Pay-as-you-go or PTU: when should an SME switch AI capacity?
Choose when to move Azure AI workloads from pay-as-you-go to PTU, using a reusable matrix, regional capacity checks and a practical transition cost test.
Stanic & Co.
Perspectives for business leaders
Start with pay-as-you-go for low-volume or unpredictable AI traffic. Consider Provisioned Throughput Units (PTUs) when measured demand is stable, scale is substantial and response-time requirements justify dedicated capacity. For Australian SMEs using Azure, the transition needs a further gate: confirm capacity for the required model in an acceptable region before buying a reservation. A reservation discounts eligible usage; it does not guarantee capacity. This distinction is central to Microsoft’s provisioned throughput guidance.
The practical question is not simply which option has the lower advertised rate. It is whether your business can move from flexible consumption to dedicated capacity without creating an unaffordable commitment or an operational dependency it cannot sustain. The following proposed decision process focuses on that transition, rather than invoice reconciliation or reservation purchasing alone.
Pay-as-you-go or PTU: decide what the business needs protected
In Microsoft Foundry, Standard deployments use shared inference capacity and are billed per token. Provisioned deployments hold dedicated processing capacity, whether requests are arriving or not. Microsoft positions Standard for variable, unpredictable or low-volume workloads, and provisioned throughput for predictable, high-scale workloads needing consistent latency. These are Azure-specific distinctions, not claims about every AI provider. See the deployment comparison and billing explanation.
Separate three business objectives: keeping expenditure proportional to demand, maintaining response times, and limiting the financial consequences of a downturn. Pay-as-you-go supports the first objective. Dedicated capacity supports the second. Neither automatically achieves the third: that requires workload controls, an affordable commitment and a fallback process.
PTU and a PTU reservation are also separate decisions. A provisioned deployment can use hourly billing; a reservation applies a financial discount to eligible PTU usage. Microsoft documents one-month and one-year reservation commitments, with deployments and reservations created independently. Treat a capacity test and a financial commitment as separate approval gates, even if both eventually support the same production service.
Before comparing prices, write down the consequence of delay. Would a slower response inconvenience an internal analyst, prevent a customer completing a transaction, or block an essential operational task? Choose a measurable response-time requirement and a workable fallback. Human approval of an AI answer does not resolve slow infrastructure or unavailable capacity.
Build a transition baseline, not just a monthly total
A monthly request count is useful for budgeting but insufficient for sizing dedicated capacity. Microsoft identifies requests per minute, prompt size, response size and cache rate as relevant sizing inputs. Throughput per PTU varies by model, and minimum deployment sizes apply. Two applications with identical monthly request totals can therefore require different capacity. Use the PTU sizing guidance within Microsoft’s documentation, rather than converting requests directly into PTUs.
As an illustrative procedure, record representative operating periods, including busy intervals, quiet periods and known seasonal changes. Capture request arrival patterns, input and output token volumes, response times, retries and the proportion of work that can wait. Separate interactive work from overnight processing. Record model versions and configuration changes so that a larger prompt is not mistaken for unexplained infrastructure growth.
Build three business cases: expected demand, a downturn and a burst. For each, estimate the cost of the current Standard deployment and the proposed provisioned arrangement, including any Standard overflow and supporting services. Microsoft notes that Foundry costs are only part of total application costs; all participating Azure resources need monitoring. Its cost-management guidance also warns that some portal estimates exclude provisioned throughput and may not reflect contracted discounts.
Keep this baseline decision-focused. You need enough evidence to explain the cost and service trade-off, not a dashboard for every available metric. For the separate task of mapping observed charges back to resources and meters, use how to reconcile Azure AI costs with actual usage.
Use the AI throughput selection matrix
Copy the following entries into an approval record. This matrix is an illustrative management tool, not a Microsoft-prescribed rule. Complete these inputs: workload and owner; monthly request volume; whether volume is low, steady or burst-driven; peak requests per minute; input/output token profile; latency tolerance; predictability; permitted processing locations; fallback; and downside budget. Set numerical pass limits locally before testing.
Entry A — Low volume or low predictability, with tolerable delays: retain pay-as-you-go. Example: an internal assistant used irregularly by a small team. Pass if observed service performance meets the business requirement and the stressed consumption estimate is affordable. Stop the PTU proposal if its minimum viable deployment would leave substantial capacity unused or the business cannot explain next period’s demand.
Entry B — High-scale, predictable demand with mission-critical response times: test PTU. Example: a consistently busy interactive workflow where delays interrupt operations. Pass only when representative testing meets the agreed response-time requirement, the downside cost case is affordable, and capacity is deployed successfully in an approved location. Fail the transition if any of those conditions remains unresolved.
Entry C — Predictable base demand with manageable bursts: evaluate PTU plus Standard spillover. Example: steady daytime traffic with occasional campaign peaks. Microsoft documents spillover for supported Azure OpenAI provisioned models, routing overflow to a corresponding Standard deployment in the same Foundry resource. Pass only after testing both the overflow path and its cost. Do not assume support for every model provider; check the documented spillover limits.
Final gate — Record the selected option, evidence, unresolved assumptions, budget owner, technical owner, review trigger and rollback method. An unanswered critical field means “not ready”, rather than implied approval. For example: “Retain Standard; demand remains campaign-dependent; operations manager owns the decision; reassess after the next representative operating period.”
Prove regional capacity before making the commitment
For an Australian SME that requires processing in a particular Australian region, location is a deployment constraint, not merely a pricing preference. Microsoft distinguishes regional, data-zone and global provisioned deployment types with different routing arrangements. Do not treat a globally routed deployment as interchangeable with a regional one. Check the actual model, version, deployment type and processing-location requirements before approving a substitute.
Quota is not capacity. Microsoft states that quota is a policy limit, while capacity is the model-specific processing capacity available to deploy. Having sufficient quota does not guarantee deployment success. Its provisioned throughput documentation advises creating deployments to confirm availability before purchasing a reservation. This article makes no claim that a particular Australian region currently has capacity.
Use a controlled transition proposal: approve a bounded test expenditure, confirm the required deployment can be created, test representative traffic, then decide whether a reservation is justified. Hourly PTU billing starts when the deployment is created and stops when it is deleted, so the test itself needs an owner and a cleanup deadline. Availability checks should not become unattended paid deployments.
The reversal plan needs equal attention. Microsoft warns that deleting or downscaling a deployment releases capacity and that the same capacity may not be available later. Therefore, “we can recreate it tomorrow” is not an adequate recovery assumption. Retain a tested Standard or operational fallback, and document what service degradation the business will accept if dedicated capacity cannot be restored.
Keep the reservation approval separate from this technical gate. For the financial commitment checklist, use Azure PTU reservations: when is an SME ready to commit?.
A hypothetical failure: the cheaper forecast becomes idle capacity
Consider a wholly hypothetical Australian distributor with an AI-assisted order-processing workflow. All figures below are illustrative AUD planning assumptions, not Azure prices, quotes or savings forecasts. Finance assumes Standard would cost $6,000 a month at expected demand. A proposed reserved PTU arrangement is allocated $4,000 a month, with another $500 allowed for Standard overflow. Supporting application costs are assumed equal and excluded from both sides.
At expected demand, the apparent monthly advantage is $6,000 − ($4,000 + $500) = $1,500. Management approves a one-year commitment after a busy-season pilot, without testing a seasonal downturn. Orders then fall sharply. Under the revised planning assumptions, Standard would cost $1,800 a month, while the committed PTU cost remains allocated at $4,000 with no overflow.
The downside difference is now $4,000 − $1,800 = $2,200 a month against the provisioned option. If that condition lasted six months, the illustrative excess would be $13,200. The failure is not that PTU cannot perform the workload. It is that management treated a peak-period consumption forecast as a dependable annual baseline and ignored the cost of unused capacity.
Detect the mistake by comparing demand, delivered service and total cost against the approved downside case—not simply checking whether the monthly invoice matches the reservation forecast. Falling utilisation combined with a shrinking Standard counterfactual should trigger review. Keep the counterfactual based on comparable workload and pricing assumptions; it is an estimate, not a bill that was actually incurred.
Possible responses include stopping further commitments, reviewing renewal, changing the future deployment mix or assessing whether other eligible demand can use the purchased benefit. Do not assume deletion cancels the financial commitment. The supplied evidence does not establish refund or exchange entitlements; verify the applicable terms before treating recovery as available. Avoid claiming the commitment is either automatically refundable or categorically non-refundable.
Choose a simpler route when immediacy has little value
Not every AI workload needs interactive capacity. Microsoft describes batch deployments as discounted, per-token processing with asynchronous results and no latency service-level agreement in its deployment comparison. If a task can wait, compare an eligible batch option before purchasing dedicated throughput. Confirm suitability for the actual model and workflow rather than assuming all API requests can be moved unchanged.
A non-AI alternative may be better still. For a low-frequency task, propose a structured form, deterministic validation rules and a staff processing queue. For example, collecting required order fields through a form may remove the need for AI interpretation altogether. Estimate the staff effort and delay alongside the technical cost. This is a proposed alternative, not a claim that manual processing is always cheaper.
Use the same decision standard across alternatives: acceptable completion time, total operating cost, exception handling and recovery. A manual queue can become overloaded; a rules-based process may reject legitimate exceptions; batch processing cannot satisfy an immediate-response requirement. The point is to choose the least complicated arrangement that meets the business need, not to avoid AI at any cost.
Do not use PTU to compensate for an undefined workflow or unmeasured demand. Equally, do not retain Standard solely because its billing is flexible if measured performance cannot meet an essential requirement. Infrastructure choice should follow a defined service requirement and evidence, with a fallback for work that cannot safely proceed.
Assign ownership and make the decision reversible where possible
Make the business process owner accountable for the service requirement and fallback. Give the technical owner responsibility for sizing, regional deployment, traffic routing and cleanup. Finance should approve the downside exposure and commitment terms. One named decision owner should sign the final record; a collection of dashboards and email recipients is not a substitute for authority to pause or change the service.
Use budgets as monitoring, not as a hard spending cap. Microsoft states that Azure budget thresholds trigger notifications but do not stop consumption or affect resources. Cost and usage data is typically available within 8–24 hours, budgets are evaluated every 24 hours, and emails are normally sent within an hour of evaluation. Those delays make budget alerts unsuitable as the sole protection against a rapid spending surge. See Azure budget behaviour and notification timing.
As an illustrative control procedure, pair financial alerts with tested application limits, bounded retry behaviour and an explicit pause or queueing action. Decide what happens if the person receiving an alert is unavailable. For spillover, approve a separate consumption allowance and an action when it is exhausted; otherwise the hybrid design leaves the variable-cost exposure unresolved.
Human review has limits here. A finance reviewer cannot establish regional capacity from a spreadsheet, and a technical reviewer cannot guarantee future demand or capacity availability. Reviewers need deployment evidence, representative traffic measurements and clear assumptions. Likewise, staff checking AI outputs may add processing delay and labour cost; their presence does not guarantee correctness or infrastructure reliability.
Finish with a dated decision record and event-based review triggers: a material change in demand, model, prompt size, response-time performance, processing-location requirements or renewal exposure. Keep Standard until the PTU case passes both the service and downside tests. Move only the justified baseline to dedicated capacity, preserve an affordable fallback, and treat any reservation as a separate financial commitment.
Sources
General information only; verify sources before making business decisions.