ROI & Product Reference

Complete ROI, product configurations, tariffs, and market analysis. All numbers from actual quotes (manufacturing partner quote (March 2026), Supermicro January 2026, OpenRouter April 2026).

GPU Comparison

Spec	NVIDIA B200	NVIDIA H100	AMD MI300X Best Value	AMD MI325X	Huawei Ascend 910C
Price per GPU	~$55,000	~$25,000	~$12,000	~$18,000	~$1,800
Inference (tok/s, 70B)	10,000	3,000	3,500	5,000	1,800
tok/s per Watt	10.0	4.3	4.7	5.0	4.5
TDP per GPU	1,000W	700W	750W	1,000W	400W
Memory	192 GB HBM3e	80 GB HBM3	192 GB HBM3	256 GB HBM3e	128 GB HBM2e
Memory Bandwidth	8.0 TB/s	3.35 TB/s	5.3 TB/s	6.0 TB/s	1.8 TB/s
Cost per TB/s bandwidth	$6,875	$7,463	$2,264	$3,000	$1,000
Software Stack	CUDA, TensorRT-LLM, vLLM, SGLang	Same	ROCm, vLLM, SGLang	Same	MindSpore (limited)
Open-source models	All	All	All	All	DeepSeek, Qwen
Cooling	Liquid (DLC)	Liquid (DLC)	Liquid (DLC)	Liquid (DLC)	Air OK at 400W
When to choose	$5M+ budget, max throughput	Legacy, cheaper than B200	$1-5M budget, best $/performance	More memory than MI300X	<$1M, emerging markets

Memory bandwidth is the key metric for LLM inference. Every token requires reading the full model from memory. More bandwidth = more tokens/second. MI300X is 3x cheaper per TB/s than B200 — the value play for open-source models. Ascend is 7x cheaper per TB/s but limited model support.

Product Configurations

Server Specs (Supermicro ARS-121GL-NB2B-LCC)

Spec	Value
Form factor	1U rackmount
Dimensions	438mm W × 44mm H × 766mm D (17.3" × 1.7" × 30.2")
Weight	32 kg (70.5 lbs) per server
GPUs	4× NVIDIA B200 (NVL4, liquid cooled)
CPUs	2× NVIDIA Grace 72-core (liquid cooled)
Memory	960 GB LPDDR5X
Networking	4× 800G ConnectX-8 SuperNIC
Power	5.4 kW per server (48-54V DC busbar)
Cooling	Direct-to-chip cold plate

Rack Configurations (from Supermicro Saqrion quote, Jan 2026)

Rack	Servers	GPUs	IT Power	Rack at Wall	Price	Per Server	Per GPU
C1 (SRS-48LB26-NVL4-C1)	26	104 B200	140 kW	~150 kW	$5,717,770	$219,914	$54,978
C2 (SRS-48LB26-NVL4-C2)	17	68 B200	92 kW	~100 kW	$3,871,812	$227,754	$56,938

Each rack includes: 8× power shelves (33kW Delta/LiteOn AC→48V DC), 1× in-rack CDU, NVLink switch rails, cabling, shipping crate, rack integration service. 3yr labor warranty, 1yr parts. EXW San Jose.

Full System Configurations — Corrected for Total Site Power

Total site power = GPU IT load + power shelf losses (3-5%) + in-rack CDU + external CDU pumps + inrow cooling fans + chiller compressor + UPS losses + monitoring. PUE ~1.4-1.5×.

Config	Racks	GPUs	IT Load	Total Site Power	CDUs	Chiller Use	UPS Use	Containers	Footprint
Starter	1× C2	68 B200	92 kW	140 kW	1	28%	18%	20ft + 20ft	36 m²
Standard	1× C1	104 B200	140 kW	205 kW	2	41%	26%	40ft + 20ft + 20ft	72 m²
Full	C1 + C2	172 B200	232 kW	340 kW	2-3	68%	43%	40ft + 20ft + 20ft	72 m²
Max	C1 + 2×C2	240 B200	324 kW	479 kW	3	96%	60%	40ft + 40ft + 20ft	108 m²
2×C1 + C2	—	276 B200	373 kW	535 kW	3	107%	67%	—	Exceeds chiller

2×C1 + C2 is not viable with a single 500 kW chiller container — exceeds capacity at 535kW total heat rejection. Max viable config is C1 + 2×C2 (240 GPUs, 479 kW, 96% chiller). To go higher, add a second chiller unit (~$50-80K).

Non-NVIDIA Configurations

Config	GPUs	IT Load	Total Site Power	Cooling	Containers
AMD MI300X (104 GPUs)	104× MI300X	78 kW	115 kW	DLC (1 CDU)	20ft + 20ft
AMD MI325X (104 GPUs)	104× MI325X	104 kW	150 kW	DLC (1-2 CDU)	20ft + 20ft
Huawei Ascend (128 GPUs)	128× 910C	51 kW	72 kW	Air-cooled OK	20ft + 20ft

Huawei Ascend at 400W/GPU can be air-cooled — no DLC required. Simplifies the container significantly (no chiller needed in temperate climates).

Space: What Fits Where

Container	Internal Space	Max Racks	Notes
20ft IT	5,500 × 2,300 × 2,600mm	2 racks	2 racks (1,200mm) + 900mm aisle + 200mm clearance = 2,300mm. Works.
40ft IT	11,500 × 2,300 × 2,600mm	3 racks	3 racks (1,800mm) + 500mm aisle. Hot/cold aisle containment possible.
40ft Power	11,500 × 2,300 × 2,600mm	—	UPS, batteries, PDU, room cooling, monitoring, fire suppression.
20ft Chiller	5,500 × 2,300 × 2,600mm	—	Chiller unit, cold storage tank, pumps, water treatment.

Each ORV3-style rack: ~600mm W × 1,200mm D × 2,133mm H (48U). Loaded weight: ~700-1,000 kg per rack.

ROI Models — All GPUs × All Pricing

Revenue Per GPU Per Year (70% utilization, 24/7)

GPU	Model Served	tok/s/GPU	Annual Tokens	@ $0.20/M	@ $0.50/M	@ $1.00/M	Power Cost ($0.05/kWh)	Net @ $0.50
B200	Llama 70B	10,000	221B	$44K	$110K	$221K	$4.4K	$106K
B200	Llama 4 Maverick 400B	15,600	344B	$69K	$172K	$344K	$4.4K	$168K
B200	DeepSeek-R1 671B	3,750	83B	$17K	$41K	$83K	$4.4K	$37K
B200	Qwen 3.5 27B	11,500	254B	$51K	$127K	$254K	$4.4K	$123K
MI300X	Llama 70B	3,500	77B	$15K	$39K	$77K	$3.3K	$35K
MI300X	DeepSeek-R1 671B	1,500	33B	$7K	$17K	$33K	$3.3K	$13K
MI300X	Qwen 3.5 27B	4,000	88B	$18K	$44K	$88K	$3.3K	$41K
Ascend 910C	DeepSeek-R1 671B	800	18B	$4K	$9K	$18K	$1.8K	$7K
Ascend 910C	Qwen 3.5 72B	1,800	40B	$8K	$20K	$40K	$1.8K	$18K

Full System ROI — All Tiers

Config	GPU Cost	Saqrion Infra	Total Investment	Annual Revenue @ $0.50/M	Annual OpEx	Annual Net Profit	Payback	3-Year Return
Ascend 128 GPUs	$230K	$399K	$629K	$762K	$92K	$670K	11 mo	219%
MI300X 104 GPUs	$1,250K	$399K	$1,649K	$4,056K	$172K	$3,884K	5.1 mo	607%
B200 Starter (68)	$3,872K	$399K	$4,271K	$7,480K	$386K	$7,094K	7.2 mo	398%
B200 Standard (104)	$5,718K	$420K	$6,138K	$11,440K	$567K	$10,873K	6.8 mo	432%
B200 Full (172)	$9,590K	$440K	$10,030K	$18,920K	$885K	$18,035K	6.7 mo	440%
B200 Max (240)	$13,462K	$482K	$13,944K	$26,400K	$1,222K	$25,178K	6.6 mo	442%

Saqrion infra cost scaled with batteries: Starter/MI300X/Ascend use 2 cabs ($399K), Standard uses 3 ($420K), Full uses 4 ($440K), Max uses 6 ($482K). Revenue at $0.50/M blended tokens, 70% utilization, serving Llama 70B class models on B200, equivalent models on others. OpEx includes power (total site draw at $0.05/kWh), internet, maintenance, water treatment, monitoring.

Why MI300X has the best 3-year return (607%): GPU cost is $1.25M vs $5.7M for B200, but per-GPU revenue at $0.50/M is still $35K/yr (vs $106K for B200). The much lower upfront cost drives faster payback and higher percentage returns. Absolute dollar returns are higher on B200, but %-returns favor MI300X.

Own vs Rent — Cloud GPU Comparison

Cloud GPU Rental Rates (B200, April 2026)

Provider	Type	B200 $/hr	H100 $/hr	Annual Cost (104 GPU, 70% util)	vs Owning
Vast.ai	Marketplace/spot	$4.40	$1.49-4.00	$28.1M	10.4x more
RunPod	Hybrid	~$3.50	$1.99-2.69	$22.3M	8.3x more
DataCrunch	Managed	$3.99	~$2.50	$25.5M	9.4x more
Lambda Labs	Managed	$4.99	$2.99	$31.9M	11.8x more
CoreWeave	Enterprise	$5.50	$6.16	$35.1M	13x more
GCP	Hyperscaler	$6.69 (spot)	$3.00	$42.7M	15.8x more
AWS	Hyperscaler	$9.36	$3.90	$59.8M	22.1x more
Azure	Hyperscaler	—	$6.98	—	—
Own (container)	Owned hardware	$0.37	$0.17	$2.7M	Baseline

Own = $6.4M investment amortized over 3 years ($2.13M/yr) + $567K annual OpEx = $2.7M/yr. Effective hourly rate = $2.7M / (104 GPUs x 8,760 hrs x 0.70 util) = $0.37/GPU-hr.

Profitability Comparison — Serving Llama 70B at $0.50/M tokens

	Vast.ai	RunPod	Lambda	CoreWeave	AWS	Own
Token revenue	$11.4M	$11.4M	$11.4M	$11.4M	$11.4M	$11.4M
GPU cost	-$28.1M	-$22.3M	-$31.9M	-$35.1M	-$59.8M	-$2.7M
Net profit	-$16.7M	-$10.9M	-$20.5M	-$23.7M	-$48.4M	+$8.7M
Profitable?	No	No	No	No	No	Yes
Min token price to break even	$2.46/M	$1.96/M	$2.80/M	$3.08/M	$5.25/M	$0.24/M

You cannot profitably serve tokens at market rates ($0.30-1.25/M) on rented GPUs. Every cloud provider loses money at competitive pricing. Only owning the hardware makes inference profitable.

Breakeven Utilization — At What Point Does Owning Win?

vs Provider	Their $/hr	Breakeven Utilization	That's...
Vast.ai	$4.40	6.7%	1.6 hrs/day
RunPod	$3.50	8.5%	2.0 hrs/day
Lambda	$4.99	5.9%	1.4 hrs/day
CoreWeave	$5.50	5.4%	1.3 hrs/day
AWS	$9.36	3.2%	0.8 hrs/day

Any real inference workload runs 70%+ utilization. The breakeven is so low (1-2 hrs/day) that owning wins in every realistic scenario. Renting only makes sense for short bursts or prototyping.

3-Year Total Cost of Ownership

Provider	3-Year Cost (104 B200, 70% util)	vs Owning	Excess Paid
Own (container)	$8.1M	baseline	—
Vast.ai	$84.2M	10.4x	$76.1M wasted
RunPod	$67.0M	8.3x	$58.9M wasted
Lambda	$95.6M	11.8x	$87.5M wasted
CoreWeave	$105.4M	13x	$97.3M wasted
AWS	$179.5M	22.1x	$171.4M wasted

Power Cost Doesn't Change the Answer

Hardware depreciation ($2.1M/yr) dominates total cost. Power at 205 kW is only $54-269K/yr depending on rate — a rounding error.

Power Rate	Annual Power Cost	Effective $/GPU-hr	Annual Profit ($0.50/M tokens)	vs Vast.ai ($4.40/hr)	Profitable up to...
$0.00 (free)	$0	$0.35	$9.16M	12.6x cheaper	$5.10/kWh
$0.03 (flare gas)	$54K	$0.36	$9.10M	12.2x cheaper
$0.05 (industrial)	$90K	$0.37	$9.07M	12.0x cheaper
$0.08 (US grid avg)	$144K	$0.37	$9.01M	11.8x cheaper
$0.12 (expensive)	$216K	$0.39	$8.94M	11.2x cheaper
$0.15 (very expensive)	$269K	$0.39	$8.89M	11.2x cheaper
$0.50 (absurd)	$898K	$0.49	$8.28M	9.0x cheaper

Going from free power to $0.15/kWh only changes effective GPU cost by $0.03/hr. Cloud providers charge $3.50-9.36/hr regardless of your power cost. Owning stays profitable up to $5.10/kWh — literally 12x more expensive than the most expensive power on earth.

When Renting Still Makes Sense

Rent when:

Short-term training runs (days/weeks, not months)

Prototyping — figuring out which model to serve before committing

Burst overflow — owned GPUs at base, rent for spikes

No upfront capital available

Don't rent when:

Running 24/7 inference (our primary use case)

Need > 3-4 months of GPU time (owning is cheaper)

Serving tokens at competitive market rates ($0.30-1.25/M)

Deploying at stranded energy sites (cloud can't do this)

Pure Compute — Cost Per ExaFLOP

Forget tokens for a moment. What does raw compute cost across different approaches?

GPU Compute Specs

GPU	FP8 (PFLOPS)	FP4 (PFLOPS)	GPUs per ExaFLOP (FP8)	Buy Price/GPU	Rent/GPU/hr (cheapest)
B200	72	144	14	$55K	$3.50 (RunPod)
H100	30	—	34	$25K	$1.49 (Vast)
MI300X	~46	—	22	$12K	~$1.50
Ascend 910C	~16	—	63	$1.8K	N/A
tinybox pro (4x 6000 Pro)	~5.2	—	192	$50K/box	N/A

1 ExaFLOP = 1,000 PFLOPS. FP8 is the standard precision for LLM inference.

Cost to Build 1 ExaFLOP (FP8) — Own the Hardware

Platform	GPUs Needed	Hardware Cost	Container Infra	Total Buy	Power Draw	Annual OpEx
B200 (Supermicro racks)	14	$770K	~$400K (1 container)	$1.17M	~28 kW	$35K
H100	34	$850K	~$400K	$1.25M	~24 kW	$30K
MI300X	22	$264K	~$400K	$664K	~17 kW	$22K
Ascend 910C	63	$113K	~$400K	$513K	~25 kW	$28K
tinybox pro	192 GPUs (48 boxes)	$2.4M	$0 (desktop)	$2.4M	~100 kW	$80K

Cost to Rent 1 ExaFLOP-Hour

Provider	GPU	GPUs for 1 ExaFLOP	$/hr per GPU	Cost per ExaFLOP-hr	Annual (24/7)
Vast.ai	B200	14	$4.40	$61.60/hr	$539,616
RunPod	B200	14	$3.50	$49.00/hr	$429,240
Lambda	B200	14	$4.99	$69.86/hr	$611,974
CoreWeave	B200	14	$5.50	$77.00/hr	$674,520
AWS	B200	14	$9.36	$131.04/hr	$1,147,910
Vast.ai	H100	34	$1.49	$50.66/hr	$443,782
Own (B200)	B200	14	$0.37	$5.18/hr	$45,377
Own (MI300X)	MI300X	22	$0.12	$2.64/hr	$23,126

Owned B200: $5.18 per ExaFLOP-hour. Rented B200 (cheapest): $49/hr. That's 9.5x more expensive for the same compute. Owned MI300X is even cheaper at $2.64/ExaFLOP-hr — best value per FLOP on the market.

Scaling Up: Cost to Build Multi-ExaFLOP

Scale	B200 GPUs	Racks	Containers	Own Cost	Rent Cost/yr (RunPod)	Power
1 ExaFLOP	14	<1	2 (IT+Chiller)	$1.2M	$429K/yr	28 kW
7 ExaFLOPS (1 C1 rack)	104	1	3	$6.4M	$3.2M/yr	205 kW
17 ExaFLOPS (3 racks)	240	3	4	$16.3M	$7.4M/yr	479 kW
72 ExaFLOPS (1,000 B200s)	1,000	10	12+	$62M	$30.7M/yr	~2 MW
1,000 ExaFLOPS (1 ZettaFLOP)	13,889	134	~140	$830M	$426M/yr	~28 MW

Own cost = GPU hardware + container infrastructure. Rent cost = cheapest B200 cloud rate ($3.50/hr) at 100% util. Power at ~205W per GPU total site draw.

tinygrad ExaBox Comparison

	tinybox pro	Our MI300X Container	Our B200 Container
Price	$50K/box	$1.9M (104 GPUs)	$6.4M (104 GPUs)
Compute (FP8)	5.2 PFLOPS	4,784 PFLOPS	7,488 PFLOPS
$/PFLOP (buy)	$9,615	$397	$855
RAM bandwidth	16 TB/s	551 TB/s	832 TB/s
$/TB/s bandwidth	$3,125	$3,448	$7,692
Power + cooling included?	No (office/lab)	Yes (full DC)	Yes (full DC)
Deployable at energy site?	No	Yes	Yes
Target	Developers, $10K-$10M	Energy providers, enterprise	Max performance, enterprise

tinybox wins on $/TB/s bandwidth (the metric that matters for LLM inference speed per user). MI300X container wins on $/PFLOP (raw compute per dollar). Different products for different scales — tinybox is a desktop, our container is deployable infrastructure at energy sites.

Sensitivity Analysis

Impact of Token Pricing on Standard Config (104 B200)

Token Price ($/M)	Annual Revenue	Annual OpEx	Annual Net	Payback	Viable?
$0.10 (wholesale)	$2,288K	$567K	$1,721K	42.8 mo	Marginal
$0.20 (discount)	$4,576K	$567K	$4,009K	18.4 mo	Yes
$0.50 (competitive)	$11,440K	$567K	$10,873K	6.8 mo	Strong
$0.80 (blended realistic)	$18,304K	$567K	$17,737K	4.2 mo	Excellent
$1.25 (premium API)	$28,600K	$567K	$28,033K	2.6 mo	Excellent

Impact of Power Cost on Standard Config (104 B200, $0.50/M tokens)

Power Source	$/kWh	Annual Power Cost	% of Revenue	Annual Net	Payback
Free / surplus	$0.00	$0	0%	$10,873K	6.8 mo
Flare gas / stranded	$0.02	$36K	0.3%	$10,837K	6.8 mo
Industrial rate	$0.05	$90K	0.8%	$10,783K	6.8 mo
US grid average	$0.08	$144K	1.3%	$10,729K	6.9 mo
Expensive / peak	$0.15	$269K	2.4%	$10,604K	7.0 mo
Very expensive	$0.25	$449K	3.9%	$10,424K	7.1 mo

Power cost is almost irrelevant. Even at $0.25/kWh (very expensive), it's only 3.9% of revenue. The payback moves by less than 2 weeks. Token pricing is the only variable that matters.

Impact of Utilization on Standard Config (104 B200, $0.50/M tokens)

Utilization	Annual Revenue	Annual Net	Payback	Notes
30% (ramp-up, month 1-3)	$4,903K	$4,336K	17.0 mo	Realistic during customer acquisition
50% (established)	$8,171K	$7,604K	9.7 mo	Steady state with some idle capacity
70% (target)	$11,440K	$10,873K	6.8 mo	Industry standard target
90% (fully loaded)	$14,709K	$14,142K	5.2 mo	Near capacity — need to scale

Saqrion Infrastructure Costs

Full Component Breakdown (from manufacturing partner quotation, March 2026, ¥6.9 = $1)

Component	Model	Qty	Unit (RMB)	Total (USD)
UPS 400kVA	HPM3300E-800	2	¥156,548	$45,377
PDU cabinet	KPM-1250/380 (1250A)	1	¥65,183	$9,446
LFP battery cabinet	PU100 (563V/104Ah = 58.6 kWh)	16	¥143,465	$332,672
Room cooling	MT025UA (26.9kW)	2	¥45,051	$13,058
Outdoor condenser	KCS036HNA	2	¥16,163	$4,685
CDU 160kW	LX160WIDAW (cold plate, DI water)	3	¥49,215	$21,397
Inrow cooling	FS030FC (30.5kW, chilled water)	3	¥36,919	$16,052
Pipeline integration	Liquid + chilled water piping	1	¥115,084	$16,679
DCIM monitoring	PE-HP-50D (10" touch, 3D)	2	¥22,839	$6,620
Sensors + security	Temp, smoke, leak, cameras, access	2 sets	Various	$21,180
40ft container	ISO, insulated, anti-static floor	1	¥429,311	$62,219
20ft containers	ISO, insulated	2	¥269,085 ea	$78,030
Fire suppression	NOVEC 1230 / Heptafluoropropane	2	Various	$17,844
Classification cert	Marine shipping	2	¥33,200 ea	$9,624
TOTAL (16 batteries)				$689,962

Battery Scaling — The Biggest Cost Lever

Each PU100 cabinet: 563.2V × 104Ah = 58.6 kWh. Cost: $20,792 each. 16 cabinets = 937 kWh = $333K (48% of total infra cost).

Backup Duration	@ 140kW (Starter)	@ 205kW (Standard)	@ 340kW (Full)	@ 479kW (Max)	Battery Cost	Total Infra
15 min (graceful shutdown)	1 cab	1 cab	1 cab	2 cabs	$21-42K	$378-399K
30 min (generator startup)	2 cabs	2 cabs	3 cabs	4 cabs	$42-83K	$399-440K
1 hour	3 cabs	4 cabs	6 cabs	9 cabs	$63-187K	$420-544K
2 hours	5 cabs	7 cabs	12 cabs	16 cabs	$104-333K	$461-690K
4+ hours (off-grid)	16 cabs	16 cabs	16 cabs	16 cabs	$333K	$690K

Grid-connected sites: 2-3 cabinets ($42-63K) for 15-30 min bridge power. Saves $270-291K vs full spec.
Off-grid / flare gas: Keep 16 cabinets for multi-hour backup, or pair with diesel generator + ATS.

Tariffs & Import Duties

Container Infrastructure (no chips) — Shipping to USA

4× 40ft containers, $10M EXW value, significant steel/aluminum structure, no semiconductors.

Duty Layer	China → USA	India → USA	UAE → USA	Notes
MFN Base Duty	$300K (3%)	$250K (2-3%)	$300K (3%)	Standard rate
Section 301 (China-specific)	$2,500K (25%)	$0	$0	China only — this is the killer
Section 122 (global)	$1,000K (10%)	$1,000K (10%)	$1,000K (10%)	Expires ~July 2026
Section 232 (steel/aluminum)	$750K	$750K	$750K	25% on ~30% of value
Fees (MPF + HMF)	$47K	$47K	$47K	Fixed
Total Duties	$4,597K (46%)	$2,047K (21%)	$2,097K (21%)
Logistics	$80K	$85K	$70K
Landed Cost	$14,677K	$12,132K	$12,167K
Savings vs China	—	$2,545K	$2,510K

GPU strategy: Semiconductors for data centers are exempt from Section 232 tariffs. Install GPUs at destination — avoids duty on the highest-value component entirely.

Section 122 (10% global tariff) expires ~July 2026. If not renewed, India/UAE landed cost drops by $1M, saving $3.5M+ vs China.

Importing into India (domestic market)

Scenario	BCD Rate	IGST (18%)	Effective Duty	Notes
Finished module from China	10-15%	18% (reclaimable)	10-15% sunk + working capital	Higher duty on finished goods
Components from China → assemble in India	0-7%	18% (reclaimable)	0-7% sunk	Lower duty on components
SEZ assembly	0%	0%	0%	Special Economic Zone — full exemption

India Government Incentives

Incentive	Value	Notes
PLI for IT Hardware	4-6% of incremental sales	$30-45K per module at $750K selling price
SEZ Tax Holiday	100% income tax exemption, 5 years	Register entity in SEEPZ/MIHAN
SEZ Import Duty	0% BCD, 0% IGST on inputs	For export-oriented units
DC Infrastructure Status	0.5-1.5% lower interest rates	Access to cheaper long-term financing
State DC Policy (Maharashtra)	Electricity duty exemption, stamp duty waiver	$5-15K/year savings
National Green Hydrogen Mission	Free interstate transmission for renewables	$2.4B budget, 500 GW target by 2030

Selling to Other Markets

Destination	Import Duty	Gross Margin (Turnkey)	Certification Needed	Rating
Middle East (UAE/Saudi)	0-5%	154%	ECAS/ESMA ($10-30K)	Best
Europe (UK/Germany)	0-3.7%	124%	CE/EN 50600 ($30-80K)	Good
India (domestic)	10-15% BCD	117%	BIS ($15-30K)	Good
USA (via India assembly)	~21%	119%	UL/NFPA ($50-150K)	OK
USA (from China direct)	~46%	88%	UL/NFPA ($50-150K)	Worst
Africa (most countries)	0-10%	130-150%	Varies ($10-30K)	Good
Sri Lanka	5-15%	110-130%	Minimal	OK

Container Sales Margins (Equipment Sales Business)

Module Type	COGS (FOB China)	India Selling Price	US Selling Price	ME Selling Price	Margin Range
Basic (edge, air-cooled)	$70K	$150-210K	$200-250K	$160-210K	88-154%
Mid-range (enterprise)	$165K	$320-440K	$400-550K	$320-460K	90-133%
Full (AI/DLC, liquid-cooled)	$230K	$520-720K	$650-950K	$530-770K	119-141%
Turnkey (+commissioning)	$250K	$650-850K	$800K-1.15M	$650-950K	119-154%

Compare: Supermicro reports 9.5-11.8% gross margin on GPU servers. NVIDIA takes 70-80% of server COGS. Container infrastructure margins are 10x better because no single component dominates cost.

Target Markets

United States

Half of planned DCs canceled. $650B Big Tech spend can't find power.

Targets: Permian Basin flare gas, ERCOT curtailed wind/solar (negative pricing), Georgia Power 3 GW solar+BESS, behind-the-meter deployments.

Precedents: Crusoe Energy, Giga Energy, MARA Holdings (114 MW wind farm).

Tariff note: India assembly saves $2.5M/shipment vs China direct.

India

$30B+ DC CAPEX through 2030. 500 GW renewables target.

50+ identified customers: NTT, Reliance, HDFC, Sun Pharma, Tata, ONGC.

PLI incentives (4-6%), SEZ benefits (0% duty), National Green Hydrogen Mission ($2.4B).

CN Water Systems (Mumbai): existing facility, 300+ engineers, assembly partner.

Middle East (UAE/Saudi)

0-5% duties. 154% gross margin. Fastest growing DC market.

Dubai free zone (JAFZA/DAFZA) as tariff-free assembly and re-export hub.

Vision 2030, NEOM driving massive investment. Our manufacturing partner already has ME presence.

Africa

223 DCs across 38 countries — 0.02% of global. Greenfield opportunity.

Green Flare Nigeria: 50+ MW flare gas DCs. Teraco SA: 120 MW solar for DCs.

Huawei Ascend at $0.6M = lowest entry point. Dubai hub for serving East Africa.

2Africa + Equiano submarine cables enabling global latency for first time.

Sri Lanka

Akbar Brothers — largest private renewable energy company. Existing power assets.

70% renewable target by 2030. Solar (16 GW potential), hydro (1,800 MW), wind (50 GW).

Indian Ocean location: low latency to India, ME, SE Asia. English workforce.

Container at existing Akbar power site = zero site development cost.

Europe

0-3.7% duties. 124% margin. Strong demand, CE certification required.

Entry via Netherlands or Ireland as EU distribution hub.

Nordic countries have cheap hydro/wind power — natural fit for inference containers.

Cooling & Water Specs

Two-Loop Architecture

	Facility Loop (Loop 1)	Server Loop (Loop 2)
Path	Chiller → Cold Storage → CDU (facility side)	CDU (IT side) → GPU Cold Plates → CDU
Coolant	Standard treated water or glycol mix	Deionized water (strict!)
Temperature	20°C supply / 36.5°C return	~25°C supply / ~50°C return
Filtration	50-100 micron (standard HVAC)	25µm inline + <5µm sidestream
Resistivity	N/A	> 1.0 MΩ·cm (minimum 0.1)
pH	Standard	8.0-10.5
Additives	Standard corrosion inhibitor	Di-potassium phosphate ≥2,200 ppm + biocide
Glycol	30-50% if outdoor pipes freeze	25-35% propylene glycol only if freeze protection needed

Water Treatment Costs

Item	Initial Cost	Annual Cost
DI water system (RO + DI cartridge polisher)	$3-8K	$500-1K (cartridge replacement)
Inline filter (25µm, stainless housing)	$500-1K	$200-400 (cartridges)
Sidestream filter (<5µm, 10% flow bypass)	$1-3K	$200-400
Corrosion inhibitor (di-potassium phosphate)	$500-1K	$300-500
Biocide (glutaraldehyde 50%)	$200-500	$200-400
Monitoring (pH + conductivity meters)	$300-800	$100-200 (calibration)
Initial fill + 3x flush (DI water, 12-16 hrs)	$100-300	—
Total	$8-15K	$2-5K/year

CDU Specs

Spec	Value
Model	LX160WIDAW
Cooling capacity	160 kW per unit
Coolant	Deionized water, cold plate
Pump	Variable frequency
Form factor	Rack-mounted, 8U
Dimensions	440 × 354 × 850mm
Power input	220V/50Hz
Display	4.3" screen
Cost	~$7,133 each

Chiller

Spec	Value
Chiller container capacity	500 kW (20°C supply / 36.5°C return)
Chiller unit	Not included in container — source separately
Estimated chiller cost	$20-50K
Max config on single chiller	C1 + 2×C2 (479 kW total site, 96% chiller capacity)
Exceeds single chiller	2×C1 + C2 (535 kW — need second chiller or larger unit)

Contact

Engage with Saqrion

Signal · kryptos.31337

Scan the QR or search the username on Signal.

hello@saqrion.org →

Data sources: Manufacturing partner quotation (March 2026) · Supermicro Saqrion quote (January 2026) · OpenRouter API pricing (April 2026) · NVIDIA/AMD/Huawei benchmarks (2025-2026) · OCP Liquid Cooling Guidelines · US trade data (Section 301/122/232)

Modular AI Inference Containers