video_OCP2026_Microsoft_主題演講_前沿AI基礎設施_20261015

Good morning. >> [applause] >> It's great to be back at OCP. I've been coming to this stage for a while now, and every year the work this community does matter. And this year is no different. It matters a lot. Let me walk you through the why, and I'll be talking about scaling AI for the frontier era. Every major technology transition begins with a breakthrough. But it creates lasting value only when that breakthrough becomes broadly available. AI is now reaching that moment. For the past several years, much of the conversation has centered around frontier models. Larger models, more parameters, and more intelligence. And now as we stand at the cusp of AI reaching broader utility, the long-term opportunity in front of us is not just building models, but broad diffusion, delivering AI through the products that people already use and trust. So, it is now up to us to turn AI into something uniquely its own. By redesigning our organization's work, scale rapidly with greater agility, uh generate and generate value by combining machine intelligence and human judgment. Every company has human capital, the knowledge, judgment, relationships, and ingenuity of its people. And every company is now also building token capital. It's that AI capability that captures that knowledge, applies it, and gets better with every interaction. So, the opportunity is in bringing the human capital and token capital together in a continuous learning loop within a frontier firm, in which people contribute judgment, creativity, and purpose, while AI helps make that expertise repeatable, scalable, and continuously improving. This will enable intelligence to flow into the workflows and applications that we use every day. The ultimate promise of such a frontier ecosystem isn't intelligence for its own sake, but useful intelligence that helps people live healthier, learn faster, create more, and solve complex problems. You know, as uh Satya said, if the AI we build isn't helping humanity and under human control, it isn't worth pursuing. So, this is our design goal. But, that vision rests on on a physical reality. There can be no frontier firm without a frontier infrastructure. Intelligence has to feel instantaneous, but it's not infrastructure-free. Every model trained, every agent deployed, every token generated rests on this complex physical infrastructure. Intel- intelligence no longer comes from a a single chip or a single component, but it comes from thousands of these components operating as one coordinated system. So, what the scaling for scaling AI for the frontier era actually means it isn't simply about building more infrastructure. It's about producing useful intelligent useful intelligence from every watt, every byte, every chip, and every dollar. The next phase of AI will be defined defined by how effectively we turn an increasingly complex infrastructure stack into a unified engine for useful intelligence. So, what I'm going to be doing is talking about building that how we build this foundation together, and how we turn this capability into useful intelligence, and scale it efficiently. At Microsoft, we believe the answer starts with full stack co-design. Rather than optimizing each layer, the silicon, the software, the systems models, we look at designing it all together and co-optimizing it. The goal is simple. Every layer works together and contributes to what customers ultimately care about. And the customers are looking for lower operating cost where you get more from every dollar spent and every token. Faster deployment that's made possible through the standards that we create to scale and interoperate operate together. Stronger governance primarily around security where it's built right right from day one versus an afterthought. And dependable performance at scale. So, this moves from isolated AI experiments to an AI-driven operation that goes end-to-end. So, let me start with how we build the Azure infrastructure starting with purpose-built silicon. We're excited to say that yesterday uh Microsoft announced Maia 300, our next generation AI accelerator in the Maia family to advance the inference efficiency frontier. It delivers up to 7x gen over gen inference performance. And and we will be scaling Maia systems to multi-gigawatts uh in 2027. So, it's quite exciting. Now, Maia 200, our current generation AI accelerator, is production proven and outperforming other GPUs in our fleet today. Maia Maia 200 is designed alongside our infrastructure where it's optimized for real workloads rather than the technical or theoretical benchmarks. It demonstrates Microsoft's system of co-design philosophy. The this efficiency is only possible by co-designing models, application frameworks, reinforcement learning environments uh together as one unified stack. At the same time, Maia 200, a purpose-built silicon, is flexible to support a broad ecosystem of models. And it's already scaling globally. Now, AI is not all about accelerators. Agentic workloads create new demands for CPUs. Every agent interaction generates orchestration, data retrieval, memory management, and service calls that must occur quickly and efficiently. This is why we built 200 as an agent native CPU. Uh Cobalt with higher memory and bandwidth and improved power efficiency helps lower the infrastructure cost while improving the user experience. This is just another example of designing infrastructure around emerging workloads uh as it's a very dynamic in the space. And this is what truly differentiates Microsoft. It's this tight feedback loop of of the of the workloads because Microsoft runs a very global fleet. We get the information about the workloads and we co- and we and that informs us into our development across the full stack. And the next opportunity for us is driving efficiency on the storage stack. Storage operations require efficient data compression, networking transactions, and security functions, which all consume valuable CPU resources. Azure Boost DPU offloads these operations to a dedicated silicon. And by moving these storage services onto a purpose-built platform, we can deliver about 4x more performance and 3x less power. Also, the Azure integrated HSM makes security part of the hardware foundation with our in-house chip in all our servers in the data center. It also brings cryptographic operations closer to the source so that so that the keys are well protected within the hardware boundary. So, these same advances in AI as we build this infrastructure, we also do use AI quite a bit in in our development of of of the full hardware stack all the way from silicon to system. We use AI we we use AI to augment our development to to drive innovations faster. So, now compute creates a capability in the intelligence. The network is what makes it usable at scale. So, the scale of AI is fundamentally changing what we ask of the networking. Today, we are building AI systems that connect hundreds of accelerators into a massive supercomputer. And increasingly that scale now extends even between connecting across data centers. And one of the ways we are doing this is by working with the industry to build foundations and and standards. Microsoft is working with our industry partners and have developed multiple multi-path reliable connection or MRC. Uh MRC fundamentally rethinks how Ethernet networks operate for AI. MRC is multi-plane network architecture enables Azure to scale beyond 100,000 GPUs using just two network tiers, reducing network power conversion reducing power consumption um and and helping minimize congestion and and and and improving reliability. Networking is especially important for synchronous synchronous model training uh where the where the job is constrained by the lowest connection or even worse disrupted by a link failure or congestion. MRC intelligently distributes its traffic across the multiple network path. This way this reliability and performance is foundational to frontier scale synchronous model training. And we remain committed to this to be be to the ecosystem so that it can be built on so it can be implemented and built on. And we'll continue to lead in this area. Now, we are also building AI fabrics and advancing open infrastructure for scale up. Um you know, at the high level network, you know, there's three levels of network. There's network connections that are high memory and low latency for the scale up network. High performance for the for the scale out network and then network connections across across clusters and data centers. Our work around scale up and the foundation of OCI MSA is helping advance the high bandwidth power efficient connectivity required for the next generation of AI platforms. OCI is the first optical specification for multi-rack scale up systems. This establishes the open interoperable optical specifications and secures a scalable multi-vendor supply chain. The first spec was released this year and then we invite the industry to provide feedback to it. Now, Microsoft is also working across OCP and the ecosystem to create open open specs for the Ethernet scale up network. This was first announced at Global Summit last year. And now glad to say that this extends to more than 175 organizations. Because ultimately the challenge isn't simply about connecting more GPUs. It's about turning massive distributed compute into one efficient resilient system capable of delivering useful intelligence at scale. Now, let's talk about power delivery. Through OCP and part of open systems for AI, Microsoft is helping lead the industry's transition to 800-volt power infrastructure. Mind Diablo is Microsoft's disaggregated architecture for high voltage separating power conversion from the compute rack uh that their compute rack and the power can sort of scale independently. Its modular design uh delivers plus minus 400 V or 800 V output uh differential outputs scaling from uh 1 MW to beyond uh to beyond uh sorry, scaling from 100 kW to beyond 1 MW. And we've contributed this back uh to OCP. And our vision extends beyond the rack. We are re- We are rethinking power delivery from the data center, all the way from the data center to the silicon. As we move into the future, solid-state uh transformer-based architecture uh is is being advanced. Microsoft and industry partners exploring new ways to perform intelligent power conversion closer to the source while reducing the complexity of multiple AC-DC conversions. With the recent spec release, Microsoft is collaborating with OCP, utilities, and vendor partners to help define requirements and reference architecture for the utility solid-state deployment uh supporting the next generation of power requirements. We're also working with industry partners to define a common uh grid interface for uh for uh specifications for data centers. The first spec was just released last month, and will enable utilities and operators to to design to the same requirements uh and and evolve architectures evolve. Now, every watt delivered to a chip eventually becomes heat. As AI infrastructure density continues to increase cooling is becoming a first-class design consideration as we design the infrastructure. And here too, Microsoft is approaching cooling as a data center-wide challenge uh driving innovation across the stack. Our third-generation Azure heat exchanger unit, Mount Atlas, nearly doubles the cooling capacity while staying within the same operational footprint. At the data center level, Microsoft is advancing zonal cooling strategy in its next generation data centers to handle the growing mix of liquid cool racks and air cool racks. So, by tailoring cooling to the specific needs of each system and introducing multiple independent water loops, each supplying coolant at different temperature, our goal is to increase efficiency and performance. Security has always been fundamental to Microsoft Cloud, and we believe the strongest security foundations are built on openness, transparency, and industry collaboration. And Microsoft has helped advance this. We've been founding members of the Confidential Computing Consortium. Uh through OCP, we will help establish Caliptra as the open source silicon root of trust. Uh and we've also contributed to OCP safe as a com for a common security uh conformance framework for devices. But, the security landscape is rapidly changing. AI is increasing the speed of speed, scale, and sophistication of cyber attacks, uh amplifying risks such as credential theft, key infiltration, and signing abuse. At the same time, advances in quantum computing uh are changing the cryptographic horizon risk. Both these transformations at the center of these are keys and credentials. And these are among the most valuable assets, and we have to rethink how we protect them. Today, we are taking the next step in the journey. Microsoft, together with Marvell and Oracle, is introducing Tessera. Tessera evolves Azure's integrated HSM into an open source initiative for for integrated hardware security modules. The goal is straightforward. Create a more open, interoperable, and transparent foundation for protecting keys and credentials. Tessera builds on the workload-based architecture, bringing cryptographic operations closer to the workload. This strengthens key protections while improving cryptographic performance and simplifying the operating model. And this is already moving from architecture to implementation. Marvell HSM adopted the open protocols and interoperate with Azure's integrated HSM in a unified security foundation, where the keys don't have to be exported across HSMs. So, we invite the community to build with us and see Tessera interoperability in action at the Microsoft booth later today. And at the pace we're scaling, it's very important to keep the fleet healthy. Uh the fleet is continuing to grow in scale and complexity. Uh and so, on this just want to highlight a couple of things that we are doing at Microsoft. We are working with the industry to standardize through OCP common RAS, debug, diagnostics, firmware updates, and manage manageability specifications. We're also working with our partners and silicon vendors to build these into the designs from day one. The other thing I want to highlight is we're working on what we call a self-managed fleet. This is where the fleet given uh given all the capabilities that is there, this fleet self-monitor, self-detect, self-diagnose, and self-repair uh by itself, but leveraging years and years of uh data that we have that now we're leveraging AI to kind of use reasoning behind it. And what's even more exciting is that now with all this data and as we've kind of continued to advance this frontier, we're starting to see the fleet actually self-predict when an issue can put a it will potentially happen and mitigate it before a customer gets interrupted. It's quite exciting here. Because what's key over here is avoiding customer disruptions and and keeping a uptime high. Now, sustainability runs through everything I've talked about today. Microsoft is committed to bring carbon negative, water positive, and zero waste by 2030. As AI demands accelerate, they're rethinking every part of the data center ecosystem, from the computing communities that host our infrastructure to the hardware that powers it. It start with the places where we build. We design data center for local environmental conditions while it's also supporting ecosystem restoration efforts. From there, we focus on our facility the facilities are built and operated. Through OCP, we are helping advance low-carbon construction materials, including green steel and low-carbon concrete with pathways to reduce concrete emission by more than 50% per cubic yard. And as you know, water is another critical path of the equation. Today, more than 90% of our data center capacity runs on a closed-loop liquid cooling system, allowing water to be recirculated with no evaporative loss. And power And what are we doing about power? We've contracted tens of gigawatts of renewable energy across 24 countries, helping advance cleaner energy markets. And then even within the DC through intelligent power management, we can reduce server energy consumption by 35% when idle. Now, as infrastructure scales, reliable power during grid grid dis- reliable power during grid disturbances is becoming more critical. Through the OCP clean backup power initiative, Microsoft and industry partners are helping define open technology agnostic requirements for the next generation backup power power systems that can deliver DC-grade reliability while reducing our dependence on diesel fuels. And we equally focus on minimizing waste. We're finding ways to extend the useful life and material through our circular centers, more than 90% of our servers are are recycled or reused when they complete their first deployment. And we're exploring new opportunities beyond efficiency. We're actively contributing to the waste heat reuse subproject within OCP. And Microsoft as a company invest in skills, partnerships, and infrastructures within the communities we operate so that communities grow alongside us and as part of key of how we operate and move into the future and scale of AI. So let's bring it all together. We began with a vision where every organization becomes a frontier firm. Human capital and token capital in the continuous learning loop that turns knowledge, judgment, and ingenuity into scalable useful intelligence. And the goal is to maximize efficiency and productivity through full-stack co-design. We build an enterprise platform on the top of this infrastructure that gives everyone choice which would then lead to the creation of the frontier firm itself. The operating model that connects knowledge, people, and data and deploy govern agents learn from real usage and scale what works. Through this capability, intelligence becomes accessible and affordable to everyone. So with So I just want So with that, we must continue to innovate across company lines, agree on open interfaces wherever the industry needs to scale together. And so my ask of all of you is to continue to collaborate and this is why the conversation at OCP is very important. I look forward to learning from the rest of the whole global summit here and I encourage all of you all to visit the Microsoft booth to hear a little bit more to see a little bit more about what I've talked about today. Thank you. >> [applause]