As a software practitioner, “192 million requests per second” or “213 million messages per second ” are not numbers I can really understand. Yes, they are big. But how big? And more usefully: how did we get there?
Did the architecture change over the years? Did end users buy more of everything year on year? Did we see some trends?
AWS has published a Prime Day infrastructure post almost every year since 2016, with 2018 as the only gap. Or at least I could not find the post. If you read them back to back, they’re one of the few public datasets tracking the same (or at least similar workload) at the same company over a decade. You can find the reference to the ten posts at the end of this article.
Some Recent Numbers to Start
No surprise that the backbone combines CloudFront as a CDN, EC2/Lambda for CPU, ElastiCache, and DynamoDB and Aurora as databases.
According to the latest article, this year DynamoDB served over 59 trillion requests across the four-day event, peaking at 192 million per second. Spread evenly, that’s about 171 million requests per second sustained for 96 hours. The peak is only ~12% above the average, a narrow gap. However, this comparison says little about short-term traffic variability: averaging across 96 hours can conceal substantial differences between individual hours, regions, and shopping periods. That’s a substantial increase from the “56 billion extra requests” reported for Prime Day a decade ago.
Let’s understand the scale first: DynamoDB’s default per-table read throughput quota is 40,000 read capacity units per second. Although requests and capacity units are not directly comparable, this provides a rough reference point for the scale of Amazon’s reported peak. But how did those numbers change over the years?
Peak Per Second
Totals are hard to compare because the event changed shape: 30 hours in 2017, 66 hours in 2020 and 2021, two days from 2022 to 2024, and four days in 2025 and 2026. But peak throughput per second is still probably the easiest way to compare usage, even if the distribution of purchases is clearly different and flatter the longer the event lasts. According to the available posts, these are the peak throughput per second in millions for DynamoDB, SQS, and Kinesis. Not every number is available every year.
| peak/sec (M) | 2017 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
| DynamoDB | 12.9 | 45.4 | 80.1 | 89.2 | 105.2 | 126 | 146 | 151 | 192 |
| SQS | 47.7 | 70.5 | 86 | 166 | 213 | ||||
| Kinesis | 807 | 988 |
How has throughput grown? DynamoDB grows at roughly 35% a year across the full series, but between 2017 and 2020, the reported peak increased approximately sixfold; subsequent annual increases were generally smaller. Missing years and possible changes in reporting definitions limit the precision of these comparisons. Interestingly, this year jumped 27%. A real jump or a mismatch in the numbers? SQS tells a similar story and overtakes DynamoDB in raw message rate.
Volume and footprint
Other metrics are equally interesting, but some are not available for the first few years. What was the peak in EBS I/O ops? How much data has been transferred every day? What about serverless usage?
Numbers in brackets are normalized by day or derived from other available metrics.
| Metric | 2019 | 2021 | 2023 | 2024 | 2025 | 2026 |
| EBS I/O ops (peak/day) | 2.1T | 11.1T | 15.35T | 20.3T | 24.8T | |
| EBS data transferred/day | 185 PB | 614 PB | 764 PB | ~1 EB | >1 EB | |
| Lambda invocations/day | (1.3T) | 1.7T | 2.3T | |||
| ECS tasks on Fargate/day | (62.8M) | (77.2M) | 107.1M | 158.3M | ||
| ElastiCache requests/day | 1 Q | 1.5 Q | 2.3 Q | |||
| Aurora data stored | (1,595 TB) | 2,140 TB | 2,978 TB | 4,071 TB | 5,491 TB | |
| CloudTrail events | 830B | 976B | 2.5T | 3.6T | ||
| GuardDuty log events/hr | 6T | 8.9T | 14.08T |
Two things jump out. ElastiCache is the real monster, with 2.3 quadrillion requests a day, which works out to about 27 billion per second (think about that number carefully), two orders of magnitude above DynamoDB and Aurora. Caching can reduce pressure on backend databases, but the published figures alone do not establish how much database traffic it absorbs or how the two tiers compare in cost.
Operational telemetry is substantial in its own right: CloudWatch processed 2.15 quadrillion metric observations a day in 2026, while GuardDuty’s reported event volume also increased significantly.
What Really Changed Over the Years
Over the years, the architecture changed in multiple directions. Let’s identify some major trends:
Graviton: Jeff Barr started to mention Graviton in 2021, when it “supported 12 core retail services”. By 2026, it carried up to 49% of the Amazon.com EC2 compute. The 2022 post shows the payoff most directly: 12% more normalized instances for only 7% more server-equivalent footprint. And numbers get better in the following years.
Serverless: while AWS talks less about serverless computing, it has become a significant part of Amazon’s infrastructure. AWS reported 1.7 trillion Lambda invocations per day in 2025 and 2.3 trillion in 2026, an increase of approximately 35%. That is roughly 27 million invocations per second. For comparison, the default Lambda concurrency quota is 1,000 per region. Fargate tasks grew significantly too.
Databases: Aurora picked up DSQL alongside PostgreSQL and MySQL in 2026, but no specific numbers are shared regarding the new kid on the block. Instance counts have quietly disappeared from the posts, with AWS no longer reporting Aurora instance counts after 2024, when 6,311 instances running the PostgreSQL-compatible and MySQL-compatible editions processed more than 376 billion transactions, stored 2,978 terabytes, and transferred 913 terabytes.
Resilience: Fault Injection Service experiments went from 733 to 6,800 and then 44,000 in two years, a 60x increase, likely because FIS moved into CI/CD pipelines: the numbers suggest that deliberate failure testing has become a more prominent part of Amazon’s resilience practices.
AI inference: the latest two posts put generative AI in the opening paragraphs rather than the metrics list, mentioning Rufus, Alexa+, and AI shopping guides. SageMaker inference requests went from 145 billion in 2024 to 626 billion in 2025, a 4.3x jump. Amazon deployed over 87,000 Inferentia and Trainium chips to power Rufus, up from over 80,000 the year before.
Caveats and Challenges
Let’s remember that these are super interesting yearly posts but still marketing posts, not benchmarks. There are clear trends, but there are also well-defined definitions that shift between years, while some numbers clearly do not add up.
Some metrics appear once and vanish, and some numbers contradict themselves: for example, CloudFront reported over 3 trillion requests in 2025 but 2.1 trillion in 2026, describing this as a 5% increase. Quoting verbatim from the two articles, bold is mine:
2025
Amazon CloudFront – Amazon CloudFront, a content delivery network (CDN) service that securely delivers content with low latency and high transfer speeds, delivered over 3 trillion HTTP requests during the global week of Prime Day 2025, a 43 percent increase in requests compared to Prime Day 2024.
2026
Amazon CloudFront – Amazon CloudFront delivered over 2.1 trillion HTTP requests during the global week of Prime Day 2026, a 5 percent increase in total requests compared to Prime Day 2025.
The discrepancy illustrates why these figures should be treated as indicators of broader trends rather than audited, directly comparable measurements.
Conclusions
Over the last five years, DynamoDB’s peak grew 2.2x and EBS I/O 2.2x, while SQS grew 4.5x. Fast, but not the exponential curve the headline numbers imply. Amazon’s infrastructure has evolved alongside growing workloads: Graviton accounts for a larger share of EC2 compute, Lambda usage has increased, and fault-injection testing has expanded.
References
- 2016 – How AWS Powered Amazon’s Biggest Day Ever
- 2017 – Prime Day 2017 – Powered by AWS
- 2019 – Amazon Prime Day 2019 – Powered by AWS
- 2020 – Amazon Prime Day 2020 – Powered by AWS
- 2021 – Prime Day 2021 – Two Chart-Topping Days
- 2022 – Amazon Prime Day 2022 – AWS for the Win!
- 2023 – Prime Day 2023 Powered by AWS – All the Numbers
- 2024 – How AWS powered Prime Day 2024 for record-breaking sales
- 2025 – AWS services scale to new heights for Prime Day 2025: key metrics and milestones
- 2026 – All the numbers: Amazon Prime Day 2026 powered by AWS