N04 THE REALITY LAYER
The Difference Between Compute Capacity and Usable Capacity
GPU count is an input. The customer buys a workload that runs.
IN THIS NOTE · DECEMBER 2024
The market loves counting GPUs because the number is simple. Customers experience something else: clusters, queues, failures, data movement, software environments and delivery dates.
A capacity ladder
Capacity moves through distinct states. Announced capacity may describe intent. Reserved capacity may reflect procurement or site planning. Delivered hardware may still be waiting for power, networking or commissioning. Installed systems may not have passed acceptance. Only then does customer-ready capacity emerge.
Collapsing those states is not merely imprecise. It destroys planning. A buyer with a product launch cannot hedge an operational dependency using a press release.
The cluster is the unit
A modern AI workload depends on topology, interconnect, storage throughput, scheduler behavior and the ability to recover from component failure. The nominal accelerator can be excellent while the system underperforms. That is why measured workload behavior matters more than a component list.
Usable capacity also includes organizational readiness. Someone must own security, observability, support, maintenance windows and escalation.
Sell the acceptance test
The cleanest commercial language defines what the customer will be able to do, by when, and how both parties will verify it. That can include delivered topology, software image, benchmark suite, burn-in period, service levels and remediation path.
The more scarce the market appears, the more tempting it is to sell ambiguity. Durable operators do the opposite.
Write a capacity acceptance matrix
Usable capacity should be tested against the workload it is meant to serve. The matrix starts with physical facts: accelerator model, quantity, power state, cooling envelope and network fabric. It then adds system behavior: collective performance, storage throughput, job-start latency, fault recovery, observability and software compatibility. Finally it adds operational facts: access controls, support ownership, maintenance policy and evidence that the cluster can sustain load rather than merely complete a demonstration.
This matrix prevents a familiar category error in which a component benchmark becomes a service promise. A rack may pass a vendor diagnostic while a distributed training job stalls on network congestion. A cluster may deliver peak throughput while repeated node failures destroy useful utilization. Acceptance therefore needs both capability and endurance. The relevant question is not whether the hardware exists, but whether the complete system can repeatedly execute the customer's defined work inside agreed limits.
Availability has a denominator
Providers often describe capacity with a numerator and omit the denominator. Ten thousand GPUs sounds precise, but over what period, in which topology, with what reservation status and after which exclusions? Customer planning needs a time-bounded denominator: accelerator-hours offered, accelerator-hours schedulable and accelerator-hours that completed accepted work. The gaps between them expose commissioning loss, maintenance, fragmentation, queueing and application-level failure.
The same discipline improves commercial conversations. A buyer can separate immediately usable clusters from future pipeline, then value each according to start-date certainty and switching cost. A provider can stop treating every installed device as equivalent revenue inventory. Capacity becomes legible when its state, configuration and time window are explicit. That makes fewer numbers look impressive, but it makes the remaining numbers useful for engineering, finance and procurement.
- Ask whether capacity is ordered, delivered, commissioned or accepted.
- Evaluate the cluster and operating layer, not the chip in isolation.
- Make the workload acceptance test part of the commercial offer.
I would change my view if nominal GPU inventory proved to be a reliable proxy for customer-ready performance across providers.
Primary and institutional sources used as the grounding layer. Interpretation and synthesis are Luca's.
01