EC2 Instance Family Decision Tree 2026

AWS EC2 has 50+ instance types. The decision tree that picks the right family in seconds.

By workload

EC2 instance family selection determines both performance and cost for any workload. AWS publishes many families optimized for different workload types; choosing the right family is one of the highest-leverage performance and cost decisions. The 2026 landscape includes Graviton (ARM) variants for most families; the family decision now also implies an architecture decision.

What the family-by-workload mapping looks like:

• Web/app servers: M family.: General-purpose workloads use the M family (m7i, m7g). Balanced CPU, memory, and network resources match the typical web/app server profile. The M family is the default for most application workloads.
• Compute-heavy: C family.: CPU-bound workloads (encoding, scientific computing, batch processing) use the C family (c7i, c7g). Higher CPU-to-memory ratio fits these workloads; the cost per CPU is lower than M family.
• Memory-heavy: R or X family.: Memory-bound workloads (in-memory caches, large databases, ML inference) use R or X. R is the standard memory-optimized family; X provides the highest memory-to-CPU ratio for extreme memory needs.
• Storage-heavy: I or D family.: Workloads with high local storage requirements use I (NVMe SSD) or D (HDD). Database workloads, big-data processing, and similar workloads fit; the local storage avoids the latency and cost of EBS for every IO.
• Specialty families.: P (GPU), G (graphics/GPU), F (FPGA), Inf (inference), Trn (training) for specialized workloads. The specialty families fit narrow use cases; the general-purpose families cover most workloads.

The family-by-workload mapping is the foundation. Most workloads fit cleanly into one family; outliers are evaluated specifically.

Pick newest generation

Within a family, the newest generation is almost always the best choice. AWS prices new generations at parity or below older generations while delivering significantly better performance. Migrating to newer generations is one of the easiest cost-performance wins.

• M5 to M6i to M7i.: Each new generation delivers better performance per dollar. M7i instances have roughly 15% better performance than M6i, at similar or better pricing. M6i instances have similar improvements over M5. The trajectory is consistent.
• Newer is faster and cheaper per unit.: The combination of better performance and equivalent or lower price means newer generations deliver more value per dollar. Sticking with old generations leaves money on the table.
• Migration cost is low.: Moving from M6i to M7i is usually a launch template change. The application typically does not need code changes; the new instance is binary-compatible. The migration is mechanical.
• The savings compound.: A 15% improvement applied to every instance in the fleet adds up. Across a large fleet, the savings are substantial. The discipline of staying on the newest generation produces continuous cost optimization.
• Graviton variants are the next step.: The Graviton (g-suffix) variants offer additional savings (typically 20% better price-performance than equivalent Intel/AMD variants). Workloads compatible with ARM should consider Graviton; the cost-performance is significantly better.

Generation selection is the second-highest-leverage decision after family. Newest generation should be the default unless a specific reason argues otherwise.

Avoid

Some instance choices are recurring sources of regret. The team is better served avoiding them; the savings or simplicity that motivated the choice often disappears under sustained operations.

• T family for production.: The T family (t3, t3a, t4g) uses CPU credits: instances accumulate credits during idle periods and spend them during burst. The credit system creates unpredictable performance. Production workloads with sustained CPU usage hit credit exhaustion and slow dramatically.
• Bursty CPU; unpredictable for sustained workloads.: The bursty model fits truly idle workloads (dev environments, low-traffic services). Production workloads usually have sustained baseline traffic; the credit model produces incidents.
• Old generations: M3, M4.: Several generations old. The price has not kept up with newer generations; the performance is significantly worse. Workloads stuck on old generations are paying more for less.
• Cost more for less.: Old generations are sometimes priced equivalently to new ones. Even at price parity, the performance difference means the older generation costs more per unit of work delivered.
• Specialty families when general-purpose fits.: Some teams reach for specialty families (memory-optimized, compute-optimized) when M family would be cheaper and fit the workload. Specialty families have premium prices; only use them when the workload genuinely needs the optimization.

EC2 instance family 2026 decision is one of the persistently rewarded performance and cost optimizations available. Nova AI Ops integrates with EC2 inventory, surfaces workloads on old generations or wrong families, and produces the migration queue that the platform team uses to drive continuous optimization.