UK Power Networks publishes half-hourly smart-meter data for the data centers on its grid. Across 96 sites from January 2023 to April 2026, they drew just over 20% of the capacity they had booked, and the largest sites, the hyperscale campuses and new AI builds, drew about 12% (Computer Weekly, 2026). Part of that gap is capacity reserved for GPUs still on order. It still shows how far booked power runs ahead of what the equipment actually draws.
GPUs are hungry, so I assumed AI racks would be the exception. In the AI build-outs we work on, they are not. The GPU racks hit their circuit limit with most of the rack still empty, and the power that is left over is too small for a second GPU server.
For an operator serving neoclouds and physical AI startups, that leftover power and space is the cheapest place in the building to put the thing every one of those tenants needs next to their GPUs: storage.
In this post I walk through the numbers, a reference design on Lenovo, Celestica and Seagate hardware, and how it turns into new revenue per rack.
What is stranded power in a data center?
Stranded power is capacity that has been built and provisioned, and usually paid for, but that no equipment is drawing. It shows up at the utility feed, the UPS, the PDU and the individual rack circuit.
The industry has been open about how much of it there is:
- UK grid data from 96 data centers shows average use of just over 20% of booked capacity between January 2023 and April 2026 (UK Power Networks via Computer Weekly, 2026).
- Meta reported in May 2026 that in its own 100 MW AI cluster, uneven load across power distribution alone can strand up to 10% of the power (arXiv, 2026).
- Uptime Institute reported in December 2025 that poor utilization can leave 10% or more of a facility's UPS capacity stranded.
- [Placeholder for operator interviews: share of contracted power actually drawn in GPU cabinets, with a named quote.]
In AI deployments a lot of that stranded capacity sits inside the GPU rack.
How many GPU servers fit in one rack?
In most air-cooled colocation today, the answer is one.
An 8-GPU server draws a lot of power. NVIDIA rates the DGX H100 at about 10.2 kW and the DGX B200 at about 14.3 kW (IntuitionLabs). For the B300 generation, Lenovo's configurator puts a ThinkSystem SR680a V4 with an NVIDIA HGX B300 board at about 12.3 kW per node.
A common high-density colo feed is 60A at 208V 3-phase. At the usual 80% continuous load that gives about 17.3 kW per rack, with the B feed held in reserve for redundancy (Leviathan Systems). Two B300 servers need close to 25 kW. One fits, two don't.
So the rack gets one GPU server. It is 8U tall, which leaves 30U or more of a 42U or 48U rack empty, plus about 5 kW of contracted power the GPU server won't use at its rated draw. On 415V feeds the kW figures are higher, and the same thing happens: one or two servers use up the circuit long before they use up the rack.
Measured draw tells the same story. Microsoft's 2024 study of its own GPU fleet found inference clusters peaked at 79% of their provisioned power, while training clusters ran close to the limit at 97%. Plan on contracted kW, not on measured draw. A storage block at a steady 1.65 kW fits inside the contract either way.
Why infra operators serving neoclouds and physical AI should care
Every one of these tenants needs storage close to the GPUs.
Neoclouds sell GPU hours, and their customers bring datasets, checkpoints and model weights that have to live somewhere. Physical AI startups push this further. One robotics team we work with collects up to 70 TB of video a week from its data collection devices. That is more than 3.5 PB a year from one customer, and it has to land next to the GPUs that train on it.
Today that storage usually ends up in a separate storage cabinet with its own circuit, PDUs and floor space, which is slow to provision. Or it sits in a public cloud and gets pulled across the WAN every time a job needs it, with egress fees and latencies on every pull. The GPU racks themselves rarely enter the conversation, even though they already have the power and the space.
What fits in the space a GPU server leaves behind?
A storage block is a 2U server attached to a 4U JBOD full of hard drives. This is the Akave reference block:
That is 2.59 PB in 6U at about 1.65 kW, well inside the 5 kW a single B300 server leaves behind.
A reference design: three GPU racks, one storage pool
Take three GPU racks built to Lenovo's B300 design, each on a 60A 208V 3-phase circuit, and add one storage block to each. Lenovo's design puts the 800G leaf switch in one of the GPU racks (about 15.1 kW with the switch); here we move it to a network rack so all three GPU racks have the same headroom.
With Akave Cloud the three blocks run as one pool, with erasure coding spread across the three racks. That gives 7.8 PB raw and 5.2 PB usable, and the pool keeps serving data if an entire rack goes down. Each rack still has 3.3 kW spare after its block, room for a second block per rack as the data grows.
How the operator earns from it
There are two ways this works, depending on who holds the rack.
If you run the GPU racks yourself and sell GPU capacity, the headroom is yours. You own the storage blocks and sell capacity to the customers already renting your GPUs.
If a tenant holds the cabinet, storage becomes a service you offer that tenant inside the power they already contract. They get petabytes next to their GPUs without signing for another cabinet, and you get a new line on their invoice. Agree up front how the block's 1.65 kW is credited or passed through, and when your remote hands can open the cabinet.
Either way the cost side is short: the storage hardware, the software and management to run it, and about 1.65 kW of power per block. You do not build or lease another cabinet for it. We will not publish hardware pricing here because it depends on your market and terms, but we have modeled this three-rack design across a range of storage prices and sell-through rates. The design brings in well over $100K a year in incremental storage revenue at conservative prices. Once the pool is more than 85% sold, the operator clears six figures a year after the Akave software fee and hardware depreciation. For scale, a single physical AI tenant collecting 70 TB a week would get the pool there in about 15 months.
To size it for your own floor, you need three answers:
- How many of your GPU racks have 3 kW or more of contracted headroom?
- What do your tenants pay for storage today, and where does it sit?
- How fast would they fill 5 PB once it is next to their GPUs?
It also makes the floor easier to run
GPU load swings with the job. A training run pulls the circuit hard and inference comes and goes. Storage draws a steady 1.65 kW per block around the clock, which puts the headroom you already provisioned to work without new grid commitments.
It also changes the tenant relationship. When a tenant's data sits on the same floor as its GPUs, training jobs read it over the local network and nothing leaves the building, so there is no egress bill. Moving petabytes to another provider takes weeks and real money, which makes a tenant whose data lives with you far more likely to renew its compute with you too.
What to check before you put storage in a GPU rack
This works for most air-cooled GPU deployments. Before you commit, check:
- Measured draw. Lenovo's numbers are configurator estimates. Meter your actual peak, especially for tenants who train around the clock.
- Start-up peak. A full JBOD spins up over a hundred drives at power-on. Use staggered spin-up and leave room on the circuit for that moment.
- Vibration and inlet temperature. Mount the JBOD at the bottom of the rack, away from the GPU fan wall, and confirm drive inlet temperature stays within Seagate's spec.
- Weight. A full block weighs around 170 kg. Check the rack and floor rating together with the GPU server.
- Network. Each storage server needs two 25GbE ports to the leaf or top-of-rack switch. It does not touch the GPU fabric.
- Operations. Akave monitors and runs the storage remotely. Drive swaps go through your remote hands, a handful of times a year across a three-rack pool.
- Who owns what. Put it in writing: Akave runs the storage software and the pool, you own power, space and physical access, and the tenant contract says who can touch the cabinet.
- Liquid-cooled racks. A 100 kW+ GB300-class rack is full by design. This is for air-cooled 8-GPU servers, which is still most of what lands in colocation today.
How we think about it at Akave
At Akave we build object storage for AI infrastructure, and most of our time is spent inside deployments like these. Storage almost always gets planned last. The GPUs get ordered and powered, and then someone asks where the data goes. By that point the obvious answer, the headroom already sitting in the GPU racks, has usually been missed.
Akave Cloud runs on standard servers and JBODs like the block above. It is S3-compatible, so tenants point their existing tools at it, and there are no egress fees between the storage and the GPUs beside it. We operate the storage layer remotely, 24x7, so you do not need to hire a storage team to offer it. Akave also keeps a verifiable, tamper-evident record for every object, so a tenant can prove where its training data came from and that it has not changed. Physical AI teams working with real-world video and sensor data are starting to get asked exactly that.
We deploy rack by rack. Start with three GPU racks, prove the numbers on your own floor, and add blocks as tenants fill them.
Plan storage into the build-out
Every air-cooled GPU rack going in today is likely to ship with a few kilowatts and thirty-odd rack units that nobody uses, and that power is already built and billed. If you want to see what storage in that space looks like across your GPU footprint, talk to the Akave team. We will walk your racks with you and size it.
In the next post we look at the other half of the physical AI problem: why moving that much data over the network breaks down, and what that means for where storage has to live.
Sources
- Computer Weekly, "Data dive: Power grid data shows birth of AI in UK datacentres", 2026 (UK Power Networks smart-meter data, Jan 2023 to Apr 2026)
- Meta, "Provisioning to Runtime Optimization of a 100 MW-Scale AI Cluster", arXiv, May 2026
- The Register, "Datacenters are hoarding grid power just in case, says Uptime Institute", December 2025
- Microsoft Research, "Characterizing Power Management Opportunities for LLMs in the Cloud", ASPLOS 2024
- IntuitionLabs, "NVIDIA HGX Platform: Data Center Physical Requirements Guide"
- Leviathan Systems, "Power Requirements for GPU Racks"
- Lenovo Data Center Solution Configurator estimates for ThinkSystem SR680a V4 with NVIDIA HGX B300