Commitments: what RIs, SPs, CUDs really save

Outcome

By the end of this lesson, you will be able to calculate the effective discount of a reservation and recognize the over-commitment trap that destroys ROI.

1. Concept

Cloud providers will charge you less per hour if you promise to keep buying.

The deal is always the same shape. Commit to a baseline of usage for one year or three, pay a lower rate on it, keep the difference. The idea is sound. Almost all the money is lost in how it is carried out.

The four commitment instruments

INSTRUMENT          PROVIDER     SCOPE                       1-YR / 3-YR (ILLUSTRATIVE)
─────────────────────────────────────────────────────────────────────────
Reserved Instance   AWS          Specific instance family    ~40% / ~60%
Savings Plan        AWS          Compute-broad ($/hr commit) ~30% / ~55%
Committed Use       GCP          Specific machine family     ~37% / ~57%
Reservation         Azure        Specific VM SKU             ~40% / ~60%
Spot / Preemptible  All three    Stateless workloads only    50-90%

Each trades flexibility for a bigger discount.

The strictest kinds tie you to one machine family for the whole term. The most flexible ask only for a dollar amount per hour and apply it to whatever you run. Interruptible capacity is a different bargain entirely: no commitment at all, very large discount, and the provider can take the machine back with as little as 30 seconds to two minutes of warning.

The math: effective discount

The discount is advertised as a percentage off the list price, say 40%. What you actually realise is always less, for two reasons:

Two levers turn a rate-card discount into a realized one:

  coverage    = (eligible hours covered by the commitment) / (total eligible hours)
  utilization = (committed hours used) / (committed hours purchased)

Coverage sets how much of your eligible spend earns the discount.
Utilization decides whether the commitment wins or loses money at all.

A worked example with a 1-year RI on EC2 m5.large at 40% rate-card discount:

SCENARIO A, healthy
  Utilization: 95% (you use 95% of the hours you committed to)
  Coverage:    80% (80% of eligible usage runs at the RI rate)
  95% is above the 60% break-even, so every covered hour saves near 40%.
  The uncovered 20% just runs on-demand. Net: a clear win.

SCENARIO B, under-committed (savings left on the table)
  Utilization: 100% (every committed hour is used)
  Coverage:    40% (only 40% of eligible usage is covered)
  Still a win on every covered hour (100% is well above break-even).
  The other 60% pays rack rate. No loss, just unclaimed savings.

SCENARIO C, over-committed (the real failure)
  Utilization: 45% (workload moved; most committed hours go unused)
  45% is below the 60% break-even for a 40% RI.
  You keep paying the committed rate on hours nobody uses, and that
  overpayment now exceeds the discount earned. Net cost INCREASES.

The “disaster” case is Scenario C. An RI bought for a workload that gets re-architected or migrated mid-term keeps billing the committed rate until expiry, on capacity nobody uses. That is over-commitment, the textbook commitment mistake, and it is driven by low utilization, not low coverage.

Break-even math

Break-even is a utilization threshold, not a coverage one. Under-coverage never loses money: eligible hours the commitment does not cover simply run on-demand at rack rate. Under-utilization is what loses money, because you paid for committed hours you did not use.

Break-even utilization = 1 - discount

  40% RI:   1 - 0.40 = 60%    use < 60% of what you bought and you lose vs on-demand
  30% SP:   1 - 0.30 = 70%
  57% CUD:  1 - 0.57 = 43%

This is why commitments are bought on the post-schedule floor, not the current peak. The floor is the always-on overlap that survives your schedules. Buy for the peak and utilization drops below break-even the moment the peak subsides.

Spot: the asterisk

Spot instances offer the deepest discount (50-90%) without a commitment but introduce eviction risk: the provider can reclaim the capacity at short notice (2 minutes on AWS, about 30 seconds on GCP and Azure). Spot is a fit for:

• Stateless batch workloads
• K8s nodes that can drain gracefully
• Build and CI workloads
• Data processing jobs that checkpoint

Spot is not a fit for:

• Single-replica databases
• Stateful workloads without checkpointing
• Anything where 2-minute eviction is unacceptable

Spot is its own discipline. It is not a substitute for commitments on the floor: it complements them on the variable layer above.

How commitments compose with scheduling

LAYER                                  PRICING
───────────────────────────────────────────────────────────
Peak above the floor (bursty)          on-demand
Steady floor (always-on)               commitment (RI / SP / CUD)
Stateless / batch above the floor      spot, where safe
Non-prod (scheduled off-hours)         on-demand only: no commit

Each layer gets the right instrument. The discipline is identifying which layer a workload sits in, not picking one instrument and forcing it.

2. Demo

An illustrative commitment-design exercise after scheduling fires:

WORKLOAD ANALYSIS: Production EC2, post-schedule
─────────────────────────────────────────────────────────
Always-on floor (steady)             40 vCPU
Peak above floor                     +15 vCPU (afternoon, ~6 hrs/day)
Burst above peak                     +8 vCPU (event-driven, ~2 hrs/day)
─────────────────────────────────────────────────────────
COMMITMENT DESIGN
1-yr Savings Plan covering 40 vCPU       commits the floor
On-demand for the 15-vCPU peak           absorbs daily variance
Spot for the 8-vCPU burst (batch only)   absorbs event-driven

EXPECTED EFFECTIVE DISCOUNT
  Floor (Savings Plan): 1 - 0.70 / 0.95 = ~26% (at 95% utilization)
  Peak (on-demand):     0%
  Burst (Spot):         85%
  Weighted by vCPU-hours (960 floor, 90 peak, 16 burst per day):
                        ~25% effective discount on production EC2

The team’s previous plan had been to buy a 1-yr RI on the full 63 vCPU peak at 40% off. That RI would be used for 1,066 of its 1,512 vCPU-hours a day, about 70% utilization. Paying 60% of on-demand for every committed hour then works out to 0.60 / 0.705 = 0.85 of on-demand per hour used: only ~15% effective discount, with commitment locked in that they could lose in a year of change. The post-schedule, layered approach realizes ~25% with no over-commitment risk and full flexibility on the burst layer.

3. Hands-on (6 min)

For one of your steady-state production workloads:

1. Identify the always-on floor (the vCPU count or instance count
   that is running 100% of the hours, after any scheduling)
   FLOOR: ____________

2. Identify the peak (additional capacity at busy hours)
   PEAK: ____________

3. Identify burst (event-driven spikes)
   BURST: ____________

4. The commitment should cover the FLOOR, not the peak or burst.
   For the floor, model a 1-yr Savings Plan or RI:

   Committed hours per year   = FLOOR × 8760
   Expected coverage          = ____%  (target >85%)
   Expected utilization       = ____%  (target >95%)
   Saving on covered hours    = 1 - (1 - published) / utilization
                              = 1 - (1 - ____) / ____ = ____%
   Effective discount         = coverage × saving on covered hours
                              = ____ × ____ = ____%

5. If the effective discount is below ~20%, the commitment is not
   worth the flexibility loss. Stay on-demand on this workload.

4. Knowledge check

Q1

A team buys a 3-yr RI at 50% rate-card discount, but uses it only 65% of hours. Effective discount:

A. 32.5%
B. 50%
C. ~23%
D. 17.5%

Q2

A FinOps Analyst proposes buying a 1-yr Savings Plan calibrated to current peak compute usage. The defensible counter-proposal is:

A. Buy two SPs
B. Wait for next year
C. Use Spot instances for everything instead of committing to anything
D. Calibrate to the post-schedule, post-rightsizing floor: not current peak

Q3

A 1-yr RI at 40% discount needs roughly what utilization to break even against pure on-demand?

A. 60%
B. 40%
C. 71.4%
D. 100%

5. Apply

ZopNight’s Costs → Flow, set to a layout that includes Purchase, shows the current mix of on-demand / commitment / spot across the estate. The Sankey view splits any service column by purchase type.

ZopNight ships advisory commitment recommendations for Reserved Instances and CUDs. Each offers 1-year and 3-year No-Upfront options priced from your real on-demand and committed rates, and a commitment is suppressed unless the saving clears a break-even margin. They advise; they do not buy anything. Savings Plan and upfront options are not covered today; for those, AWS Cost Explorer’s RI / SP recommendations or managed commitment services such as ProsperOps (a Flexera company) are the right tools.

Related lessons

• L3: The non-prod fallacy (next)
• L4: When scheduling wins, when commitments win
• T5.M5.7: Commitments demystified

Glossary terms touched

Reserved Instance · Savings Plan · Committed Use Discount · Spot · Effective discount · Over-commitment