AWS Graviton5 GA: 192 cores on TSMC 3nm, 25% perf gain at 9% price premium; Meta commits tens of millions of cores
AWS has released M9g and M9gd instances based on Graviton5, doubling core count to 192 per chip across four chiplets on TSMC 3nm. Specs include 192 MB of L3 cache (5x Graviton4), DDR5-8800 (fastest DDR5 in any cloud instance), and PCIe Gen 6. The most noteworthy addition is the first formally verified hypervisor isolation in a production cloud environment—mathematical proof of VM isolation boundaries instead of testing-based assurance.
Pricing is 9% higher than M8g across all sizes, but with 25% better compute performance, that yields ~15% better price-performance per dollar. Customer benchmarks from six-month preview period demonstrate concrete gains: ClickHouse saw 36% boost vs M8g zero code changes; Honeycomb achieved 36% better throughput per core; HubSpot saw MySQL query duration drop up to 60%; Airbnb called them "some of the fastest EC2 instances we have ever tested."
Meta has committed to deploying tens of millions of Graviton5 cores for agentic AI workloads—a major validation and one of the largest Graviton customer commitments to date. Uber and Snowflake are also adopting Graviton5 for agent tasks. AWS positions the chip squarely at agent workloads due to their CPU-intensity, though the architecture improvements (core count, cache, memory bandwidth, interconnect latency) also benefit databases, web servers, and batch processing.
The Nitro Isolation Engine, built into the sixth-generation Nitro System, provides mathematically proven isolation between VMs—the first major public cloud instance to carry formally verified hypervisor isolation in production. Compute-optimized C9g and memory-optimized R9g variants are planned for later in 2026.
For practitioners, this matters because Graviton5 offers strong price-performance for CPU-bound workloads and provides a genuinely stronger security boundary for multi-tenant execution than competing cloud instance types. Instances are live now in US East (Virginia, Ohio), US West (Oregon), and EU (Frankfurt).
Sources
- Primary source
- infoq.com
“M9g instances are uniformly 9% more expensive than M8g across all sizes, with 25% compute performance improvement, translating to roughly 15% better price-performance per dollar”