Summary

  • Microsoft's defining sentence calls a provisioned throughput unit "a fixed amount of model processing capacity" and never says fixed in what; the amount sits one page over as a per-model figure running from 750 to 14,900 input tokens per minute across the six rows reviewed.
  • At the printed 15-unit floor, the same purchase buys 11,250 input tokens a minute on gpt-6.2-astra and 223,500 on gpt-4.1-mini, a spread of about 19.9 times that no Microsoft page states and that is my division rather than theirs.
  • Capacity becomes computable once the buyer supplies an output-to-input mix; the hourly rate is the term the reviewed documents leave to a pricing calculator, and the concrete dollars are 2024 and 2025 vintage.

The sentence that defines the unit

Start with the sentence Microsoft uses to describe the thing it bills by the hour. On the page that introduces provisioned throughput for Foundry Models, it reads: "Provisioned throughput units (PTUs) are the unit of measure for provisioned throughput. A PTU represents a fixed amount of model processing capacity."

Read it twice. It fixes two things and refuses a third. The unit measures provisioned throughput; the unit stands for a fixed amount of processing capacity; and the sentence declines to say how much. Capacity is not a synonym for tokens here, and the same page explains why. The throughput a given number of units delivers depends on the model, each model carries a minimum unit count for a deployment, and output tokens consume more processing capacity than input tokens, with a model-specific ratio expressing how many input tokens one output token is worth. The page that withholds a number also explains why one number would be false.

That is not a dodge. It is a division of labour, and the quotient lives one address away.

The amount is a row in a table

On the sizing page, one table cell does the work a definition sentence did not: input TPM per PTU is "the maximum input tokens per minute (TPM) that one PTU supports", and the buyer is told to use that value when estimating units. The same page prints the arithmetic. Units required equals normalized TPM divided by the model's input TPM per PTU, and normalized TPM is input TPM multiplied by one minus the cache rate, plus the output-to-input ratio multiplied by output TPM. For GPT-6 and newer models the normalized figure also carries cached input and cache writes, and a short-context input token counts as one.

Six rows came back when I read that page, ordered here by capacity:

The amount is a row in a table
Model Input TPM per PTU Minimum units, Global or Data Zone Minimum units, Regional
gpt-6.2-astra 750 15 25
gpt-4o 2,500 15 50
gpt-6.2-sol 2,950 15 25
gpt-4.1 3,000 15 50
gpt-5 4,750 15 50
gpt-4.1-mini 14,900 15 25

The floor is 15 units in every Global or Data Zone row and 25 or 50 in the Regional column. Multiply the smallest legal purchase by the printed coefficient and the fixed amount stops looking fixed: 15 x 750 is 11,250 input tokens a minute, 15 x 14,900 is 223,500, and 223,500 divided by 11,250 is 19.87. That last division is mine. The page prints the multiplicands and not the product, and the product is the number a buyer actually needs.

Then the buyer adds a term of their own. Because capacity is counted in normalized tokens, two tenants moving identical token volumes through the same model get different answers if their output shares differ, since one output token costs the model's printed ratio in input-token capacity and, on the newer rows, caching adjusts the numerator again. The fixed amount is fixed in a currency the buyer mints: the mix of prompt and completion in their own traffic. The unit belongs to the vendor. Its size in tokens is a joint product.

The price is a rate, a reservation, and a calculator

The billing page is explicit about the basis: "Provisioned deployments (Regional, Data Zone, and Global) are charged at an hourly rate ($/PTU/hr) based on the number of PTUs deployed." Deployed capacity is billed in full whether or not it is used, partial hours are prorated, deployments above a reservation quantity are billed at the hourly rate, and reservations run one month or one year. The same page says the reservation discount varies by model family and term, which retires the idea of a single price for a single unit.

The current pricing page lays out the columns a buyer wants: minimum units, per-unit hourly pricing, monthly reservation pricing, yearly reservation pricing. It repeats the model minimums the sizing page prints, including 15 units for GPT-4.1 Global and 50 for GPT-4.1 Regional, and it states that its prices are estimates while directing the reader to the pricing calculator or to sales. That is the architecture of the price: a table of quantities with the money delegated.

The concrete dollars in Microsoft's own documents are dated, and only two are close enough to quote. A Microsoft Azure blog post from 28 October 2024 announced that Provisioned Global pricing would move from $2.00 to $1.00 per hour on 1 November 2024, with Data Zone at $1.10, a one-month reservation at $260 per unit per month and a one-year reservation at $221. A Community Hub post from 9 October 2025 put a monthly reservation for a GPT-4o unit at $260 for global deployments, and cautioned that prices change. That post's own illustrative token figures do not reconcile with the sizing table's 2,500 input TPM per unit for gpt-4o, so the sizing page governs every capacity figure in this article and the post is used here only for its dated reservation dollar.

Now do the arithmetic those two documents permit, and label it as mine. At $1.00 per unit-hour the 15-unit floor costs $15.00 an hour. One hour at the printed coefficient buys 60 minutes of input tokens: 13,410,000 on gpt-4.1-mini, 2,700,000 on gpt-4.1, 2,655,000 on gpt-6.2-sol, 2,250,000 on gpt-4o, 4,275,000 on gpt-5 and 675,000 on gpt-6.2-astra. Divide $15.00 by those millions and the cost per million input tokens runs from $1.12 on gpt-4.1-mini to $22.22 on gpt-6.2-astra. The spread is 19.9 times, the same factor as the capacity gap, because one hourly rate sits on top of all six rows.

The reservation arithmetic is duller and more consequential. At $260 a month over 730 hours the effective unit-hour is $0.356; at $221 a month for a year, $2,652 a unit-year, or $0.303 against the $1.00 hourly. Those are 2024 numbers divided by my own assumption of 730 hours, and mixing them with an undated table is arithmetic rather than a quote: it shows what the relation between rate and coefficient does, not what anyone pays today. What it shows is that a commitment is a bet placed on a coefficient that arrives in a table row, and a model migration re-prices capacity without touching the rate for a second.

The ceiling is a division you perform

The operations guide closes the loop. Utilisation is defined there as the units consumed in a period divided by the units deployed in that period, and the throttling rule is blunt: "When a request is made, if current utilization is at 100%, the service returns HTTP 429 immediately, with retry-after-ms and retry-after response headers indicating how long to wait." Short bursts slightly above 100 percent are tolerated; over time, the page says, traffic is capped at 100 percent. Optional spillover can hand overflow to a standard deployment if the buyer wires one up.

So the ceiling of a provisioned deployment is a ratio whose denominator the buyer chose at purchase. Buy fewer units and the same traffic reaches the wall sooner; buy more and it reaches it later; either way the bill follows the deployed count, not the traffic. That is a materially different instrument from a metered API, where exceeding your own forecast produces a larger invoice. Here exceeding capacity produces an immediate refusal with a clock attached, and the invoice does not move. The billing page says as much in one line: the charge is based on the units deployed, not the tokens consumed.

One more printed boundary. The definition page states that quota is granted per subscription, per region and per deployment type, and that the same quota can deploy any supported model. Quota is therefore a second pool that the model row spends at different rates: the same 15 units that buy 223,500 input tokens a minute on the light row buy 11,250 on the heavy one.

What a buyer can check, and where the paper ends

Put the four pages side by side and the unit becomes legible in a way few cloud meters are. A buyer can compute capacity for an input-only workload exactly, because the coefficient and the floor are both printed. A buyer with a real workload can compute it approximately, because the output-to-input ratio is printed per model and the mix comes from their own telemetry. A buyer can even compute the wall, because utilisation is a division they can perform against their own purchase.

What no page in this set lets them finish is the rate. The basis is printed, in dollars per unit per hour on deployed units with reservations against a term, and the digits sit on a calculator while the only quoted dollars predate the model rows they would now sit beside. The fixed amount of model processing capacity, in other words, is a coefficient with a brand name: printed, floored, divided by your own verbosity, and priced by a document that asks you to sign in first.

The directory entry for Microsoft Corporation records the resource-holder behind the platform; every number above is the company's own, published on its own pages.