Define usable AI capacity before accepting delivery

An AI infrastructure project can report progress in megawatts, racks or delivered accelerators. Each measure answers a real question. None, on its own, establishes that the intended service is ready to operate under its agreed conditions.

For procurement and delivery discussions, Watts & Workloads proposes a practical definition: usable AI capacity is the configured infrastructure that has demonstrated an agreed workload within its approved operating envelope, with the evidence and responsibilities required for production handover.

This is an editorial working definition. It is not an OCP certification, a benchmark or a universal contractual test. Its value is to force an early agreement about what the project will actually accept.

Specify the service being accepted

Start with the intended workload and system configuration. A training cluster, an inference service and a mixed-use installation can have different acceptance objectives. The customer should identify the relevant workload, software version, performance measure, test conditions and constraints with the parties responsible for delivery.

Then connect that service description to the facility. Which electrical, cooling, network and physical assumptions must remain valid? Which conditions are demonstrated at handover, and which are design assumptions awaiting later verification?

OCP's Open Data Center specification, revision 0.7.0, effective August 27, 2026, provides design guidance covering structural, power and cooling infrastructure for the 2027–2028 timeframe. Its forward-looking design scope should remain distinct from evidence that a particular installation is ready today. OCP design specification

Connect the evidence across systems

Facility commissioning and computing validation generate different evidence. A proposed acceptance plan should explain how those records support the same handover decision and where they leave gaps.

For example, a project could pass an agreed facility test and still need to establish how its selected computing configuration behaves under representative operation. Equally, a successful workload run under one condition does not establish performance under every maintenance state or future hardware configuration. These are reasons to define the scope of a result precisely.

The 2026 OCP programs published by NVIDIA and Vertiv both list a computing-equipment commissioning session on October 15. Vertiv describes the approach as complementary, later-stage validation following Level 5b. That is a useful question for the Summit: what additional evidence does computing-load testing provide, and what remains outside its scope? NVIDIA program · Vertiv program

Make a result reproducible

A handover statement is more useful when another qualified team can understand how the result was obtained. The proposed record should identify configuration, test procedure, measurement points, timestamps, observed results, exceptions and the authority accepting them. It should make clear which changes trigger review or retesting.

Telemetry access belongs in this discussion. OCP's data-center telemetry work identifies the challenge of sharing facility operational data securely across operator and tenant boundaries. Acceptance planning should specify the data needed and permitted access, rather than assume a dashboard exposes everything required. OCP telemetry workstream

Give conditional acceptance a boundary

Some projects may agree to a staged handover. If so, the record should identify the accepted configuration and service scope, remaining restrictions, accountable owners and evidence needed to close each exception. A conditional decision should be understandable to the operating team as well as the procurement team.

This makes comparisons between developers, equipment suppliers and operating partners more useful across markets. It also gives capital providers a clearer account of delivery status without treating technical readiness as a forecast of utilization or financial return.

The central question is what evidence allows the customer to accept the intended service. Agreeing on that question early gives each supplier a clearer target and gives the final handover a meaning everyone can inspect.

For Chinese readers: 智算通鉴 · 中文相关分析. This is related coverage, not a word-for-word translation.