How Azure Compute is Bringing Performance, Efficiency, Flexibility, Resilience, and Trust Together for the AI Era
By Arun Kishan, Technical Fellow and Corporate Vice President, Azure Compute
Every major technology shift eventually asks whether infrastructure can turn possibility into dependable progress, and for enterprise AI, that moment is here. As organizations move beyond isolated experiments toward models, applications, and agents that run real business processes, the infrastructure beneath them must deliver more than raw scale: the right compute for each workload, scarce capacity that works harder, protection for sensitive data, support for governance and compliance, and resilience as demand and operating conditions change Azure is the connected system to run this reliably balancing cost.
At Microsoft Ignite 2026, Azure Core Compute is bringing custom silicon, a broad infrastructure portfolio, intelligent optimization, embedded security, and end-to-end resiliency together as one system, helping customers move faster with AI without trading off performance against efficiency, or innovation against enterprise control.
Purpose-built Silicon Expands What is Possible
Infrastructure innovation begins at the silicon layer—but its value is measured in the workloads it enables. As AI moves from individual prompts to long-context reasoning and sustained, multi-step agentic tasks, accelerators and CPUs must work together to deliver responsive experiences at the lowest cost, efficiently scaling to meet the meteoric rise of token demand.
Introducing Azure Maia 300, the next generation of our Maia AI accelerator family. Designed as a complete system for these emerging workloads, Maia 300 advances the inference efficiency frontier: we expect it to deliver 40% more tokens per watt than leading GPU-based systems. In a full 96-accelerator configuration, it delivers up to seven times the rack-scale token throughput of Maia 200. That combination of efficiency and throughput is central to making increasingly demanding AI experiences practical at enterprise scale.
Accelerators are only part of that equation. AI applications and agents also depend on CPUs to orchestrate tasks, retrieve data, and connect services. We designed Cobalt 200 for scale-out, agentic, and data-intensive AI workloads. Now generally available in 14 regions and expanding to 23 by the end of 2026, Cobalt 200 VMs deliver up to 50% better performance than Cobalt 100, and customers including Amadeus, Databricks, InMobi, OpenAI, and Snowflake are already seeing gains on real workloads. Custom silicon innovations like Cobalt 200 lower the overall cost of delivering Microsoft AI, data, and insights to customers. Azure Boost, included in every Cobalt 200 VM, brings accelerated network and remote-storage performance, and new Mpsv4 (high-memory) and Lpsv5 (fast access to large datasets) VM types extend Cobalt’s price-performance to more workloads.
That same workload-first approach extends across our broader compute portfolio. AMD v8 Venice VMs bring high-density x86 compute and memory for agents, retrieval-augmented generation (RAG), and preprocessing. AMD Lasv5/Laosv5 local storage optimized VMs provide local NVMe storage for RAG, indexing, Spark shuffle, and staging. Together with Maia 300 and Cobalt 200, these complementary options reflect a common principle: enterprise AI needs infrastructure matched to each stage of the workload, from inference to the services and data pipelines that support it.
Efficiency Must Become Continuous
Enterprise AI must scale efficiently, has to be continuous, with the available capacity must turn into useful work as workloads, demand, and economics shift. Azure’s broad virtual machine portfolio lets customers match architecture, processor, and performance profile to each workload, whether it’s an agent, a data pipeline, or a business-critical application. Flexible pricing options, including Pay-As-You-Go, Azure savings plans, Azure Reservations, and Azure Spot Virtual Machines, build cost into the architecture from the start. Azure Virtual Machine Scale Sets and Azure Compute Fleet then let those choices respond to production conditions, balancing capacity, availability, and price-performance across large deployments.
Alongside these offerings, we're expanding our portfolio to help customers turn workload flexibility into efficiency gains, with VM adaptive placement, compute templates, Spot Plus and DT VMs.
VM adaptive placement brings that flexibility to the individual VM. Instead of tying a VM to one fixed size, customers preferred SKU for deployment along with an opt-in setting for adaptive placement, to further simplify deployment flow, you can include these requirements in a compute template to reuse across deployments. At creation, adaptive placement deploys the VM on an eligible size with available capacity. The result is fewer failed deployments and retries, faster provisioning during demand spikes and less effort maintaining SKU lists, and customers are billed only for the size actually deployed.
Spot plus is the newest addition to the Azure Spot portfolio. With longer runtimes and an extended eviction notice, it lets customers move a broader range of workloads to interruptible capacity and capture deep savings compared with pay-as-you-go pricing. DT VMs are a new subfamily of general-purpose VMs that offer latest-generation compute performance at lower prices. They make it easy to optimize costs for workloads that don’t need sustained peak performance. Sign up for DT here.
We’re bringing this to the workloads where demand swings the most. Our integrations of IBM Spectrum Symphony and IBM Spectrum LSF with Azure Compute Fleet let clusters provision Azure compute as demand rises and release it when work completes, so customers can extend to Azure for financial simulations, quantitative modeling, and engineering simulation. Virtual desktop infrastructure is another example: a Monday-morning login surge can require thousands of VMs at once. Azure Bulk Actions starts, deallocates, and hibernates VMs at scale while Azure handles throttling, throughput limits, and transient failures. Hibernate and resume saves a VM’s state so compute can be released while desktops sit idle, with a fallback to deallocate or a clean boot if an operation fails. Together, compute choice, economics, and orchestration help customers get more from their infrastructure as applications evolve and give enterprise AI a stronger foundation to scale.
Trust is Being Engineered into the Infrastructure
As AI reaches sensitive records, proprietary models, and business-critical processes, customers need control over who can access that information while it is being used. Azure Confidential Computing adds hardware-backed protection during processing, with attestation providing evidence that an environment meets the conditions for releasing sensitive data and keys. At Ignite, we are advancing that foundation with a confidential bare-metal technology showcase on NVIDIA Vera Rubin NVL72, a gated preview of AMD Turin v7 confidential VMs, and a protected-I/O showcase combining Intel TDX Connect with Azure Boost Confidential Device. Together, these investments extend protection across the processors, memory, and connections that sensitive workloads depend on.
Making that protection practical also means meeting customers where their applications run. We are expanding confidential AKS worker-node options and making existing workloads easier to move with generally available migration and large OS-disk encryption capabilities. Azure Confidential Clean Rooms, in limited preview, enables organizations to analyze shared datasets while protecting each participant’s raw data. Looking further ahead, Project Nova explores a complementary approach: using zero-knowledge proofs to verify specific claims without revealing the underlying records. The direction is clear: help customers put sensitive information to work, with stronger protection during execution and less unnecessary disclosure.
Resilience is a Design Principle, not a Recovery Plan
AI-powered services will become part of the operational core of the enterprise. That raises the standard for availability and recovery. Resilience must be designed across architecture, operations, data protection, and organizational readiness.
At Microsoft Ignite, we’re taking another step toward making resilience continuous, application-centric, and increasingly AI-powered. Azure Infrastructure Resiliency Manager reaches general availability, while the resiliency agent in Azure Copilot helps customers move from insight to action across the resiliency lifecycle: understanding application dependencies, assessing zonal resiliency, prioritizing recommendations, generating resilient infrastructure as code, and validating readiness through recovery plans and drills. We’re also bringing Azure Backup into the agent experience, making it easier to identify protection gaps, strengthen backup security, and act through a conversational experience. At the same time, we’re extending the application-centric foundation of Azure Infrastructure Resiliency Manager beyond availability zones to regional resiliency, while new cyber recovery capabilities help organizations identify suitable recovery points and restore safely at scale. Together, these investments bring us closer to one connected, AI-powered resiliency experience across infrastructure, regions, data, and cyber recovery.
The Platform is the Multiplier
These innovations are most powerful when they reinforce one another across the connected system of Azure. Custom silicon can improve efficiency; flexible compute options can adapt to changing demand; intelligent optimization can make resources go further; confidential computing can expand the set of data that can be used safely; and resilient architecture can keep critical systems available. Working cohesively together, they create an infrastructure foundation that is ready for enterprise AI at scale.
This is the deeper message behind Azure Compute at Ignite. The next era will not be defined by a single chip, virtual machine, security control, or recovery service. It will be defined by how effectively the platform turns those capabilities into reliable customer outcomes.
Recommended links
Continue the conversation at Microsoft Ignite with these Azure infrastructure sessions:
Executive and platform perspective
- BRK542 — Inside Azure innovations with Mark Russinovich
- BRK452 - AI-powered resiliency for modern workloads with Mark Russinovich
- BRK429 — The evolution of Azure infrastructure for AI models, apps, and agents
Compute innovation and efficiency
- BRK440 — Cobalt 200: Delivering Azure’s best performance and efficiency
- BRK430 — Architecting your workloads for maximum compute flexibility
- BRK432 — Running AI workloads efficiently on Azure general compute
- THR895 — AI for FinOps, continuous optimization for every workload
Confidential computing and security
- BRK431 — Secure by default: Azure’s multilayered embedded security capabilities
- BRK443 — From models to infrastructure: Confidential AI at hyperscale on Azure
- LTG1288 — Where sensitive data meets trusted execution
- LTG1289 — Can competitors collaborate without sharing data?
- LTG1290 — The future of trust: Proof without compromising privacy
Resilience and recovery
- BRK437 — Learning from disruption: Resilience against force majeure
- BRK445 — Architecting for end-to-end resiliency on Azure
- LAB721 — Hands-on with resiliency in Azure: Assess, recommend, recover
- LTG1284 — Data resiliency and cyber recovery with Azure Backup
- LTG1467 — From resilience gaps to confident recovery