In conversations about AI compute, chips have long taken centre stage: whoever secures more and newer GPUs holds the advantage. Our view at Qovane is that power is replacing chips as the first constraint on AI compute expansion.
From "no cards" to "no connection"
A modern AI data center draws far more power than a traditional facility. Dense GPU clusters consume large amounts of electricity, and the cooling that keeps them running needs plenty more. When a large compute site is planned, the first question is often not where the chips will come from, but where the power will come from, whether the grid can carry it, and how long the connection will take.
In many mature markets, grid upgrades involve long approval and construction cycles, and the best sites and power capacity have already been spoken for. Even with chips in hand, capacity can sit idle while it waits for a connection.
Power prices set compute prices
Power decides not only whether compute comes online, but at what price it is delivered. Over a GPU's working life, electricity is the most persistent operating cost. For long training runs and always-on inference services, differences in power prices compound hour by hour and end up in the price of every GPU-hour.
That is why compute capacity is moving toward regions with abundant, lower-cost power:
- Energy-rich regions offer lower generation and consumption costs.
- Emerging markets often have more room to expand land and power capacity.
- Favourable climates reduce the energy spent on cooling.
Distributed supply is the more resilient answer
Concentrating compute in one region concentrates power, policy and operational risk with it. We believe the more resilient approach is a distributed supply network that sources capacity where power is cost-effective and schedules it centrally, so customers can use it as a single service.
The value of this model:
- Cost: draw first on capacity in lower-cost power regions to bring unit prices down.
- Resilience: power or policy disruption in one region doesn't interrupt overall supply.
- Elasticity: capacity in different regions complements each other and smooths demand peaks.
How Qovane puts this into practice
Headquartered in New York, Qovane's compute supply network spans data center partners in Libya in North Africa, Xinjiang in China and the United States. We favour regions with an advantage in power and land costs and improving network links, and allocate capacity across them through our own scheduling system. Every partnership must clear restricted-party screening and export-control review.
As demand for compute keeps growing, what becomes truly scarce is usable, affordable, reliable power. Whoever can turn a power advantage into deliverable compute services will hold the edge in the next phase of competition.