qwen 3.5 match GPT/Claude at 90-95%. Anyone with GPUs can serve enterprise-grade AI. Self-hosted margins: 94-96%. This happened in the last 6 months.

3. Hyperscalers are coming

Google, Amazon, Microsoft building their own container programs. When they arrive with unlimited capital, margins compress. First movers who lock in energy sites and customer contracts now have the defensible position.

Where the Opportunities Are

Market	Why Now
US	Permian flare gas, ERCOT curtailed wind/solar, Georgia Power 3 GW solar+BESS with off-peak excess
India	$30B DC CAPEX by 2030, 500 GW renewables, PLI incentives, 50+ customers. Assembly saves $2.5M/shipment to US
Middle East	0-5% duties, 154% gross margin on container sales
Africa	223 DCs across 38 countries — greenfield. Green Flare building 50+ MW flare gas DCs in Nigeria
Sri Lanka	Largest private renewable company ready to deploy at existing sites

Crusoe & Giga proved container-at-wellhead with BTC at $0.08/kWh. We do AI at $1.50/kWh. Same playbook. 10-50x revenue.

6 / 8

The Team

Three domains. One product.

🧠

M.

Technologist & AI Expert

AI models, GPU selection, inference software (vLLM, TensorRT), token economics, ROI modeling

Silicon Valley

⚡

V.

Energy Systems Expert

Power infrastructure, tariff & trade analysis, manufacturing economics, India/ME business development

Dubai

💧

O.

Cooling Systems Expert

Liquid cooling, CDU/chiller engineering, water treatment, Chinese manufacturing & factory relationships

Lisbon / Guangzhou

Partners: KSTAR (container infrastructure) · Supermicro (GPU server racks) · CN Water Systems (India assembly)

7 / 8

Next Step

Tell us your power and budget.
You're earning in 10 weeks.

Start with any GPU tier

GPU	Investment	Annual Profit	Payback	Power
Huawei Ascend	$0.6M	$0.65M	11 mo	72 kW
AMD MI300X	$1.9M	$3.2M	7.1 mo	115 kW
NVIDIA B200	$6.4M	$9.9M	7.8 mo	205 kW

Then scale — each additional rack is cheaper

Rack	Add Cost	Why Cheaper	Cumulative
1st	$6.4M	Full infra (power + chiller + IT)	$6.4M
2nd	$5.85M	Shared power & chiller — save $550K	$12.3M
3rd	$4.0M	Shared infra + smaller rack — save $2.4M	$16.3M

Every month you wait is a month of margin someone else captures.

8 / 8