The Reservation Trap

Executive Summary

The conversation in every cloud engineering team goes the same way. The CFO sees Reserved Instance pricing and says "3-year commitment, 40% off — sign it." The CTO sees Spot pricing and says "65% off, no commitment — why would we ever pay more?" Both are right. Both are wrong. The answer is not in the discount percentages. It's in a single number neither of them has computed: the breakeven interruption rate at which RI dollars equal Spot dollars after accounting for recovery from every 2-minute termination warning.

An RI is not a discount. It is a financial derivative — a 36-month cash advance to AWS in exchange for a guaranteed unit price. You are betting that over those 36 months, the spot market will not move in your favor enough to make the RI premium a net loss. The bet has a price tag. For a 100-instance fleet on m6i.xlarge at $0.192/hr on-demand, 3-year Standard RI costs approximately $369,000 over the term versus $591,000 on-demand — $222,000 in "savings." But the same fleet on spot at 65% discount, with 2.5 interruptions per 1,000 instance-hours at 10 minutes recovery each, costs approximately $213,000 — $156,000 less than the RI. The RI didn't save $222,000. It cost $156,000 more than the alternative.

This briefing deconstructs the RI vs Spot decision into its component variables: the breakeven interruption rate, the commitment term's embedded interest rate, the Convertible RI liquidity premium, the Savings Plan structural arbitrage, and the mixed-fleet allocation math that sophisticated teams use to extract 60%+ savings without betting the farm on a single pricing model. Because picking RI or Spot is not a binary choice. It's a portfolio allocation problem. And portfolios get rebalanced when market conditions change.

The RI Is Not A Discount — It's A Loan To AWS

When you buy a 3-year Standard RI, you are not "saving 40%." You are advancing AWS 36 months of cash flow. AWS gets your money upfront (or in monthly installments at a slightly worse rate). You get a lower per-hour price over the term. The economic substance of this transaction is a loan from you to AWS, collateralized by EC2 capacity, with an interest rate embedded in the discount spread between 1-year and 3-year terms.

Here is the IRR decomposition. An m6i.xlarge on-demand costs $0.192/hour = $1,681.92/year. A 3-year Standard RI at 40% off costs $0.1152/hour = $1,009.15/year — you save $672.77/year per instance. A 1-year Standard RI at 31% off costs $0.1325/hour = $1,160.70/year — you save $521.22/year. The incremental discount for locking in 2 extra years is $151.55/year per instance. For a 100-instance fleet, locking in 2 extra years earns you $30,310 in additional discount. The question is: what would that $100,910 in extra commitment capital earn if deployed elsewhere? And more importantly — what happens to spot pricing over those 2 years?

AWS is not selling you a discount. AWS is buying certainty about its capacity planning. Your RI commitment tells AWS exactly how many instances to provision in each AZ for the next 1-3 years. The discount is the price AWS pays for that demand signal. The longer the commitment, the more valuable the signal — and the deeper the discount. But the calculus on your side is different. You are not being paid for demand forecasting. You are locking in a unit price in a market where the unit price has historically trended down. The question is whether the guaranteed discount exceeds the expected decline in spot prices plus the cost of managing spot interruptions over the same period.

The Breakeven Surface — Why One Number Decides Everything

The breakeven interruption rate is the spot interruption frequency at which RI cost equals Spot cost. Below that rate, spot is cheaper. Above it, RI is cheaper. The formula:

breakevenRate = (8760 × (1 − riDisc) / (annualHours × (1 − spotDisc)) − 1) × 1000 / recoveryHours

Four variables control the outcome. Two of them are market conditions (RI discount, spot discount). Two are workload characteristics (monthly runtime hours, recovery time per interruption). Change any one and the breakeven rate shifts. The sensitivity is not linear. A 10% change in spot discount moves the breakeven more than a 10% change in RI discount because the denominator is smaller.

For a concrete example: 10 instances of m6i.xlarge, running 730 hours/month (24/7), on-demand $0.192/hr. 3-year Standard RI at 40% off. Spot at 65% off. Recovery cost: 10 minutes per interruption (stateless web tier behind load balancer, in-flight request drain).

The breakeven interruption rate for this scenario computes to ~2.5 interruptions per 1,000 instance-hours. At 2.5 int/1K hrs, spot still beats RI. At a real-world interruption rate of 5.0/1K hrs (typical for stateful workloads in contested instance families), the breakeven flips — RI becomes cheaper. The number is not 2.5 for every workload. It's 2.5 for this workload with these parameters. Change the monthly hours from 730 to 400 (batch processing), and the breakeven jumps to ~8.7 — RI loses across almost all realistic interruption rates. Change the spot discount from 65% to 50%, and the breakeven drops to ~1.2 — RI wins at nearly any interruption rate above zero. The breakeven is a function, not a constant.

The Spot Pool Depth Illusion — Why "Only 5% Interruption Rate" Is A Misleading Number

AWS publishes spot interruption frequency data at the instance-family level. The reported interruption rate for m6i in us-east-1 has been under 5% for most of 2026. This number creates a false sense of safety for three reasons.

1. The rate is an average across AZs and instance sizes. The spot pool in us-east-1a for m6i.xlarge can be empty while us-east-1d for m6i.2xlarge is 80% available. If your autoscaling group is pinned to specific AZs (which most are, for cross-AZ data transfer cost avoidance), your effective interruption rate is the rate in your specific AZ, not the regional average. An AZ-level spot drought — triggered by a large on-demand customer provisioning in that AZ — produces a 100% interruption rate for your fleet while the regional average stays at 3%.

2. Interruptions are correlated, not independent. The probabilistic model that says "2.5 interruptions per 1,000 hours" implies each instance-hour is an independent draw. It is not. When a spot pool exhausts, every instance in that pool gets the 2-minute warning simultaneously. A 10-instance fleet does not experience 10 independent interruption events. It experiences 1 correlated event that terminates all 10 instances at once. The cost of a correlated termination is not 10 × recovery_time. It is recovery_time for all 10 instances happening in parallel — and the operational chaos of cold-starting an entire fleet simultaneously. Your recovery process has a throughput limit. If you can restart 3 instances per minute and all 10 get terminated, your effective recovery time is 3.3 minutes of degraded capacity, not 10 individual 30-second restarts.

3. GPU spot pools experience structural scarcity unmodeled by historical averages. NVIDIA's H100 allocation to cloud providers is fixed and oversubscribed. When a hyperscale customer places an on-demand order for 128 p5.48xlarge instances — and this happens regularly — the spot pool drains to zero in that region within seconds. The historical average interruption rate might show 10% because most days are quiet. The actual experience is: 29 days of 0% interruption, 1 day of 100% interruption. Averages are useless for tail events. The metric that matters is the maximum interruption rate during your workload's critical windows, not the mean.

For GPU training workloads, the practical defense is not forecasting interruption rates. It's checkpointing every 100-500 training steps so that a termination at any moment loses at most 5 minutes of work. The checkpoint itself takes time (writing 40-80 GB of model weights + optimizer state to S3). If you checkpoint every 500 steps and a 500-step batch takes 8 minutes, the checkpointing overhead is 10-20% of training time. That overhead is the real cost of spot — not the per-hour price, but the fraction of GPU time spent writing checkpoints to survive interruptions that may or may not happen. For an 8-GPU training run at $3.06/instance-hour (p4d.24xlarge on-demand), checkpointing overhead of 15% costs $3,221/month in "wasted" GPU time. The net spot savings after checkpoint overhead = spot discount − checkpoint tax. And the checkpoint tax is paid regardless of whether an interruption actually occurs. It's insurance, not a marginal cost.

The Mixed-Fleet Allocation — Not RI vs Spot, But RI% As A Continuous Variable

The teams that win at cloud pricing do not argue about RI versus Spot. They argue about what percentage of the fleet goes into each bucket. The decision is not binary. It's a portfolio allocation problem: given the workload's minimum steady-state traffic, its peak-to-base ratio, its tolerance for interruption, and the market pricing for RI and Spot, what's the optimal RI percentage?

The allocation framework works like this:

For a workload with a minimum of 6 instances (never drops below, verified by 6 months of CloudWatch data), a peak of 18 instances (daily 2-hour spike), this framework allocates: 6 on 3-year Standard RI ($1,009/yr each = $6,054/yr), 8 on spot (0.7 × (18 − 6) = 8.4, rounded down), 2 on on-demand (10% × 18). Total annual cost with spot at 65% off and 2.5 int/1K hrs, 10-min recovery: $6,054 (RI) + $3,270 (spot) + $3,364 (on-demand) = $12,688. Pure on-demand: $30,275. Pure 3-year RI: $18,165. Pure spot (same parameters): $11,033. The mixed-fleet approach costs $1,655 more than pure spot — a 15% premium — in exchange for eliminating the risk of a correlated spot termination taking down the entire fleet simultaneously. That $1,655 is the insurance premium. Whether it's worth paying depends on your service's SLO and the cost of a full-fleet cold start during peak traffic.

Convertible RI — The Liquidity Premium Nobody Prices Correctly

A 3-year Standard RI locks instance family, size, AZ, OS, and tenancy for 36 months. A 3-year Convertible RI at 34% off (vs Standard's 40%) gives you the right to change instance family, size, OS, and tenancy during the term. The 6-percentage-point discount gap is the liquidity premium — the price you pay for the option to change your mind.

Is 6 percentage points worth it? The answer depends on your expected instance evolution over 3 years. AWS releases a new instance generation roughly every 18 months. A 3-year Standard RI signed in July 2026 on m6i locks you into m6i until July 2029. When m7i launches in late 2026 at 20% better price/performance, your m6i RI is still running at the old price. With a Convertible RI, you exchange the m6i reservation for m7i — the discount percentage stays the same (34%), but it applies to the new instance's on-demand price. If m7i costs the same as m6i ($0.192/hr) but delivers 20% more throughput, your effective cost per unit of work drops by 20%. With a Standard RI, you stay on m6i at 40% off, paying $0.1152/hr for 20% less throughput — effectively $0.144/hr per unit of work. Convertible at 34% off on m7i = $0.1267/hr for the same unit of work — 12% cheaper in throughput-adjusted terms despite the nominally smaller discount.

The liquidity premium is not a cost. It's an option. Options have value when the underlying asset (instance performance) is volatile. Over the last 3 years, instance performance per dollar has improved by approximately 40% cumulatively (m5→m6i: 15% perf/$ improvement, Graviton3→Graviton4: 25% improvement). A Convertible RI lets you capture those performance gains during the commitment term. A Standard RI locks you out of them. The question is not "is 40% off better than 34% off?" The question is "what is the expected performance/$ improvement over the next 3 years, and does the Convertible's flexibility to capture it exceed the 6-point discount gap?"

For CPU instances (m6i/c6i/r6i), the expected improvement is moderate — maybe 15-20% over 3 years. The Convertible premium may not pay back. For GPU instances (p4d/p5), the expected improvement is massive — NVIDIA's roadmap has H200→B100→B200 over the same period, each generation delivering 2-4× performance. A Standard RI on a GPU instance family is a bet against Moore's Law. In GPU land, Convertible is not optional. It's the only way to avoid locking into an instance generation that will be obsolete in 12 months.

Savings Plans vs RIs — The Tax Difference That Matters In Mixed Fleets

Compute Savings Plans and Reserved Instances look equivalent on a pricing table. A 3-year Compute Savings Plan gives ~37% discount versus a 3-year Standard RI's ~40%. The 3-percentage-point gap seems small. It's not. The structural difference is that Savings Plans apply to a dollar commitment ($X/hour of spend) while RIs apply to a capacity commitment (N instances of type T in AZ Z). In a mixed-fleet strategy, this structural difference creates a tax arbitrage.

Here is the scenario that exposes the gap. You run a mixed fleet: 6 m6i.xlarge RIs (the base), 8 spot instances (the burst), 2 on-demand (the buffer). You also run 4 c6i.2xlarge instances for a batch processing tier — entirely on spot. Under a pure RI strategy, you would need separate RI purchases for m6i and c6i — two reservation types, managed separately, with no fungibility. Under a Savings Plan, the dollar commitment covers both instance families automatically. The $X/hour of Savings Plan spend applies first to your highest on-demand rate instances (c6i.2xlarge at $0.3392/hr), then spills to m6i.xlarge ($0.192/hr). This automatic optimization means your Savings Plan discount always applies to your most expensive instances first — no manual RI exchange or modification required.

The tax difference emerges when you change your fleet composition. RIs are bound to specific instance types. If you migrate from m6i to m7i, the m6i RIs become unused — you keep paying for 8,760 hours/year on instances you no longer run. You can sell them on the RI Marketplace, but the resale market is thin and you typically recover 50-70% of the remaining value. A Savings Plan automatically applies to m7i at the same discount rate — no resale, no unused commitment, no friction. The 3 percentage points of extra discount from a Standard RI are compensation for bearing the instance-evolution risk. Whether that compensation is adequate depends on how stable your instance fleet is over 3 years. Teams running a single instance family with predictable steady-state capacity do well with RIs. Teams running heterogeneous fleets with ongoing optimization cycles do better with Savings Plans — the flexibility is worth more than the 3-point discount gap.

The killer feature of Savings Plans that nobody talks about: they apply across regions at no penalty. An RI in us-east-1 cannot be used in us-west-2. A Savings Plan dollar commitment can be consumed by instances in any region. For multi-region architectures, this eliminates the regional RI fragmentation problem — instead of buying RIs in each region and hoping your capacity distribution stays stable, one Savings Plan covers all regions and follows your capacity wherever it moves.

Cross-Region Spot Arbitrage — When It Works And When Data Transfer Kills It

Spot prices vary by region and AZ. An m6i.xlarge spot instance might cost $0.0576/hr in us-east-1a (70% off $0.192 on-demand) and $0.0768/hr in us-west-2a (60% off $0.192 on-demand). Moving a workload from us-west-2 to us-east-1 saves $0.0192/instance-hour in compute — $168/instance-year. For a 10-instance fleet, $1,680/year. The question is whether the data transfer cost of moving the workload eats the compute savings.

Inter-region data transfer costs $0.02/GB (for traffic leaving us-east-1 to us-west-2). A stateless web tier serving 200 rps at 50 KB average response size generates approximately 864 GB/day of egress. Moving that tier from us-west-2 to us-east-1 adds $17.28/day in inter-region transfer costs for the requests that still need to reach us-west-2 users. That's $6,307/year. The compute savings of $1,680/year are wiped out 3.7× over by data transfer. Cross-region spot arbitrage only works for compute-heavy, data-light workloads where the workload can be relocated without dragging its data dependency along.

Cross-AZ spot arbitrage within the same region is free for data transfer. And spot pools vary by AZ. A fleet spread across 3 AZs, each drawing from its own spot pool, has a lower probability of simultaneous full-fleet termination than a fleet concentrated in 1 AZ. This is not arbitrage — you cannot simultaneously buy in one pool and sell in another. It's diversification. By intentionally spreading spot instances across multiple AZs and multiple instance types within the same family (m6i.xlarge + m6i.2xlarge + m6i.4xlarge), you reduce the probability that all spot pools exhaust at the same instant. The trade is: slightly higher operational complexity (managing a heterogeneous spot fleet) for a significantly lower effective interruption rate. The cost is zero additional dollars — just the engineering effort to configure your autoscaling group with multiple instance types and AZs. A single CloudFormation parameter change that costs 30 minutes of engineering time can reduce your spot interruption risk by 50-70%. The portfolio diversification that financial markets figured out in 1952 applies to cloud capacity markets too.

Concrete Steps: What To Do Before Signing The Next RI Commitment

1. Compute your specific breakeven interruption rate. Do not use a rule of thumb. Use your actual instance type, your actual monthly runtime hours, your actual spot discount in your region and AZ, and your actual recovery time per interruption. The RI vs Spot Breakeven Calculator produces this number from your real parameters. If your actual spot interruption rate for this instance family in your AZ is below the breakeven, spot is cheaper — period. If it's above, RI is cheaper. If it's close (within 20%), the decision is a risk question, not a pricing question.

2. If you are not running 8,760 hours/year per instance, do not buy an RI. RIs bill for 24/7/365 regardless of your actual usage. If your workload runs 400 hours/month (batch processing), you are paying for 330 hours/month of idle capacity. The on-demand equivalent of those 330 idle hours is $63.36/instance-month at $0.192/hr. The RI discount would need to be >33% just to break even with on-demand for a 400-hour workload — and spot at 65% off pays only for the hours you use, making it functionally impossible for RI to win at low utilization. An RI at less than ~70% utilization is always worse than on-demand.

3. Diversify spot instances across AZs and instance types. If you run spot, run it across at least 3 AZs and 2 instance types within the same family. This is the cheapest insurance against spot pool exhaustion. The operational overhead is negligible (autoscaling groups support mixed instances natively) and the risk reduction is significant. For a workload that cannot tolerate any interruption, do not run spot — period. No amount of diversification eliminates the 2-minute warning. It only reduces the probability of simultaneous termination.

4. Match the commitment instrument to your instance evolution risk. Standard RI for stable, predictable, single-family fleets that will not change for 3 years. Convertible RI for GPU instances or fleets where you expect to upgrade instance generations within the term. Compute Savings Plan for heterogeneous, multi-family fleets with ongoing optimization — the flexibility premium costs 3 points of discount and is worth it for any fleet that changes instance types more than once every 18 months.

5. Build the mixed-fleet model, not the binary choice. Allocate base capacity to RIs/Savings Plans (the minimum you will always run), burst capacity to spot (the variable you can afford to lose temporarily), and a small buffer to on-demand (insurance against coincident spot drought + traffic peak). The percentages are specific to your workload shape. A stateless web tier might run 30% RI / 65% spot / 5% on-demand. A stateful database cluster should run 100% RI. There is no one-size-fits-all allocation. The framework is what matters — not the specific percentages.

Frequently Asked Questions

Methodology & Disclosure

Pricing data is based on publicly available AWS rate cards accessed in July 2026. On-demand pricing: m6i.xlarge at $0.192/hr (us-east-1). RI discounts: 3-year Standard ~40%, 3-year Convertible ~34%, 1-year Standard ~31% (actual discounts vary by instance family and payment option — upfront vs partial-upfront vs no-upfront). Spot discounts: modeled at 50-70% based on trailing 3-month average for m6i family in us-east-1; actual spot prices fluctuate continuously. Compute Savings Plan rates: modeled at 37% (3-year) and 28% (1-year) based on published AWS rates. Inter-region data transfer: $0.02/GB (standard rate for most region pairs). All scenario computations use us-east-1 pricing. Regional multipliers are not applied unless explicitly stated.

Breakeven interruption rate calculations use the formula published in our RI vs Spot Breakeven Calculator. The formula assumes spot interruptions are independent events — in practice, interruptions are correlated within an AZ and the calculator's output should be interpreted as a directional breakeven, not a probabilistic guarantee. Workloads with high tail-risk sensitivity should apply a correlation multiplier to the breakeven rate or run a mixed-fleet strategy with on-demand buffer capacity.

Disclosure: jslet is an independent research project. This analysis was produced using our own RI vs Spot Breakeven Calculator and publicly available AWS pricing data. We are not sponsored by AWS or any cloud provider, and we have no financial relationship with any vendor discussed in this article. No AWS employees reviewed or contributed to this analysis.

References & Further Reading

1. AWS (2026). "Amazon EC2 Reserved Instances Pricing." aws.amazon.com
2. AWS (2026). "Amazon EC2 Spot Instances Pricing." Spot Advisor and spot price history by instance type, region, and AZ. aws.amazon.com
3. AWS (2026). "Savings Plans — Compute and EC2 Instance Savings Plans." Pricing model comparison, discount rates, and eligibility rules. aws.amazon.com
4. AWS (2026). "Spot Instance Interruptions." Spot interruption frequency data, the 2-minute warning mechanism, and instance rebalance recommendations. docs.aws.amazon.com
5. AWS (2026). "RI Marketplace — Selling Reserved Instances." Resale mechanics, pricing, and the secondary market for unwanted RI commitments. docs.aws.amazon.com
6. AWS (2026). "EC2 Instance Type Specifications." Per-instance-family on-demand pricing, vCPU/memory/network specs for all instance families referenced in this article. aws.amazon.com
7. AWS (2026). "Data Transfer Pricing." Inter-region, inter-AZ, and internet egress pricing tables. AZ→AZ within region is free; inter-region is $0.02/GB; internet egress tiers from $0.05-0.09/GB. aws.amazon.com
8. NVIDIA (2026). "Data Center GPU Roadmap — H200, B100, B200." GPU architecture cadence, performance projections, and cloud instance availability timelines. nvidia.com
9. jslet (2026). "GPU Training Cost Estimator — Compare AWS, GCP, Azure, Lambda Labs." Spot vs on-demand vs reserved GPU pricing across 4 clouds. jslet.com/gpu-training-cost

📜 Copyright & Attribution

📡 Enjoyed this? When a "40% discount" is actually a 36-month cash advance to AWS with an embedded interest rate, the pricing page is a derivative contract. RSS covers one pricing-model reality check per week. No vendor sponsors. No tracking. RSS Feed → | More options →