From the GPU Race to the Power Wars: Why Electricity Is the New Battleground for Neoclouds
In 2023, the question every AI infrastructure buyer asked was how many H100s they could get. In 2026, the question has changed: who can actually turn them on. GPU allocation is no longer the scarce input. Power is.
CoreWeave co-founder Brannin McBee has been blunt about it: the binding constraint is now the "powered shell," not chips or HBM. That single reframe changes how neoclouds raise capital, pick sites, hire crews, and compete on price.
The Numbers Behind the Power Wars
The scoreboard is now measured in megawatts. CoreWeave is running roughly 1.5 GW of active power while holding contracted capacity above 4 GW, with a stated 2030 ambition near 8 GW. Nebius has pushed contracted power past 3.5 GW as well. SoftBank's SB Neo is aiming at 10 GW with on-site generation.
The macro picture explains the urgency. Morgan Stanley estimates a global data center power gap in the neighborhood of 55 GW, and the IEA projects data center electricity consumption will more than double by 2030, approaching 945 TWh. Jensen Huang put it plainly this October: capital, not demand, is what is slowing neoclouds down.
Why a Powered Shell Is Harder Than a GPU Rack
A powered shell is not an empty building. It is a facility with completed substations, switchgear, distribution, and cooling, ready to accept load. Transformers and high-voltage gear carry one- to two-year lead times, and certified electricians and high-voltage crews are in short supply everywhere that matters.
That creates a timing mismatch the industry is only now pricing in. GPUs ship in weeks. A powered shell ships in quarters. Meanwhile rack densities keep climbing: an NVIDIA GB200 NVL72 rack draws about 120 kW, and next-generation designs push well beyond that. Traditional air cooling tops out near 20 kW per rack, which makes liquid cooling a requirement rather than an option.
The Three Layers of the New Power Moat
The advantage now compounds across three layers. First, grid access and site selection: proximity to a substation, a favorable PPA, and a fast interconnection queue. Second, on-site power architecture, including the shift toward 800 VDC distribution to cut conversion losses between the meter and the accelerator. Third, thermal engineering, where direct-to-chip liquid cooling, rear-door heat exchangers, and immersion are the only credible paths above roughly 30 kW per rack.
This is also where strategies diverge. Nebius is buying land and generation and building itself. CoreWeave is moving from colocation into self-build to control delivery dates. Bitcoin miners converted to AI compute are monetizing something no one valued two years ago: an existing grid connection.
The Fit-Out Gap Is the Real Execution Risk
The unglamorous truth is that commissioned power and deployed compute rarely arrive on the same day. Busway, intelligent PDUs, CDUs, and coolant manifolds have to be specified before the GPUs land, or expensive silicon idles in a warehouse. Vendors that handle electrical fit-out, liquid cooling deployment, load-bank testing, and phased commissioning as a single scoped engagement are becoming the connective tissue between a signed power contract and revenue. ByteBridge's data center solutionspractice, spanning electrical fit-out, thermal management, and global delivery, sits squarely in this gap.
Megawatts Beat GPUs
At roughly $35 billion of capex per gigawatt, power is now the largest line on the balance sheet, and cost per kilowatt-hour is a margin driver, not a utility bill. Neoclouds that treat power as a first-class engineering discipline will set the terms for the next decade. The era began with a GPU shortage. It will be won in a power.