Forty-four thousand. That is the number of controlled failure experiments Amazon ran on its own infrastructure before Prime Day 2026. More than six times as many as the year before.
Shutting down components. Cutting connections. Overloading services on purpose. All with one goal: finding where the system breaks before millions of shoppers do.
It sounds like a paradox, but it makes sense. Prime Day 2026 ran from June 23 to 26. When a platform packs several weeks' worth of traffic into four days, the question is no longer whether something will fail. The question is what will happen when it does, and how long the system will take to recover without anyone noticing.
What happens behind a single click
For the shopper, the sequence is simple: search, compare, add to cart and pay. For the platform, each of those steps triggers dozens of independent services.
One checks the catalog. Another checks real-time stock across different warehouses. Another calculates the price, which can change with demand. Another validates the payment and runs it through fraud detection models. Another reserves the item so nobody else can buy it. And another decides which warehouse the package will ship from and when it will arrive.
All of that has to respond before the user loses patience. And it has to do so without contradictions. If the system sells the last unit twice, or charges for an order that is never recorded, the problem stops being technical. It becomes a customer who never comes back.
Scale helps explain the pressure. During Prime Day 2026: •
- The DynamoDB database handled more than 59 trillion requests in four days, peaking at 192 million per second with millisecond response times.
- The Lambda service ran more than 2.3 trillion functions a day.
- The CloudWatch monitoring system recorded more than 2 quadrillion measurements a day.
One detail is easy to overlook. That same database doesn't just support the website: it also serves all of Amazon's fulfillment centers. The click and the package leaving the warehouse depend on the same infrastructure. If that layer stops, it's not just a web page that goes down. The whole chain stops.
It's worth remembering something the word "cloud" tends to hide. Every one of those requests runs on a physical server, inside a specific building, connected to a power supply and a cooling system. The cloud is data centers. When traffic multiplies, so do the power they draw and the heat they generate.
The last figure on the list is the most revealing. The company that best masters scale doesn't just add servers: it measures absolutely everything. More than 2 quadrillion measurements a day means every component, every service and every resource is under constant observation. Without that visibility, the 44,000 experiments would be useless, because there would be no way of knowing what failed or why.
Demand peaks: the challenge of a few days a year
The other big challenge is that load isn't stable. Prime Day, Black Friday, Cyber Monday and seasonal sales create peaks that multiply normal traffic within hours.
Infrastructure is sized for those peaks, not for the average. That means reserving power, cooling capacity and space for just a few days a year. And knowing precisely how much headroom is left before the surge arrives, because finding out during the campaign is too late.
Efficiency has become part of the answer. In 2026, up to 49% of the compute Amazon.com used in its cloud ran on Graviton, its own in-house processors. The logic is simple: if each operation uses less energy, the same data center can handle more operations without increasing available power or cooling.
AI adds another layer. Conversational shopping assistants are already part of the experience. To run its Rufus assistant during Prime Day 2025, Amazon deployed more than 87,000 of its own AI chips. Each query to an assistant like this uses far more compute than a traditional search. And that compute translates into something very physical: higher power density per rack, more heat and greater demands on cooling.
Your Black Friday is critical too
Most companies are not Amazon. But their Black Friday is just as critical to their business. And many face it with on-premises or hybrid infrastructure that doesn't have thousands of engineers behind it.
The difference lies in what they know about that infrastructure before the peak:
- How much power is still available in each room.
- Which racks are close to their thermal limit.
- What happens if a circuit fails mid-campaign.
- How long it takes to add capacity if forecasts fall short.
Amazon answers those questions by rehearsing failure 44,000 times and measuring every second of operation. Many organizations only find out the answers on the day of the campaign itself.
Could your team say today how much capacity your data center has left if demand doubled tomorrow?
Sources
– AWS News Blog, “All the numbers: Amazon Prime Day 2026 powered by AWS” (6 de octubre de 2026).
– AWS News Blog, “AWS services scale to new heights for Prime Day 2025: key metrics and milestones” (agosto de 2025).