Advances in multi-modal generative models are enabling new applications, from storytelling to automated media synthesis. Most current workloads generate simple outputs (e.g., image generation from a prompt) in batch mode, often requiring several seconds even for basic results. Serving real-time multi-modal workflows at scale is costly and complex, requiring efficient coordination of diverse models (each with unique resource needs) across language, audio, image, and video, all under strict latency and resource constraints. We tackle these challenges through the lens of real-time podcast video generation, integrating LLMs, text-to-speech, and video-audio generation. To meet tight SLOs, we design an adaptive, modular serving system, StreamWise, that dynamically manages quality (e.g., resolution, sharpness), model/content parallelism, and resource-aware scheduling. We leverage heterogeneous hardware to maximize responsiveness and efficiency. For example, the system can lower video resolution and allocate more resources to early scenes. We quantify the trade-offs between latency, cost, and quality. The cheapest setup generates a 10-minute podcast video on A100 GPUs in 1.4 hours (8.4x slower than the real-time) for less than \$25. StreamWise enables high-quality real-time streaming with a sub-second startup delay under $45.

As artificial intelligence (AI) inference scales to billions of queries, estimates of per-query energy use are increasingly important for capacity planning, efficiency interventions, and policy. Yet many public estimates assume non-production settings, leading to systematic overestimation. We introduce a bottom-up framework estimating inference energy from token throughput, node power, and overhead under large-scale deployment assumptions. For frontier-scale models (>200B parameters) on H100 nodes, we estimate a median energy of 0.31 Wh/query (interquartile range [IQR] 0.16–0.60), indicating that widely cited estimates are overstated by 4–20×. In test-time scaling scenarios 15× longer than typical queries, the median energy rises 13× to 3.91 Wh (IQR 2.15–7.05). Across models, serving systems, and hardware, we estimate 8–20× line-of-sight energy reductions. At data-center scale, serving 1 billion queries/day requires 0.7 GWh; if 10% are long queries, demand rises to 1.7 GWh/day. With efficiency interventions, it falls to 0.8 GWh/day, mitigating the energy impact of test-time scaling.

The rapid rise of large language models (LLMs) has driven an enormous demand for AI inference infrastructure, mainly powered by high-end GPUs. While these accelerators offer immense computational power, they incur high capital and operational costs due to frequent upgrades, dense power consumption, and cooling demands, making total cost of ownership (TCO) for AI datacenters a critical concern for cloud providers. Unfortunately, traditional datacenter lifecycle management (designed for general-purpose workloads) struggles to keep pace with AI's fast-evolving models, rising resource needs, and diverse hardware profiles. We rethink the AI datacenter lifecycle scheme across three stages (building, IT provisioning, and operation) highlighting how power, cooling, and networking decisions affect long-term TCO. We focus on hardware refresh strategies aligned with evolving hardware trends and evaluate operational software optimizations that further reduce cost. While these optimizations at each stage yield benefits, unlocking the full potential requires rethinking the entire lifecycle. We present a holistic lifecycle management framework that optimizes decisions across all three stages, accounting for workload dynamics, hardware evolution, and system aging. Our approach reduces TCO by 40% compared to traditional methods and offers guidelines for managing AI datacenter lifecycles in the future.

The efficiency of a datacenter rests on its control plane policies. Designing these policies is increasingly hard: the hardware-software stack grows fast, the design space is vast and interdependent, and prototyping a single policy takes months. Agentic AI promises to automate this search. Off the shelf, however, it falls short on three fronts. It is not formal: with no structured, searchable statement of the problem, the search has little structure to exploit and hard constraints are not guaranteed. It is not transferable: each task is solved from scratch, so nothing learned on one task carries to the next. Finally, it is not systematic: relying on the LLM as the sole source of candidates, it explores a narrow slice of the design space and settles into local optima. We introduce AtumAI, a framework that generates datacenter control-plane policies with agentic AI, making the process formal, transferable, and systematic. From a goal stated in plain language, AtumAI autonomously proposes, tests, and refines candidate policies until one satisfies the request. It does so through two components. The Datacenter Task Compiler automates problem formulation: it compiles the request into a formal, machine-checkable, and searchable specification of the task's objectives, constraints, decision variables, and evaluation methodology. The Evolutionary Design Discovery Loop then searches this specification, expanding the search beyond the LLM itself via a diffusion model, an evolutionary algorithm, and a surrogate model. Together, they reduce onboarding a new task from months of engineering to writing its description. We evaluate AtumAI on three control-plane tasks with distinct problem scopes, design spaces, and trade-offs: workload placement, resource scaling, and power management. Across all tasks, the policies generated by AtumAI consistently outperform expert-engineered baselines.

Demand for AI accelerators is rapidly increasing rack power density, with projections approaching 1MW per deployment by 2027. This poses a major challenge for datacenter power delivery designers. As power densities increase, a datacenter designed for a different target density may strand power, i.e., may be unable to use all the power that its delivery hierarchy has provisioned. Designs must remain efficient over long datacenter lifetimes and multiple hardware generations. Power utilization is particularly important as grid power capacity is a scarce resource in the AI era. Designing an efficient power delivery hierarchy for the long run is difficult because rack placement feasibility, workload impact, and cost depend jointly on electrical topology, deployment granularity, placement policy, power oversubscription, and workload mix. Moreover, each of these factors evolve over time, have inter-dependencies across multiple resource dimensions, and generally do not lend themselves to closed-form analysis. To address this challenge, we develop a framework for evaluating datacenter power delivery designs using throughput, power, and cost metrics over realistic arrival, oversubscription, and decommissioning sequences. The framework combines projection models for GPU, compute, and storage deployments with operational factors grounded in production data from Microsoft Azure. Our results show that multi-resource stranding materially changes deployable capacity, effective capital expenditure, and delivered performance, and quantify how rising density from rack- and pod-scale AI systems shapes these outcomes. For AI datacenter design, the relevant planning objective is not installed megawatts, but deployable capacity over time.

Datacenters consume many megawatts of power, and regularly encounter scenarios that require modulating their power draw. These scenarios include datacenter infrastructure failures, power grid failures, grid services, and more, spanning a diverse range of requirements in terms of the power magnitude, the scope of the reduction, the notice time, and other dimensions. To address these scenarios, we have built Slasher, a general system for modulating the power of datacenters to handle scenarios ranging from individual racks to regional multi-datacenter grid events. Slasher coordinates datacenter resources with the goal of meeting power targets while minimizing negative impact on hosted workloads. In this paper, we review the main power modulation scenarios, characterize the power reduction levers using data from production cloud datacenters, describe Slasher's system architecture, and formulate the cloud datacenter power modulation control problem. We also develop a high-fidelity datacenter simulator and propose a workload impact model, using them to design and evaluate power control algorithms.

AI coding agents like GitHub Copilot, Claude Code, and Codex interleave multi-step LLM inference with tool execution, creating a workload different from chatbots. We present the first production-scale characterization of this workload using sampled GitHub Copilot traces from June 2026, comprising 3.2M users, 13M sessions, 761M LLM calls, and 95T tokens. Our analysis reveals distinctive workload properties with important systems implications. For example, agentic coding sessions consist of sparse user-initiated turns, each unfolding into an autonomous agent loop of LLM calls almost always coupled with tool execution. This structure yields KV cache hit rates averaging 90

Cloud platforms remain underutilized despite multiple proposals to improve their utilization (e.g., disaggregation, harvesting, and oversubscription). Our characterization of the resource utilization of virtual machines (VMs) in Azure reveals that, while CPU is the main underutilized resource, we need to provide a solution to manage all resources holistically. We also observe that many VMs exhibit complementary temporal patterns, which can be leveraged to improve the oversubscription of underutilized resources. Based on these insights, we propose Coach: a system that exploits temporal patterns for all-resource oversubscription in cloud platforms. Coach uses long-term predictions and an efficient VM scheduling policy to exploit temporally complementary patterns. We introduce a new general-purpose VM type, called CoachVM, where we partition each resource allocation into a guaranteed and an oversubscribed portion. Coach monitors the oversubscribed resources to detect contention and mitigate any potential performance degradation. We focus on memory management, which is particularly challenging due to memory's sensitivity to contention and the overhead required to reassign it between CoachVMs. Our experiments show that Coach enables platforms to host up to ~26% more VMs with minimal performance degradation.

The rising demand for generative large language models (LLMs) poses challenges for thermal and power management in cloud datacenters. Traditional techniques often are inadequate for LLM inference due to the fine-grained, millisecond-scale execution phases, each with distinct performance, thermal, and power profiles. Additionally, LLM inference workloads are sensitive to various configuration parameters (e.g., model parallelism, size, and quantization) that involve trade-offs between performance, temperature, power, and output quality. Moreover, clouds often co-locate SaaS and IaaS workloads, each with different levels of visibility and flexibility. We propose TAPAS, a thermal- and power-aware framework designed for LLM inference clusters in the cloud. TAPAS enhances cooling and power oversubscription capabilities, reducing the total cost of ownership (TCO) while effectively handling emergencies (e.g., cooling and power failures). The system leverages historical temperature and power data, along with the adaptability of SaaS workloads, to: (1) efficiently place new GPU workload VMs within cooling and power constraints, (2) route LLM inference requests across SaaS VMs, and (3) reconfigure SaaS VMs to manage load spikes and emergency situations. Our evaluation on a large GPU cluster demonstrates significant reductions in thermal and power throttling events, boosting system efficiency.

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

Compound AI Systems, integrating multiple interacting components like models, retrievers, and external tools, have emerged as essential for addressing complex AI tasks. However, current implementations suffer from inefficient resource utilization due to tight coupling between application logic and execution details, a disconnect between orchestration and resource management layers, and the perceived exclusiveness between efficiency and quality. We propose a vision for resource-efficient Compound AI Systems through a declarative workflow programming model and an adaptive runtime system for dynamic scheduling and resource-aware decision-making. Decoupling application logic from low-level details exposes levers for the runtime to flexibly configure the execution environment and resources, without compromising on quality. Enabling collaboration between the workflow orchestration and cluster manager enables higher efficiency through better scheduling and resource management. We are building a prototype system, called Murakkab, to realize this vision. Our preliminary evaluation demonstrates speedups up to similar to 3.4x in workflow completion times while delivering similar to 4.5x higher energy efficiency, showing promise in optimizing resources and advancing AI system design.

Large Artificial Intelligence (AI) training workloads spanning several tens of thousands of GPUs present unique power management challenges. These arise due to the high variability in power consumption during the training. Given the synchronous nature of these jobs, during every iteration there is a computation-heavy phase, where each GPU works on the local data, and a communication-heavy phase where all the GPUs synchronize on the data. Because compute-heavy phases require much more power than communication phases, large power swings occur. The amplitude of these power swings is ever increasing with the increase in the size of training jobs. An even bigger challenge arises from the frequency spectrum of these power swings which, if harmonized with critical frequencies of utilities, can cause physical damage to the power grid infrastructure. Therefore, to continue scaling AI training workloads safely, we need to stabilize the power of such workloads. This paper introduces the challenge with production data and explores innovative solutions across the stack: software, GPU hardware, and datacenter infrastructure. We present the pros and cons of each of these approaches and finally present a multi-pronged approach to solving the challenge. The proposed solutions are rigorously tested using a combination of real hardware and Microsoft's in-house cloud power simulator, providing critical insights into the efficacy of these interventions under real-world conditions.

With the increasing adoption of large language models (LLM), agentic workflows, which compose multiple LLM calls with tools, retrieval, and reasoning steps, are increasingly replacing traditional applications. However, such workflows are inherently error-prone: incorrect or partially correct output at one step can propagate or even amplify through subsequent stages, compounding the impact on the final output. Recent work proposes integrating verifiers that validate LLM output or actions, such as self-reflection, debate, or LLM-as-a-judge mechanisms. Yet, verifying every step introduces significant latency and cost overheads. In this work, we seek to answer three key questions: which nodes in a workflow are most error-prone and thus deserve costly verification, how to select the most appropriate verifier for each node, and how to use verification with minimal impact to latency? Our solution, Sherlock, addresses these using counterfactual analysis on agentic workflows to identify error-prone nodes and selectively attaching cost-optimal verifiers only where necessary. At runtime, Sherlock speculatively executes downstream tasks to reduce latency overhead, while verification runs in the background. If verification fails, execution is rolled back to the last verified output. Compared to the non-verifying baseline, Sherlock delivers an 18.3% accuracy gain on average across benchmarks. Sherlock reduces workflow execution time by up to 48.7% over non-speculative execution and lowers verification cost by 26.0% compared to the Monte Carlo search-based method, demonstrating that principled, fault-aware verification effectively balances efficiency and reliability in agentic workflows.

The rapid rise of large language models (LLMs) has been driving an enormous demand for AI inference infrastructure, mainly powered by high-end GPUs. While these accelerators offer immense computational power, they incur high capital and operational costs due to frequent upgrades, dense power consumption, and cooling demands, making total cost of ownership (TCO) for AI datacenters a critical concern for cloud providers. Unfortunately, traditional datacenter lifecycle management (designed for general-purpose workloads) struggles to keep pace with AI's fast-evolving models, rising resource needs, and diverse hardware profiles. In this paper, we rethink the AI datacenter lifecycle scheme across three stages: building, hardware refresh, and operation. We show how design choices in power, cooling, and networking provisioning impact long-term TCO. We also explore refresh strategies aligned with hardware trends. Finally, we use operation software optimizations to reduce cost. While these optimizations at each stage yield benefits, unlocking the full potential requires rethinking the entire lifecycle. Thus, we present a holistic lifecycle management framework that coordinates and co-optimizes decisions across all three stages, accounting for workload dynamics, hardware evolution, and system aging. Our system reduces the TCO by up to 40% over traditional approaches. Using our framework we provide guidelines on how to manage AI datacenter lifecycle for the future.

Power dissipation in the latest accelerator chips for artificial intelligence has exceeded 1 kW and is increasing with each successive generation, while simultaneously introducing new thermal resistances and spatially varying temperature requirements due to heterogeneous integration. In contrast, cooling design has remained largely homogeneous, relying predominantly on straight channels or arrays of fins which are not tailored to the underlying chip. As data center operators target to increase coolant supply temperatures of data centers to reduce electricity and water consumption, more power needs to be extracted from increasingly complex chips with reduced temperature differentials between the junction and coolant. Hence, cooling is becoming the main driver for both chip performance and data center efficiency. Microfluidic cooling, with coolant directly flowing through channels embedded inside the silicon chip, has been proposed as a candidate to overcome the performance limitations of standard cold plates, as it aims to minimize thermal resistance by eliminating interfaces between the chip and the coolant. In addition, the close coupling between the heat source and cooling, combined with mature silicon-based microfabrication methods, enables more design freedom for cooling structures beyond straight channels and fins. However, while prior demonstrations of microfluidic cooling on function integrated circuits have shown the feasibility of this cooling integration, they suffered from high pressure drops, large temperature gradients and localized hot-spots, in addition to packaging and integration concerns. In this work, we show that the chip-level spatial power distribution is the source of these discrepancies, and that microfluidic topology optimization can simultaneously address the temperature and pressure drop challenges with microfluidic cooling, by creating a hierarchical network of channels that balance heat transfer and pressure drop mimicking the arteries, veins and capillaries in the circulatory system. Our result demonstrates that, by utilized a bio-inspired chip-aware microfluidic cooling architecture, up to 18 ⁰C reduction in hot-spot temperature rise on functional CPUs can be achieved, or a reduction in pressure drop by over 67% compared to a canonical design of pin fins. We demonstrate a 3x reduction in core temperature spread versus unoptimized designs, and furthermore observe that average thermal resistance can be reduced by up to 55% compared to standard cold plate cooling. These results validate that including microfluidic cooling as an integral part of the chip design process may be a viable path to allow for a further extension of the roadmap of silicon CMOS chips with increasing power levels, while simultaneously addressing the sustainability concerns at the data center scale.

Agentic workflows commonly coordinate multiple models and tools with complex control logic. They are quickly becoming the dominant paradigm for AI applications. However, serving them remains inefficient with today's frameworks. The key problem is that they expose workflows as opaque sequences of model and tool calls that tightly couple agent logic with model and hardware choices. Often, these workflow components are fragmented across different entities, preventing systems from reasoning about trade-offs across accuracy, latency, energy, and cost. This leads to resource waste and degraded service-level objectives (SLOs). We present Murakkab, a resource-efficient serving system for agentic workflows. Murakkab introduces a declarative abstraction that decouples workflow specification from execution configuration. A profile-guided optimizer and adaptive runtime jointly manage the full stack: orchestrating workflow components, mapping them to models and hardware, and dynamically reconfiguring execution to satisfy user-defined SLOs. By exposing the internal structure of agentic workflows, Murakkab enables cross-layer optimization that existing frameworks and cloud schedulers cannot achieve. Our evaluation on diverse workflows shows that reduces GPU usage by up to 2.8×, energy consumption by 3.7×, and cost by 4.3× while maintaining SLOs.

Addressing climate change requires accelerating the development of sustainable alternatives to energy- and water-intensive technologies, particularly for rapidly growing infrastructure such as data centres and cloud1. Here we present a life cycle assessment study examining the impacts of advanced cooling technologies on cloud infrastructure, from virtual machines to server architecture, data centre buildings and the grid. Life cycle assessment is important for early-stage design decisions, enhancing sustainability outcomes alongside feasibility and cost analysis2. We discuss constructing a life cycle assessment for a complex cloud ecosystem (including software, chips, servers and data centre buildings), analysing how different advanced cooling technologies interact with this ecosystem and evaluating each technology from a sustainability perspective to provide adoption guidelines. Life cycle assessment quantifies the benefits of advanced cooling methods, such as cold plates and immersion cooling, in reducing greenhouse gas emissions (15-21%), energy demand (15-20%) and blue water consumption (31-52%) in data centres. This comprehensive approach demonstrates the transformative potential of life cycle assessment in driving sustainable innovation across resource-intensive technologies.

To combat climate change, we must reduce carbon emissions from hyperscale cloud computing. Compute servers cause the majority of a general-purpose cloud's emissions. Thus, we are motivated to design carbon-efficient compute server stock keeping units (SKUs), or GreenSKUs, using recently available low-carbon components. We built three GreenSKU prototypes, integrating energy-efficient CPUs, reusing old dynamic RAM via compute express link, and reusing old solid-state drives. We reveal challenges that limit GreenSKUs' carbon savings at scale and may prevent their adoption by cloud providers. To address these challenges, we developed a novel framework, GSF (GreenSKU Framework), that enables cloud providers to systematically evaluate GreenSKUs' carbon savings at scale. By implementing GSF within Microsoft Azure's production constraints, we demonstrate that GreenSKUs reduce net cloud emissions by 8%, which is globally significant. This work is the first to demonstrate and quantify how carbon-efficient server designs translate to measurable cloud-scale emissions reductions, enabling meaningful contributions to cloud sustainability goals.

Today, cloud workloads are largely opaque to the cloud platform. Typically, the only information the platform receives is the virtual machine (VM) type and possibly a decoration to the type (e.g., the VM is evictable). Similarly, workloads receive minimal information from the platform; generally, only telemetry from their VMs or occasional signals (e.g., just before a VM is evicted). The narrow interface between workloads and platforms has several drawbacks: (1) a surge in VM types and decorations in public cloud platforms complicates customer selection; (2) key workload characteristics (e.g., low availability requirements) are often unspecified, hindering platform customization for optimized resource usage and cost savings; and (3) workloads may be unaware of potential optimizations or lack sufficient time to react to platform events. To resolve these issues and improve cloud efficiency, we propose Workload Intelligence (WI), a framework for enabling dynamic bi-directional communication between cloud workloads and cloud platform.

The growing demands for computational power in cloud computing have led to a significant increase in the deployment of high-performance servers. The growing power consumption of servers and the heat they produce is on track to outpace the capacity of conventional air cooling systems, necessitating more efficient cooling solutions such as liquid immersion cooling. The superior heat exchange capabilities of immersion cooling both eliminates the need for bulky heat sinks, fans, and air flow channels while also unlocking the potential go beyond conventional 2D blade servers to three-dimensional designs. In this work, we present a computational framework to explore designs of servers in three-dimensional space, specifically targeting the maximization of server density within immersion cooling tanks. Our tool is designed to handle a variety of physical and electrical server design constraints. We demonstrate our optimized designs can reduce server volume by 25--52% compared to traditional flat server designs. This increased density reduces land usage as well as the amount of liquid used for immersion, with significant reduction in the carbon emissions embodied in datacenter buildings. We further create physical prototypes to simulate dense server designs and perform real-world experiments in an immersion cooling tank demonstrating they operate at safe temperatures. This approach marks a critical step forward in sustainable and efficient datacenter management.