The real constraint on AI compute is power, not high-bandwidth memory. That is the message from a company identified as an Nvidia partner, in remarks reported by the Seoul Economic Daily. The claim cuts against the industry's dominant narrative, which has treated HBM supply as the scarce resource of the AI buildout.

For two years, the memory stack has absorbed most of the blame for hardware shortages. HBM capacity has dictated how many accelerators ship. That framing now looks incomplete, according to the partner's assessment.

What does this change for AI hardware planning?

It moves the bottleneck one layer down the stack. If power delivery, not memory supply, sets the ceiling on deployment, then the binding limit sits in the data centre itself: grid capacity, cooling budgets and rack-level electricity allocation. Chips can exist on a roadmap and still stay undeployable.

The argument is structural, not rhetorical.

• HBM is a component problem. Suppliers can add capacity over time.
• Power is an infrastructure problem. Grid connections and generation take years to commission.
• A fab can respond to demand in quarters. A utility cannot.

That asymmetry matters for anyone forecasting AI capacity. Component shortages resolve on silicon timelines. Electricity shortages resolve on civil-engineering timelines.

Why the memory framing dominated until now

HBM sat at the centre of the story for a simple reason. Every accelerator shipped with a fixed memory allocation, and memory makers could not keep pace with Nvidia's order book. The shortage was visible, measurable and easy to attribute.

Power scarcity looks different. It shows up as stalled data centre projects, capped rack densities and regions that cannot host new sites at all. The symptom is delay, not absence. That makes it harder to track in quarterly component data.

The partner's statement reframes where buyers should look for risk. The question is no longer only "can memory vendors supply enough stacks?" It is also "can the site take the load?"

What follows from the claim

If power is the binding constraint, several planning assumptions shift:

• Procurement teams gain less from bidding early on memory allocation than from securing energy contracts.
• Efficiency per watt becomes a purchase criterion on par with raw throughput.
• Regions with surplus generation gain a durable siting advantage.

The statement comes from a single partner, and it reflects one player's reading of the market. It does not settle the question of whether HBM supply concerns have faded. It does reframe them as secondary.

The Seoul Economic Daily report gives the argument a specific weight because of who is making it. A Nvidia partner sits close to the demand signal. Partners see order patterns, deployment schedules and the reasons deployments slip. When a partner names power rather than memory, that reflects observations from the field.

For data centre operators, accelerator buyers and grid planners, the practical takeaway is narrow. Watch megawatts, not just memory gigabytes. The next constraint on AI capacity growth may already be wired into the wall.