Most storage design conversations start with cost per terabyte and capacity growth. That is the right pair of questions when floor space and power are effectively unlimited and only the budget binds.
In the Australian market in 2026, they are not. CBRE reported Sydney data centre vacancy at 4.5 per cent in the first quarter of 2026, with availability across major Asia Pacific markets down 43 per cent year on year and Sydney asking rents holding around US$188 per kW per month. CBRE also flagged the specific reasons Sydney supply is slow to arrive: power procurement timelines and planning approvals constraining new development, and continuing supply chain bottlenecks for transformers and switchgear stretching construction completion.
When you cannot simply take another cabinet in a colocation hall, and the contract you signed is denominated in kilowatts, storage design changes shape. The scarce resource is not money. It is rack units and watts, and every watt a storage tier consumes is a watt you cannot give to compute.
Storage now competes with compute for the same power envelope
The reason the envelope has tightened is on the compute side. Gartner analysis reported in Australia has data centre GPU deployments moving from roughly 10 to 26 kW per rack in 2020 to around 150 kW currently, with some designs heading toward 600 kW by 2027. CBRE made the same point from the other direction when describing Hong Kong: most existing facilities there were designed for 5 to 15 kW per rack, which is insufficient for the 40 kW-plus densities AI workloads require.
Legacy halls were built for a density profile that no longer matches the equipment going into them. That produces the awkward situation many operators are now in: floor space available, power already committed. A cabinet you cannot energise is not capacity.
Storage feels the squeeze indirectly. If a GPU pool takes the power, the storage that feeds it has to fit into whatever remains. So the sizing question becomes: how much usable capacity can I land per rack unit, and what does it cost me in watts to keep it online?
Design in two units before you look at price
Terabytes per rack unit answers the space constraint. It sets how much of a fixed cabinet your capacity target consumes, and therefore how much room is left for the compute and networking that justify the site.
Watts per terabyte answers the power constraint. Take the drive’s rated draw, divide by its capacity, and do it separately for active and idle states, because a bulk tier spends most of its life idle and a hot tier does not.
Published specifications have moved these numbers a long way. Solidigm’s 122 TB D5-P5336 is rated at 25 W maximum and under 5 W idle. That is roughly 0.2 W per terabyte at full draw and under 0.04 W per terabyte at idle. The vendor puts the achievable figure at up to 4 petabytes per rack unit, states the drive delivers 3.4 times more terabytes per watt than 30 TB TLC alternatives, and claims 84 per cent lower energy consumption than a comparable hard disk plus TLC hybrid arrangement. Treat the comparative figures as vendor claims and validate them against your own workload. Treat the rated wattage and capacity as what they are: the two numbers your power budget actually needs.
The point is not that one product wins. It is that a capacity plan built on a three-year-old terabytes-per-rack-unit assumption will overbuy both space and power, and in a 4.5 per cent vacancy market that error is expensive and slow to correct.
What this does to tiering decisions
Classical tiering optimises for access latency and cost: hot data on fast media, warm on cheaper media, cold on the cheapest available. Under a power cap, add two more axes.
Idle draw per terabyte. A cold tier is defined by how little it costs to hold data you are not reading. If the media draws meaningful power while idle, a cold tier is not actually cold, it is just slow. Compare idle watts per terabyte across candidate tiers before comparing capital cost.
Recall behaviour. Media with spin-up, staging or rehydration costs converts a read into a power spike and a delay. That is acceptable for an archive with a monthly restore test. It is not acceptable for a compliance dataset your regulator can ask you to produce within a defined window. Set the tier by the retrieval obligation, not by the age of the file.
Three practical consequences follow.
Fewer, larger tiers usually beat many small ones, because each tier carries a fixed overhead in controllers, enclosures and network ports, and those consume power regardless of how full the tier is.
Consolidation is often the cheapest capacity increase available. Replacing a densely populated low-capacity tier with fewer, higher-capacity drives can return both rack units and watts inside an existing contract, with no new procurement and no waiting on a transformer.
Deletion has a power value. Retention policy is usually treated as a compliance artefact. Under a power cap it is a capacity lever, and it is the only one that costs nothing to pull.
Rules that survive a power-constrained site
- Size to the power contract first, then to the capacity forecast. Write the kW budget for storage at the top of the plan and treat it as fixed. If the capacity forecast does not fit inside it, the density assumption has to change, not the envelope.
- Reserve headroom for the compute you have not bought yet. Rack density expectations have shifted repeatedly over five years. A storage design that consumes the entire remaining envelope leaves nowhere for the next accelerator generation.
- Measure at the power distribution unit, not from the specification sheet. Rated draw is a ceiling. Real draw depends on workload mix, ambient conditions and how full the drives are. Instrument it before the next refresh, so the next design starts from measurement.
- Count the supporting equipment. Controllers, switches, enclosures, fans and cooling overhead all sit inside the same envelope. A storage tier’s power number is the whole shelf, not just the media.
- Assume lead times, not availability. With transformer and switchgear supply constrained, assume a long interval between deciding you need more power and having it. Design the current build so it can absorb one more growth cycle without a new power request.
The sovereignty layer sits on top of all of it
Density and power decisions determine how much data you can hold in a given facility. Jurisdiction determines which law reaches it once it is there. Both are properties of the site, and both are usually decided in the same procurement, which makes it worth settling them together rather than sequentially.
Amaze operates Australian regions in Sydney (ap-syd-2) and Melbourne (ap-mel-1) as an Australian company under Australian law, with no foreign parent entity and no CLOUD Act exposure, and is ISO 27001 certified. Pricing is AUD-denominated, which matters more than it sounds when the alternative is a capacity contract quoted per kW in a foreign currency across a five-year term.
Related reading: How sovereign data centres support AI compliance and What is sovereign AI? A guide for Australian businesses.
Frequently asked questions
What is a realistic storage density per rack unit in 2026? It depends entirely on media choice and enclosure design, so use published drive specifications rather than a rule of thumb. As an upper marker, one vendor states up to 4 petabytes per rack unit using 122 TB QLC drives. Work out your own figure from the drives you can actually procure, and recheck it every budget cycle because the numbers keep moving.
How do I calculate watts per terabyte? Divide the drive’s rated power by its usable capacity, and do it twice: once at maximum draw and once at idle. A drive rated at 25 W with 122 TB of capacity is roughly 0.2 W per TB at full draw. Then add the shelf overhead, because controllers, fans and networking draw power whether the media is busy or not.
Why is data centre capacity so tight in Sydney? CBRE reported a 4.5 per cent vacancy rate in Q1 2026 and named the causes: power procurement timelines and planning approvals slowing new development, and supply chain bottlenecks for transformers and switchgear extending construction schedules. Strong demand from AI and neocloud deployments is absorbing new supply as it lands.
Should I move cold data to tape when power is the constraint? It can be the right answer for genuine archives, because idle power per terabyte is very low. Check the retrieval obligation first. If a regulator or a contract requires production within a fixed window, or if the dataset feeds model retraining, the recall time and staging cost usually rule it out regardless of how good the power figure looks.