Amazon Web Services flipped the switch on its fifth generation of custom cloud silicon on August 31, 2026, making R9g and R9gd memory-optimized EC2 instances built on AWS Graviton5 generally available. The launch lands quietly compared to AWS’s usual re:Invent fireworks, but the numbers behind it explain why database teams, cache operators, and finance departments running massive AWS bills are paying attention: up to 25% better compute performance per vCPU than the Graviton4-based R8g generation, up to 35% faster web application throughput, and up to 30% faster database performance, according to AWS’s own R9g instance page.
The timing matters. AWS made this move while GPU shortages keep squeezing budgets across the industry, and while Microsoft, Google, and a growing list of AWS’s own customers race to build custom Arm chips that reduce dependence on Nvidia. R9g and R9gd are not AI accelerators. They are ordinary memory-optimized compute instances for databases, in-memory caches, and real-time analytics. But their arrival says something bigger about where cloud economics are headed in late 2026, and it’s worth pulling apart what changed, what didn’t, and who should care.
What AWS Actually Shipped on August 31
The official AWS “What’s New” announcement confirms R9g and R9gd instances became generally available on August 31, 2026, with R9gd distinguished from R9g by local NVMe SSD instance storage, per the EC2 Instance Types Guide document history. R9g relies purely on Amazon EBS for storage, while R9gd bolts on fast ephemeral NVMe drives for workloads that need low-latency local disk alongside memory headroom.
Both families run on Graviton5, AWS’s fifth-generation Arm-based processor. According to the AWS Graviton entry on Wikipedia, Graviton5 packs 192 cores across a four-chiplet design, built on Arm’s Neoverse V3 core and the Armv9.2-A instruction set, manufactured on a 3-nanometer process. That’s a jump from Graviton4’s architecture in raw density, and it shows up directly in the memory subsystem: DDR5 speed climbs to 8,800 MT/s from 5,600 MT/s on Graviton4, and L3 cache grows roughly fivefold, per technical analysis of the chip cited alongside AWS’s own blog material. Graviton5 also adds PCIe Gen 6 support, which is what enables the network and EBS bandwidth gains AWS is advertising on R9g.
R9g and R9gd launched in four regions: US East (N. Virginia), US East (Ohio), US West (Oregon), and EU (Frankfurt), matching what AWS listed in its official announcement and what independent testing confirmed by checking live availability across seven regions. That’s a narrower footprint than AWS’s typical global rollout cadence, which usually signals the company wants real production load on the new silicon before it opens the floodgates to every region.
The Performance Numbers, Workload by Workload
AWS doesn’t publish a single blanket performance figure for R9g. Instead, its instance page breaks gains down by workload type relative to the R8g generation:
- Up to 25% better overall compute performance per vCPU
- Up to 35% faster performance for web applications
- Up to 35% faster machine learning inference
- Up to 30% faster database performance
- Higher networking bandwidth and more EBS bandwidth than R8g, with some large instance tiers seeing throughput roughly double
Those figures track a pattern AWS has repeated with every Graviton generation: publish workload-specific ranges rather than a single synthetic benchmark score, then let customers validate against their own traffic. It’s a defensible approach given how differently a Redis cache, a PostgreSQL cluster, and a Java monolith respond to more cache and faster memory. A database with a large working set that spills constantly out of L3 cache benefits enormously from a fivefold cache increase; a CPU-bound batch job might barely notice.
Context from the previous generation transition helps calibrate expectations. Per the Wikipedia Graviton entry, Graviton4-based R8g instances delivered up to 30% better overall performance than Graviton3-based R7g, up to 40% faster database performance, and up to 45% faster large Java application performance when AWS launched that generation. Measured against that jump, R9g’s 25-35% range looks like a smaller, more incremental step for compute and web workloads, though the database gain (up to 30%) sits in a similar range to the prior cycle’s database improvement once you account for the different baseline.
Graviton Generation-Over-Generation: A Five-Year Scorecard
Look at Graviton’s climb since AWS started publishing comparative figures, and a pattern emerges: each generation trades one dominant improvement area for another, while price-performance keeps compounding across the stack.
| Generation | Instance Family | vs. Prior Gen Compute | vs. Prior Gen Database | Key Architecture Change |
|---|---|---|---|---|
| Graviton3 (R7g) | R7g | Baseline for this comparison | Baseline | Arm Neoverse V1, DDR5 memory introduced |
| Graviton4 (R8g) | R8g | Up to 30% vs. Graviton3 | Up to 40% vs. Graviton3 | Arm Neoverse V2, more cores per socket |
| Graviton5 (R9g/R9gd) | R9g, R9gd | Up to 25% vs. Graviton4 | Up to 30% vs. Graviton4 | Arm Neoverse V3, 192 cores, 3nm process, PCIe Gen 6 |
Source: AWS EC2 R9g instance page, AWS EC2 Instance Types Guide document history, and the AWS Graviton entry on Wikipedia.
Graviton5 isn’t limited to memory-optimized R9g and R9gd. The same chip underpins general-purpose M9g and M9gd instances, which reached general availability earlier in the summer, and compute-optimized C9g and C9gd instances that followed shortly after. That staggered rollout, general-purpose first, then compute-optimized, then memory-optimized, mirrors how AWS sequenced Graviton4’s launch and suggests the company is deliberately spacing out validation across workload types before pushing the newest silicon everywhere at once.
Why AWS Is Betting Harder on Arm Right Now
The backdrop to this launch is a memory and AI-accelerator market that has been unusually tight through most of 2026. DRAM and HBM shortages have pushed AI server component costs up across the board, and Nvidia’s flagship data center GPUs remain both scarce and expensive for many buyers outside the largest AI labs. Against that backdrop, a CPU family that AWS controls end to end, built on Arm licenses rather than x86 royalties, and manufactured on a leading-edge 3nm process gives Amazon a lever it doesn’t have with third-party silicon: it can tune supply, pricing, and availability without waiting on an external vendor’s roadmap.
AWS is not alone in this strategy, and that’s arguably the more important market story here. Google has pushed its own custom silicon path with Ironwood TPUs, positioning them explicitly as a lower-cost alternative to Nvidia GPU rental pricing for inference workloads. Microsoft has been ramping its Maia AI accelerator line, reportedly ordering hundreds of thousands of units from TSMC to reduce its own Nvidia dependency. What Graviton5 adds to that picture is the reminder that the custom-silicon push isn’t confined to AI accelerators, it now spans the general-purpose and memory-optimized CPU instances that run the unglamorous but massive workloads: the Postgres clusters, the Redis fleets, the Kafka brokers, and the Java services that make up the bulk of most companies’ actual cloud bills.
That distinction matters for anyone trying to read AWS’s silicon strategy correctly. Graviton5 does claim up to 35% faster machine learning inference over R8g, which puts it in indirect competition with GPU-based inference for lighter models. But R9g and R9gd are still CPU instances, not accelerators, and AWS is not positioning them as a wholesale replacement for GPU-backed training or large-model serving. The more accurate framing is that AWS is squeezing cost and performance out of every layer of its stack at once: custom CPUs for general compute and memory-bound work, and a separate, much more contested fight over GPU and AI-accelerator capacity for the heaviest AI workloads.
Competitive Landscape: Arm CPUs Across the Big Three Clouds
AWS was first to bring custom Arm server silicon to broad general availability with the original Graviton in 2018, and it has kept that lead through five generations. Google and Microsoft have each built out their own Arm CPU lines in response, though neither has matched AWS’s breadth of instance families running on custom Arm chips.
| Cloud Provider | Custom CPU Silicon | Current Generation Status | Primary Target Workloads |
|---|---|---|---|
| AWS | Graviton (5 generations since 2018) | Graviton5 shipping across M9g, C9g, R9g/R9gd families | General compute, databases, caching, analytics, light ML inference |
| Google Cloud | Axion (Arm CPU) plus Ironwood (AI accelerator/TPU) | Axion in general availability; Ironwood TPU positioned against Nvidia for inference | General compute (Axion); large-scale AI training and inference (Ironwood) |
| Microsoft Azure | Cobalt (Arm CPU) plus Maia (AI accelerator) | Cobalt available in select regions; Maia 300 ramping via TSMC production | General compute (Cobalt); AI training/inference (Maia) |
Note: Competitor generation details reflect publicly reported roadmaps as of September 2026 and are not independently benchmarked against Graviton5 in this article.
The pattern across all three hyperscalers is the same: separate the general-purpose CPU silicon story from the AI-accelerator silicon story, but pursue both aggressively. AWS’s advantage is tenure. Five generations of Graviton means five generations of customer migration experience, tooling maturity, and workload compatibility testing that competitors are still building out on their own custom CPU lines.
Who Should Actually Care About R9g and R9gd
AWS’s own guidance points R9g and R9gd at databases, in-memory caches, and real-time analytics, per the AWS News Blog’s Graviton and EC2 category summaries. In practice, that means teams running things like PostgreSQL or MySQL on EC2 rather than fully managed RDS, self-hosted Redis or Memcached clusters, Elasticsearch or OpenSearch nodes with large heap requirements, and real-time analytics engines that keep large working sets in memory.
R9gd’s addition of local NVMe storage extends that further to workloads that need fast scratch space alongside memory headroom: write-heavy time-series databases, streaming pipelines with local buffering, or search indexes that spill to disk under load. For any of these, the arithmetic is straightforward: AWS’s price-performance framing implies lower cost per unit of throughput than the equivalent R8g fleet, which for large, steady-state database workloads can translate into real savings even before accounting for the underlying compute, network, and EBS bandwidth gains.
The instances are available through the full range of standard EC2 purchasing models, including On-Demand, Spot Instances, Savings Plans, Dedicated Instances, and Dedicated Hosts, according to reporting on the launch. That’s a meaningful detail: AWS isn’t gating the newest Graviton generation behind a beta program or reserved-only access, it’s treating R9g and R9gd as mainstream EC2 inventory from day one.
The Migration Calculus: Should You Move From R8g Now?
Migrating between Graviton generations on Arm-native workloads is typically far less painful than the original x86-to-Arm migration many teams did years ago. If an application already runs cleanly on R8g, moving to R9g or R9gd should mostly be a matter of testing against the new instance type, watching for any dependency on architecture-specific binaries, and validating memory and network behavior under production load before cutting over.
The bigger question is whether the 25-35% workload-specific gains justify a migration project right now versus waiting for broader regional availability. With R9g and R9gd confined to four regions at launch, teams outside US East (N. Virginia), US East (Ohio), US West (Oregon), and EU (Frankfurt) simply can’t test yet. For teams inside those regions running memory-intensive, steady-state workloads at meaningful scale, the case for at least benchmarking R9g against current R8g fleets is straightforward given the price-performance claims AWS is making. For smaller or bursty workloads, the incremental gain over R8g is unlikely to outweigh the operational cost of a migration cycle in the near term.
Historical Context: Five Generations in Eight Years
AWS launched the first Graviton processor in 2018 as a modest, somewhat experimental Arm offering aimed at cost-sensitive, non-critical workloads. Graviton2 arrived in 2019 as the first generation taken seriously at scale, offering meaningful price-performance gains that pulled real production traffic onto Arm. Graviton3 in 2021 added DDR5 memory support and better per-core performance. Graviton4 in 2024 pushed core counts and memory bandwidth further, delivering the 30-45% gains over Graviton3 cited above. Graviton5’s 2026 arrival, spread across M9g, C9g, and now R9g/R9gd, continues that roughly two-year cadence AWS has settled into for major Graviton generations.
That cadence is itself a competitive signal. A two-year major-generation cycle, with staggered instance-family rollouts within each generation, gives AWS a steady drumbeat of reasons to stay on AWS without requiring it to leapfrog Nvidia or Intel on any single release. It’s an incremental, compounding strategy rather than a moonshot one, and eight years in, it has produced a chip that technical breakdowns describe as having a fivefold larger cache and roughly 57% faster memory speed than Graviton4.
Market Impact: What This Means for Cloud Spend in Late 2026
For enterprise finance and platform teams tracking cloud spend, R9g and R9gd matter less as a headline event and more as one more data point in a broader shift: custom silicon is steadily eating share from general-purpose x86 instances across memory-optimized and general-purpose categories at all three hyperscalers. Every workload that moves from an Intel- or AMD-based instance to a Graviton-based one is revenue AWS effectively keeps in-house rather than paying out to a third-party chip vendor, and AWS has consistently used part of that margin to offer better price-performance to customers, which is exactly the best-price-performance claim it’s making for R9g.
The knock-on effect touches Intel and AMD’s cloud instance revenue, though neither company depends on AWS’s memory-optimized instance category alone, and both continue to win share in categories where Arm compatibility, specialized instruction sets, or existing x86 toolchains matter more than raw price-performance. It’s also worth noting that AWS’s Graviton push, and the parallel Axion and Cobalt efforts at Google and Microsoft, haven’t slowed those same companies’ enormous capital spending on Nvidia GPUs. Custom CPU silicon and GPU accelerator silicon are complementary bets, not substitutes, and the R9g launch doesn’t change that calculus for AI training or large-model inference workloads.
What Comes Next in the Graviton Roadmap
With M9g/M9gd, C9g/C9gd, and now R9g/R9gd all shipped, Graviton5 has completed its rollout across AWS’s three core general-purpose, compute-optimized, and memory-optimized instance families, the same pattern Graviton4 followed in its own generation. Based on that pattern, the most likely next moves are broader regional expansion for R9g and R9gd beyond the initial four regions, and eventually specialized variants aimed at storage-heavy or network-heavy workloads, mirroring how earlier Graviton generations eventually spawned network-optimized and storage-optimized variants after their initial launch wave.
A sixth Graviton generation is not expected before 2028 at the earliest if AWS holds to its roughly two-year major-generation cadence, though AWS has occasionally shipped mid-cycle refinements to existing generations rather than waiting for a full node-process jump.
Five Predictions for the Next 12 Months
- Regional expansion within six months. AWS’s historical pattern with new Graviton instance families is to expand from an initial 3-5 region launch to double-digit region availability within two to three quarters, so expect R9g and R9gd in Asia-Pacific and additional EU regions well before mid-2027.
- Managed database services follow within a year. Graviton-based EC2 instances have historically preceded matching Graviton support in RDS, ElastiCache, and OpenSearch Service by several months to a year; expect Graviton5-backed managed database options to appear through 2027.
- Price pressure on x86 memory-optimized instances. As R9g adoption grows, expect AWS to adjust pricing on comparable Intel- and AMD-based memory-optimized instances to stay competitive, continuing a pattern seen with prior Graviton launches.
- Competitors accelerate their own memory-optimized Arm instances. Google’s Axion and Microsoft’s Cobalt lines are currently weighted toward general-purpose compute; expect both companies to push memory-optimized Arm variants faster in response to R9g’s positioning.
- Custom silicon keeps expanding while GPU dependency persists. None of AWS, Google, or Microsoft show signs of slowing their Nvidia GPU purchasing despite growing custom CPU fleets, because CPU workloads and AI accelerator workloads remain fundamentally different problems requiring different silicon.
The Bottom Line for Engineering Teams
R9g and R9gd aren’t a dramatic reinvention of cloud compute. They’re the fifth iteration of a strategy AWS has run consistently since 2018: ship incrementally better custom Arm silicon roughly every two years, target it first at the workloads with the clearest price-performance case, and let adoption compound over time. For teams already comfortable running production workloads on Graviton4’s R8g, the move to R9g should be a relatively low-friction evaluation, gated mainly by whether your workloads live in one of the four launch regions yet. For teams still on x86 memory-optimized instances, R9g adds one more reason to test the Arm migration path, especially with database and web application workloads where the claimed 30-35% gains are largest.
The bigger story sitting underneath this specific launch is the one worth tracking through the rest of 2026: all three major hyperscalers are now running parallel silicon strategies, custom CPUs to control general compute economics, and custom AI accelerators to chip away at Nvidia’s grip on the highest-value workloads. Graviton5’s arrival on memory-optimized instances is a small, quiet data point in that much larger shift, but it’s a real one, backed by AWS’s own published performance claims rather than speculation.
Frequently Asked Questions
What is AWS Graviton5?
Graviton5 is AWS’s fifth-generation custom Arm-based server processor. According to the AWS Graviton Wikipedia entry, it uses a 192-core, four-chiplet design built on Arm’s Neoverse V3 core and Armv9.2-A instruction set, manufactured on a 3-nanometer process, with DDR5 memory running at 8,800 MT/s and PCIe Gen 6 support.
When did R9g and R9gd instances become available?
AWS made R9g and R9gd instances generally available on August 31, 2026, according to the official AWS “What’s New” announcement and the EC2 Instance Types Guide document history.
What’s the difference between R9g and R9gd?
R9gd adds local NVMe SSD instance storage on top of the same Graviton5 processor and memory-optimized profile as R9g, which relies solely on Amazon EBS for storage, per the EC2 Instance Types Guide.
Which AWS regions support R9g and R9gd at launch?
At general availability, R9g and R9gd are available in US East (N. Virginia), US East (Ohio), US West (Oregon), and EU (Frankfurt), based on AWS’s official announcement and independent region-availability checks.
How much faster is R9g than R8g?
AWS states R9g delivers up to 25% better overall compute performance per vCPU than R8g, with up to 35% faster web application performance, up to 35% faster machine learning inference, and up to 30% faster database performance, according to the AWS R9g instance page.
What workloads is R9g designed for?
AWS positions R9g and R9gd for memory-intensive workloads including databases, in-memory caches, and real-time analytics, per the AWS News Blog’s Graviton and EC2 category descriptions.
Does Graviton5 compete with Nvidia GPUs?
Not directly. Graviton5 is a general-purpose and memory-optimized CPU architecture, not an AI accelerator. It claims meaningful gains in lighter machine learning inference tasks, but AWS continues to rely heavily on Nvidia GPUs for large-scale AI training and heavy inference workloads.
Can I use R9g instances with Spot pricing or Savings Plans?
Yes. R9g and R9gd are available through all standard EC2 purchasing models, including On-Demand, Spot Instances, Savings Plans, Dedicated Instances, and Dedicated Hosts, according to reporting on the launch.
