Infrastructure for the frontier era: Scaling intelligence through open industry collaboration

At Microsoft, we’re on a mission to create what we call a “Frontier Firm”: a place where individuals and technology collaborate seamlessly, blending human skills, insights, and judgment with advanced technology to enhance intelligence in everyday tasks. This has immense potential: empowering people to innovate, tackle intricate challenges, and address some of the most pressing issues facing society.

To unlock this potential, we need advanced infrastructure. The next stage of Super Intelligence requires a cohesive infrastructure engine that offers increased intelligence and efficiency, making AI more accessible, dependable, and sustainable across areas like science, healthcare, education, and the economy. This ambitious endeavour involves our entire ecosystem and aligns with the vision we’ve been promoting through our active participation in the Open Compute Project Foundation.

Designing Silicon for Real Workloads

As we advance into the next era of Super Intelligence, the focus on inference as the primary workload is becoming clearer. Here, efficiency, latency, and costs are crucial. Our next-gen AI accelerator, the Azure Maia 300, is set to push the boundaries of efficiency in inference. This complete system is crafted for the novel AI workloads of today, such as long-context reasoning, latency-sensitive conversations, and complex multi-step tasks. With an expected 40% increase in tokens per watt compared to current GPU systems, the Maia 300 will provide up to seven times the token throughput when set up with 96 accelerators as compared to the Maia 200.

The Maia project showcases our approach to system co-design. It’s developed in tandem with Microsoft’s infrastructure and production workloads, integrating models, application frameworks, kernels, communication libraries, systems software, and custom silicon in a single optimization loop. The Maia 200 is being rolled out across the US, Europe, and Australia, achieving up to 1.7 times more tokens per watt for built-in AI functions in Microsoft Teams, Copilot, and Cowork experiences compared to leading GPU-based systems.

AI isn’t solely about accelerators; agentic workloads introduce new requirements for CPUs. Every interaction involving an agent necessitates quick orchestration, data retrieval, memory handling, and service calls—these all need to happen effectively. This is why we developed the Cobalt 200, a CPU specifically designed for agentic tasks. It boasts a 50% performance improvement over the Cobalt 100 and reduces the latency of agent calls by 33%, while also delivering 23% higher throughput for agent-related workloads.

Maia and Cobalt are engineered in conjunction with our major enterprise workloads. This gives Microsoft a unique advantage in optimising performance, reliability, and cost to meet customer needs in real-world applications.

We are applying the same design principles to storage and security. The Azure Boost Data Processing Unit effectively transfers tasks like compression, networking operations, and security functions away from CPUs to specially designed silicon. As a result, this approach enhances performance by up to four times while significantly lowering power consumption.

Additionally, Azure Integrated HSM makes security more robust by embedding it within the hardware foundation, using our proprietary security chip in all new datacentre servers. This strategy ensures that cryptographic tasks occur closer to workloads while keeping sensitive keys protected inside a secure hardware context—limiting exposure in application memory.

Creating an Open, Resilient AI Network

Computing drives intelligence, and networks make that intelligence usable at scale.

In collaboration with industry partners, Microsoft introduced the Multipath Reliable Connection (MRC). This innovative solution redefines Ethernet for AI, boasting a multi-plane architecture that can support over 100,000 GPUs. By having two network tiers, MRC reduces power consumption, complexity, potential failure points, and costs. It smartly distributes traffic across different paths and employs packet spraying to reduce congestion, ensuring reliable performance during large-scale training sessions.

Moreover, we are leading the charge in open scale-up connectivity with the Optical Compute Interconnect Multi-Source Agreement (OCI MSA), the first optical specification aimed at multi-rack scale-up systems that facilitates high-bandwidth, energy-efficient, multi-vendor infrastructure. We look forward to fostering industry participation in the OCI MSA going forward.

Additionally, the Ethernet for Scale-Up Networking (ESUN) workstream is uniting over 175 organisations to outline requirements for Ethernet-based scale-up networking. The group has released version 1.0 of the specifications and is currently working on 1.1 for later this year. Find out more about our network innovations within the open community.

Rethinking Power from Grid to Silicon

AI infrastructure is prompting a re-evaluation of how we deliver power. Through our collaboration with the Open Compute Project (OCP) and the Open Systems for AI initiative, Microsoft is leading the industry’s shift towards 800-volt direct current (VDC) infrastructure.

Our Mt. Diablo architecture, which separates power conversion from computation, allows each to scale independently. It’s designed to support approximately 100 kilowatts (kW) to over one megawatt (MW) with a modular set-up that includes redundant power feeds and telemetry, plus optional energy storage. Microsoft, in conjunction with OCP partners, has shared the Mt. Diablo specifications v0.7 and v1.0 with OCP.

Together with our partners, we’re working on solid-state transformer architectures to enhance power conversion efficiency, reaction speed, and scalability for next-gen 800 VDC datacentres. In conjunction with OCP’s Clean Backup Power initiative, we are establishing open requirements for dependable backup systems to minimise reliance on diesel fuels. We are also collaborating to create a universal grid-interface specification that aligns utilities, operators, and equipment vendors as hyperscale power architectures evolve. Read more about Microsoft’s strategy for the 800VDC transition and sustainable power infrastructure scaling.

Prioritising Cooling in Design

At Microsoft, we treat cooling as a comprehensive datacentre challenge, spearheading innovations from overall facility design to rack-level liquid cooling technologies.

Our latest Azure Heat Exchanger Unit, Mt. Atlas, nearly doubles the cooling capacity of its predecessor without altering its footprint. We’re also sharing public reference architecture through OCP to facilitate quicker industry adoption.

At the datacentre level, we’re implementing a zonal cooling strategy tailored for our new AI datacentres, managing the balance between liquid-cooled accelerators and traditional air-cooled systems. This approach aims to boost efficiency and performance by customising cooling solutions to meet the unique needs of each workload while implementing multiple independent water loops at varied temperatures, aligning with Microsoft’s energy and environmental goals.

Innovating Open Hardware Security

We believe that robust cloud security thrives on openness, transparency, and collaboration throughout the industry. This ethos is evident in our foundational participation in the Confidential Computing Consortium, our efforts to develop Caliptra as an open-source silicon root of trust, and our contributions to the OCP S.A.F.E. security conformance framework. As AI reshapes the cybersecurity landscape and quantum computing introduces new risks, securing the keys and credentials that underpin digital trust is more critical than ever.

In response to these challenges and as part of our dedication to security transparency and industry collaboration, Microsoft, along with Marvell and Oracle, is introducing the Tessera Integrated HSM. This open-source endeavour focuses on integrated hardware security modules (HSM). Tessera carries forward the direct attached HSM architecture introduced with Azure Integrated HSM, making its architecture, specifications, protocols, drivers, software, firmware, and silicon accessible to the broader community. Discover more about Tessera Integrated HSM.

Ensuring Fleet Resilience at Scale

As AI infrastructure expands, ensuring fleet reliability becomes vital for safeguarding customer workloads, capacity, and service quality. Through the OCP, Microsoft has released standard specifications for RAS, debugging, telemetry, and diagnostics and is contributing to efforts surrounding hardware fault management and fleet memory issue management. We’re leveraging years of fleet data, applying AI to predict and drive actions that prevent faults while enabling more autonomous issue resolution. You can learn more about how Microsoft is enhancing fleet reliability through industry partnerships.

Promoting Sustainability Across the Infrastructure Stack

Scaling AI in a responsible manner means taking the entire datacentre lifecycle into account as we strive to meet our ambitious 2030 goals of becoming carbon negative, water positive, and waste-free. Microsoft is dedicated to embedding sustainability into the design, construction, and operation of our datacentres, extending our commitment to the communities in which we’re based. Through OCP, we’re collaborating with partners to utilise low-carbon steel and concrete, aiming to reduce concrete emissions by more than 50% per cubic yard.

We are enhancing waste-heat reuse and developing Cloud Services Product Category Rules for consistent and auditable lifecycle-carbon reporting. Within our datacentres, practices such as closed-loop liquid cooling, smart power management, and agreements for renewable energy help decrease resource consumption, while Circular Centres are reusing or recycling over 90% of servers and components. Beyond the datacentre, we’re investing in local skills, partnerships, and infrastructure to ensure that the communities supporting our growth can also thrive alongside us. Read Microsoft’s 2026 Environmental Sustainability report.

Collaborating to Build the Frontier

Our aim is to empower every organisation to evolve into a Frontier Firm, integrating human knowledge with AI in an ongoing learning loop that transforms intelligence and creativity into scalable solutions. Achieving this vision necessitates comprehensive co-design across the entire stack: silicon, computation, memory, networking, power, cooling, models, and applications—all functioning together as a secure, reliable, efficient, and sustainable platform while ensuring customers have choices without overcomplicating their experience.

Realising such a future isn’t a task for one entity alone. Through OCP, we are promoting open interfaces, sharing insights from our global operational infrastructure, and collaborating across the industry to expedite innovation, ensuring that we don’t all tackle the same problems repeatedly. Together, we can transition frontier innovation into infrastructure that makes intelligence accessible and affordable for every organisation, in every sector, and across every part of the globe.

At the OCP Global Summit 2026, we’re inviting the community to join us in building this future.