All commercial tools, reference data, and deployment planning in one place. Use the sidebar to navigate between tools.
Platform stage
Phase 1
JCU PoC active
DTA deadline
1 Jul 2026
Cloud Marketplace app
Target ARR (1,000 nodes)
$2.32M
Blended rate model
AU AI gap by 2028
700MW–1.7GW
M3 Property Nov 2025
Quick access tools
Phase 1 milestones
JCU Ideas Lab PoC contractIn progress
DTA Cloud Marketplace applicationDue 1 Jul 2026
AusTender registrationPending
NVIDIA Australia partnershipApproach Q3 2026
MIST Ergon testing — Tier 1Phase 1 months 2–4
IRAP assessment — PROTECTEDPhase 2 months 7–18
AEMO VPP registrationPhase 2 months 7–18
Offshore leak: An estimated 87% of Australian enterprise AI inference currently routes through US-owned hyperscalers (AWS Sydney, Azure) — both subject to US CLOUD Act subpoena regardless of physical location. GridMind is the only provider with deployable sovereign AI hardware on Australian soil across all tiers.
Hardware Requirements Calculator
Define your workload and we calculate the exact hardware tier, rack count, MCPU requirement, and total cost. Based on the 5-step demand calculation framework from the GridMind Commercial Whitepaper.
Step 1 — Workload type
Step 2 — Scale and usage
Recommended configuration
GridMind Starter
2 nodes
4× RTX 4090 per node · 8× GPUs total
Hardware tierStarter (RTX 4090)
Nodes required2
Total GPUs8
Total VRAM192 GB
Total power draw3.2 kW
MCPU requiredNone
Enclosure typeIP55 outdoor unit × 2
IRAP compliantStandard tier
Hardware cost
Hardware cost (est.)$40,000
MCPU / coolingIncluded
Total installed cost$40,000
Electricity/month (QLD)-$529/mo
Value prop A — GSOL idle revenue
What is GSOL idle revenue?
During hours when the customer is NOT using their AI hardware, GSOL automatically sells that spare GPU capacity on the GridMind marketplace. The customer earns money passively — like renting out a car when they are not driving it.
Who this matters for: Schools (idle nights and weekends = 82% of hours), small businesses, universities — any organisation that does not run AI workloads 24/7.
This is NOT the same as savings vs AWS. It is additional income on top of any savings they make.
How calculated: GPUs x GSOL rate x idle hours/day x 30 days x 85% efficiency, minus electricity, minus GridMind 15% platform fee.
Idle hours/day (24 minus peak hours)16 hrs
GSOL gross revenue/month-
Platform fee (15%)-
Electricity cost-
Net idle income/month-
Payback period (idle revenue)-
Value prop B — savings vs hyperscaler
What are hyperscaler savings?
If the customer currently pays AWS, Azure, or Google for GPU compute, GridMind replaces that spend. The saving is the difference between what they pay the cloud provider vs what GridMind costs to run.
Who this matters for: Enterprise, government, banks, hospitals already spending on AWS/Azure AI. They pay $8-10/GPU-hr on AWS. They could own the same silicon and run it for $0.26-0.35/GPU-hr in electricity.
This is NOT idle revenue. It is cost avoidance - money they stop losing every month.
How calculated: AWS rate x GPUs x peak hours/day x 30 days = current spend. GridMind electricity = opex. Saving = difference.
AWS equivalent rate$8.00-$10.60/GPU-hr
Current AWS monthly spend (est.)-
GridMind electricity only-
Monthly saving vs AWS-
Payback period (savings only)-
3-year ROI vs staying on AWS-
Why this configuration
Select your workload parameters to see the recommendation rationale.
Virtual Rack Builder
Visually build your rack configuration. Select a node tier and see the exact rack unit layout inside the MCPU enclosure. Drag the floor plan to see placement.
Starter · SMB
UNIT-A1 · Starter Plus
UNIT-A2 · Enterprise
Cluster · Sovereign
Starter (4× RTX 4090)
Starter Plus (8× RTX 4090)
Enterprise (H100/H200)
Enterprise
Enterprise Plus H200
Enterprise Plus B200
Sovereign NVL72
Rack unit diagram
Legend
Enclosure floor plan
Idle Capacity Revenue Calculator
During hours when your primary AI workloads are not running, your hardware is idle. GSOL automatically sells that idle capacity on the GridMind marketplace — generating passive revenue for your organisation. Calculate your idle revenue potential here.
Your hardware
Your usage pattern
Idle revenue potential
Idle hours/day
16 hrs
Idle hours/year
4,340 hrs
Monthly idle revenue
$1,240
Annual idle revenue
$14,880
Your primary AI use (33%)Operating cost
GSOL idle revenue (67%)$1,240/mo
Total GPU-hours available/year—
GMV generated (full year)—
GridMind platform fee—
Your net idle income/year—
How GSOL manages idle capacity
GSOL (GridMind Sovereign Orchestration Layer) monitors your hardware continuously. When your primary workloads finish or drop below 20% GPU utilisation, GSOL automatically offers the remaining capacity to the GridMind marketplace.
Priority rules you define: You can set minimum response time guarantees (your jobs always preempt marketplace jobs within 30 seconds), working-hours lockout (no marketplace jobs 9am–5pm on weekdays), and minimum batch size (no jobs under X GPU-hours).
Revenue appears in your GSOL dashboard and is settled monthly via PayID direct to your nominated account. Fully automated — no manual marketplace management required.
Unit Economics — Queensland
GSOL idle-revenue economics for every GridMind module. 18 idle hrs/day assumed (6 hrs peak use). Electricity at $0.35/kWh (Energex standard tariff), idle-hours share only. GridMind platform fee 15%. Rates: RTX 4090 $0.53–$1.53/GPU-hr · RTX PRO 6000 $0.75–$2.00 · H100 NVL $2.00–$5.00 · H200/B200 SXM5 $2.50–$10.00/GPU-hr.
📐 How these numbers are calculated — methodology & rate sources
💡 What is GMV?
GMV (Gross Merchandise Value) is the total revenue generated on the GSOL marketplace before any fees or costs are taken out — the raw top-line number showing what buyers pay in total for GPU compute time on your node. It is not what the hardware owner takes home. The owner receives Gross revenue (GMV minus GridMind's 15% platform fee), then Net income after electricity costs. Payback is calculated from Net income only.
The formula
Step 1 — Cost saving (operational use)
Cost saving = GPUs × rate × 10 op hrs × 30 days (no fee)
10 hrs/day operational use = owner using their own AI instead of paying AWS/Azure. No GridMind fee applies — this is the cost-saving stream, not a marketplace sale.
10 hrs/day idle on GSOL marketplace. GridMind takes 15% on marketplace sales only — not on cost-saving hours. 4 hrs/day held as buffer, not counted in either stream.
Step 3 — Electricity (20 operational hours)
Electricity = kW × 20 hrs × 30 days × $0.35
Full 20 hrs of electricity (10 op + 10 idle) counted as a cost. The 4 hrs buffer is not counted. $0.35/kWh = Energex standard business tariff (QLD, 2026).
Step 4 — Combined monthly benefit
Combined net = Cost saving + GSOL idle net − Electricity
Step 5 — Payback period
Payback = installed cost ÷ net monthly income
Installed cost includes hardware, enclosure, concrete pad, licensed electrician, and GSOL commissioning — everything needed to go live.
Rate sources — what the market actually pays
RTX 4090 — $0.53–$1.53/GPU-hr
Floor $0.53: Vast.ai live RTX 4090 listing, June 2026. Spot market — variable, this is the observed floor. $0.78–$1.03: Blended/reserved rate per SemiAnalysis GPU Rental Price Index (H1 2026); CoreWeave AU pricing. $1.28–$1.53: Sovereign premium justified by DTA Hosting Certification Framework (Jul 2026); IRAP uplift per ASD guidance.
RTX PRO 6000 Server Ed. — $0.75–$2.00/GPU-hr
Premium over RTX 4090: ECC memory, NVIDIA AI Enterprise licence, ISV certifications (Ansys, Autodesk, VMware), and passive server-grade cooling justify a 40–100% rate premium. Reference: NVIDIA RTX PRO 6000 Server Edition launch pricing (Apr 2026). APRA/IRAP uplift: Financial and government workloads requiring certified hardware attract further premium per CoreWeave enterprise contracts.
H100 NVL — $2.00–$5.00/GPU-hr
Floor $2.00: Vast.ai H100 NVL listing June 2026; Lambda Labs H100 PCIe at $2.49/hr. $3.00–$4.00: AWS p4d.24xlarge A100 equivalent at $32/hr ÷ 8 GPUs = $4.00/GPU-hr (gives market ceiling context); CoreWeave H100 reserved at $2.95/hr. $5.00 sovereign: SemiAnalysis H100 sovereign premium estimate for IRAP-rated infrastructure (40% uplift on open market).
H200 SXM5 / B200 — $2.50–$10.00/GPU-hr
H200 floor $2.50: Vast.ai H200 SXM5 listing June 2026; Lambda Labs H200 at $3.29/hr. B200 floor $3.00: Early market estimates from Bain Technology Report 2025 and Omdia GPU Pricing Index Q1 2026. B200 supply constrained — prices expected to rise. $6.00–$10.00 sovereign: AWS p5e.48xlarge H200 equivalent at ~$98/hr ÷ 8 = $12.25/GPU-hr ceiling; sovereign AU rate est. at 65–80% of AWS equivalent with data residency premium.
Key assumptions & caveats: All rates are indicative ranges based on publicly available marketplace data as of June 2026. GSOL idle utilisation is not guaranteed — the 18 hrs/day idle assumption is an upper bound; actual utilisation will vary by customer usage pattern and GSOL marketplace demand. Electricity rate is Energex standard business tariff for QLD; other states will differ. Payback periods are a financial planning tool, not a guarantee. GridMind's 15% platform fee covers marketplace matching, compliance framework, SLA monitoring, billing, and support — it is not a fixed fee and may be subject to change. For IRAP/APRA-regulated workloads, the sovereign premium rates require GridMind's compliance certification pathway (Phase 2, est. Q4 2026).
Spark — GB10 Grace Blackwell · 128 GB unified
Installed cost
$9,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$94
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$0.53/GPU-hr
Cost saving/mo+$159
Idle revenue/mo+$135
Electricity/mo−$94
Combined net/mo$200
Annual benefit$2,400
Payback3.8 yrs
📈 Medium case
$1.03/GPU-hr
Cost saving/mo+$309
Idle revenue/mo+$263
Electricity/mo−$94
Combined net/mo$478
Annual benefit$5,736
Payback1.6 yrs
🚀 Best case
$1.28/GPU-hr
Cost saving/mo+$384
Idle revenue/mo+$326
Electricity/mo−$94
Combined net/mo$616
Annual benefit$7,392
Payback1.2 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$0.53
+$159
+$159
−$24
+$135
−$94
$200
3.8 yrs
$0.78
+$234
+$234
−$35
+$199
−$94
$339
2.2 yrs
$1.03
+$309
+$309
−$46
+$263
−$94
$478
1.6 yrs
$1.28
+$384
+$384
−$58
+$326
−$94
$616
1.2 yrs
Cost saving = 1 GPU × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 1 GPU × rate × 10 idle hrs × 30 days × 85%.
Electricity = 0.45 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
Starter — 4× RTX 4090 · 96 GB VRAM
Installed cost
$20,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$441
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$0.53/GPU-hr
Cost saving/mo+$636
Idle revenue/mo+$541
Electricity/mo−$441
Combined net/mo$736
Annual benefit$8,832
Payback2.3 yrs
📈 Medium case
$1.03/GPU-hr
Cost saving/mo+$1,236
Idle revenue/mo+$1,051
Electricity/mo−$441
Combined net/mo$1,846
Annual benefit$22,152
Payback10.8 mo
🚀 Best case
$1.53/GPU-hr
Cost saving/mo+$1,836
Idle revenue/mo+$1,561
Electricity/mo−$441
Combined net/mo$2,956
Annual benefit$35,472
Payback6.8 mo
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$0.53
+$636
+$636
−$95
+$541
−$441
$736
2.3 yrs
$0.78
+$936
+$936
−$140
+$796
−$441
$1,291
1.3 yrs
$1.03
+$1,236
+$1,236
−$185
+$1,051
−$441
$1,846
10.8 mo
$1.28
+$1,536
+$1,536
−$230
+$1,306
−$441
$2,401
8.3 mo
$1.53
+$1,836
+$1,836
−$275
+$1,561
−$441
$2,956
6.8 mo
Cost saving = 4 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 4 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 2.1 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
Starter Plus — 8× RTX 4090 · 192 GB VRAM
Installed cost
$37,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$861
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$0.53/GPU-hr
Cost saving/mo+$1,272
Idle revenue/mo+$1,081
Electricity/mo−$861
Combined net/mo$1,492
Annual benefit$17,904
Payback2.1 yrs
📈 Medium case
$1.03/GPU-hr
Cost saving/mo+$2,472
Idle revenue/mo+$2,101
Electricity/mo−$861
Combined net/mo$3,712
Annual benefit$44,544
Payback10.0 mo
🚀 Best case
$1.53/GPU-hr
Cost saving/mo+$3,672
Idle revenue/mo+$3,121
Electricity/mo−$861
Combined net/mo$5,932
Annual benefit$71,184
Payback6.2 mo
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$0.53
+$1,272
+$1,272
−$191
+$1,081
−$861
$1,492
2.1 yrs
$0.78
+$1,872
+$1,872
−$281
+$1,591
−$861
$2,602
1.2 yrs
$1.03
+$2,472
+$2,472
−$371
+$2,101
−$861
$3,712
10.0 mo
$1.28
+$3,072
+$3,072
−$461
+$2,611
−$861
$4,822
7.7 mo
$1.53
+$3,672
+$3,672
−$551
+$3,121
−$861
$5,932
6.2 mo
Cost saving = 8 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 8 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 4.1 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
Pro — 4× RTX PRO 6000 Server Ed. · 192 GB ECC
Installed cost
$78,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$504
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$0.75/GPU-hr
Cost saving/mo+$900
Idle revenue/mo+$765
Electricity/mo−$504
Combined net/mo$1,161
Annual benefit$13,932
Payback5.6 yrs
📈 Medium case
$1.50/GPU-hr
Cost saving/mo+$1,800
Idle revenue/mo+$1,530
Electricity/mo−$504
Combined net/mo$2,826
Annual benefit$33,912
Payback2.3 yrs
🚀 Best case
$2.00/GPU-hr
Cost saving/mo+$2,400
Idle revenue/mo+$2,040
Electricity/mo−$504
Combined net/mo$3,936
Annual benefit$47,232
Payback1.7 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$0.75
+$900
+$900
−$135
+$765
−$504
$1,161
5.6 yrs
$1.00
+$1,200
+$1,200
−$180
+$1,020
−$504
$1,716
3.8 yrs
$1.50
+$1,800
+$1,800
−$270
+$1,530
−$504
$2,826
2.3 yrs
$2.00
+$2,400
+$2,400
−$360
+$2,040
−$504
$3,936
1.7 yrs
Cost saving = 4 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 4 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 2.4 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
Pro Plus — 8× RTX PRO 6000 Server Ed. · 384 GB ECC
Installed cost
$129,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$1,008
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$0.75/GPU-hr
Cost saving/mo+$1,800
Idle revenue/mo+$1,530
Electricity/mo−$1,008
Combined net/mo$2,322
Annual benefit$27,864
Payback4.6 yrs
📈 Medium case
$1.50/GPU-hr
Cost saving/mo+$3,600
Idle revenue/mo+$3,060
Electricity/mo−$1,008
Combined net/mo$5,652
Annual benefit$67,824
Payback1.9 yrs
🚀 Best case
$2.00/GPU-hr
Cost saving/mo+$4,800
Idle revenue/mo+$4,080
Electricity/mo−$1,008
Combined net/mo$7,872
Annual benefit$94,464
Payback1.4 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$0.75
+$1,800
+$1,800
−$270
+$1,530
−$1,008
$2,322
4.6 yrs
$1.00
+$2,400
+$2,400
−$360
+$2,040
−$1,008
$3,432
3.1 yrs
$1.50
+$3,600
+$3,600
−$540
+$3,060
−$1,008
$5,652
1.9 yrs
$2.00
+$4,800
+$4,800
−$720
+$4,080
−$1,008
$7,872
1.4 yrs
Cost saving = 8 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 8 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 4.8 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
Enterprise — 8× H100 NVL · 640 GB HBM2e
Installed cost
$543,000
Op. hours/day
10 hrs
Idle hrs/day
10 hrs
Electricity/mo (20 hrs)
$966
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$2.00/GPU-hr
Cost saving/mo+$4,800
Idle revenue/mo+$4,080
Electricity/mo−$966
Combined net/mo$7,914
Annual benefit$94,968
Payback5.7 yrs
📈 Medium case
$3.00/GPU-hr
Cost saving/mo+$7,200
Idle revenue/mo+$6,120
Electricity/mo−$966
Combined net/mo$12,354
Annual benefit$148,248
Payback3.7 yrs
🚀 Best case
$5.00/GPU-hr
Cost saving/mo+$12,000
Idle revenue/mo+$10,200
Electricity/mo−$966
Combined net/mo$21,234
Annual benefit$254,808
Payback2.1 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$2.00
+$4,800
+$4,800
−$720
+$4,080
−$966
$7,914
5.7 yrs
$3.00
+$7,200
+$7,200
−$1,080
+$6,120
−$966
$12,354
3.7 yrs
$4.00
+$9,600
+$9,600
−$1,440
+$8,160
−$966
$16,794
2.7 yrs
$5.00
+$12,000
+$12,000
−$1,800
+$10,200
−$966
$21,234
2.1 yrs
Cost saving = 8 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 8 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 4.6 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$2.50/GPU-hr
Cost saving/mo+$6,000
Idle revenue/mo+$5,100
Electricity/mo−$1,764
Combined net/mo$9,336
Annual benefit$112,032
Payback8.0 yrs
📈 Medium case
$4.00/GPU-hr
Cost saving/mo+$9,600
Idle revenue/mo+$8,160
Electricity/mo−$1,764
Combined net/mo$15,996
Annual benefit$191,952
Payback4.7 yrs
🚀 Best case
$8.00/GPU-hr
Cost saving/mo+$19,200
Idle revenue/mo+$16,320
Electricity/mo−$1,764
Combined net/mo$33,756
Annual benefit$405,072
Payback2.2 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$2.50
+$6,000
+$6,000
−$900
+$5,100
−$1,764
$9,336
8.0 yrs
$4.00
+$9,600
+$9,600
−$1,440
+$8,160
−$1,764
$15,996
4.7 yrs
$6.00
+$14,400
+$14,400
−$2,160
+$12,240
−$1,764
$24,876
3.0 yrs
$8.00
+$19,200
+$19,200
−$2,880
+$16,320
−$1,764
$33,756
2.2 yrs
Cost saving = 8 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 8 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 8.4 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
💰 Stream 1 — Cost saving (10 hrs/day)
Owner uses their own GPU instead of paying AWS/Azure. Every operational hour saves the market rate — no GridMind fee applies.
📡 Stream 2 — GSOL idle revenue (10 hrs/day)
Unused GPU capacity sold on the GSOL marketplace. GridMind takes 15% platform fee. 4 hrs/day held as buffer.
Scenario summary — Base / Medium / Best
📊 Base case
$3.00/GPU-hr
Cost saving/mo+$7,200
Idle revenue/mo+$6,120
Electricity/mo−$2,037
Combined net/mo$11,283
Annual benefit$135,396
Payback10.0 yrs
📈 Medium case
$5.00/GPU-hr
Cost saving/mo+$12,000
Idle revenue/mo+$10,200
Electricity/mo−$2,037
Combined net/mo$20,163
Annual benefit$241,956
Payback5.6 yrs
🚀 Best case
$10.00/GPU-hr
Cost saving/mo+$24,000
Idle revenue/mo+$20,400
Electricity/mo−$2,037
Combined net/mo$42,363
Annual benefit$508,356
Payback2.7 yrs
Full rate breakdown — all market tiers
Rate/GPU-hr
Cost saving/mo
GSOL gross/mo
GridMind 15%
GSOL net/mo
Electricity
Combined net/mo
Payback
$3.00
+$7,200
+$7,200
−$1,080
+$6,120
−$2,037
$11,283
10.0 yrs
$5.00
+$12,000
+$12,000
−$1,800
+$10,200
−$2,037
$20,163
5.6 yrs
$7.00
+$16,800
+$16,800
−$2,520
+$14,280
−$2,037
$29,043
3.9 yrs
$10.00
+$24,000
+$24,000
−$3,600
+$20,400
−$2,037
$42,363
2.7 yrs
Cost saving = 8 GPUs × rate × 10 op hrs × 30 days (no fee — own use).
GSOL gross = 8 GPUs × rate × 10 idle hrs × 30 days × 85%.
Electricity = 9.7 kW × 20 hrs × 30 days × $0.35.
4 hrs/day buffer not counted.
MCPU — Modular Cooling & Power Unit
The GridMind MCPU is the external cooling and power distribution module for Enterprise-tier nodes. It handles liquid cooling (direct liquid cooling, dry coolers, or full fluid pods) and serves as the primary interface between the compute pod and site utilities.
MCPU-S — Small
Up to 15 kW cooling · Outdoor dry cooler wall-mount · Suits Starter Plus, Pro, Pro Plus, Enterprise H100 NVL
~$8,000–$14,000 installed
MCPU-M — Medium
Up to 60 kW cooling · Dry cooler array 2–4 panels · Suits Enterprise H100 SXM5, H200, B200
~$35,000–$65,000 installed
MCPU-L — Large
Up to 140 kW cooling · Full outdoor fluid pod · Suits Sovereign NVL72, multi-pod clusters
~$120,000–$200,000 installed
What the MCPU does: GPU compute generates intense heat — H100/H200/B200 servers run at 5–10 kW per server, far beyond what standard air conditioning can handle. The MCPU is a dedicated outdoor liquid cooling unit that absorbs this heat via a coolant loop, transfers it to an outdoor dry cooler or fluid circuit, and rejects it to the atmosphere. It also provides clean, conditioned power distribution and can include UPS bypass and generator interfaces for enterprise deployments.
MCPU-S — Small · up to 15 kW
Specification
Cooling capacityUp to 15 kW
Cooling typeDry cooler · wall-mount
Footprint600×400mm wall bracket
Power required32A 3-phase dedicated circuit
RefrigerantGlycol/water loop · no refrigerant
Noise~58 dBA at 1m
IP ratingIP55 · C4 cyclone rated
Suited to
Starter Plus (8× RTX 4090, 4.1 kW) · Pro Plus (8× RTX PRO 6000, 4.8 kW) · Enterprise H100 NVL (4.6 kW) · Campus M configurations
Installed cost estimate
$8,000–$14,000
Includes unit, coolant pipes, pump station, commissioning
Includes unit, coolant loop, CDU, manifold, commissioning
MCPU-L — Large · up to 140 kW
Specification
Cooling capacityUp to 140 kW
Cooling typeFull outdoor fluid pod
Footprint3.0×2.0m ground pad + pump skid
Power required125A 3-phase dedicated circuit
IP ratingIP55 · C4 cyclone rated
Suited to
Sovereign NVL72 (120 kW) · Multi-pod UNIT-A2 clusters · Full campus deployments
Installed cost estimate
$120,000–$200,000
Includes fluid pod, pump skid, coolant loop, UNIT-A2 manifold, commissioning
Enclosure Guide — Australian Conditions
Physical housing for every GridMind module tier. Three enclosure types covering SMB outdoor units, single-storey UNIT-A1 Kingspan modular pods, and dual-storey UNIT-A2 enterprise pods. All rated to cyclone C4, designed for Queensland tropical conditions.
SMB Outdoor Enclosure — IP55-rated anodised aluminium, CNC-engraved. Installs beside any building like a split-system AC unit. Half a day, no DA required, no MCPU. Covers Spark (desktop), Starter, and Starter Plus.
Form factor
Outdoor unit
IP rating
IP55
Cyclone rated
C4
Install time
4–6 hours
DA required
No
Rack configuration diagrams — Starter and Starter Plus
GridMind Starter — 4× RTX 4090
GridMind Starter Plus — 8× RTX 4090
Physical specs (outdoor enclosure): 680mm × 860mm × 600mm (Starter) / 950mm × 860mm × 600mm (Starter Plus) · IP55 dust + water jet · C4 cyclone rated · 20A single-phase (Starter) / 32A 3-phase (Starter Plus) · no DA required — same regulatory category as outdoor AC unit · installs in 4–6 hours.
Pro Outdoor Enclosure — Same IP55 outdoor enclosure as Starter, housing RTX PRO 6000 Server Ed. cards. ECC memory, NVIDIA AI Enterprise certified, passive server-grade cooling. Upgrade from Pro to Pro Plus by adding 4 cards — same enclosure, no new pad.
Form factor
Outdoor unit
IP rating
IP55
ECC memory
Yes
AI Enterprise
Certified
Upgrade path
Pro → Pro Plus
Rack configuration diagrams
GridMind Pro — 4× RTX PRO 6000 Server Ed.
GridMind Pro Plus — 8× RTX PRO 6000 Server Ed.
RTX PRO 6000 Server Ed. vs RTX 4090: 96 GB ECC GDDR7 (vs 24 GB GDDR6X no ECC) · passive server-grade cooling · dual-slot 267mm · NVIDIA AI Enterprise certified · ISV certifications (Ansys, Autodesk, VMware) · 24/7 validated. Suits APRA CPS 234, IRAP SENSITIVE, and clinical AI workloads.
UNIT-A1 — Single-storey modular pod — Kingspan KS1000 PIR 100mm insulated panel construction. 6×8m external footprint, 48m² internal floor area, 3,600mm clear height. NCC Class 8. DA required 6–14 weeks. Hot/cold aisle separation. Raised access floor 400mm void.
Raised floor void400–600mm for cold air distribution
H100 NVL vs H100 SXM5: NVL is PCIe Gen5 form factor — fits standard server. SXM5 requires a dedicated SXM5 host baseboard and commands higher per-GPU throughput for LLM training workloads.
UNIT-A2 — Dual-storey enterprise pod — Upper floor: compute rack room with raised access floor. Lower floor: power distribution, UPS, switchboard, GSOL control node. Rooftop solar + HVAC array. MCPU-M coolant manifold on rear wall. Side-mounted access stair. ~96–120m² total.
Total area
~96–120m²
Upper height
3.4m clear
Lower height
3.2m clear
Wall insulation
Kingspan 150mm PIR
Floor loading
20 kN/m² upper
Cyclone
C4 · engineer certified
Rack configuration diagrams — Enterprise Plus tiers
Enterprise Plus H200 — 8× H200 SXM5
Enterprise Plus B200 — 8× B200 Blackwell
UNIT-A2 upper floor — Compute rack room
Raised access floor 400mm void for cold air distribution · under-floor cable trays · hot-aisle containment · 6–8 racks · InfiniBand NDR400/800 fabric · MCPU-M coolant loop manifold connection on rear wall.
UNIT-A2 lower floor — Power and operations
250–400A 3-phase main switchboard · N+1 UPS rated to full rack load · generator input terminals · GSOL control node · operations console · battery room · cable tray (power separated from data 300mm).
Compliance Framework
Every GridMind commercial deployment operates within Australian compliance frameworks. Understanding the compliance tier your customer needs determines both the hardware selection and the rate premium they can command through GSOL.
Standard Commercial
GSOL rate tierOpen market
Hardware requiredAny tier
Data residencyAU-based recommended
Who uses thisSMBs, startups, universities
Rate premiumNone — market rate
IRAP SENSITIVE
GSOL rate tierReserved premium
Hardware requiredRTX PRO 6000 min.
Data residencyAU sovereign required
Who uses thisAPS agencies, banks, health
Rate premium~40–60% over open market
IRAP PROTECTED
GSOL rate tierSovereign
Hardware requiredH100+ · Campus M min.
Data residencyAir-gapped or sovereign pod
Who uses thisDefence, ASD, ASIS, classified
Rate premium~80–150% over open market
APRA CPS 234: Financial institutions (banks, insurers, super funds) must comply with APRA CPS 234 for information security. GridMind nodes with RTX PRO 6000 Server Ed. (NVIDIA AI Enterprise certified) satisfy the hardware requirements for financial workloads. GSOL provides audit logs, access controls, and data residency certification letters on request.
AI Demand Stack — What the Forecasts Miss
All current AU demand forecasts (AEMO, M3 Property, Oxford Economics) are based on Wave 1 workloads only. The unmodelled next wave — physical robotics, autonomous vehicles, personal AI agents, defence, and biotech — will dwarf current projections from 2028 onwards.
Australian AI compute demand stack 2022–2035, showing current forecasted workloads and the unmodelled next-wave demand from robotics, autonomous vehicles, personal AI agents, defence, and biotech — illustrating how all current forecasts systematically undercount future demand.
What current forecasts capture vs what they miss
Modelled by AEMO / M3 / Oxford
Enterprise LLM inference · Cloud AI APIs · Government digital services · Hospital / clinical AI · University research compute · Financial AI (fraud, credit)
Not in any current forecast
Physical robotics inference (humanoid, industrial) · Autonomous vehicles (V2X, real-time perception) · Personal AI agents (always-on, device + cloud hybrid) · Defence / sovereign AI (classified workloads) · Autonomous biotech labs · AGI-class frontier training
Sources: AEMO 2025 ISP, M3 Property Nov 2025, Oxford Economics Jul 2025, Deloitte Insights Nov 2025, McKinsey Technology Report 2025, RAND AI Power Requirements 2025, Bain Technology Report 2025. Next-wave layers are author projections based on known deployment trajectories — not yet in any published AU forecast model.
Power & Infrastructure — Founder Reference
Everything a founder needs to know about power, connectivity, and physical infrastructure requirements for GridMind deployments. Use this panel to answer customer site assessment questions on the spot.
Module
Total draw
Circuit required
Phase
Meter upgrade?
Spark
0.45 kW
10A single-phase
Single
No — standard outlet
Starter
2.1 kW
20A single-phase dedicated
Single
Usually no
Starter Plus
4.1 kW
32A 3-phase
3-phase
Usually no
Pro
2.4 kW
20A single-phase dedicated
Single
Usually no
Pro Plus
4.8 kW
32A 3-phase
3-phase
Check with Energex
Enterprise (H100 NVL)
4.6 kW
63A 3-phase
3-phase
Likely — contact Energex
Enterprise Plus H200
8.4 kW
125A 3-phase
3-phase
Yes — dedicated supply
Enterprise Plus B200
9.7 kW
125A 3-phase
3-phase
Yes — dedicated supply
QLD electricity tariff: Energex standard business tariff is $0.35/kWh (2026). Large deployments (>10 kW) may qualify for ToU (time-of-use) tariffs which can reduce costs during off-peak GSOL hours. GridMind recommends customers speak to their energy retailer about ToU pricing before installation.
Minimum requirements
Connection typeNBN or fibre preferred
Upload speed100 Mbps minimum
Download speed100 Mbps minimum
Latency<50ms to AU IX
Static IPRequired for GSOL
Port forwardingGSOL agent handles automatically
Recommended
Enterprise nodes1 Gbps fibre dedicated
GSOL backup4G/5G SIM failover included
ManagementOut-of-band via GSOL SBC
On-site switch10GbE for rack nodes
SMB site checklist
☐ Flat concrete area ≥1.2×0.8m for unit
☐ 600mm clearance in front of fan face
☐ Existing 20A circuit nearby (or licensed electrician to install)
☐ Cat6A cable run to nearest switch (<100m)
☐ NBN or fibre connection with static IP
☐ No DA required (same as outdoor AC unit)
Enterprise site checklist
☐ DA lodged (6–14 week lead time)
☐ RC slab 200mm poured and cured
☐ 3-phase power connection confirmed with Energex
☐ MCPU-S/M/L pad and clearance allocated
☐ 1 Gbps fibre to site confirmed
☐ RPEQ engineer engaged for cyclone certification
☐ QBCC licensed builder engaged
Infrastructure item
Who arranges
Typical cost
Notes
Concrete pad (SMB)
Customer
$300–$800
100mm, 1.2×0.8m. Standard concretor.
Concrete slab (UNIT-A1)
GridMind / customer
$8,000–$18,000
200mm RC slab, N12 mesh, engineer-designed.
Electrical circuit (SMB)
Customer
$400–$1,200
Licensed electrician. 20A dedicated run.
3-phase power connection
Customer + Energex
$2,000–$12,000
Energex connection fee + internal wiring.
NBN static IP upgrade
Customer
$20–$50/mo extra
Most ISPs offer static IP as add-on.
GSOL commissioning
GridMind
Included
Remote — typically 2 hours after power-on.
DA (Development Application)
Customer + GridMind support
$3,000–$8,000
Council fees + private certifier. Enterprise only.
GridMind Module Designer
AI-assisted architecture for both unit types. Configure your requirements, generate a full design brief, materials schedule, compliance checklist, production-line assembly sequence, and dimensional drawings — ready to hand to a builder, engineer, or manufacturer.
Step 1 of 3 — configure your module requirements
Tell the AI designer what you need. It will generate a complete design brief.
Select your unit type, hardware tier, site location, and special requirements. The AI will generate a full specification including structure, insulation, cooling integration, electrical, floor loading, and NCC compliance notes.
Unit type and hardware
Module type
Primary hardware tier
Number of racks / servers in this module 2
2
MCPU cooling tier required
Site and environment
Site location (wind/cyclone region)
Ambient temperature range
Site classification (NCC)
Special requirements
Live configuration preview
Unit type
UNIT-A1
Internal floor area
48 m²
Total IT load
15.4 kW
Structural type
SHS steel frame
Kingspan KS1000 75mm PIR
Wall R-value
R4.0
Floor loading spec
12 kN/m²
Cyclone rating
C4
NCC class
10a
No DA typically required
Configure your requirements above to see the design brief. The AI will generate a full specification including materials schedule, compliance notes, and production sequence.
Generating design brief...
AI-generated design brief
Materials specification — Australian suppliers · NCC compliant
Standard materials schedule for GridMind modular units
Every material is specified for Australian conditions: cyclone regions C2–C4, tropical/subtropical humidity, UV exposure, coastal salt air, and seismic zones. All products are available from Australian suppliers with standard lead times.
Span tables AS/NZS 4600. Max 1.2m centres in cyclone C4.
Bracing
Flat strap 75×6mm G350 + turnbuckles
AS 4100
OneSteel / InfraBuild
Diagonal wall bracing for racking resistance. Engineer-designed in C3/C4.
Connections
Grade 8.8 M16/M20 bolts + structural cleats
AS 4100
Hobson / TFC
All connections engineer-certified. No site welding required — bolt-together for production line.
Base frame
RHS 200×100×8 G350 perimeter + levelling feet
AS/NZS 1163
OneSteel
Factory-welded base frame. Hot-dip galv after fabrication. Anchor to 200mm slab.
Wall panels — Kingspan KS1000 AWP insulated panel system
Why Kingspan KS1000 AWP: PIR foam core (polyisocyanurate — thermosetting, forms fire-resistant char, does not melt or drip). Highest R-value per mm of any commercial insulation panel. Australian-manufactured (Kingspan AU, CodeMark certified). Factory tongue-and-groove joint — zero site-cut thermal bridging. Available in Colorbond colour range. 30-year warranty. Used in pharmaceutical cold stores, food processing, and data centres globally.
Application
Panel type
Thickness
R-value
Wind rating
Fire (FRR)
Notes
External walls — Tier 1 (SMB, Starter Plus)
Kingspan KS1000 AWP PIR
75mm
R4.0
C3 standard · C4 with eng. fix
–/–/– (non-rated) or 30/30/30 with lining
Adequate for ≤10 kW nodes. Class 10a sufficient.
External walls — Tier 2 (H100, H200)
Kingspan KS1000 AWP PIR
100mm
R5.3
C4 with eng. fix
60/60/60 with GypRoc Fyrchek lining
Required for Class 8. Provides thermal mass for 40–50 kW load.
External walls — Tier 3 (B200, Sovereign)
Kingspan KS1000 AWP PIR
150mm
R8.0
C4 engineer-certified
90/90/90 with 2× GypRoc lining
Maximum insulation for high-heat nodes. Class 8 mandatory.
Roof panel — all tiers
Kingspan KS1000 RW PIR
100mm
R5.3 (roof)
AS/NZS 1170.2 C4
60/60/60
Low-slope (≥3°) or flat with falls. Concealed fix Kliplok profile. Solar PV-mount-compatible.
Internal partition — hot/cold aisle
Kingspan KS1000 inner
50mm
R2.5
Internal only
30/30/30
Hot-aisle containment wall. Non-structural — bolts to server rails or ceiling track.
Panel dimensions: 1,000mm wide × up to 12,000mm length (factory-cut to requirement). Steel skins: 0.5mm Colorbond Ultra (coastal XRW grade for salt air). Panel weight: ~12–18 kg/m² depending on thickness. Colour: Woodland Grey external / White internal (standard).
Floor system
Tier
Floor spec
Load rating
Notes
SMB (Starter / Spark / Pro)
IP55 outdoor enclosure — no floor slab required
N/A
Unit mounts to concrete pad via M12 anchor bolts. No raised floor.
Roof-mount on factory-installed Colorbond standing seam rails. 6–12 panels standard (2.5–5 kW offset). Isolator per AS/NZS 5033.
Solar inverter
SolarEdge / Fronius 3-phase 5–10 kW
Grid-tied, export-limited. AS/NZS 4777 compliant. Mounted internally in electrical room.
Rooftop HVAC unit
Daikin / Mitsubishi commercial packaged unit, 10–25 kW
Building envelope heating only — not compute cooling. Maintains ambient ≤28°C in electrical room and operations area.
MCPU dry cooler (rooftop option)
Vertiv CoolChip or nVent CX121 roof-mount variant
UNIT-A1 only. Connects to CDU manifold via insulated copper pipes through roof penetration. Engineer-designed penetration with flashing.
Australian building compliance — NCC 2025 · QBCC · AS/NZS standards
Compliance requirements by unit type and NCC class
Building classification determines what approvals you need before you can build. The difference between Class 10a and Class 8 is significant — it determines whether you need a Development Application (DA), a structural engineer, fire engineering, and access compliance. Choose the right class at the start.
NCC building class decision guide
Class 10a Fastest path
Non-habitable structure. Sheds, garages, utility buildings. No one works inside for extended periods. No sanitary facilities required.
Development Application (DA)Often exempt (state rules vary)
Building certifierRequired but low-bar
Structural engineerRequired for C3/C4 cyclone
Fire engineeringNot required typically
Access (DDA)Not required
Electrical certLicensed electrician cert required
GridMind unitsStarter · Spark · Pro · Starter Plus
Typical approval time2–6 weeks
Class 8 Standard commercial
Laboratories, workshops, and buildings where hazardous processes occur. Commercial data centre infrastructure where people regularly work inside classifies here.
Development Application (DA)Required — Council lodgement
Building certifierPrivate or council certifier
Structural engineerMandatory — certified drawings
Fire engineeringPerformance solution likely needed
Access (DDA)Required — path of travel to entry
Electrical certQBCC licensed + test & tag
GridMind unitsUNIT-A1 · UNIT-A2 · all enterprise
Typical approval time6–14 weeks
Class 5 / 6 If customer-facing
Office buildings or retail. Only applies if GridMind operates a shared-access facility where members of the public or third-party tenants access the space directly and regularly.
Development ApplicationMandatory — full DA + EIS
Fire engineeringFull FER + sprinklers likely
Access (DDA)Full DDA + accessible toilet
Energy efficiencyNCC Section J compliance
When this appliesCo-lo / shared access facility only
Typical approval time4–12 months
Queensland-specific requirements (QBCC)
Requirement
Authority
When required
Notes
Building permit
QBCC / Private certifier
All Class 8 and most Class 10a >10m²
File Form 5 (building permit application) with local council or private certifier. Include structural drawings.
Electrical safety certificate
Electrical Safety Office (ESO)
All new electrical work >50V AC
Licensed electrician issues ESO Form 1. Required before energisation. RPEQ for installations >100A.
Plumbing permit
QBCC Plumbing
Only if sanitary facilities installed
Not required for compute-only UNIT-A1. Required if UNIT-A2 includes amenities or floor drain.
Cyclone tie-down certification
Structural engineer (RPEQ)
All C2–C4 wind regions
Engineer must certify connection of panel to frame, frame to slab, slab to ground. Mandatory in North QLD.
QBCC builder licence
QBCC
All construction >$3,300 value
Builder must hold QBCC licence. Subcontractors (electrician, plumber) must hold own licences.
AS/NZS 3000 electrical inspection
Licensed electrical inspector
All Class 8 switchboard installations
Third-party inspection of switchboard and distribution. Required for insurance.
Key Australian standards referenced
Standard
Covers
AS/NZS 1163
Cold-formed steel hollow sections (SHS, RHS)
AS 4100
Steel structures — design and construction
AS/NZS 4600
Cold-formed steel structures (purlins, girts)
AS/NZS 1170.2
Wind actions — cyclone loading design
AS 1170.1
Structural loads — dead, live, snow loads
AS 3600
Concrete structures — slab design
Standard
Covers
AS/NZS 3000
Electrical wiring rules (the "Wiring Rules")
AS/NZS 4509
Stand-alone power systems (UPS)
AS/NZS 5033
Solar PV system installation
AS/NZS 1768
Lightning protection
AS/NZS 1530
Fire resistance testing for building materials
AS 1851
Fire protection system maintenance
Reference drawings — not for construction use without engineering sign-off
Dimensional drawings — UNIT-A1 and UNIT-A2
These are schematic reference drawings showing key dimensions, structural grid, panel layout, and service zone allocations. They are not stamped construction drawings. For construction, a licensed RPEQ structural engineer must certify site-specific versions of these drawings.
Select drawing to view
UNIT-A1 — Single-storey modular compute pod — Floor plan (schematic)Not for construction · RPEQ certification required
Key dimensions reference
Factory production — off-site manufacturing for consistent quality
GridMind module production line — from factory to site in 5 steps
Designed for mass production. Every module is built to the same engineering drawings with the same certified materials. Off-site manufacturing means weather delays don't stop production, quality is consistent, and the site installation time is minimal — 1 to 5 days depending on unit type.
Production sequence — UNIT-A1 (single storey)
1
Factory — base frame fabrication (Day 1–2)
RHS 200×100×8 perimeter frame factory-welded and jig-drilled on flat steel table. All connection holes CNC-punched to drawings. Hot-dip galvanise after fabrication. Weld inspected to AS/NZS 1554. Corner post sockets welded and certified. Levelling feet or forklift pockets welded. This is the only welding in the entire build — everything else is bolted.
2
Factory — wall panel and roof prefabrication (Day 2–4)
Kingspan KS1000 AWP panels cut to length on factory saw — zero site cutting. All openings (door frames, MCPU penetrations, cable entry points, louvre openings) pre-cut and frame-fitted in factory. Panels labelled and stacked per installation sequence drawing. Roof panels pre-drilled for solar mounting rails. HVAC curb pre-welded to roof panel at factory.
3
Factory — electrical and mechanical pre-build (Day 3–5)
Switchboard built, wired, tested, and certified off-site. Cable tray sections pre-cut. PDU units pre-wired and tested. UPS unit checked and firmware-loaded. Rooftop HVAC unit assembled and pressure-tested. MCPU manifold (if applicable) pre-piped with pressure test cert. All components packed in numbered crates matching installation sequence.
4
Site — concrete slab and anchor installation (Day 1–3 on site, concurrent with factory steps 1–3)
200mm reinforced concrete slab poured by local concrete contractor. Cast-in M20 anchor bolts per engineer's layout drawing. Base frame delivered and set on slab — level checked. Base frame bolted down. All anchor bolt positions match factory-jig-drilled base frame holes exactly — no site drilling. Slab cure 7 days before structural frame loaded.
5
Site — frame erection and panel installation (Day 4–6 on site)
Corner posts bolted to base frame — 2 person crew + telehandler. Roof purlins bolted. Wall panels installed starting from corner — tongue-and-groove joints engaged, factory screws at 300mm centres. Door frames (pre-hung, factory-fitted) installed. Roof panels installed. All penetrations sealed with factory-supplied PIR foam backer rod and Sika Flex 11FC sealant. Cyclone bracing straps installed and torqued to specification.
6
Site — electrical and mechanical connection (Day 6–8)
Switchboard lifted into electrical room — pre-wired, bolt-to-wall. Cable tray installed overhead. Server racks positioned and bolted to floor anti-seismic rails. Power feeds connected from switchboard to rack PDUs. MCPU outdoor module positioned on pad and coolant loop connected (if applicable). Solar PV panels mounted on pre-installed roof rails. All systems energised and tested. ESO electrical certificate issued.
7
Site — commissioning and GSOL registration (Day 8–10)
GSOL agent software installed on management node. Hardware discovered and registered. First inference job dispatched as commissioning test. MIST Ergon type test certificate (from Phase 1 testing) referenced for compliance record. Building certifier final inspection. QBCC compliance certificate issued. GridMind dashboard live — node registered, earnings commence within 30 days.
Key manufacturing principle — zero site welding: Every structural connection in the GridMind module is bolted, not welded. This is critical for production-line manufacturing: bolt connections require no certified welders on site, no weld inspections, no NDT (non-destructive testing) in the field. The only welding is the factory base frame fabrication and any custom brackets — all done in a controlled environment with certified welders and jigs, inspected before shipment. This reduces site risk, speeds erection, and enables any licensed builder to assemble the module without specialist structural steel expertise.
Hardware Expertise — Computer & AI Engineer Reference
Everything a computer engineer and AI software engineer would know about matching hardware to workloads — written for founders without that background. Use this to understand what hardware is needed, why, and how to explain it confidently to any technical customer or procurement team.
Computer engineer perspective — the most important question
What type of AI work is the customer actually doing?
Every hardware decision flows from this one question. The same GPU cluster that runs inference on a live chatbot is the wrong choice for training a new model. Get this wrong and either the hardware is too small to do the job, or the customer is massively overpaying for capacity they'll never use.
The four fundamental AI workload types
1. Inference — "serving" a trained model
The model is already trained. A user sends a question, the GPU processes it and returns an answer. This is what 95% of GridMind customers need. Low latency matters. High concurrency matters. Exact precision doesn't — you can use FP8 or INT4 quantisation to fit larger models in less VRAM.
Customer examples: Hospital chatbot answering doctor queries · Bank fraud alert system · Government document summariser · School AI tutor
Optimal hardware
RTX 4090 / H100 NVL / H200
VRAM per request
Low — model fits in VRAM once
Precision
FP16 / FP8 / INT4 — all fine
GridMind fit
Excellent — all node tiers
2. Training — teaching a model from scratch
The model does not yet exist. You feed it millions of examples over days or weeks and the GPU computes the weight adjustments. Needs massive VRAM. Needs high-bandwidth interconnect between GPUs (NVLink) so all GPUs share the model. Very few GridMind customers will do this — only large universities and research organisations.
Customer examples: University AI lab training a climate model · Research institute training on Indigenous language data · Large hospital training a radiology model on Australian patient data
Optimal hardware
H100 SXM5 / H200 SXM5 / B200
VRAM required
Very high — 640 GB – 1.5 TB
Interconnect
NVLink mandatory (not PCIe)
GridMind fit
Enterprise+ only
3. Fine-tuning — adapting an existing model
Take a large pre-trained model (e.g. Llama 3.1) and continue training it on your specific data so it learns your domain. Requires less compute than full training but still needs significant VRAM. Most common use case: a bank wants a model that understands Australian financial regulations, or a hospital wants a model that knows their specific clinical protocols.
Customer examples: Government agency fine-tuning on policy documents · Legal firm adapting Llama for Australian law · Hospital adapting a base model for their drug formulary
Optimal hardware
H100 NVL / H200 / RTX PRO 6000
VRAM required
Medium — 96 GB – 640 GB
Time
Hours to days (not weeks)
GridMind fit
Good — Pro, Enterprise, H100
4. Embedding / RAG — search and retrieval AI
Convert documents into numeric vectors (embeddings) so an AI can search through thousands of documents and find relevant context before answering. Very low VRAM per document. Can run on smaller GPUs. Often combined with inference in a RAG (Retrieval-Augmented Generation) pipeline — search for relevant docs, then feed them to the LLM to generate an answer.
Customer examples: Law firm searching 50,000 case documents · Government querying all APS policy documents · Hospital searching clinical guidelines
Optimal hardware
RTX 4090 / RTX PRO 6000 / GB10
VRAM per model
Low — 2–8 GB
Bottleneck
CPU / storage speed, not GPU
GridMind fit
Excellent — any tier
Most customers say "training" when they mean "inference" — this is the most common misunderstanding
When a CEO or CTO says "we want to train our own AI", they almost always mean one of two things: (1) they want to run a pre-trained model that already exists (inference), or (2) they want to fine-tune an existing model on their documents (fine-tuning). Full training from scratch is extremely rare and extremely expensive. Clarify this early — it determines whether the customer needs a $20,000 Starter node or a $400,000 Enterprise.
The most important hardware constraint — VRAM determines everything
VRAM: why it matters and how to calculate how much you need
VRAM (Video RAM) is the memory on the GPU where the AI model lives during operation. The entire model must fit in VRAM — if it doesn't fit, it either fails to load or runs at a crawl using system RAM. This single constraint determines which GPU tier the customer needs.
VRAM required by model size (inference, most common use case)
Model
Parameters
FP32 (full)
FP16 (half)
INT8 (quant.)
INT4 (quant.)
Minimum GPU tier
Notes
Llama 3.2 1B
1 billion
4 GB
2 GB
1 GB
0.5 GB
Any — even GB10
Entry-level — customer service bots, simple Q&A
Llama 3.1 7B
7 billion
28 GB
14 GB
7 GB
4 GB
RTX 4090 (24 GB) at INT8
Good quality general assistant. Most common inference workload.
Llama 3.1 13B
13 billion
52 GB
26 GB
13 GB
7 GB
2× RTX 4090 or RTX PRO 6000
Better quality. Needs GPU bridging at FP16.
Llama 3.1 70B
70 billion
280 GB
140 GB
70 GB
35 GB
RTX PRO 6000 Server (96 GB INT8) or 4× RTX 4090
High quality. VRAM is the constraint. 4× RTX 4090 = 96 GB covers INT8.
Llama 3.1 405B
405 billion
1,620 GB
810 GB
405 GB
202 GB
8× H100 (640 GB HBM3) at INT4
Frontier quality. Minimum 8× H100 at INT4. H200 or B200 more comfortable.
GPT-4 class (est.)
~1.76 trillion
7,040 GB
3,520 GB
1,760 GB
880 GB
Multiple Sovereign Racks
Not practical on any single deployment. Frontier AI providers only.
What is quantisation and why does it matter?
A model trained in FP32 (32-bit floating point) stores each weight as a 32-bit number. FP16 halves that to 16 bits, INT8 halves again to 8 bits, and INT4 halves once more to 4 bits. Each step roughly halves the VRAM needed. INT4 quantisation (using tools like llama.cpp, GGUF format, or NVIDIA TensorRT-LLM) allows a 70B parameter model to run on 4× RTX 4090 GPUs (96 GB total) that would otherwise need 280 GB at FP32. Quality loss from INT4 is typically less than 2% on most benchmarks — imperceptible in practice. This is why the RTX 4090 and RTX PRO 6000 are so powerful at the SMB tier — they can serve very large models at acceptable quality using quantisation.
VRAM for concurrent users — the key multiplier
VRAM is shared — the model loads once, all concurrent users share it
This is a common misunderstanding. If a 7B model needs 14 GB of VRAM, it uses 14 GB regardless of whether 1 user or 50 users are querying it simultaneously. What changes with more concurrent users is the KV cache (key-value cache) — temporary memory used to track each active conversation. Each concurrent user typically uses 0.5–2 GB of KV cache, depending on context length. Rule of thumb: VRAM needed = model size + (concurrent users × 1 GB KV cache). So 8 concurrent users on a 7B model = 14 GB + 8 GB = 22 GB. Just fits in an RTX 4090.
GridMind VRAM capacity by node
Node
Total VRAM
Max model (INT4)
Max concurrent users (7B model)
Max concurrent users (70B model)
Starter (4× RTX 4090)
96 GB
~190B parameters (INT4)
~80 users
~60 users (70B INT4 = 35 GB + 60 GB KV)
Spark (GB10)
128 GB unified
~250B parameters (INT4)
~110 users
~90 users
Pro (4× RTX PRO 6000)
192 GB ECC
~380B parameters (INT4)
~170 users
~155 users
Starter Plus (8× RTX 4090)
192 GB
~380B parameters (INT4)
~170 users
~155 users
Enterprise (8× H100 NVL)
640 GB HBM2e
Llama 3.1 405B (INT4)
~600 users
~550 users
Enterprise (8× H100 SXM5)
640 GB HBM3
Llama 3.1 405B (INT4)
~600 users
~550 users
Ent H200 (8× H200 SXM5)
1.1 TB HBM3e
Multiple 405B instances
~1,050 users
~1,000 users
Ent B200 (8× B200)
1.5 TB HBM3e
Multiple large models
~1,450 users
~1,400 users
Sovereign NVL72
13.8 TB HBM3e
Multiple frontier models
13,000+ users
13,000+ users
Software AI engineer perspective
Throughput vs latency — two different performance requirements
A hospital emergency system needs answers in <500ms (latency matters). A law firm processing 10,000 contracts overnight just needs high throughput — speed per document doesn't matter as much as volume. The GPU choice differs significantly between these two requirements.
Latency-critical (real-time) workloads
The user is waiting. Every 100ms matters. These workloads need fast time-to-first-token (TTFT) — the delay before the first word appears. HBM3e memory (H200, B200) has much higher bandwidth than GDDR6X (RTX 4090), which means faster initial response.
ICU patient monitoring — alert latency ≤ 500ms
Real-time fraud detection — decision in < 1 second
Voice AI assistants — response in < 200ms for natural feel
Trading / algorithmic decisions — sub-millisecond in some cases
Hardware priority: HBM3e memory bandwidth. H200 (4.8 TB/s) or B200 (8.0 TB/s) for smallest models at highest speed. RTX 4090 acceptable for <13B models.
Throughput-critical (batch) workloads
The user submits a job and comes back later. Speed per item doesn't matter — total volume per hour does. These can run at lower priority during off-peak hours (GSOL idle time). The GPU should process as many tokens per second as possible for maximum throughput.
Hardware priority: Raw TFLOPS. B200 at 9,000 TFLOPS FP4 is ideal. RTX 4090 clusters are cost-effective. GSOL idle time can fulfil batch jobs from other organisations.
Tokens per second by GPU (Llama 3.1 70B, INT4 quantisation)
GPU configuration
Tokens/sec (70B INT4)
Words/sec (approx)
User experience
4× RTX 4090 (96 GB)
~18–25 tok/s
~13–18 words/sec
Readable streaming — slightly slow for heavy users
8× RTX 4090 (192 GB)
~35–50 tok/s
~25–35 words/sec
Good streaming experience
4× RTX PRO 6000 Server (192 GB)
~40–55 tok/s
~28–40 words/sec
Good — better memory bandwidth than 4090
8× H100 NVL (640 GB)
~120–150 tok/s
~85–107 words/sec
Excellent — near-native typing speed
8× H100 SXM5 (640 GB)
~150–200 tok/s
~107–143 words/sec
Excellent — fast even for complex prompts
8× H200 SXM5 (1.1 TB)
~200–280 tok/s
~143–200 words/sec
Very fast — HBM3e 4.8 TB/s memory bandwidth
8× B200 (1.5 TB)
~400–600 tok/s
~285–430 words/sec
Exceptional — 8 TB/s memory bandwidth
Figures are estimates for single-user inference. Throughput per user decreases with concurrency but total system throughput increases. Real-world figures depend on context length, prompt complexity, and batch size.
Why some GPUs need special cabling and some don't
GPU interconnect — NVLink vs PCIe, and when it matters
When a model is too large to fit on a single GPU, it must be split across multiple GPUs. How fast those GPUs can communicate with each other determines how efficiently the distributed model runs. The wrong interconnect can negate the benefit of having multiple GPUs.
PCIe (standard motherboard bus)
The standard connection between GPU and motherboard. PCIe 4.0 × 16 = 64 GB/s in each direction. Adequate for inference where GPUs don't need to talk much, but too slow for training where GPUs constantly synchronise gradients.
Bandwidth
64 GB/s (PCIe 4.0 × 16)
When sufficient
Inference · Embedding
When insufficient
Training · Large fine-tuning
GridMind nodes
RTX 4090 · RTX PRO 6000 · H100 NVL
NVLink (NVIDIA proprietary high-speed)
Direct chip-to-chip connection between NVIDIA GPUs. 10–28× faster than PCIe. Makes a cluster of 8 GPUs behave almost like a single giant GPU. Essential for training and large-model inference where the model must be split across multiple GPUs.
NVLink 4.0 (H100 SXM5)
900 GB/s bidirectional
NVLink 5.0 (B200)
1,800 GB/s bidirectional
NVLink-C2C (GB200)
900 GB/s CPU–GPU
GridMind nodes
H100 SXM5 · H200 SXM5 · B200 · GB200
InfiniBand — connecting multiple servers together
NVLink connects GPUs within one server. When you need multiple servers to cooperate on the same training job (distributed training across server nodes), you need a high-speed network. Standard 1 GbE or 10 GbE Ethernet is far too slow. All GridMind Enterprise nodes include InfiniBand HDR (200 Gb/s) or NDR (400 Gb/s) NICs. InfiniBand provides ~25 GB/s per port vs ~1.25 GB/s on 10 GbE — it's effectively the "PCIe" of the network layer. For inference-only GridMind deployments, standard 10 GbE is typically sufficient. InfiniBand only becomes critical when you're splitting a single training job across multiple servers.
AI software engineer perspective
The software stack — what runs on top of the hardware
Hardware is useless without the software that talks to it. Understanding the AI software stack lets you have informed conversations with technical teams and know what GSOL provides vs what the customer needs to bring themselves.
Layer by layer — from silicon to application
Customer application
Web interface, API, chatbot, document pipeline — what end users actually interact with
vLLM, Ollama, NVIDIA Triton, llama.cpp — manages the model and handles concurrent requests via batching
GSOL deploys standard options. Customer can bring their own.
AI framework
PyTorch, JAX, TensorRT-LLM — the Python library that talks to the GPU driver and runs tensor operations
Pre-installed on all GridMind nodes via GSOL.
CUDA / ROCm
NVIDIA's parallel computing platform. Every AI framework uses it. Requires specific driver versions. All GridMind nodes run CUDA 12.x.
Managed by GridMind. Automatic updates via GSOL.
GPU driver + firmware
NVIDIA driver 535+ for Hopper, 560+ for Blackwell. Controls how the OS talks to the GPU silicon.
GridMind manages. Locked by GSOL. Not customer-accessible.
GPU silicon
RTX 4090 · H100 · H200 · B200 · GB200 — the physical chip doing the matrix multiplication
GridMind owns and maintains.
Key software the customer will likely ask about
Software
What it does
Customer uses it when
GridMind stance
vLLM
High-throughput inference server. PagedAttention means very efficient KV cache management — 2–4× more concurrent users per GPU than naive approaches.
They need to serve many users from one GPU node simultaneously
Pre-deployed via GSOL. Default inference backend.
Ollama
Simple local model runner. One command to download and run any open model. No coding required.
Small team needs to run Llama/Mistral without DevOps complexity
Available on Starter/Spark nodes. Customer self-service.
NVIDIA TensorRT-LLM
NVIDIA's optimised inference compiler. Converts model to GPU-specific format for 2–4× speed increase vs raw PyTorch.
Maximum performance on H100/H200/B200
Available. Requires brief setup — GSOL can configure.
LangChain / LlamaIndex
Python libraries that help build RAG pipelines — connecting LLMs to databases, documents, APIs.
Customer is building a RAG system over their documents
Customer installs themselves — no GPU interaction.
Hugging Face Transformers
Library with 500,000+ pre-trained models and standard interface to load and run them.
Customer needs a specific model for their use case
Pre-installed on all nodes. Standard interface.
NVIDIA AI Enterprise (NVAIE)
NVIDIA's commercial support subscription. Includes security updates, SLAs, and enterprise support for AI frameworks.
Government/APRA-regulated customers needing vendor support SLA
Available on RTX PRO 6000 Server, H100, H200, B200 tiers.
The quick reference guide — for customer conversations
Hardware decision guide — from customer requirement to GridMind node
Use this guide in customer meetings. Ask these questions in order. The answers lead directly to the correct hardware tier without needing any engineering background.
The four questions to ask every customer
1
"Are you running an existing model, or training a new one?"
Existing model (inference) → almost any tier works. Training from scratch → Enterprise+ minimum. Fine-tuning → Enterprise or Pro minimum.
2
"Which AI model are you planning to use? Do you know its size in parameters?"
7B → Starter or Pro. 70B → Pro or Starter Plus. 405B → Enterprise or Enterprise. Bigger → H200, B200, Sovereign. If they don't know: ask "What task — answering questions, writing reports, analysing images?" and match from there.
3
"How many people will be using it at the same time, at peak?"
1–20 concurrent → Starter or Spark fine. 20–100 → Starter Plus or Pro. 100–500 → Enterprise or Enterprise. 500–2,000 → Ent H200 or B200. 2,000+ → multiple nodes or Sovereign.
4
"Do you have compliance requirements — IRAP, APRA, health data?"
Standard commercial → any tier. IRAP SENSITIVE / APRA → RTX PRO 6000 minimum (NVIDIA AI Enterprise). IRAP PROTECTED → Enterprise or H100 minimum (Phase 2 pathway). Defence classified → contact GridMind directly — sovereign configuration.
Quick lookup matrix
If the customer says...
Workload type
Recommended node
Why
"We want a chatbot for our 200 staff"
Inference, 7B–13B model
Starter or Pro
200 staff rarely all concurrent. 20–50 peak → 96 GB VRAM sufficient for 13B INT8.
"AI tutor for our 1,000-student school"
Inference, 7B model, many concurrent
7–8× Starter Plus
280 peak students × 1 GPU/8 users = ~35 GPUs. 7 Starter Plus = 56 GPUs + headroom.
"Replace our AWS Bedrock spend"
Inference, Claude-class model
Enterprise or H200
AWS Bedrock runs 70B+ class models. Needs 640 GB+ VRAM. H100 or H200.
Fraud models are typically 7B–13B. VRAM is low. Latency is the driver — multiple GPUs for parallel queries.
"We process 50,000 documents overnight"
Batch embedding or summarisation
Starter or Starter Plus
Batch, not real-time. RTX 4090 at 18–25 tok/s × 4 GPUs × overnight = millions of tokens. More than enough.
"Sovereign AI for ADF logistics"
Inference + fine-tuning, classified
Sovereign Rack (NVL72)
Defence workloads require IRAP PROTECTED, maximum performance, air-gapped capability.
AI-powered — acts as your senior computer and AI software engineer
Ask the AI hardware advisor
Describe your customer's situation in plain English. The AI advisor — drawing on the expertise of a senior computer engineer and AI software engineer — will specify the correct hardware, explain why, identify any risks, and give you a script for the customer conversation.
Describe the customer and their AI requirement
Consulting computer engineer and AI software engineer...
Hybrid Agentic Inference — The Inevitable Architecture
This is not a product pitch. It is a description of where every regulated Australian business is heading — whether they know it yet or not.
The fundamental argument
Your competitive advantage is not the AI model you chose. It is the tacit knowledge your organisation has built over years — and your ability to train AI on it.
The commodity
The frontier AI model
GPT-4, Claude, Llama, Gemini — available to every one of your competitors at the same price. The model itself confers no advantage. It is infrastructure, like electricity.
The moat
Your tacit knowledge
The clinical protocols refined over decades. The underwriting intuitions from thousands of claims. The engineering solutions from years of failure. The customer patterns no competitor can see. This cannot be bought. This is yours.
The advantage
AI trained on your tacit knowledge
A model fine-tuned on your proprietary data outperforms any frontier model on your specific domain tasks — because it knows your business from the inside. This runs on infrastructure you own. It never leaves.
Your competitive advantage is your knowledge. Your knowledge lives in your data. Your data must stay on your hardware. That is why you own your AI infrastructure.
Why your data must never leave your infrastructure
The moment you send your proprietary data to a hyperscaler for AI processing, you have handed your most valuable asset to a company whose infrastructure also serves your competitors. You have no visibility into how that data is used. You have no guarantee it is not used to train their next model. And you have US CLOUD Act exposure on top of all of that.
A hospital's patient outcomes data is its most valuable research asset. A bank's fraud pattern data is its most defensible competitive moat. A manufacturer's quality control data is its accumulated engineering intelligence. These are not just compliance obligations — they are strategic assets that belong on infrastructure the organisation controls.
The direction of travel — legislation always follows practice
We legislated data sovereignty because nations recognised that sensitive national information should not be stored on foreign infrastructure. The same logic is now being applied to AI inference — because inference on data is equivalent to transmitting that data to whoever runs the compute.
GDPR came after data misuse was already happening. APRA CPS 230 came after the operational risk was already being taken. Privacy Act 2024 penalties are already law. The AI inference enforcement action hasn't landed yet — but when it does, every organisation sending sensitive data to offshore AI will need a solution in weeks. GridMind deploys in 12.
The hybrid architecture — two inference paths, one AI interface
Sensitivity routing — how every query is handled
Cloud frontier model (AWS / Azure / OpenAI)
General research · Public document summaries · Marketing content · Code assistance · Staff training · Any query with no sensitive content
Best capability · Lowest cost · Fine for public data
US CLOUD Act applies regardless of server location
AI Agent
Classifies sensitivity · Routes query · Combines result
GridMind on-premises (your hardware)
Patient records · Financial data · Employee files · Legal advice · Proprietary IP · Commercial negotiations · Classified briefs · Anything that cannot legally leave
Open source 70B model fine-tuned on your tacit knowledge
Australian hardware · Australian law · Your model · Your advantage
Do you need frontier models? — the open source question
Open source is sufficient for most sensitive workloads
Llama 3.1 70B, Mistral Large, Qwen 2.5, DeepSeek — running on a GridMind H100 cluster — handles clinical documentation, legal contract review, fraud alert analysis, policy drafting, and financial reporting at a quality that most enterprise users cannot distinguish from GPT-4 in practice.
For these workloads, you don't need a frontier model. You need a model that understands your domain — and a 70B open source model fine-tuned on your institutional data outperforms any general frontier model on your specific tasks, because it knows your business from the inside.
Where frontier models genuinely outperform open source
Complex novel reasoning chains · Multi-step scientific analysis · Broad creative tasks requiring world knowledge · Very long-context planning across many steps
Critically: these tasks almost never involve sensitive data. A lawyer asking for novel legal argumentation on a public case can safely use a cloud frontier model. The overlap between "needs frontier capability" and "contains sensitive data" is smaller than most people assume — which makes the hybrid architecture work cleanly.
The GSOL advantage — two benefits most customers don't expect
Benefit 1 — Idle capacity earns passive revenue
Your GridMind hardware runs your AI workloads during business hours. But outside those hours — nights, weekends, and quiet periods — the GPUs are idle. GSOL automatically sells that idle capacity on the GridMind sovereign marketplace, generating passive revenue for your organisation with zero management overhead.
A hospital with 7× Starter Plus nodes (56 GPUs) running AI during clinical hours has approximately 82% of total annual GPU-hours available for GSOL. At blended marketplace rates, that generates $67,000–$185,000 per year in passive income — offsetting electricity costs and partially recovering hardware investment.
You didn't buy the hardware just to run your own AI. You bought infrastructure that works for you 24 hours a day — your workloads during the day, GSOL revenue through the night.
Benefit 2 — Burst to sovereign hardware when demand spikes
Your own hardware is sized for your typical peak load — not for rare demand spikes. When your annual report drops and 800 staff hit the AI system simultaneously, or when a major clinical event requires intensive compute, your local nodes may reach capacity.
GSOL gives you access to the wider GridMind network — other organisations' idle sovereign hardware across Australia, all IRAP-pathway compliant, all on Australian soil. Your overflow queries route to the nearest available GSOL node automatically. You get burst capacity on demand, with the same sovereignty guarantees as your own hardware.
This is the key difference from cloud bursting. Cloud burst capacity goes to AWS or Azure — back to foreign-owned infrastructure, back to CLOUD Act exposure. GSOL burst capacity stays on Australian sovereign hardware, owned by other Australian organisations, governed by Australian law.
Pitching this in a customer conversation
Opening — reframe from compliance to strategy
"The question isn't whether you'll run AI on local sovereign infrastructure. That's already being legislated — just like data residency was legislated after Snowden. The question is whether you build that infrastructure now, as a strategic asset, or scramble to build it when an enforcement action forces your hand."
The tacit knowledge argument
"Your competitive advantage isn't the AI model you chose — every one of your competitors can buy the same model at the same price. Your advantage is the tacit knowledge your organisation has built over years. The clinical protocols, the risk models, the customer patterns, the institutional memory. The organisation that trains AI on that knowledge, on infrastructure they own, builds a compounding advantage that no competitor can replicate — because they don't have your data."
The question that closes the conversation
"Can you show me your current data governance policy for what categories of information you're allowed to send to AWS or Azure for AI processing?" — Most organisations either don't have one, or their policy explicitly prohibits the categories they're already sending. This creates the moment of realisation. GridMind is the solution that was already required.
The economics close
"And while your hardware is idle — nights, weekends, quiet periods — GSOL is selling that spare capacity and depositing revenue into your account. If you hit a demand spike, GSOL bursts you to sovereign Australian hardware from other organisations on the network. You're not just buying AI infrastructure. You're buying an asset that generates income and scales with you — on Australian soil, under Australian law, serving your competitive advantage."
Hardware Pricing — AUD cost reference
Complete bill of materials for every GridMind node tier. All prices in AUD inc. GST. Exchange rate: AUD/USD 0.70 (June 2026). Prices are indicative ranges — confirm with distributors before quoting customers.
SMB tiers (Starter, Spark, Pro, Starter Plus) — component assembly RTX 4090, RTX PRO 6000 Blackwell Server Ed., and GB10 are PCIe cards or appliances — buy individually and assemble in a validated server chassis. Saves 25–35% vs pre-built. Sourceable via Ingram Micro AU, Arrow AU, Scorptec, PCCaseGear.
Enterprise tiers (H100, H200, B200, Sovereign) — complete Supermicro / Dell system only H100/H200/B200 SXM GPUs are NOT sold individually — only as HGX baseboard assemblies inside OEM servers. You cannot buy the GPU chip alone. Supermicro SYS-821GE-TNHR and SYS-A21GE-NBRT-G1 are the required vehicles.
💻 Who this product is for
Individual developer · sole trader · startup · proof of concept. Zero installation — sits on a desk, plugs into a standard 10A power point. Connect to GSOL and earn passive revenue from idle capacity. The entry point — when they see it work, they upgrade to a Starter node.
Upgrade path
Upgrades to: GridMind Starter (outdoor unit)
Spark
NVIDIA Spark (GB10)
Hardware only
$6K–$9K
Enclosure (optional)
$2K–$3K
Total excl. install
$8K–$12K
Total incl. install
$9K–$14K
Build approach: Buy the NVIDIA Spark appliance complete — the GB10 superchip is not sold separately. USD $3,000–$4,000 estimated (AUD $4,300–$5,700 + 10% duty + GST ≈ $5,200–$7,000 landed). Source: NVIDIA direct or Ingram Micro AU. Lead times 8–14 weeks.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
NVIDIA Spark (GB10 superchip)
128 GB unified · 450W · integrated system
1
$5,200–$7,000
$5,200–$7,000
NVIDIA AU / Ingram Micro AU
UPS 1 kVA
APC Back-UPS Pro 1500VA
1
$400–$600
$400–$600
APC AU
GSOL management SBC
Pi CM4 out-of-band
1
$120–$200
$120–$200
Core Electronics AU
10GbE managed switch
TP-Link TL-SG2210P or equiv.
1
$250–$400
$250–$400
Amazon AU
IP55 outdoor housing (optional)
Compact wall-mount enclosure
1
$1,500–$2,500
$1,500–$2,500
Custom AU fabricator
Total excl. installation
$7,470–$10,700
Licensed electrician (10A dedicated)
Circuit + cert · 2 hrs
1
$400–$700
$400–$700
Local electrician
Installation labour (1 hr)
Mounting + network
1
$150–$300
$150–$300
GridMind
TOTAL INCL. INSTALLATION
$8,020–$11,700
🏪 Who this product is for
Small business (10–50 staff) · GP clinic · law firm (small) · accountant · real estate agency · retail chain HQ. General AI workloads — email drafting, document summaries, client Q&A, compliance checking. No DA, no builder, no MCPU. Half-day installation.
Upgrade path
Upgrades to: GridMind Starter Plus (same enclosure, add 4 GPUs)
RTX 4090
NVIDIA RTX 4090 24 GB
Hardware only
$18K–$22K
Enclosure + power
$3K–$5K
Total excl. install
$21K–$27K
Total incl. install
$24K–$31K
Build approach: Component assembly. RTX 4090 stock is thinning (discontinued). Buy in bulk from Ingram Micro AU or Scorptec now. Threadripper PRO 5955WX platform for the motherboard. Price: AUD $2,400–$3,200/card (retail), potentially rising as stock depletes.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
NVIDIA RTX 4090 24 GB
Ada Lovelace · 450W · GDDR6X
4
$2,400–$3,200
$9,600–$12,800
Scorptec / PCCaseGear / Ingram Micro
AMD Threadripper PRO 5955WX CPU
16C · 280W · sTRX4
1
$1,800–$2,400
$1,800–$2,400
MSY / PBTech AU
WRX80 workstation motherboard
sTRX4 · PCIe 4.0 · 4× x16 slots
1
$1,200–$1,600
$1,200–$1,600
ASUS Pro WS WRX80E-SAGE
DDR5 ECC RDIMM 128 GB
6× 32 GB DDR5-4800 ECC
6
$180–$240
$1,080–$1,440
Kingston / Samsung via Ingram
Samsung 990 Pro NVMe 4 TB
PCIe 4.0 · 7,400 MB/s
2
$350–$450
$700–$900
Amazon AU / Scorptec
2× 1,200W 80+ Titanium PSU
Seasonic / EVGA Supernova
2
$380–$500
$760–$1,000
Scorptec / PBTech
PCIe 4.0 riser cables × 4
200mm right-angle ribbon
4
$80–$120
$320–$480
AliExpress / custom AU
GSOL ARM SBC
Raspberry Pi CM4 management
1
$120–$200
$120–$200
Core Electronics AU
10GbE NIC
Intel X550-T1
1
$220–$300
$220–$300
Server Parts AU
Hardware subtotal
$15,800–$21,120
IP55 outdoor aluminium enclosure
3mm 6061-T6 · powder-coated · C4
1
$2,500–$4,000
$2,500–$4,000
Custom AU fabricator
400mm axial fan + PWM controller
IP55 · variable speed · Ebm-papst
1
$300–$500
$300–$500
Ebm-papst AU
IP55 weatherproof socket + conduit
Clipsal 56 series 20A
1
$150–$250
$150–$250
Clipsal / Rexel AU
Total excl. installation
$18,750–$25,870
Concrete pad (150mm · 1.2×0.8m)
20 MPa · M12 anchors
1
$400–$800
$400–$800
Local concreter
Licensed electrician (20A circuit)
Circuit + socket + ESO cert
1
$800–$1,400
$800–$1,400
Local electrician
Site labour (2 people · 4 hrs)
Delivery + anchor + network
1
$400–$600
$400–$600
GridMind
TOTAL INCL. INSTALLATION
$19,950–$28,670
🏥 Who this product is for
Mid-market enterprise · hospital department · law firm (mid-size) · financial services · government agency. Enterprise-grade ECC memory, ISV certifications, 24/7 server-validated passive cooling. Same outdoor enclosure as Starter/Starter Plus — different cards, higher compliance capability.
Upgrade path
Upgrades to: GridMind Pro Plus (same enclosure, add 4 RTX PRO 6000 cards)
RTX PRO 6000 Server Ed.
NVIDIA RTX PRO 6000 Blackwell Server Ed.
Hardware only
$65K–$84K
Enclosure + power
$4K–$6K
Total excl. install
$69K–$90K
Total incl. install
$72K–$97K
Build approach: Component assembly in validated 4U server chassis (Supermicro or HPE platform for passive PCIe GPUs). RTX PRO 6000 Server Ed. sourced via Ingram Micro AU, Arrow AU, or HPE enterprise — not available at retail. USD $8,000–$9,200/card = AUD $11,500–$13,200/card before GST and distributor margin. Confirm availability — not yet in mainstream AU stock as of June 2026.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
NVIDIA RTX PRO 6000 Blackwell Server Ed.
96 GB GDDR7 ECC · 450W passive · PCIe 5.0
4
$11,500–$13,500
$46,000–$54,000
Ingram Micro AU / Arrow AU / HPE AU
AMD Threadripper PRO 7995WX
96C · 350W · sTR5 · Zen 4
1
$8,000–$12,000
$8,000–$12,000
PBTech AU / PCCaseGear (price volatile)
WRX90 server motherboard
sTR5 · 8× PCIe 5.0 slots · 12 DIMM
1
$1,800–$2,800
$1,800–$2,800
ASUS Pro WS TRX90-E2-SAGE SE
256 GB DDR5 ECC RDIMM
12× 32 GB DDR5-5200 registered ECC
12
$200–$280
$2,400–$3,360
Kingston / Micron via Ingram AU
NVMe 4 TB × 2
Samsung 990 Pro PCIe 4.0
2
$350–$450
$700–$900
Scorptec
4U server chassis (passive GPU airflow)
Front-to-back · Supermicro SC847 equiv.
1
$800–$1,400
$800–$1,400
Supermicro AU (ASI)
2× 2,000W 80+ Titanium PSU (redundant)
Server-grade hot-swap
2
$600–$900
$1,200–$1,800
Seasonic / FSP via Server Parts AU
10GbE dual-port NIC
Intel X710-T2L
1
$400–$600
$400–$600
Server Parts AU
GSOL ARM SBC
Raspberry Pi CM4
1
$120–$200
$120–$200
Core Electronics AU
Hardware subtotal
$61,420–$77,060
IP55 outdoor enclosure (Pro — larger unit)
3mm 6061-T6 · dual PSU bay · fan array
1
$3,500–$5,500
$3,500–$5,500
Custom AU fabricator
Redundant axial fan array (2× 400mm)
IP55 · Ebm-papst
1
$600–$1,000
$600–$1,000
Ebm-papst AU
Total excl. installation
$65,520–$83,560
Concrete pad (1.5×1.0m · 150mm)
20 MPa · M12 anchors
1
$600–$1,100
$600–$1,100
Local concreter
Licensed electrician (32A 3-phase)
3-phase circuit + switchboard + cert
1
$1,500–$2,500
$1,500–$2,500
Local electrician
Site labour (4 hrs)
Delivery + positioning + network
1
$600–$900
$600–$900
GridMind
TOTAL INCL. INSTALLATION
$68,220–$88,060
🏢 Who this product is for
Growing SMB (20–100 staff) · school · medium professional firm · Starter customers who have outgrown 4 GPUs. Same outdoor enclosure as Starter — upgrade by adding 4 GPU cards to the existing unit. No new pad, no new electrician, no new installation.
Upgrade path
Upgrades to: GridMind Pro (enterprise-grade ECC cards)
Hardware only
$35K–$50K
UNIT-A1 SIP pod
$55K–$95K
Total excl. install
$90K–$145K
Total incl. install
$110K–$175K
Build approach: Component assembly (8× RTX 4090) in Supermicro 8-GPU server chassis (AS-4124GS-TNRT2 or 4124GO-NART). UNIT-A1 SIP panel pod for the enclosure. Procure hardware and pod separately — both ship to site and assemble independently on the concrete slab.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
NVIDIA RTX 4090 24 GB
Ada Lovelace · 450W · GDDR6X
8
$2,400–$3,200
$19,200–$25,600
Scorptec / Ingram Micro bulk
Supermicro AS-4124GS-TNRT2 chassis + board
8-GPU · AMD EPYC · 8× PCIe 4.0 x16
1
$4,000–$6,500
$4,000–$6,500
Supermicro AU (ASI)
AMD EPYC 9554 or 2× Xeon Platinum 8480+
Server CPU · 360W TDP · dual socket
2
$3,500–$5,500
$7,000–$11,000
Ingram Micro AU
256 GB DDR5 ECC RDIMM
Server platform memory
1
$1,800–$2,600
$1,800–$2,600
Kingston / Micron
NVMe 4 TB × 2 · switch · rack + PDU
Storage + networking + rack infrastructure
1
$2,500–$4,000
$2,500–$4,000
Various AU
Hardware subtotal
$34,500–$49,700
UNIT-A1 SIP panel pod (48m²)
SIPs Industries WA · R3.8 · C4 · Class 10a
1
$40,000–$67,200
$40,000–$67,200
SIPs Industries WA / SIPs QLD
200mm RC slab (6×8m)
N12 mesh · 32 MPa · 12 kN/m²
1
$8,000–$14,000
$8,000–$14,000
Local concreter
Total excl. installation labour
$82,500–$130,900
Electrician (125A 3-phase) + pod erection + commissioning
4-person crew · 2 days · GSOL live
1
$10,000–$18,000
$10,000–$18,000
QBCC builder + electrician
TOTAL INCL. INSTALLATION
$92,500–$148,900
NVIDIA RTX PRO 6000 Blackwell × 8
Hardware only
$128K–$165K
Enclosure + power
$4K–$7K
Total excl. install
$132K–$172K
Total incl. install
$136K–$180K
Build approach: Same as Pro — 8× RTX PRO 6000 Blackwell Server Edition in validated 4U passive GPU server chassis. This is the upgraded Pro enclosure — same outdoor unit, full GPU bay populated. Source via Ingram Micro AU or Arrow AU enterprise channels. Confirm stock availability — Server Edition is enterprise channel only.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
NVIDIA RTX PRO 6000 Blackwell Server Ed.
96 GB GDDR7 ECC · 450W passive · PCIe 5.0
8
$11,500–$13,500
$92,000–$108,000
Ingram Micro AU / Arrow AU / HPE AU
AMD Threadripper PRO 7995WX
96C · 350W · sTR5 · Zen 4
1
$8,000–$12,000
$8,000–$12,000
PBTech AU / PCCaseGear
WRX90 server motherboard + 256 GB DDR5 ECC
sTR5 · 8× PCIe 5.0 · memory
1
$4,500–$6,500
$4,500–$6,500
ASUS Pro WS / Kingston via Ingram
4U server chassis + 2× 2,000W PSU
Passive GPU airflow · Supermicro equiv.
1
$2,500–$3,800
$2,500–$3,800
Supermicro AU (ASI)
10GbE dual-port NIC + GSOL SBC + NVMe
Management + storage
1
$700–$1,000
$700–$1,000
Server Parts AU
Hardware subtotal
$107,700–$131,300
IP55 outdoor enclosure (Pro size)
3mm 6061-T6 · dual PSU bay · fan array
1
$3,500–$5,500
$3,500–$5,500
Custom AU fabricator
Redundant axial fan array
2× 400mm · Ebm-papst
1
$600–$1,000
$600–$1,000
Ebm-papst AU
Total excl. installation
$111,800–$137,800
Concrete pad + licensed electrician (32A 3-phase) + site labour
Full installation
1
$3,000–$5,000
$3,000–$5,000
Local trades
TOTAL INCL. INSTALLATION
$114,800–$142,800
🏗 Who this product is for
Large enterprise · university · hospital network · government department running frontier-class inference. Requires dedicated UNIT-A1 building pod — separate structure, DA required, 6–14 weeks. Supermicro or Dell complete validated system. MCPU-S liquid cooling.
Upgrade path
Upgrades to: GridMind Enterprise Plus H200 (UNIT-A2 dual-storey pod)
Complete server
$300K–$460K
Enclosure + MCPU
$70K–$115K
Total excl. install
$370K–$575K
Total incl. install
$420K–$650K
Build approach: H100 NVL PCIe cards can be individually sourced (USD $25,000–$30,000 each = AUD $35,700–$42,900) but a validated Supermicro or Dell server chassis is strongly recommended for 8-GPU NVLink configuration. At this tier the firmware integration complexity makes pre-validated systems worth the 15–20% premium over component assembly.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
Validated 8× H100 NVL server (Dell XE8545 / Supermicro)
8× H100 NVL · dual EPYC · 1TB DDR5 · NVLink
1
$300,000–$455,000
$300,000–$455,000
Dell AU enterprise / Supermicro AU (ASI)
InfiniBand HDR switch (40-port)
Mellanox QM8790 · 200G HDR
1
$18,000–$28,000
$18,000–$28,000
Ingram Micro AU
42U rack + PDU (3-phase 32A)
APC NetShelter + PDU
1
$3,000–$5,000
$3,000–$5,000
APC AU
GSOL management server (1U)
Out-of-band · IPMI · GSOL agent
1
$2,000–$3,500
$2,000–$3,500
Supermicro AU
Hardware subtotal
$323,000–$491,500
UNIT-A1 Kingspan PIR pod (48m²)
100mm PIR · R5.3 · Class 8 · C4
1
$52,800–$86,400
$52,800–$86,400
Kingspan AU
200mm RC slab · RPEQ cert
N12 · 32 MPa · 12 kN/m²
1
$12,000–$20,000
$12,000–$20,000
RPEQ structural contractor
MCPU-S dry cooler (rooftop)
Vertiv CoolChip · up to 15kW
1
$8,000–$14,000
$8,000–$14,000
Vertiv AU
Total excl. installation labour
$395,800–$611,900
DA + electrician (200A 3-phase) + pod erection + commissioning
Class 8 approval · RPEQ · 5-day crew
1
$28,000–$50,000
$28,000–$50,000
QBCC builder + RPEQ electrician
TOTAL INCL. INSTALLATION
$423,800–$661,900
Supermicro SYS-821GE-TNHR
Supermicro SYS-821GE-TNHR
Complete server
$453K–$570K
Enclosure + MCPU
$75K–$130K
Total excl. install
$528K–$700K
Total incl. install
$580K–$790K
Mandatory: buy complete Supermicro SYS-821GE-TNHR system. H100 SXM5 GPUs are not sold as individual cards — only as part of HGX H100 baseboard assemblies inside OEM servers. USD $317,495 list = AUD ~$453,500. Source via Supermicro AU distributor: ASI (Australian Scientific Instruments) or contact Supermicro AU directly.
DA + RPEQ cert + electrician (250A 3-phase) + pod erection + commissioning
Class 8 approvals · RPEQ · 6-day crew
1
$38,000–$62,000
$38,000–$62,000
QBCC builder + RPEQ licensed trades
TOTAL INCL. INSTALLATION
$603,800–$815,400
🏢 Who this product is for
Large enterprise · national institutions · defence agencies. UNIT-A2 dual-storey pod. Direct liquid cooling mandatory.
Upgrades to: GridMind Enterprise Plus B200
Complete server (H200)
$453K–$640K
UNIT-A2 + MCPU
$150K–$260K
Total excl. install
$603K–$900K
Total incl. install
$680K–$1.02M
Mandatory: Supermicro SYS-821GE-TNHR configured with H200 HGX baseboard. Same chassis as H100 — different GPU baseboard (~15–20% premium). DLC mandatory. UNIT-A2 dual-storey pod required for H200 deployment (higher floor load rating, MCPU-M integration).
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
Supermicro SYS-821GE-TNHR (H200)
8× H200 SXM5 · 1.1TB HBM3e · DLC mandatory
1
$453,500–$640,000
$453,500–$640,000
Supermicro AU (ASI)
IB NDR switch + rack + PDU + GSOL mgmt
Networking + infrastructure
1
$36,000–$60,000
$36,000–$60,000
Ingram Micro AU / APC AU
MCPU-S or MCPU-M CDU
Vertiv · up to 60kW · 45°C coolant loop
1
$18,000–$30,000
$18,000–$30,000
Vertiv AU
Hardware subtotal
$507,500–$730,000
UNIT-A2 Kingspan dual-storey pod
150mm PIR · R8.0 · Class 8 · C4 · RPEQ
1
$110,000–$192,000
$110,000–$192,000
Kingspan AU + SHS fabricator
300mm PT RC slab (upper floor)
N16 · 40 MPa · 20 kN/m²
1
$22,000–$38,000
$22,000–$38,000
RPEQ structural contractor
Total excl. installation labour
$639,500–$960,000
DA + RPEQ + 400A 3-phase switchboard + pod erection + commissioning
Full Class 8 dual-storey works · crane
1
$76,000–$120,000
$76,000–$120,000
QBCC builder + RPEQ licensed trades
TOTAL INCL. INSTALLATION
$715,500–$1,080,000
🏛 Who this product is for
National AI agencies · top-tier research institutions · large defence. Maximum Blackwell performance. UNIT-A2 dual-storey pod.
Maximum capability in this product line
Complete server (B200)
$715K–$1.0M
UNIT-A2 + MCPU-M
$210K–$380K
Total excl. install
$925K–$1.38M
Total incl. install
$1.05M–$1.58M
Mandatory: Supermicro SYS-A21GE-NBRT-G1 (10U B200 Gold Series). USD $500,000+ = AUD $715,000+. Lead times 8–20 weeks. B200 HGX baseboard not sold as individual GPUs. DLC-2 mandatory — integrated into the 10U chassis.
Component
Spec
Qty
Unit (AUD)
Total (AUD)
AU supplier
Supermicro SYS-A21GE-NBRT-G1 (B200)
8× HGX B200 · 1.5TB HBM3e · 10U · DLC-2 · 400GbE
1
$715,000–$1,000,000
$715,000–$1,000,000
Supermicro AU (ASI) · 8–20 wk lead time
IB NDR400 switch + rack + PDU + GSOL
NVIDIA QM9790 + 42U rack + 63A PDU
1
$48,000–$80,000
$48,000–$80,000
Ingram Micro AU / APC AU
MCPU-M CDU outdoor (60kW)
Vertiv MegaMod HDX · outdoor dry cooler
1
$45,000–$75,000
$45,000–$75,000
Vertiv AU
Hardware subtotal
$808,000–$1,155,000
UNIT-A2 Kingspan dual-storey pod (120m²)
150mm PIR · R8.0 · Class 8 · C4 · RPEQ
1
$132,000–$216,000
$132,000–$216,000
Kingspan AU
300mm PT slab + civil + MCPU-M pad
RPEQ designed · 20 kN/m²
1
$55,000–$90,000
$55,000–$90,000
RPEQ structural contractor
Total excl. installation labour
$995,000–$1,461,000
DA + RPEQ + 400A switchboard + crane + pod erection + DLC-2 + commissioning
Full Class 8 · crane · 10-day crew
1
$97,000–$165,000
$97,000–$165,000
QBCC builder + RPEQ licensed trades
TOTAL INCL. INSTALLATION
$1,092,000–$1,626,000
GB200 NVL72
NVIDIA GB200 NVL72
NVL72 rack hardware
$17M–$23M
UNIT-A2 + MCPU-L
$300K–$550K
Total excl. install
$17.3M–$23.6M
Total incl. install
$17.7M–$24.0M
NVIDIA AU direct or authorised partner only. USD $12M–$16M/rack = AUD $17.1M–$22.9M. Lead time 6–18 months. Requires NVIDIA country export approval. For Australian government or defence: contact NVIDIA AU + DTA for sovereign pathway. MCPU-L (LiquidStack GigaModular, 14MW scalable CDU) is the external cooling system.
DA + Class 8 RPEQ + 1,000A switchboard + NVL72 specialist install + MCPU-L + commissioning
NVIDIA certified install team · crane · 2-week programme
1
$225,000–$500,000
$225,000–$500,000
NVIDIA cert team + QBCC + RPEQ trades
TOTAL INCL. INSTALLATION
$17.7M–$24.0M
For investor conversations
Is buying hardware and depreciating over 5 years better than paying hyperscalers forever?
Short answer: yes, decisively — and the advantage compounds over time. Here is the analysis a VC or CFO needs to see.
Configure scenario
Hardware tier
Peak usage hours/day 10 hrs
Number of GPUs in deployment 8
Depreciation period (years)
Annual hardware value decline
AWS monthly cost
$0
at peak usage hours
GridMind monthly cost
$0
depreciation + electricity only
Monthly saving
$0
vs staying on AWS
Payback period
0 yrs
hardware cost ÷ monthly saving
5-year AWS spend
$0
total outlay, zero residual
5-year GridMind TCO
$0
hardware + power (net of GSOL)
5-year net advantage (GridMind)
$0
savings vs AWS over 5 years
Annual depreciation (straight-line)
$0
P&L charge per year (non-cash after year 1)
Configure the scenario above to see the analysis.
Year-by-year comparison
Year
AWS cumulative spend
GridMind hardware cost (depreciated)
GridMind electricity
GridMind GSOL revenue (idle)
GridMind net TCO
Cumulative advantage
Answering the VC / CFO questions
"Isn't it risky to spend $700K upfront when you could just use AWS and scale as needed?"
The risk framing is backwards. Every dollar spent on AWS is gone — zero residual value, zero asset, zero balance sheet contribution. Every dollar spent on GridMind hardware is a depreciating asset that appears on the balance sheet, generates GSOL revenue when idle, reduces monthly opex by the AWS equivalent, and retains some resale value. The AWS "flexibility" argument only holds if the organisation is uncertain whether they will need AI compute at all — and at the scale of an enterprise H100 deployment, that uncertainty has already been resolved. The risk of AWS is not flexibility — it is permanent operating leverage with no asset to show for it.
"What about hardware obsolescence — will a $700K H100 server be worthless in 5 years?"
Partially, but this cuts both ways. The H100 will still run inference on 70B models competently in 2030 — the models are not going to suddenly require more compute for the same tasks. What changes is that newer models may require more. The correct response is a refresh cycle, which every tech company already budgets for. More importantly: the comparison is not "H100 in 2030 vs B200 in 2026." The comparison is "H100 depreciated over 5 years vs the equivalent AWS spend over 5 years." The AWS equivalent spend is $0 in residual value and $0 in balance sheet contribution — the H100 at year 5 still has a resale market and still runs inference workloads that existed at year 1.
"How is the depreciation structured — is it deductible?"
In Australia, server hardware qualifies for depreciation under the general depreciation rules (Division 40 ITAA 1997). The ATO effective life for computer hardware is 5–6 years, making a 5-year straight-line schedule appropriate. Under the instant asset write-off provisions (current threshold AUD $20,000 for small business, full expensing for eligible businesses), a business may be able to deduct the full cost in Year 1 rather than depreciating — which significantly improves the Year 1 cash position. For enterprise-scale deployments ($700K+), standard Division 40 depreciation applies. The depreciation charge is a non-cash P&L item from Year 2 onwards — the cash outflow is at acquisition. Confirm treatment with your tax adviser. Note: this is information, not tax advice.
"What's the GSOL revenue assumption — is it realistic?"
The GSOL revenue in the model uses Vast.ai floor rates (H100: ~$2.50–$3.50/GPU-hr) as a conservative proxy for sovereign GPU marketplace pricing. The idle assumption (24 hrs minus peak usage hrs per day) is conservative — organisations typically use AI infrastructure for 8–12 hours of peak demand, leaving 12–16 hours per night available. GSOL revenue is presented as a partial offset rather than the primary financial case — even at zero GSOL revenue, the depreciation model substantially outperforms AWS over 5 years at enterprise scale. GSOL improves the case further but is not required for the model to work.
All tiers — AUD cost summary
Node
GPUs / VRAM
Build approach
Excl. install
Incl. install
Key constraint
Spark
GB10 · 128 GB unified
Buy Spark appliance
$7K–$11K
$8K–$12K
Spark only via NVIDIA/Ingram
Starter
4× RTX 4090 · 96 GB
Component assembly
$19K–$26K
$20K–$29K
RTX 4090 stock depleting
Starter Plus
8× RTX 4090 · 192 GB
Component assembly
$93K–$145K
$110K–$175K
RTX 4090 stock + UNIT-A1 pod lead time
Pro
4× RTX PRO 6000 Server · 192 GB
Component assembly
$66K–$84K
$68K–$88K
Server Ed. not in retail — enterprise channels only
Pro Plus
8× RTX PRO 6000 Server · 384 GB ECC
Component assembly
$112K–$138K
$115K–$143K
Same enclosure as Pro — upgrade by adding 4 cards
Enterprise
8× H100 NVL · 640 GB
Validated server recommended
$400K–$615K
$425K–$665K
Firmware complexity — use Supermicro/Dell
Enterprise H200
8× H200 SXM5 · 1.1 TB
Complete Supermicro ONLY
$640K–$960K
$716K–$1.08M
H200 SXM5 not sold individually · DLC mandatory
Enterprise B200
8× B200 · 1.5 TB
Complete Supermicro ONLY
$995K–$1.46M
$1.09M–$1.63M
8–20 week lead time · DLC-2 mandatory
Sources: AUD/USD rate 0.70 (June 2026, Trading Economics). Supermicro SYS-821GE-TNHR USD $317,495 (Supermicro store). RTX PRO 6000 Blackwell USD $8,000–$9,200 (Thunder Compute, June 2026). H200 server +15–20% premium over H100 (industry consensus). B200 server USD $500,000+ (Supermicro/Viperatech). GB200 NVL72 USD $12M–$16M (NVIDIA). All AUD prices include 10% GST and estimated distributor margin. Installation costs are Queensland estimates — vary by site and region.
Customer Discovery — What hardware does this customer need?
Two paths to a hardware recommendation. Use whichever matches the conversation. Both paths output a tier recommendation, payback period, and a ready-to-use customer pitch.
How to use this tool
1
Ask the customer: "Are you currently paying for any AI services — ChatGPT Enterprise, AWS Bedrock, Azure OpenAI, or similar?" If yes, ask: "Could you share your last invoice or last three months of billing?"
2
Enter their monthly spend by provider on the left. Select the model they primarily use — this matters because pricing per token varies enormously (GPT-4 is 20× more expensive per token than GPT-3.5).
3
Tick the sensitive data checkboxes based on what you know about their business. This determines whether compliance is an argument — and with health, finance, or government data it is a legal requirement, not a preference.
4
The tool instantly generates a hardware recommendation, payback period, 5-year cost comparison, and a pitch script with real numbers filled in. Read it directly in the meeting.
Why the billing statement is so powerful
It tells you everything instantly. A $4,200 Bedrock invoice means they processed roughly 200–400 million tokens last month. You know the tier immediately. No lengthy discovery required.
It surfaces the compliance issue visually. When you can point at a line item and say "this line here is your patient data being processed by an American company on American servers" — that lands harder than any abstract sovereignty argument.
It reframes the conversation to cost. The customer is already paying. GridMind is not an additional expense — it is a replacement with a payback period and a residual asset on the balance sheet.
Small spend = Starter conversation. Large spend = Enterprise conversation. A company spending $500/month on ChatGPT is a $25K Starter deployment. A company spending $15,000/month on Bedrock is a $700K H100 conversation. You qualify the deal size in 30 seconds.
Enter monthly AI spend by provider
Monthly spend (AUD)
OpenAI (ChatGPT Enterprise / API)
Monthly spend $AUD
Primary model
AWS Bedrock (Claude / Titan / Llama)
Monthly spend $AUD
Primary model
Azure OpenAI / Microsoft 365 Copilot
Monthly spend $AUD
Type
Google Vertex AI / Gemini
Monthly spend $AUD
Primary model
Other (GPU rental / other API)
Total monthly spend
$0
Est. tokens/month
—
Annual AI spend
$0
Does any of this spend involve sensitive data?
Enter monthly spend figures on the left to see the hardware recommendation and payback analysis.
How to use this tool
1
Ask: "How many people work at the company, and what do they mostly do?" You don't need exact numbers — a rough split is enough.
2
Enter headcount by role type below. The tool estimates token volume, AWS equivalent cost, and recommends the correct GridMind tier.
3
Tick any sensitive data categories. Clinical or government flags automatically adjust the tier recommendation and compliance language in the pitch.
Why this works
No AI spend yet doesn't mean no AI risk. Staff may be using personal ChatGPT — company data in consumer tools with no audit trail.
On-premises removes the per-token barrier. Usage typically increases 3–5× when the token counter disappears.
Start small, scale with GSOL. Starter or Pro node, sized for current staff. GSOL earns idle revenue while they build AI capability.