Amazon says cloud capacity is sold out through 2027
On November 5, Reuters reported that Amazon CEO Andy Jassy said the company's cloud capacity is effectively sold out through 2027. The Hacker News thread was small but the comments were sharp.
Two things can be true at once. First, AI labs have genuinely exhausted available compute, memory is strained, and data center construction is running at full tilt. Capacity scarcity is real. Second, scarcity is also a convenient thing for a cloud vendor to announce, because it justifies price discipline and makes customers commit earlier and longer.
One commenter noted out-of-capacity errors in dynamic, non-AI workloads over the last six months. That is the part that should worry normal customers. If you are not an AI lab and your workload still gets capacity errors, the scarcity is not just a frontier-lab problem — it is a scheduling problem that has leaked into ordinary compute.
What this means if you run on AWS
- Reserved instances and savings plans become more valuable, not less. Commit early.
- Multi-region and multi-cloud stop being nice-to-haves and become capacity strategy.
- Spot and on-demand for bursty workloads get riskier. Have a fallback region.
- If you self-host models, your own GPU fleet is suddenly a competitive asset, not a cost center.
The cynical read — "used car salesman tactics" — is not entirely wrong either. But it does not matter whether the announcement is marketing or reality if the result is the same: capacity is tight, commit early, design for scarcity.
Source: Reuters / Hacker News (2026-11-05).
Further context
Tagging is boring and essential. Without consistent tags, cost attribution is guesswork.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Tagging is boring and essential. Without consistent tags, cost attribution is guesswork.
Backups are not disaster recovery. Test your restores or you do not have backups.
Cloud cost is the new technical debt. Teams that do not track spend by service and team end up with bills that nobody can explain and nobody can reduce.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Backups are not disaster recovery. Test your restores or you do not have backups.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Cloud cost is the new technical debt. Teams that do not track spend by service and team end up with bills that nobody can explain and nobody can reduce.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Tagging is boring and essential. Without consistent tags, cost attribution is guesswork.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
IAM is the foundation of cloud security. Get it wrong and nothing else matters. Get it right and most attacks become much harder.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Backups are not disaster recovery. Test your restores or you do not have backups.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
Cloud cost is the new technical debt. Teams that do not track spend by service and team end up with bills that nobody can explain and nobody can reduce.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Serverless is great until it isn't. Cold starts, concurrency limits, and vendor lock-in are real costs that show up at scale.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
IAM is the foundation of cloud security. Get it wrong and nothing else matters. Get it right and most attacks become much harder.
IAM is the foundation of cloud security. Get it wrong and nothing else matters. Get it right and most attacks become much harder.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
The shared responsibility model means the cloud secures the infrastructure and you secure everything else. Most breaches are in the everything else.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
Backups are not disaster recovery. Test your restores or you do not have backups.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Backups are not disaster recovery. Test your restores or you do not have backups.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
The cloud is just someone else's computer, and that someone else is very good at charging you for convenience. The question is always whether the convenience is worth the markup.
Cloud cost is the new technical debt. Teams that do not track spend by service and team end up with bills that nobody can explain and nobody can reduce.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Spot instances are free money if your workload can tolerate interruption. Most batch and async workloads can.
Networking is where cloud bills hide. Data transfer charges between regions and out to the internet add up fast.
Observability is the difference between knowing something is broken and knowing why. Invest in it before you need it.
Reserved instances and savings plans reward commitment. If your workload is stable, not using them is leaving money on the table.
Infrastructure as code is not optional at scale. Clicking in the console is fine for learning and unacceptable for production.
Backups are not disaster recovery. Test your restores or you do not have backups.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.
Managed databases trade operational burden for cost and flexibility. Know what you are giving up.
Multi-region is not multi-cloud. Running in two AWS regions is redundancy. Running on AWS and GCP is strategy, and it is expensive.
Tagging is boring and essential. Without consistent tags, cost attribution is guesswork.
Tagging is boring and essential. Without consistent tags, cost attribution is guesswork.
Autoscaling is not a substitute for capacity planning. It reacts to load; it does not predict it.