Optimized Rack Space (Tiered)¶
Service ownership
Owner: dc-operations (colo-pm@clouddigit.ai) — Status: GA — Last audited: 2026-05-11
Density-tiered rack pricing for HPC, GPU, or other high-power-density footprints.
Why tiered¶
A standard 5 kW full rack works for normal workloads. GPU-dense racks (16x H100, GPU bare metal pods) easily pull 30+ kW. Pricing per RU stops being meaningful — what matters is delivered kW and matching cooling. Tiered pricing aligns to that.
Tiers¶
| Tier | Power per rack | Cooling | Best for |
|---|---|---|---|
| T1 | up to 8 kW | Standard CRAC | Standard production (use racks.md instead) |
| T2 | 8–15 kW | In-row cooling | Mid-density compute |
| T3 | 15–25 kW | In-row + chimney containment | GPU compute, dense storage |
| T4 | 25–40 kW | Rear-door heat exchanger | GPU bare metal pods |
| T5 | 40+ kW | Direct liquid cooling (rear or DLC) | High-density training clusters |
Pricing model¶
Tier-based per-rack-month rate; commitment plans available.
T4 / T5 require a feasibility check (cooling, distribution, ground loading) before contract — open a ticket with the planned SKUs (e.g., "8x bm-gpu-h100-8 over 4 racks") and we'll respond in ≤ 5 BWD with capacity availability per region.
Region availability¶
| Region | T1–T3 | T4 | T5 |
|---|---|---|---|
bd-dha-1 | ✓ | ✓ | ✓ |
bd-ctg-1 | ✓ | ✓ | — |
bd-syl-1 | ✓ | — | — |
Related¶
Operate this service¶
Pre-engineered rack templates optimized for specific workloads — lower BDT per RU than traditional colo for predictable footprints.
Tiers¶
| Tier | RU + Power | Use case |
|---|---|---|
opt-storage | 42U / 8 kW | Storage-heavy (JBOD arrays) |
opt-compute | 42U / 14 kW | Compute-dense (1U/2U servers) |
opt-gpu | 24U / 30 kW | GPU racks; aisle with extra cooling |
opt-network | 42U / 5 kW | Network gear (switches, routers) |
Each tier has fixed physical/cooling/power baseline — you fit your workload to the tier, not the other way.
When this fits over traditional racks¶
- Your workload matches one of the tier profiles
- You want to scale fleet predictably (10 storage racks → 10× the tier)
- You don't need bespoke power/cooling
For irregular workloads, traditional racks give more flexibility.
IAM¶
Same as traditional racks — colo.* roles.
Pricing¶
~15-25% lower BDT per RU than equivalent traditional rack. The savings come from CD's standardized engineering.
Provisioning lead time¶
- Storage tier: 7 BWD (rack ready, you ship gear)
- Compute tier: 7 BWD
- GPU tier: 14 BWD (specialized cooling provisioning)
- Network tier: 7 BWD
Capacity reservation¶
For multi-rack deployments (3+ racks of the same tier in the same DC): pre-reserve to lock pricing and ensure capacity:
bash cd colo opt-rack reserve --tier opt-compute --quantity 10 --dc bd-dha-1 --term 24-months
Related¶
Metrics¶
Same shape as traditional racks (colo.rack.*) — temp, humidity, power, uplinks.
Additional metrics per tier:
| Tier | Tier-specific metrics |
|---|---|
opt-storage | colo.disks.health_pct, colo.disks.smart_warnings (when integrated) |
opt-compute | Standard compute envelope |
opt-gpu | colo.aisle.gpu_temp_c (separate alert tier) |
opt-network | colo.network.switch_uplinks_state |
Tier-fit operations¶
The platform monitors whether your actual usage fits the tier:
bash cd colo opt-rack fit-report --rack-id <id>
If you're running 14kW gear in an 8kW storage tier → migrate to compute tier (or accept overage charges).
Capacity planning¶
For fleet operators (10+ racks of the same tier):
- Standardize on the tier; refuse non-standard gear that doesn't fit
- Auto-deploy via golden images and rack-build SOPs
- Track per-rack utilization; rebalance proactively
Hardware refresh¶
Old gear → new gear within the same rack: schedule a 4-8h maintenance window. Smart-hands assists with swap; uplinks stay live across the swap.
For tier upgrades (compute → GPU): typically requires a new rack (cooling/power difference).
Related¶
Tier-fit violation¶
WARN: Rack opt-storage-bd-dha-1-r042 power usage 9.4 kW, tier limit 8 kW
You've exceeded the tier envelope: - Migrate gear to a higher-tier rack (compute or GPU) - Accept overage charges (~2× per-kW over tier) - Reduce workload
CD won't trip the breaker (you bought the tier, not strict enforcement) but the overage bills.
Standardized image deployment broken¶
For fleet operators using golden images: - Image build pipeline failed (re-run) - Network boot path broken (BMC config drift) - New hardware revision incompatible with image
Diagnose per-rack: cd colo opt-rack provision-log --rack-id <id>.
GPU tier cooling alert¶
colo.aisle.gpu_temp_c rising: - A GPU's fan failed (smart-hands swap) - Cold-aisle door left open - Aisle CRAC unit degraded (CD will repair)
GPU tier has tighter thermal envelope; alerts faster than standard tiers.
Capacity reservation not honored¶
You reserved 10 racks; only 7 available when ready: - Construction / delivery delays at the DC - A previous customer over-ran on their reservation
CD reservation contracts have penalty clauses for missed capacity. Escalate to your CE.
Migration between tiers¶
You ordered the wrong tier: - Same-DC migration: 4-8 h smart-hands operation; bills as migration service - Cross-DC migration: weeks; effectively a new install
Plan tier carefully before the first install.