Amazon flipped Graviton5 to general availability in us-east-1 on Sunday, eight months after announcing the chip at re:Invent. The launch covers two families: C9g for compute-bound work and R9g for memory-heavy workloads, both with up to 192 vCPUs and DDR5-6400 across 12 channels. The headline for most buyers is not the silicon but the price sheet. On-demand rates for the R9g line sit 38 percent below the equivalent R8g configuration when measured per unit of CPU inference throughput, and reserved instance discounts stack on top of that. Availability expands to us-west-2, eu-west-1, and ap-northeast-1 through September.
What is actually new in the silicon
Graviton5 moves to Arm's Neoverse V3 core at 3.4 GHz, up from the V2 design in Graviton4, and adds a second SVE2 pipeline per core. Amazon's own numbers claim 1.7x integer throughput and 2.3x on bfloat16 matrix operations versus Graviton4 at the same core count. The more consequential change is memory: 12 DDR5-6400 channels deliver roughly 614 GB/s of theoretical bandwidth per socket, which is the number that determines whether a 70 billion parameter model can run usefully on CPU at all.
Amazon also expanded the on-die interconnect and added a hardware compression engine addressable through a new instruction set extension. That engine matters for anyone shipping large payloads through a service mesh, because gzip and zstd work moves off the general purpose cores entirely. Annapurna Labs vice president Ali Saidi said in a technical session that the compression block alone recovered about 9 percent of usable CPU on Amazon's internal DynamoDB fleet during qualification, which is the kind of internal-first validation that has become the pattern for every Graviton generation.
Independent benchmarks, with caveats
Phoronix published a first pass on Sunday afternoon covering 41 workloads on a 96 vCPU C9g instance. The integer claims broadly hold: SPEC-adjacent tests land between 1.55x and 1.8x over Graviton4. Memory bandwidth measured through STREAM Triad came in at 487 GB/s sustained, about 79 percent of theoretical, which is normal for a server part and better than the 71 percent Graviton4 achieved. Compilation benchmarks were the weakest showing at 1.3x, suggesting the improvements favor vectorizable work over branch-heavy code.
Inference results need more care in reading. Running Llama 3.3 70B in 4-bit through llama.cpp on a single R9g.48xlarge produced 14.2 tokens per second for a single stream, up from 6.1 on R8g. That is genuinely useful for batch summarization and document classification. It is not competitive with a single H200 for interactive chat, and Amazon is careful not to claim otherwise. The pitch is cost per token for workloads where 400 milliseconds of latency is fine, and on that measure the R9g undercuts GPU instances by a wide margin.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The pricing move is aimed at a specific competitor
Google Cloud has spent 18 months selling Axion instances as the cheap Arm option, and its C4A line had been undercutting Graviton4 on published rates for memory-heavy configurations. Amazon's 38 percent cut erases that gap and then some. An r9g.16xlarge at $3.19 an hour on-demand in us-east-1 compares against $4.41 for the nearest Axion configuration, and Amazon's three-year compute savings plan brings the effective rate to about $1.71. Google has not responded publicly, though two customers told us their account teams have already offered matching private discounts.
Microsoft is in a harder spot. Cobalt 200 remains in limited preview after a delay Microsoft attributed to a memory controller respin, and Azure's Arm story still leans on Ampere Altra parts that are two generations behind. A cloud economics consultant in Seattle who models spend for eight companies over $50 million in annual commitment said the Graviton5 rates change the answer on multi-cloud placement for the first time since 2023, because the cheapest place to run a Kafka cluster or a Postgres read replica is now unambiguously Amazon.
Migration is easier than it used to be, but not free
The Arm porting story has improved enormously. Every mainstream language runtime ships aarch64 builds, the container ecosystem publishes multi-arch images by default, and the Amazon Linux 2026 AMI includes the compression extension support out of the box. Kubernetes users on EKS can add a Graviton5 node group and let the scheduler place workloads by architecture label, which several teams described as a two-hour change. Karpenter's 1.6 release added Graviton5 instance types within a day of general availability.
The friction that remains is concentrated in commercial software and old dependencies. Oracle Database still has no supported aarch64 build on AWS. Several observability agents ship Arm binaries a release behind. A financial services architect in Charlotte said his blocker is a risk analytics library from a vendor that quoted a six-figure fee to produce an Arm build, which makes the 38 percent instance saving irrelevant for that particular cluster. Teams should audit their commercial dependency list before promising savings to a CFO.
What this means for capacity and for Nvidia
Amazon told analysts that Graviton now accounts for more than half of new CPU capacity added across its fleet, and Graviton5 will push that share higher. Every one of those sockets is a socket Intel and AMD did not sell. AMD's data center revenue has held up on the strength of Genoa and Turin in on-premises deployments, but the hyperscaler CPU market is visibly shifting toward in-house Arm designs, and Google, Microsoft, and Alibaba all have programs in flight.
For Nvidia the read is more nuanced. CPU inference at this price point cannibalizes the low end of GPU demand, the classification and embedding jobs that people ran on T4 and L4 cards because nothing cheaper existed. It does not touch training or interactive generation. What it does change is the calculus for companies deciding whether to buy GPU capacity for a workload that only needs to finish overnight. Several teams told us they are moving embedding pipelines to R9g and reserving GPU budget for the workloads that actually need it.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- Kubernetes
- Edge Computing
