AI capacity reservations aren't savings plans: how to compare PTUs, model units and GSUs
In brief
Provisioned AI capacity is billed every hour, used or not, and often breaks even only at high use. How to compare PTUs, model units and GSUs, with the math.
Steve Elliott · · 7 min read
On this page
An AI capacity reservation buys guaranteed throughput for a model or model family, paid for every hour whether you send traffic or not. Where providers publish enough to check, it breaks even against pay-as-you-go only at high utilization: about 75% for a one-year Gemini commitment at Google's standard list prices, and close to 100% in Microsoft's own PTU illustration, which uses an assumed price.
That makes these purchases different from savings plans, which discount spend that floats across instance types, regions and services. In our view, a FinOps team should treat Azure PTUs, Amazon Bedrock model units and Google GSUs as capacity and latency purchases first and discounts second.
How is an AI capacity reservation different from a savings plan?
A compute savings plan commits you to an hourly spend that any eligible usage draws down, so the discount follows workloads between instance families. AWS's Savings Plans guide lists EC2, Lambda, Fargate and SageMaker AI as eligible services. Microsoft's savings plan overview, updated October 29, 2026, lists compute and database services. Neither page names Amazon Bedrock or Azure OpenAI.
A provisioned throughput purchase reserves a fixed block of model capacity, in the provider's own unit, paid for around the clock. The FinOps Foundation's paper Navigating GenAI Capacity Options, updated February 4, 2026, puts it plainly: "You are paying for this capacity 24/7, whether you use it or not."
Google makes the contrast explicit. Its Flexible Savings Plans, updated November 10, 2026, give 10% off eligible spend for one year and 20% for three, and they cover first-party models such as Gemini. Provisioned Throughput SKUs draw down that commitment but get no incremental discount.
How do the main constructs compare?
Terms as of November 2026:
| Azure PTU reservation | Bedrock Provisioned Throughput | Bedrock Reserved tier | Google Provisioned Throughput | EC2 Capacity Blocks | |
|---|---|---|---|---|---|
| Unit | Provisioned throughput unit (PTU) | Model unit (MU) | 1,000 tokens per minute, input and output set separately | Generative AI scale unit (GSU) | GPU and Trainium instances |
| Terms | Hourly with no commitment, or 1-month or 1-year reservation | No commitment, 1 month or 6 months | 1 or 3 months | 1 week (Google models only), 1 month, 3 months, 1 year | 1 to 14 days, then weekly up to 182 days |
| Model lock | None: PTU quota can deploy any supported model; reservations are per deployment type | One model; custom-model capacity can switch to its base model or a sibling | Offered model by model | Can switch models within one publisher | None |
| Overflow | Optional spillover to a standard deployment | Not stated in AWS docs | Overflows to Standard tier | Spills to pay-as-you-go by default | Not applicable |
| Exit | Cancellation may carry an early termination fee | Can't delete until the term ends | Billing runs until deleted with the account team | No mid-term cancellation; decreases apply at renewal | No cancellations |
Azure's pool is the most flexible on models. It also has a waste category the FinOps Foundation paper names: PTUs reserved but not deployed. Google publishes per-GSU prices. AWS publishes MU rates only for a few older models, such as Cohere Command and Llama 2; for others, including Anthropic, it refers customers to their account team.
How do you work out break-even utilization?
Break-even utilization is the committed hourly price divided by the pay-as-you-go cost of one unit's full hourly throughput. Below it, averaged over every hour of the term, pay-as-you-go is cheaper.
The hard part is the denominator: each provider weights input, output and cached tokens differently.
Google publishes enough for a worked example. Gemini 3.8 Flash delivers 675 burndown-adjusted tokens per second per GSU, with output tokens counted at 5 times input. Its standard global pay-as-you-go price, which applies from January 1, 2027, is $1.50 per million input tokens and $7.50 per million output tokens. Because the output price is also 5 times input, one burndown-adjusted token costs $1.50 per million at pay-as-you-go. Cached input, at a tenth of the price and a tenth of the burndown, keeps the same parity.
One fully used GSU therefore processes 2.43 million burndown-adjusted tokens an hour, worth about $3.65 at pay-as-you-go. Against Google's published GSU prices:
| Term | Price per GSU-hour (global) | Monthly equivalent | Break-even utilization vs list pay-as-you-go | Break-even if pay-as-you-go has a 1-year Flexible Savings Plan |
|---|---|---|---|---|
| 1 week | $7.14 | $1,200 per week | 196%: never | Never |
| 1 month | $3.70 | $2,700 | 101%: never | Never |
| 3 months | $3.29 | $2,400 | 90% | 100%: never in practice |
| 1 year | $2.74 | $2,000 | 75% | 84% |
The break-even figures are our arithmetic from Google's published rates. 1-week and 1-month terms cost more than pay-as-you-go even at full use, so their only case is guaranteed throughput; against a 1-year savings plan, only the 1-year term can pay off. The break-even also moves whenever either price moves. Through December 31, 2026, Google gives 50% credits on this model's pay-as-you-go spend and a separate 50% credit on its Provisioned Throughput usage. Both halve their side of the comparison and end the same day, so the break-even holds today; it has to be redone if either changes without the other.
An illustrative workload: a team buys 10 GSUs on a one-year term for $20,000 a month. At 60% average utilization, the tokens used would have cost about $16,000 at pay-as-you-go: roughly $4,000 a month lost. At 85%, they are worth about $22,600: about $2,600 saved. A service that runs flat out 12 hours a day and idles overnight averages 50% and loses money on every term.
What does Microsoft's own PTU example show?
Microsoft's PTU sizing page, updated September 24, 2026, compares one GPT-6 Sol PTU with pay-as-you-go. It assumes $260 per PTU per month, labels that an example assumption rather than a quote, notes that deployment minimums still apply, and assumes 100% sustained utilization for 30 days with no discounts. On those terms the PTU works out to about $2.01 per million normalized tokens against $2.00 at pay-as-you-go. The page concludes that the two are approximately equal at full utilization and that lower utilization raises the effective PTU cost per token.
By our arithmetic, the same PTU at 70% utilization would cost about $2.87 per million normalized tokens, about 43% more than pay-as-you-go.
Microsoft also says reservations discount the hourly PTU meter but don't guarantee capacity, so deploy first and reserve after. Reservations for Global, Data Zone and Regional deployments aren't interchangeable.
What about Bedrock, where the unit isn't fully public?
Bedrock's Provisioned Throughput offers no commitment, 1-month and 6-month terms, and AWS says longer terms carry a lower hourly price. For Cohere Command, for example, AWS lists $49.50 an hour per model with no commitment, and $39.60 and $23.77 per MU-hour on 1-month and 6-month terms: 20% and about 52% below the no-commitment rate.
The public documents don't give the throughput behind an MU. AWS defines an MU as a set number of input and output tokens per minute for a given model, and refers customers to their account manager for the figures. Without them, break-even can't be computed from public information. We think a FinOps lead should get MU throughput in writing, by model, input and output, before comparing quotes.
The newer Reserved tier, announced in November 2025, is closer to a token reservation. Customers reserve input and output tokens per minute separately, for 1 or 3 months, with minimums of 100,000 input and 10,000 output tokens per minute. Excess traffic overflows to the Standard tier. AWS's capacity guidance page puts the saving at 30% to 50% for steady, high-volume load, but that page describes the tier in model units rather than tokens per minute and doesn't say what the saving is measured against. Prices aren't public, so the claim can only be checked against a quote.
EC2 Capacity Blocks answer a different question. They reserve accelerator instances for days or weeks, at a price set by supply and demand, charged up front, with no cancellation and no Savings Plans or Reserved Instance discounts. In our view, the comparison for a Capacity Block is the cost of a training run slipping, not a token rate.
When is a capacity reservation worth buying?
We think the decision runs in six steps:
- Measure the baseline, not the peak. Find the tokens-per-minute level traffic rarely drops below over 30 days. Google's guidance says you generally don't need to provision for peak demand.
- Convert to the provider's unit. Use Microsoft's normalized tokens, Google's burndown rates or the MU throughput AWS gives you.
- Compute break-even utilization for every term on offer. Recompute when pay-as-you-go prices, credits or your enterprise discount change.
- Size to the baseline and let spillover carry the rest. Traffic above the baseline runs below break-even utilization, so it is cheaper at pay-as-you-go. Check overflow support: Google and the Bedrock Reserved tier overflow by default, Azure optionally for most models, and AWS documents none for Bedrock Provisioned Throughput.
- Price the exit. Cancellation terms, renewal dates, and whether the commitment can follow a model upgrade.
- Say what you're buying. If the case is latency or guaranteed throughput, record that, and load test first. The FinOps Foundation paper warns that vendor tokens-per-minute figures are only rough estimates.
We suggest reporting reservations separately from savings plans in commitment utilization and effective savings rate figures. In our view, a PTU at 70% utilization is not a 30% waste figure in the way an idle savings plan is; it may be a reasonable price for capacity the service needs. For the wider economics of an AI initiative, including labor, see AI initiative costs: tokens vs labor.
A capacity purchase is only as sound as its link to the service it supports and the owner who answers for its cost, which is why the reservation, the product it serves and the decision to buy belong in one operating model.
Sources
- FinOps Foundation, Navigating GenAI Capacity Options, last updated February 4, 2026
- Microsoft Learn, Provisioned throughput for Foundry Models, updated July 15, 2026
- Microsoft Learn, Provisioned throughput units onboarding, updated May 26, 2026
- Microsoft Learn, Provisioned throughput sizing, updated September 24, 2026
- Microsoft Learn, What is Azure savings plans for compute?, updated October 29, 2026
- AWS, Increase model invocation capacity with Provisioned Throughput in Amazon Bedrock, accessed November 21, 2026
- AWS, Purchase a Provisioned Throughput for an Amazon Bedrock model, accessed November 21, 2026
- AWS, Amazon Bedrock pricing, accessed November 21, 2026
- AWS, Service tiers for optimizing performance and cost, accessed November 21, 2026
- AWS, Amazon Bedrock introduces Reserved Service tier, November 26, 2025
- AWS, Amazon Bedrock User Guide: capacity limits and costs, accessed November 21, 2026
- AWS, Capacity Blocks for ML, Capacity Blocks pricing and billing and Find and purchase Capacity Blocks, accessed November 21, 2026
- AWS, Services eligible for Savings Plans benefits, accessed November 21, 2026
- Google Cloud, Generative AI pricing (Provisioned Throughput and Gemini 3.8 Flash), accessed November 21, 2026
- Google Cloud, Purchase Provisioned Throughput, updated November 20, 2026
- Google Cloud, Calculate Provisioned Throughput requirements, updated November 20, 2026
- Google Cloud, Provisioned Throughput supported models, accessed November 21, 2026
- Google Cloud, Use Provisioned Throughput, updated November 20, 2026
- Google Cloud, Consumption options, updated November 20, 2026
- Google Cloud, Flexible Savings Plans, updated November 10, 2026
Frequently asked questions
Are Azure PTU reservations a discount or a capacity guarantee?
Neither alone. Microsoft describes PTU reservations as a discount applied to the hourly PTU meter and says they don't guarantee capacity; capacity comes from the deployment, which you should create before buying the reservation. Microsoft's own sizing example, assuming $260 per PTU per month, shows a PTU costing about the same as pay-as-you-go at full utilization, so, in our view, the main gain is dedicated throughput and predictable latency.
Do AWS or Azure savings plans cover AI model inference?
Not on the published eligibility lists as of November 2026. AWS lists EC2, Lambda, Fargate and SageMaker AI for its Savings Plans, and Microsoft lists compute and database services for its savings plans. Bedrock and Azure OpenAI are not named. Google's Flexible Savings Plans do cover Gemini pay-as-you-go spend, with Provisioned Throughput drawing down the commitment without extra discount.
What utilization do you need for provisioned AI capacity to pay off?
It depends on the model, term and your pay-as-you-go discount. On Google's published prices for Gemini 3.8 Flash, a one-year GSU term breaks even at about 75% average utilization against standard list prices, and 1-month terms never do. Microsoft's PTU illustration, using an assumed $260 a month, breaks even near 100%. Calculate it for each workload rather than assuming a rule of thumb.
How are EC2 Capacity Blocks different from Bedrock Provisioned Throughput?
Capacity Blocks reserve accelerator instances for 1 to 182 days, typically for training, fine-tuning or experiments, at a dynamic price paid up front with no cancellation. Bedrock Provisioned Throughput reserves inference capacity for one hosted model in model units, billed hourly over a 1-month or 6-month term or with no commitment. One buys hardware time; the other buys model throughput.
Steve Elliott
Published November 22, 2026
More on FinOps & Cloud
FOCUS, commitments, shared cost.
Keep reading
Negotiating cloud commitments when capacity is scarce
AWS and Google say demand outruns supply, and all three report backlogs over $550B. What that does to your bargaining power and which terms still move.
Effective savings rate: the commitment number finance should ask for
Coverage and utilization can both look healthy while commitments save little. Why finance should ask for effective savings rate, and how to check the math.
The Ultimate Guide to Creating Effective Action Plans
Master action plan creation. Learn to set clear goals, outline specific steps, and drive results with our comprehensive guide to effective planning.
Product Aligned vs Capacity Aligned: A Comprehensive Comparison
Product-aligned teams own a product area long term, while capacity-aligned teams are assigned work as demand shifts. Compare the trade-offs of each structure.