N08 THE REALITY LAYER
Why Power Is the First Product in AI
Every token begins upstream, with electrons that can reach the right rack at the right time.
IN THIS NOTE · MARCH 2025
A buyer asks for compute. An operator hears a chain of questions about energy, interconnection, voltage, redundancy, cooling and construction. The chip is central, but power determines whether it can become a service.
Generation is not deliverability
A region can have abundant energy while a specific site cannot secure timely interconnection. Nameplate generation, contracted supply and data-center-ready power are different states. Transformers, switchgear, substations, permits and network upgrades each carry their own schedule.
This makes geography a product decision. The cheapest theoretical megawatt is irrelevant if it cannot support the required density, reliability or expansion path.
Density changes the facility
Rack-scale AI systems concentrate heat and electrical load. That pushes cooling, distribution and service design into the compute conversation. Power and cooling must be planned with the hardware rather than added after procurement.
The result is a market where facility readiness can be more scarce than accelerators. Operators able to coordinate both can offer a more believable delivery date.
Power truth is commercial truth
Customers should ask whether power is reserved, contracted, energized and usable for the promised configuration. Providers should explain dependencies and confirm schedules in agreements. Precision builds more confidence than an enormous aggregate number.
Power is the first product because every promise downstream inherits its constraints.
Global abundance can still mean local scarcity
The International Energy Agency estimated data centres used roughly 415 terawatt-hours in 2024 and projects about 945 terawatt-hours in its 2030 base case. The global share remains modest relative to total electricity demand, but data centres are concentrated, and grids are local. A national generation surplus says little about whether a specific substation can serve a high-density site on the customer's schedule. The product bottleneck appears at the point of delivery.
That local view changes diligence. A credible capacity plan names the utility process, firm and non-firm supply, required network upgrades, substation and transformer status, backup design, commissioning sequence and assumptions about rack density. It also distinguishes a commercial agreement from the physical ability to energize. Each item may be normal on its own; together they determine whether the accelerator order becomes a customer service before the hardware and market move on.
Flexibility is a capacity source
The first response to scarcity is often to secure more megawatts. The second should be to make demand more shapeable. Training, fine-tuning, batch inference and interactive inference have different tolerances for interruption, location and time. Schedulers, storage, model checkpoints and commercial terms can move some work around grid constraints without pretending every workload is flexible. A portfolio of workloads may therefore support more useful compute than a single undifferentiated demand profile.
This flexibility must be sold honestly. A customer promised continuous dedicated capacity should not discover that reliability depends on undisclosed curtailment. But where workloads permit it, explicit tiers can align price with power conditions and create value from capacity that would otherwise be stranded. The operator's advantage is the ability to translate electrical states into compute products with clear behavior, rather than treating power as a binary input hidden behind the rack.
- Separate generation, contracted supply and energized site capacity.
- Model rack density and cooling before committing a deployment date.
- Treat interconnection and equipment lead times as product-roadmap inputs.
I would revise this view if compute availability became decoupled from local power deliverability and facility readiness.
Primary and institutional sources used as the grounding layer. Interpretation and synthesis are Luca's.
01