Amazon Web Services (AWS) continues to bolster its roster of virtual machines (VMs) powered by its custom silicon, with C9g instances for compute-intensive workloads now generally available.

The instances are powered by the hyperscaler’s Graviton5 central processing units (CPUs), which feature double the core count of its predecessor and 33% lower inter-core latency. They’re available in U.S. East, U.S. West, and Europe (Frankfurt), with additional regions set to follow.

Having fully launched EC2 M9g and M9gd instances in early June, AWS continues its CPU-centric VM ramp with a compute-optimized line it claims provides up to 25% higher performance per virtual CPU (vCPU) compared to previous-generation C8g instances.

Touted for their prowess in handling agentic workloads – as processors are more suited to complex orchestration tasks compared to the brute force of a CPU – Sébastien Stormacq, AWS’ principal developer advocate, said that the C9g instances’ faster memory and larger caches “mean your workloads spend less time waiting on data, translating into higher throughput for in-memory analytics, faster agentic loops, and more responsive real-time applications.”

“As Al shifts from answering questions to taking actions, running code, and orchestrating multistep tasks, the demand for CPU compute is growing, and C9g instances are built for this shift,” the exec wrote in a blog post.

Also available alongside the C9gs is the C9gd, which is suited for applications that require local non-volatile memory express (NVMe) solid-state drive (SSD) storage, such as scratch space during high-performance computing (HPC) simulations or local buffers for ad-serving engines. The firm claimed that high-speed, low-latency storage support is buoyed by higher throughput and input/output operations per second (IOPS) compared to previous local storage instances.

C9g and C9gd instances are the first compute-optimized AWS instances to feature its Nitro Isolation Engine, which is the hyperscaler’s security component that provides mathematically verified isolation between VMs. The recently launched M9g and M9gd instances also include this feature, which mediates all access to VM memory, CPU register state, and I/O devices through a minimal set of APIs.

The instances are available in 11 sizes ranging from medium to 48xlarge, with AWS also offering a bare metal option. The hyperscaler claims they offer up to 15% higher network bandwidth and 20% higher Elastic Block Storage (EBS) bandwidth on average across sizes compared to the previous generation.

AWS is doubling down on the number of instances powered by its custom silicon as it bids to offer cloud customers a wider variety of underlying hardware options beyond just Nvidia.

The hyperscaler penned a deal with Cerebras in March to combine its Trainium-powered servers with the chipmaker’s wafer-scale CS-3 systems. Its push into AI-optimized custom silicon, meanwhile, culminated in Project Rainier, the "ultra cluster" facility built for Claude maker Anthropic, which features more than half-a-million Trainium2 chips.

AWS' buildout efforts have, however, coincided with price rises for its EC2 Capacity Blocks for ML graphic processing unit (GPU) reservation service. Those service costs are set to jump by around 20% beginning July 1, the second increase in a little over six months.