OpenAI Adds Compute Units to Its SDK, but the Meter Has No Number Yet

OpenAI has placed a new line in the AI cost ledger, but it has not filled in the amount. Python SDK 3.6.0, released at 22:28 UTC on August 28, adds compute_units to the usage models returned by both Responses and Chat Completions. The underlying OpenAPI change defines the value as either a non-negative integer or null and repeats a decisive qualification: compute units are currently null when available. That makes this a forward interface rather than a working meter. The distinction is the lead development because a field name can quickly be mistaken for a measurable, invoice-ready unit.

The implementation is narrow and inspectable. The commit adds an optional integer to the stable ResponseUsage and CompletionUsage types, as well as the beta response type; tests are updated to expect None. The release also hardens X.509 workload-identity integration, but it does not publish a compute-unit definition, conversion formula, price, billing rule or guarantee of non-null coverage. PyPI uploaded the signed release artifacts about a minute later. Developers can safely accept the new property after upgrading, yet finance teams do not receive a new allocation key merely because the client type exists.

Why prepare a second usage dimension beside tokens? The sources do not answer that question, so any explanation must remain interpretation. Tokens describe input and output volume, while the resources consumed by a request can vary with model behavior, tools, latency targets and service architecture. A provider-defined compute measure could eventually make those differences more legible. It could also remain an internal capacity signal, a product-specific abstraction or a billing input whose meaning changes by endpoint. Until OpenAI publishes semantics and returns values, compute_units should be logged as nullable metadata, not used in budgets, chargeback or unit-economics forecasts.

Canada’s new national accounts show why the missing measurement matters beyond an SDK. Real GDP rose 0.8% quarter on quarter in the second quarter, equivalent to 3.3% at an annual rate. Business gross fixed capital formation increased 2.3%, and investment in computers and computer peripherals climbed 16.7%, mainly because of higher imports of processing units of the types used in data centres. This is unusually direct official evidence that data-centre hardware is visible in aggregate investment. It does not identify AI workloads, buyers, utilization or returns, so it should not be read as an AI revenue estimate.

The broader Canadian picture prevents a one-factor story. Exports increased 3.6%, their fastest pace since early 2023, household consumption rose 0.8% and final domestic demand advanced 1.0%. At the same time, inventory withdrawals subtracted 1.3 percentage points from GDP growth. The GDP deflator rose 2.5% in the quarter, the largest increase since the second quarter of 2022, as export prices climbed 6.5% and import prices 3.2%. Corporate incomes rose 9.6% in nominal terms, led by energy, while manufacturing operating surplus fell as energy-input costs increased. Capital formation and cost pressure are arriving together.

Germany supplies a more current input-price check, not a data-centre price index. Destatis says July import prices were 6.8% above a year earlier and 0.2% above June. Imported energy cost 26.4% more than a year earlier, intermediate goods 10.2% more, electronic components 15.0% more and capital goods 3.5% more. Excluding energy, the import-price index was still 4.8% higher year on year. These categories overlap the physical supply chain that compute-intensive services depend on, but they also cover far more than AI; the numbers cannot be converted into an AI cost-inflation rate.

The European Commission’s August survey adds a demand-side counterweight. The Economic Sentiment Indicator rose to 98.2 in the EU and 98.4 in the euro area, while the Employment Expectations Indicator reached 99.0 and 98.9 respectively. All four readings moved closer to, but remained below, their long-term benchmark of 100. Industry, services and retail confidence improved, yet consumers’ intentions to make major purchases weakened slightly. Selling-price expectations remained above long-term averages in every surveyed sector, and consumers’ price expectations increased by 3.0 points. The fieldwork ran from August 1 to 21, so it is a timely expectation survey rather than a measurement of activity after the release date.

Taken together, the records expose three different units that managers must not collapse. The SDK offers a future service-usage unit whose value and price are not yet available. Canada records real capital volumes and economy-wide output, including a sharp data-centre-linked equipment movement. Germany and the EU show price and expectation pressure across the surrounding input and demand environment. None substitutes for the others. A non-null compute count would still need a unit price; a unit price would still omit fixed capacity, networking, storage, power, depreciation, financing and idle time.

The practical response is to build the reconciliation before the new field starts speaking. Engineering teams should retain raw token counts, model and endpoint identifiers, latency, tool calls, retries and success outcomes, while accepting compute_units as nullable. FinOps teams should keep provider invoices separate from internal estimates and document any allocation rule. Infrastructure budgets should connect usage to committed capacity, utilization, electricity and the useful life of hardware. No team should impute zero cost from null, invent a token-to-compute conversion, or restate historical series until OpenAI defines coverage and stability.

What comes next is specific. Watch for the first endpoints and models that return a non-null value; a public definition of the unit; whether identical requests produce comparable values; treatment of cached input, tools, retries and failed calls; and an explicit relationship, if any, to billing. In the physical accounts, watch whether Canada’s 16.7% computer-equipment jump persists, how much capacity becomes productive, and whether imports translate into durable domestic output. In Europe, watch energy, electronic-component and capital-goods prices alongside sentiment rather than assuming improving confidence removes margin pressure.

The measured conclusion is less dramatic than a new pricing regime and more useful. OpenAI has made room for a compute quantity in a widely used client, but the official schema itself says the quantity is not populated. That is evidence of interface readiness, not evidence of cost transparency. Meanwhile, official economic data show data-centre processing equipment pulling investment higher in Canada and physical inputs carrying meaningful price pressure in Germany, while European expectations improve without fully normalizing. AI cost control is therefore becoming more important one layer before it becomes easier: the software column exists, but the finance-grade number still has to arrive and reconcile with the real assets underneath it.