Cloud Infrastructure
AWS Graviton5 Ships with 192 Cores and Formally Verified Hypervisor Isolation
AWS's new M9g and M9gd instances bring formally verified VM isolation to production cloud, doubling core count and delivering 25% compute gains—but the "agentic AI" marketing oversells what actually changed.
InfoQ
AWS has released M9g and M9gd instances powered by Graviton5, doubling the previous generation to 192 cores per chip on TSMC's 3nm process. The spec jump is substantial: 192 MB of L3 cache (5× Graviton4), DDR5-8800 memory, PCIe Gen 6, and what AWS calls "formally verified hypervisor isolation"—the first production cloud instance to carry mathematically proven VM boundaries rather than relying on testing alone.
The AI Marketing Question#
AWS is positioning Graviton5 squarely at agentic AI workloads, but the hardware community has a more grounded read. As one Reddit commenter noted, these are "normal ARM cores" that happen to excel at CPU-intensive agent tasks. The real wins—more cores, larger caches, higher memory bandwidth—benefit databases, web servers, and batch processing equally. What changed architecturally is almost nothing; what changed is capacity and speed.
The Customer Signal#
Meta's commitment matters most: it's deploying tens of millions of Graviton5 cores for agentic AI, alongside Uber and Snowflake. Real-world benchmarks from the preview period are striking. ClickHouse saw 36% gains over M8g with zero code changes. HubSpot's MySQL queries dropped by up to 60% latency. Honeycomb measured 36% better throughput per core in production observability workloads across six months of A/B testing.
Security and Economics#
The Nitro Isolation Engine—formally verified using mathematical proof techniques—gives teams deploying multi-tenant or untrusted workloads a stronger guarantee than competing cloud providers currently offer. On pricing, M9g instances run 9% higher than M8g in us-east-1, but the 25% compute gain yields roughly 15% better price-performance per dollar (workload variation applies). M9gd adds up to 11.4 TB of NVMe storage with 30% faster IOPS.
Compute-optimized (C9g) and memory-optimized (R9g) variants ship later in 2026. Instances are live now in US East, US West, and EU Frankfurt across On-Demand, Spot, Savings Plans, and Dedicated options.
For context, AWS custom silicon now represents more than half of all new CPU capacity added in three years and a $20 billion annual run rate—outpacing Google's Axion and Microsoft's Cobalt efforts.
