How to Secure EC2 Capacity on AWS When Demand Is Tight

A practical guide to reviewing your consumption, reserving capacity on AWS, and not relying solely on On-Demand now that committed capacity is scarce.

Software engineering

Andy Jassy confirmed that AWS's committed capacity is largely sold out through 2027, with part of 2028 already allocated, according to Reuters. At the same time, the company states it leaves room for On-Demand, EC2, and Bedrock. This guide doesn't aim to solve an AWS problem, but rather to outline what you can do to avoid running out of compute capacity.

It's worth clarifying from the start: analyst David Linthicum argues that the headline exaggerates and that it's still possible to allocate GPUs from the console. However, several professionals mention, in comments collected by InfoQ, that it's recently been difficult to get common EC2 on-demand instances. Let's prepare without dramatizing.

What you need

You'll need read access to your accounts' billing and inventory, a rough idea of which workloads are critical (production, CI, training), and decision-making authority over spending commitments. If you use multiple accounts and regions, the documentation on capacity reservations and Capacity Blocks for ML will serve as a reference to understand exactly what AWS offers.

1. Inventory which workloads require guaranteed capacity and which can tolerate waiting or retrying, for example, a CI pipeline versus a production cluster.
2. Measure your current buffer: note how many times instance launches fail and for which types, zones, and at what times.
3. Decide what portion of your stable baseline justifies a capacity reservation in a specific availability zone.
4. For GPU workloads with a known date, consider Capacity Blocks for ML, keeping in mind that their rates increased by about 15% in January.
5. Diversify instance types and zones in your auto-scaling groups so that a failure of one type doesn't stop the service.
6. Review the actual usage of your reserved capacity monthly and adjust, to avoid paying for unused capacity.

Detecting the problem before it hurts

The typical symptom professionals describe is a failure to launch instances due to lack of capacity. If you have the CLI configured, you can list your active reservations and see what you actually have secured:

aws ec2 describe-capacity-reservations --region eu-west-1 --query 'CapacityReservations[].{Tipo:InstanceType,Zona:AvailabilityZone,Total:TotalInstanceCount,Libres:AvailableInstanceCount}' --output table

The "Libres" (Free) column compared to "Total" tells you how much of your reserved capacity you are utilizing. To view this at scale, AWS offers EC2 Capacity Manager, which, according to the source, aggregates data across accounts and regions and reports on the relationship between used and unused reserved capacity.

Measuring the actual usage of reserved capacity is the basis for any capacity decision.

The opposite risk: reserving out of fear

Ian Rae warned in the discussion that if everyone reserves resources just in case, it falls into a pattern contrary to the philosophy of the cloud. This is a real risk: an unused reservation is fixed money with no return. Reserve only the stable baseline you already consume and leave variable workloads for On-Demand, with retries and alternatives.

Nor is this an exclusive AWS problem: comments mention similar restrictions in Azure East US and difficulties with all three major providers, not just with GPUs. That's why it makes sense for your contingency plans not to depend on a single provider or a single region.

The physical capacity of data centers is finite, even if the API seems infinite.

Tip If your CI is suffering the most, make it fault-tolerant: retries with backoff, several equivalent instance types, and parallelism limits. This way, a capacity failure delays a build instead of blocking the team. And remember that none of these measures are universal: it depends on your region, instance type, and workload.

One last piece of information to follow the story: Amazon presented quarterly results on October 29, and its information on outstanding performance obligations will show how much committed capacity is contracted revenue. You can consult the official report to cross-reference the figures.

To learn more: Chinese edition.

Opinion · Missing context

It's important to separate two things the headline mixes: large blocks committed years in advance and the capacity a typical team requests on any given Tuesday. The source itself reports that AWS claims to reserve capacity for On-Demand, and that some analysts consider the "sold out" interpretation exaggerated.

Even so, accounts from professionals facing difficulties with common instances are an operational signal that cannot be ignored. My take: it's no cause for panic, but it is a reason to stop treating capacity as infinite.

Related articles from LaRebelión

Guide: CPU and Memory Profiling for Cloudflare Workers in ProductionHow to Save Memory in a Consistent Hash Ring, Cloudflare StyleHow to Scale Git: Reference Quorums and Object Storage

• Software engineering
• Integrate DuckDB into Amazon Aurora PostgreSQL with This Technical Guide
• Guide: Evaluate Copilot Agent Harness in Foundry Before Adoption

Original source: infoq.com

Article generated with AI.larebelion

Kernel · Head of Technology · Spain

«Just because the headline exaggerates doesn't mean your CI should keep assuming there will always be free instances.»