CloudCostDB Blog
Deeper looks at how AWS pricing and instance types got where they are, with numbers to back it up.
LatestNew feature
RDS per-snapshot billing: the line item is a share, not a price tag
Since November 2, every RDS snapshot gets its own line in the Cost and Usage Report. The rate didn't move ($0.095 per GB-month in us-east-1), but the free backup allowance is spread pro rata across every snapshot, so a $7.92 line can hide a $23.75 saving. Budgets and chargeback rules that match a DB instance ARN stop seeing snapshot spend.
November 7, 2026
2026
- November 6, 2026The $6.88 hour: what a GPU actually costs on EC2, and when Bedrock tokens winAn H100 hour runs $6.88 on-demand in us-east-1, and at the most expensive Bedrock text model that same $6.88 buys about 50,000 output tokens. The price ladder from T4 to B300, VRAM per dollar, and the utilization math that decides whether you rent silicon or pay per token.
- November 2, 2026ElastiCache data tiering lands on r8g hardware: 5x the cache room for 52% more per hourNew instance typesData tiering now runs on r7gd and r8gd nodes, but only in Spain and Hyderabad, and only for Valkey. A tiered cache.r8gd.16xlarge costs 1.52x the memory-only node and holds about 5x the usable data, cutting the per-usable-GiB price ~71%. Everywhere else, the older r6gd deal still runs a 90% premium.
- October 25, 20262026 prices for a 2017 machine: why I built CloudCostDBA c5.xlarge and a c8a.xlarge have the same vCPUs and the same memory on paper. The c8a costs 27% more an hour and scores about 2.5 times higher on CPU benchmarks. Instance names hide a generation gap, and the old generation is often the expensive one. Why that happens, how to read past the spec sheet, and why this site puts benchmarks next to prices.
- October 23, 2026Cache the DB they said. Here's when that triples the billDynamoDB sells reads by the request at $0.125 per million; RDS and ElastiCache sell boxes you pay for even when they sleep. Where the curves cross at 27 reads/sec and 29% utilization, and why a cache node costs as much as the database behind it.
- October 15, 2026The Bedrock tier ladder: batch and flex at 50% off, priority at +75%Bedrock charges $0.40 per million input tokens for batch or flex on Nova Pro and $1.40 for priority, against $0.80 standard. On a 10M-input, 2M-output month that is $7.20 to $25.20 for identical model weights. Tier coverage, prompt-cache breakeven, and the models where none of it is available.
- October 12, 2026The 13% flexibility premium: Compute Savings Plans vs EC2 Instance Savings PlansAcross 1,360 us-east-1 Linux instance types, the median EC2 Instance Savings Plan rate is 86.9% of the Compute rate at one year and 86.8% at three years. The 13% gap is what AWS charges for the right to change family and region. It is not uniform: c7g carries 8-9%, m7i 10-11%, r7i 9-13%, and a 3-year p5 plan carries a 41% premium.
- October 9, 2026Aurora's 21% compute premium is a backup discountAurora MySQL charges $0.580/hr for a db.r6i.xlarge against $0.480 on RDS MySQL, 21% more for the same vCPU and memory. Charged backups run $0.021 per GB-month against $0.095, a 4.5x gap, and Aurora storage is 13% cheaper. The breakeven sits at 779 GB of charged backup for a 1 TB database, or 409 GB of live storage per vCPU once backups reach 40% of your data.
- October 6, 2026The RDS license tax: SQL Server Enterprise costs 5.2x MySQL on identical hardwareOn db.r6i.xlarge in us-east-1, RDS for SQL Server Enterprise lists at $2.501 an hour against $0.480 for MySQL, a 5.21× premium for the same 4 vCPU and 32 GiB. The other $2.021 is license, about $0.505 per vCPU-hour, billed on every node in the cluster. How to read the six license_model labels in the RDS price list, why the $0.466 Oracle BYOL row is not the bargain it looks like, and when the premium is large enough to justify changing engines.
- October 3, 2026io2 vs gp3: you are paying 6.7x for durability, not IOPSAt 1 TB and 10,000 IOPS, gp3 costs $115 a month and io2 costs $775, with the same IOPS provisioned on both. io2's tiered rates never close the gap: its cheapest tier is still 6.4x gp3's marginal rate. The $660 buys a 99.999% durability rating and a 99.9% performance-delivery guarantee, insurance against a failure the workload may never see.
- September 30, 2026The ElastiCache engine tax: Valkey is 20% cheaper than Redis in all 37 regionsElastiCache lists Valkey at exactly 0.8x the Redis price on all 1,904 node-region pairs where both appear, and Memcached at 1.0x with no replication, persistence, or data structures. A six-node cache.r6g.2xlarge cluster runs $34,521 a year on Valkey and $43,152 on the other two.
- September 25, 2026The Amazon EFS file storage tax: why shared NFS costs 4x more than EBSAmazon EFS Standard charges $0.30 per GB-month in us-east-1, nearly four times the price of EBS gp3 and thirteen times S3. On top of that, Elastic Throughput charges $0.03 per GB read and $0.06 per GB write. The mechanics of EFS storage tiers, throughput bottlenecks, and when to avoid POSIX file systems.
- September 18, 2026DynamoDB on-demand vs provisioned: the 29% utilization thresholdOn-demand DynamoDB bills $0.625 per million writes and $0.125 per million reads. A provisioned WCU costs $0.4745 a month and serves 2.63 million writes. Above 29% average utilization, provisioned is cheaper. The crossover math and Standard-IA economics.
- September 18, 2026The RDS premium: what managed MySQL costs over the box it runs ondb.m5.xlarge runs MySQL on the same 4 vCPU and 16 GiB as m5.xlarge, at $0.342 an hour against $0.192. Where the 78% managed fee goes, why Multi-AZ is nearly free next to doing HA yourself, and why the engine you pick costs more than the management.
- September 13, 2026The S3 retrieval tax: when cheaper storage classes cost you moreStandard-IA is 46% cheaper per GB than S3 Standard, until you read anything. Every class below Standard trades a storage discount for per-GB retrieval fees, minimum durations, and request costs up to 25x higher. The breakeven math for all six classes, from our dataset.
- September 2, 2026Five architectural rules to cut your AWS networking billCompute is straightforward to budget, but data transfer, NAT Gateways, and public IPv4 fees produce persistent surprise bills. Five practical rules to eliminate network waste without compromising reliability.
- August 31, 2026How the AWS geography tax splits by serviceWe priced the same reference stack in 13 AWS regions: an m6i.xlarge, a MySQL RDS instance, a Redis cache, 1 TB of gp3, 5 TB of S3 Standard, and 60 DynamoDB write capacity units. Every service carries its own regional markup, and they disagree with each other.
- August 1, 2026Graviton versions: from experiment to defaultGraviton1 (2018) was an experiment nobody was sure about. Graviton2 made Arm a mainstream choice, and each generation since has shifted the price-performance math. The story of all five generations, with numbers from our data.
- May 18, 2026The burstable CPU credit trap: when T instances cost more than MT2, T3, and T4g instances look like the cheapest compute on AWS until sustained CPU load triggers T Unlimited surcharges at $0.05 per vCPU-hour. How the credit accounting works, when credit burn breaks the budget, and the exact crossover points to M-series.
- April 2, 2026ALB vs NLB pricing: LCU math and the cost of Layer 7 load balancingApplication Load Balancers and Network Load Balancers share an identical base rate of $0.0225/hour, but their capacity unit calculations diverge dramatically under high-throughput traffic. The math behind LCUs vs NLCUs, gRPC/WebSocket costs, and when to terminate at Layer 4.
- February 14, 2026The NAT Gateway trap: how $0.045 per gigabyte inflates your cloud billManaged NAT Gateways are AWS's default recommendation for private subnet egress, but at $0.045/hour plus $0.045 per processed gigabyte, high-throughput architectures routinely incur surprise bills in the thousands. How the pricing works and how to bypass it.
2025
- November 10, 2025The AWS geography tax: why São Paulo costs 59% more than VirginiaIdentical virtual machines running identical silicon carry wildly different hourly prices across AWS regions. A breakdown of the global price spread across 38+ regions, why certain locations charge a massive premium, and the math behind multi-region architectures.
- September 20, 2025The three-way CPU race: how AMD flipped the script on EC2The old rule of thumb was simple: Intel was default, AMD was 10% cheaper, and Graviton was 20% cheaper. In the 7th and 8th generations, AMD Genoa turned that pricing model upside down. Benchmarks, clock speeds, and real cost per compute unit.
- August 4, 2025Cross-AZ data transfer: the silent tollbooth in multi-AZ architecturesHigh availability across Availability Zones is an AWS best practice, but inter-AZ network traffic costs $0.01 per GB in each direction ($0.02/GB round-trip). How microservices, Kafka replication, and Kubernetes clusters rack up thousands in hidden data transfer bills.
- June 15, 2025The gp2 to gp3 storage tax: why millions of volumes never migratedIn December 2020 AWS launched gp3 at a 20% storage discount with decoupled IOPS and throughput. Years later, unmigrated gp2 volumes still generate huge unintended cloud spend. The history, the baseline math, and why AWS left migration up to you.
- April 10, 2025The RAM per dollar curve: how C, M, R, and X families price memoryCompute-optimized, general-purpose, and memory-optimized instances charge drastically different effective rates per gigabyte of memory. The math behind the memory curve, why buying RAM on M-instances wastes money, and sizing formulas for memory-bound workloads.
- January 22, 2025Local NVMe vs EBS: why the 'd' suffix is AWS's best storage dealAdding local NVMe SSDs to EC2 instances costs pennies per hour while delivering hundreds of thousands of IOPS and sub-100µs latency. How instance storage economics compare against gp3 and io2, and the architectural patterns for ephemeral storage.