Denise Muyco,RAVEL联合创始人兼首席执行官——访谈系列

面谈

Denise Muyco,RAVEL联合创始人兼首席执行官——访谈系列

在 Google 上的首选来源中添加 Unite.AI

Denise Muyco是RAVEL的共同创始人兼CEO，她领导开发复合计算和AI基础设施的管弦软件。 一位训练有素的工程师，她的职业生涯跨越了动力基础设施，流程优化，数字媒体，数据中心和虚拟工作室。 在关注AI之前，她帮助为温哥华的视觉效果社区开发共享的高性能计算基础设施，让大约30个工作室可以使用计算和安全的连接。 这一经验决定了她如何协调昂贵、有限的资源。 在RAVEL，她将系统工程视角应用于工作量布置，基础设施利用，互操作性，以及分布式和主权AI环境日益增长的需求。

RAVEL开发软件，帮助各组织部署、管理、扩大计算和人工智能工作量，跨越云层、空间和混合基础设施。 其产品包括用于高性能虚拟工作站的RAVEL Orchestrate，用于工作量自动化和优化的RAVEL Orchestrate AI，以及用于软件许可和遥测的Boost. 公司的工具连接了基础设施，应用程序，虚拟化，以及安全，同时提供集中控制和业务能见度。 RAVEL服务于数据中心，服务提供商，企业IT团队，AI研究小组，以及创意工作室。 公司在数字内容创建的根基上，注重简化复杂环境，帮助团队更有效地利用现有基础设施。

你在工程、媒体基础设施以及分布式计算等交叉环节工作多年后，建立了RAVEL。 为何RAVEL的使命随着基因AI的发展而演变

早在今日AI繁荣之前，我便在电源、计算和基础设施的交汇处工作。 在我职业生涯初期，我在一个核电厂实习，后来我在媒体和视觉效果方面的工作使我接触到了GPU，大量的计算要求和有形基础设施的限制。

大约16年前，温哥华日益壮大的视觉效果产业正直接遇到这些限制。 工作室需要大量的高性能计算，才能为高价值项目处理(或制作)日益复杂的电影和电视内容，但很少有人能够证明建造足够的专用基础设施以满足安全环境下的高峰需求是合理的。 权力也正在成为一种制约。 我们开发了共享的基础设施，使大约30个工作室能够使用高性能计算和安全的高波段连接。

共享基础设施解决了一个问题，但又产生了另一个问题：如何在各组织和工作量之间有效地分配昂贵和有限的资源？这一经验最终导致RAVEL对智能基础设施的策划方式。

大赦国际是同一基本挑战的更大版本，其复杂性和资本风险要大得多。 工作量和基础设施已发生变化，但必须使昂贵、受制约的资源有效地共同工作，这一点已变得更加重要。

在RAVEL之前，你帮助建立了一个由大约30个温哥华工作室，声音阶段，制作公司，和视觉效果之家组成的联合体，可以共享高性能计算基础设施。 这段经历教你如何进行分布计算

我在VFX行业的工作告诉我建造和配置基础设施只是问题的一半昂贵的计算必须安全地共享和有效地使用，而更广泛的基础设施——包括能够移动大量文件的安全网络——必须共同努力。 必须安排工作量，分配能力，转移大量数据，而生产队则在严格的预算和最后期限内推动取得日益雄心勃勃的成果。 必须不断优化基础设施的利用和绩效。

大赦国际是同一基本挑战的更大版本。 各组织正在对基础设施进行大量投资，但仅仅增加能力并不能保证昂贵的资源得到有效利用。 人工智能基础设施也日益在各地、供应商和环境之间分布，其驱动力是权力、能力、安全、经济、地理、监管和主权要求。 渲染和人工智能的工作量可能大不相同，但它们都面临着重要的基础设施和生态系统挑战：两者都依赖于高性能的计算，联网，安全和存储的高效合作。

因为人工智能基础设施越来越复杂，因此，协调性和管理更广泛的基础设施生态系统的政策至关重要。 它们共同帮助各组织更好地利用昂贵、有限的资源，同时协调日益分散的环境的工作量。

RAVEL形容自己是分布式AI工厂的管弦乐层。 “AI工厂”对你们意味着什么？今天基础设施堆积中缺少什么，使得有必要建立一个管弦乐层？

我认为，人工智能工厂不仅仅是GPU和高性能集群；它是一个综合的基础设施生态系统，带来一起计算，联网，存储，安全，电力，冷却，运行人工智能工作量所需的操作层。 人工智能工厂正在从孤立的集群演变成为跨越多个地点和区域的相互关联资源的生态系统。 这种复杂性越来越难于人工管理，或用脱节点解决办法管理。 一个管弦乐层通过协调堆叠的多样而复杂的片段和抽象一些基本的复杂性来解决这个问题，使运营商更容易作为一个协调的系统来管理基础设施。

由于各组织为AI调整了现有的数据中心，许多设施最初并不是围绕当今AI工作量的需求设计的。 AI今天面临结构不匹配，分解为三个关键问题。 第一个是权力，或找到一个有足够可靠能力处理规模工作量的地点。 第二是生态系统的分裂，AI堆积物常常被分割到孤立的环境，第三是运营和所有权的总成本的能见度有限，这可以掩盖AI基础设施的实际成本。 因此，工作量可能不尽人意，硬件可能利用不足，意外的业务费用可能使所有者的总成本上升，最终导致象征性成本上升。

人工智能基础设施越来越多地分布在超尺度云层、私人数据中心、精密系统以及专门的GPU供应商中。 各组织在试图管理不同环境中的AI工作量时遇到的最大挑战是什么？

在过去几年里，AI基础设施竞赛被一个简单的假设定义：建设更大。 更大的数据中心。 较大团。 再来点GPU 增加容量。 但是，搭建者模式正日益受到物质、经济和监管方面的制约。

权力受到限制。 网格访问需要时间。 经济、地理和主权日益决定了基础设施可以而且应该建在哪里。 新的挑战不仅仅是如何建设更多的AI基础设施，而是如何在合适的地方建设正确的基础设施，同时使所有这些环境作为一个协调的系统运作。 当能力分散在超大规模云层、私人数据中心、房地基础设施和专门供应商之间时，这一点就变得特别困难，而后者不一定是共同设计的。

权力是这些制约因素中最尖锐的。 电力供应日益决定AI基础设施的建设地点。

这就是联邦变得至关重要的地方。 它将各种环境的基础设施连接到共同的业务环境中。 然后根据能力、权力、业绩、成本、政策、复原力和主权决定工作量应在何处进行。

无论是各地还是供应商，AI基础设施都进一步扩展，这将来自使这种能力共同发挥作用。 联合会为这一新的现实创建了基础设施模型，而智能的管弦乐则使该模型在规模上运作。

我们的客户、RT-One是拉丁美洲最大的AI技术园区和数据中心的开发者

GPU的稀缺性受到了极大的关注，但仅仅添加更多的加速器并不一定会转化为更有用的AI计算。 在今天的AI基础设施中

AI的基础设施竞赛主要集中于规划了多少千兆瓦的新能力，以及投入了多少资本来建设。 将AI的基础设施问题仅仅视为缺乏GPU或能力，需要指出的是，计划容量与可用容量不同。

GPU集群可以出现完全预定，但当在堆栈的其他地方出现性能瓶颈时，昂贵的GPU仍然可以闲置。 这些问题可能包括存储吞吐量、数据管道、网络拥堵、工作时间安排效率低下、安全要求和结构、应用程序、驱动器和依赖关系不匹配。 添加GPU并不一定会转化为额外的可用计算。

增加基础设施并不能自动提高效率。 在许多情况下，它造成更多的仓储，同时抬高成本。

各组织需要更大的知名度和控制，以便改进利用情况，使基础设施决策自动化，并使工作量更好地与现有资源和业务优先事项保持一致。 提高能见度和加强控制有助于将基础设施资源的收集转化为一个更协调的操作环境。 没有确切的或普遍的数字与有多少能力可以通过管弦乐解锁有关；每项工作量都有独特的要求。 但有一点是清楚的：解决被困或使用不足的问题不仅仅是购买或建设更多的能力。

RAVEL强调基础设施、虚拟化技术、身份系统和应用之间的互操作性。 为什么随着人工智能工作量变得更加复杂和分散，互操作性变得特别重要？

随着基础设施的多样化，各组织需要灵活使用不同的技术和环境，而不必建立更孤立的基础设施筒仓。 互操作性使来自不同供应商和环境的计算、存储、联网和应用程序能够共同工作，随着基础设施的发展而保持选择。 但仅靠互操作性并不能决定如何使用这些资源。

这就是协调工作量、如何转移和如何在环境中执行政策的地方。 随着人工智能基础设施的日益分布和联邦化，互操作性为操作acr提供了灵活性

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Interviews

Denise Muyco, Co-Founder and CEO of RAVEL – Interview Series

Add Unite.AI to your preferred sources on Google

Denise Muyco is the Co-Founder and CEO of RAVEL, where she leads the development of orchestration software for complex compute and AI infrastructure. A trained engineer, her career spans power infrastructure, process optimization, digital media, data centers, and virtual studios. Before focusing on AI, she helped develop shared high-performance computing infrastructure for Vancouver’s visual effects community, giving roughly 30 studios access to compute and secure connectivity. That experience shaped her approach to coordinating expensive, constrained resources. At RAVEL, she applies a systems-engineering perspective to workload placement, infrastructure utilization, interoperability, and the growing demands of distributed and sovereign AI environments.

RAVEL develops software that helps organizations deploy, manage, and scale compute and AI workloads across cloud, on-premises, and hybrid infrastructure. Its products include RAVEL Orchestrate for high-performance virtual workstations, RAVEL Orchestrate AI for workload automation and optimization, and Boost for software licensing and telemetry. The company’s tools connect infrastructure, applications, virtualization, and security while providing centralized controls and operational visibility. RAVEL serves data centers, service providers, enterprise IT teams, AI research groups, and creative studios. Building on its roots in digital content creation, the company focuses on simplifying complex environments and helping teams use their available infrastructure more effectively.

You founded RAVEL after years of working at the intersection of engineering, media infrastructure, and distributed computing. What ultimately led you to start the company, and how has RAVEL’s mission evolved as generative AI has dramatically increased demand for compute?

I’m an engineer by training, and I’ve worked at the intersection of power, compute and infrastructure long before today’s AI boom. Early in my career I interned at a nuclear power plant, and later my work in media and visual effects exposed me to GPUs, large compute demands and the constraints of physical infrastructure.

About 16 years ago, Vancouver’s growing visual effects industry was running into those constraints directly. Studios needed enormous amounts of high-performance compute to process (or render) increasingly complex film and television content for high-value projects, but few could justify building enough dedicated infrastructure to meet peak demand in a secure environment. Power was also becoming a constraint. We developed shared infrastructure that gave roughly 30 studios access to high-performance computing and secure, high-bandwidth connectivity.

Sharing infrastructure solved one problem but created another: how do you allocate expensive, constrained resources efficiently across organizations and workloads? That experience ultimately led to RAVEL’s approach to intelligent infrastructure orchestration.

AI is a much larger version of the same fundamental challenge—with far greater complexity and capital at stake. The workloads and infrastructure have changed, but the need to make expensive, constrained resources work together efficiently has only become more important.

Before RAVEL, you helped create a consortium of roughly 30 Vancouver studios, sound stages, production companies, and visual effects houses that could share access to high-performance computing infrastructure. What did that experience teach you about distributed compute, and what parallels do you see between Hollywood’s rendering challenges then and the AI industry’s infrastructure challenges today?

My work with the VFX industry taught me that building and configuring the infrastructure was only half the problem. Expensive compute had to be shared securely and used efficiently, while the broader infrastructure—including a secure network capable of moving enormous files—had to work together. Workloads had to be scheduled, capacity allocated, and massive volumes of data moved—all while production teams pushed for increasingly ambitious results within strict budgets and deadlines. Infrastructure had to be continuously optimized for utilization and performance.

AI is a much larger version of that same fundamental challenge. Organizations are investing heavily in infrastructure, but adding capacity alone doesn’t guarantee that expensive resources will be used efficiently. AI infrastructure is also becoming increasingly distributed across locations, providers and environments, driven by power, capacity, security, economics, geography, regulation and sovereignty requirements. Rendering and AI can be very different workloads, but they share an important infrastructure and ecosystem challenge: both depend on high-performance compute, networking, security and storage working together efficiently.

As AI infrastructure grows more complex, simply adding compute doesn’t solve the operational challenge of making that capacity work efficiently. This is where orchestration and the policies governing the broader infrastructure ecosystem become critical. Together, they help organizations make better use of expensive, constrained resources while coordinating workloads across increasingly distributed environments.

RAVEL describes itself as an orchestration layer for distributed AI factories. What does an “AI factory” mean to you, and what is missing from the infrastructure stack today that makes an orchestration layer necessary?

The way I see it, an AI factory is more than GPUs and high-performance clusters; it’s an integrated infrastructure ecosystem that brings together compute, networking, storage, security, power, cooling, and the operational layers needed to run AI workloads. AI factories are evolving from isolated clusters to become ecosystems of interconnected resources spanning multiple sites and regions. That complexity becomes increasingly difficult to manage manually or with disjointed point solutions. An orchestration layer addresses this by coordinating the diverse and complex pieces of the stack and abstracting some of that underlying complexity, making the infrastructure easier for operators to manage as a coordinated system.

As organizations adapt existing data centers for AI, many facilities were not originally designed around the demands of today’s AI workloads. AI today faces a structural mismatch that breaks down into three key problems. The first is power, or finding a location with enough reliable power to run workloads at scale. Second is ecosystem fragmentation, as AI stacks are often split across isolated environments and aren’t designed to work together. Third is limited visibility into total cost of operation and ownership, which can obscure the real cost of AI infrastructure. As a result, workloads can go unoptimized, hardware can be underutilized, and unexpected operational costs can drive up total cost of ownership, ultimately contributing to higher token costs.

AI infrastructure is increasingly spread across hyperscale clouds, private data centers, on-premise systems, and specialized GPU providers. What are the biggest challenges organizations encounter when trying to manage AI workloads across these different environments?

For the past several years, the AI infrastructure race has been defined by a simple assumption: build bigger. Bigger data centers. Bigger clusters. More GPUs. More capacity. But the build-bigger model is increasingly running into physical, economic and regulatory constraints.

Power is constrained. Grid access takes time. Economics, geography and sovereignty increasingly dictate where infrastructure can, and should, be built. The new challenge is not simply how to build more AI infrastructure, but how to build the right infrastructure in the right places, while making all those environments operate as a coordinated system. That becomes particularly difficult when capacity is spread across hyperscale clouds, private data centers, on-premises infrastructure and specialized providers that were not necessarily designed to operate together.

Power is the sharpest of these constraints. Increasingly, power availability is determining where AI infrastructure gets built in the first place.

That’s where federation becomes critical. It connects infrastructure across environments into a common operational environment. Intelligent orchestration then determines where workloads should run based on capacity, power, performance, cost, policy, resiliency and sovereignty.

As AI infrastructure spreads further, whether across locations or providers, the competitive advantage won’t come simply from owning more capacity. It will come from making that capacity work together. Federation creates the infrastructure model for this new reality, while intelligent orchestration makes that model operational at scale.

This is the model one of our customers, RT-One, the developer of Latin America’s largest AI technology parks and data centers, is pursuing as it plans federated infrastructure across the Americas: connecting distributed capacity so it can ultimately be managed as a coordinated environment.

GPU scarcity has received enormous attention, but simply adding more accelerators does not necessarily translate into more useful AI compute. Where do you see the largest sources of inefficiency in today’s AI infrastructure, and how much additional capacity could companies unlock through better orchestration?

The AI infrastructure race has focused heavily on how many gigawatts of new capacity are planned and how much capital is being committed to build it. It is tempting to treat AI’s infrastructure problem as simply a shortage of GPUs or capacity, but this framing misses where inefficiency truly resides. It is important to note that planned capacity isn’t the same as usable capacity.

GPU clusters can appear fully booked, but expensive GPUs can still sit idle when performance bottlenecks emerge elsewhere in the stack. These can include storage throughput, data pipelines, network congestion, job scheduling inefficiencies, security requirements and mismatches across architecture, applications, drivers and dependencies. Adding GPUs doesn’t always translate into additional usable compute.

Adding more infrastructure doesn’t automatically increase efficiency. In many cases, it creates more silos while driving up costs.

Organizations need greater visibility and control so they can improve utilization, automate infrastructure decisions and better align workloads with available resources and business priorities. Greater visibility and control can help turn a collection of infrastructure resources into a more coordinated operating environment. There is no exact or universal number tied to how much capacity could be unlocked through orchestration; each workload has unique requirements. But one thing is clear: the answer to stranded or underutilized capacity isn’t simply to buy or build more.

RAVEL has emphasized interoperability between infrastructure, virtualization technologies, identity systems, and applications. Why is interoperability becoming particularly important as AI workloads become more complex and distributed?

As infrastructure becomes more heterogeneous, organizations need flexibility to use different technologies and environments without creating more isolated infrastructure silos. Interoperability allows compute, storage, networking and applications from different vendors and environments to work together, preserving choice as infrastructure evolves. But interoperability alone doesn’t determine how those resources should be used.

That’s where orchestration comes in—coordinating where workloads run, how they move and how policies are enforced across environments. As AI infrastructure becomes increasingly distributed and federated, interoperability provides the flexibility to operate acr

Auto-fetched and translated by AI Atlas. All rights belong to the original author.