Amazon Web Services has released Graviton5, its fifth-generation custom central processing unit, now generally available on Amazon Elastic Compute Cloud M9g instances. The chip doubles core count from Graviton4's 96 to 192, moves to TSMC's 3nm process, and ships alongside the Nitro Isolation Engine, a hypervisor whose security properties are established through mathematical proof rather than audit. For enterprises planning agentic AI infrastructure, both developments affect how they should evaluate AWS compute going forward.

M9g instances deliver up to 25% higher compute performance than Graviton4. The more useful signal for enterprise planning is where that performance comes from and which workloads benefit most.

Graviton5 places 192 Arm Neoverse V3 cores in a single package, distributed across four chiplets, each with its own memory and PCIe controllers. Moving the memory controllers closer to the cores cuts inter-core latency by up to 33% compared to the previous generation. That is an architectural choice with a specific intended beneficiary: workloads where data exchange between cores is frequent and latency-sensitive, real-time databases, agentic inference loops, Electronic Design Automation.

The cache increase is equally significant. Each Graviton5 core has access to 2.6x more L3 cache than Graviton4, keeping frequently accessed data closer to the processor and reducing the cost of cache misses. For memory-intensive applications, this translates into fewer stalls and more consistent throughput under load.

Network and storage bandwidth also improved. Graviton5 instances offer up to 15% higher network bandwidth and 20% higher Amazon Elastic Block Store bandwidth on average across instance sizes, with the largest instances doubling their network capacity.

The performance claims here are not preview estimates. Several large organizations tested Graviton5 against production workloads before the general availability announcement, and their results are specific enough to be useful for planning.

Atlassian tested Jira on M9g instances and observed 30% higher performance and 20% lower latency compared to Graviton4. Airbnb tested against production search workloads and saw up to 25% improvement over same-generation alternatives and up to 20% improvement over Graviton4. SAP reported 35 to 60% improvement in online transaction processing (OLTP) query performance on SAP HANA Cloud, a range that reflects workload variance rather than a single synthetic benchmark.

Siemens Digital Industries Software tested its Calibre semiconductor verification platform and found an additional 30% performance boost on Graviton5 over Graviton4 results, where Graviton4 had already delivered 20% gains and 30% cost reductions over other instances. Synopsys reported up to 35% runtime improvements for Fusion Compiler and PrimeTime on early Graviton5 results.

Adobe is using Graviton to transform broadcasts into personalized viewing experiences, processing video streams in real time. Epic Games runs competitive gaming infrastructure on Graviton for latency-sensitive player experiences. Snowflake and Uber are deploying Graviton specifically for agentic workloads.

The more lasting development in this announcement may be the Nitro Isolation Engine. AWS describes it as the first formally verified cloud hypervisor, and that description warrants unpacking because it means something specific.

Most cloud security assurances are attestation-based. An auditor examines controls at a point in time, issues a certification, and that certification becomes the compliance anchor. Formal verification works differently. It uses mathematical proof to demonstrate that software behaves exactly as specified across all possible states, not just the scenarios covered by testing. The proof is not an opinion about the system; it is a mathematical property of it.

AWS implemented the Nitro Isolation Engine in Rust and verified it using Isabelle/HOL, a theorem-proving framework with deep roots in academic security research. The verification work comprises 330,000 lines of machine-checked proofs confirming that the engine correctly enforces confidentiality, integrity, and memory safety between virtual machines running on Graviton5 instances. AWS has made the proof available for customer review, which is an unusual level of transparency for production infrastructure at this scale.

The prior inspiration for this work is seL4, an academic separation kernel that first demonstrated large-scale operating system verification was feasible. AWS built on that lineage and shipped it in a commercial cloud environment as an always-on feature, not an optional configuration.

For regulated industries where isolation assurance is contractually and regulatorily load-bearing, healthcare, financial services, government, this shifts the conversation from certification status to mathematical proof. That is a substantively different security posture than what any hyperscaler has previously offered.

The M9g general-purpose instances are available now. C9g instances for compute-intensive workloads and R9g instances for memory-intensive workloads are planned for 2026. Enterprises evaluating Graviton5 for the most demanding inference and analytics jobs are making that assessment against a partial lineup. The architectural advantages are real, but the full instance family that would address every agentic AI use case is not yet available.

Pricing for M9g in general availability had not been published at the time of this writing. The price-performance case for agentic AI on Graviton5, compared to GPU-based inference for the same workloads, requires live pricing to model accurately. AWS's track record on Graviton pricing has been competitive across prior generations, but enterprise planning should confirm current numbers rather than extrapolate from Graviton4 pricing.

More than half of new central processing unit capacity added to AWS infrastructure is now Graviton-based, for the third consecutive year. That share reflects genuine customer adoption driven by measurable performance and cost outcomes, not just AWS internal routing decisions.