AI infrastructure is often discussed in terms of chips, models, and compute demand, but the race to scale AI depends just as much on what happens before a cluster ever becomes available to customers. Behind every production-ready GPU deployment is a complex operational process involving facility readiness, power and cooling design, hardware procurement, cabling, networking, validation, and coordination across dozens of stakeholders.
Miky Bayankin has built his career in that kind of operational complexity. He has launched and scaled new markets and products at Uber, led premium product expansion at Borzo, served as CEO of a trucking freight marketplace, co-founded a package-free grocery and delivery business, and mentored startups through TechStars. Today, as Hydra Host’s AI infrastructure leader, he oversees the supply side of AI, helping bring purpose-built data centers and large-scale NVIDIA GPU clusters online through the company’s deployment platform, Brokkr.
In this interview, Bayankin discusses why deploying AI infrastructure is not simply a hardware challenge, how his background in physical operations shapes his approach to GPU cluster launches, and what organizations often underestimate about site readiness, stakeholder coordination, testing, and long-term reliability. He also explains why the next phase of AI infrastructure will require more standardization, stronger operational discipline, and deployment models that can keep pace with increasingly powerful GPU architectures.
Your career has taken you from launching new markets at Uber to scaling logistics businesses and now leading AI infrastructure deployments at Hydra Host. How have those operational experiences shaped your approach to bringing large-scale GPU clusters online?
Improving the operations of the physical world through technology is something that I have always been passionate about. What attracted me to the infrastructure space and Hydra Host specifically is the opportunity to build a software product that helps deploy physical GPU clusters, which in turn power many of the AI tools we now use daily.
Deploying a GPU cluster is a very complex process with a lot of dependencies. But every hard problem can always be broken down into smaller pieces, solved, operationalized, and repeated. That’s how we approach cluster deployment.
The lessons that I bring from the physical operations can be summarized in a few core principles.
- Simplify first. Break down complex problems into incremental pieces. We always map out our processes in visual mediums first: network diagrams, floor layouts, deployment timeline charts. From there, we decompose the build to the individual component level to ensure compatibility and plan out the deployment sequencing.
- Process over tools. It’s very easy to get carried away by the shiny tools that are available in the age of LLM abundance. You can spend hours and days using these tools to automate a process, or even build new tools yourself to do it. But if the process is flawed, it’s just not going to work. A lot of times a simple, but well-thought out manual deployment checklist is more effective than a sophisticated llm-generated worklfow.
- Start small, increment where possible. Solve one problem at a time, eliminate as many dependencies as possible, and begin implementing as early as possible. Given the amount of dependencies in GPU deployemnt, including supply chain bottlenecks – sometimes we deploy cluster funcionality in phases and batches – all in order to shorten the time to first token.
- Get into the field. The best products are built by people who understand how the work gets done on the ground. Even though we’re mostly a software company, we always send our engineers on site to oversee the physical deployment process. This helps us find inefficiencies in our tools and make the deployment easier, faster and less error prone for the deployment crew.
These rules are mostly common sense, but consistenly applying them goes a long way. They help us build technology that doesn’t just look good on screen, but actually works in the physical world.
Many people focus on the GPUs themselves, but the reality of deploying a production-ready AI cluster involves far more than hardware procurement. What are some of the biggest operational challenges that organizations underestimate when planning large-scale deployments?
A modern GPU cluster is a very sophisticated piece of hardware. It’s not just a collection of individual servers, it is effectively an interconnected, living and breathing hardware organism. The entire cluster includes not just GPU servers but also control plane nodes, switching gear, and even hardware like PDUs (data-center grade power strip) that are connected and managed remotely. In order to bring up the cluster, you have to plan the network topology, procure all the right hardware that is all compatible (which is already hard). Then you have to methodically orchestrate racking and cabling of all the things in the right places. Each individual server has at least 8-10 fiber cables running up to it, and smaller clusters consisting of one sub-unit or 72 GPU servers (576 GPUs) will have close to 2,000 cable connections.
AI infrastructure projects often involve coordination across facility teams, hardware vendors, networking specialists, platform engineers, and operations leaders. How do you manage dependencies across so many stakeholders while keeping deployments on schedule?
Keeping the deployment on schedule is indeed one of the most important and difficult parts of the job. Even planning the deployment timeline is actually hard because many variables remain uncertain. Facility readiness dates shift, supply chains change, and it’s hard to predict exactly how long tasks such as racking, cabling, network validation, and system testing will take.
Often, we have several clusters that are in the deployment phase at the same time. With hardware delivery timelines constantly moving due to supply chain issues, we have to stay flexible with time and resource allocations between multiple deployment projects in parallel.
Keeping everyone working from the same, up-to-date information is often what keeps a deployment on track.
Our principles are
- Over-communicate
- Double and triple check
- Always have a plan B, and then a plan C
- No surprises – keep everyone updated in a transparent way, don’t hold back bad news.
Site readiness has become a major topic as organizations race to expand compute capacity. From power and cooling requirements to physical infrastructure constraints, what separates a successful deployment from one that experiences significant delays?
We work closely with data center operators and third-party vendors throughout the planning process. Our hardware engineers conduct on-site inspections, review facility designs, and collaborate directly with engineering and electrical teams to validate power, cooling, networking, and overall site readiness. Any delays or issues on the facility side are guaranteed to create blockers and delay cluster go-live.
As far as critical infrastructure, power and cooling are the areas where you can’t afford to cut corners. An inadequate power delivery system or insufficient cooling will lead to cluster instability, unplanned downtime, and even permanent damage to hardware worth hundreds of millions of dollars.
As organizations adopt increasingly powerful GPU architectures such as NVIDIA Blackwell, how have deployment requirements changed compared to previous generations of AI infrastructure?
With the recent GPU generations we’re seeing much higher power density that pushes for more sophisticated power distribution design. More and more deployments now depend on more advanced cooling solutions such as direct-to-chip liquid cooling, self-contained racks and other modern cooling systems. All these requirements are quite new for many data center operators who haven’t yet dealt with power hungry and spiky performance of GPU clusters.
You’ve overseen the process of turning newly installed hardware into production-ready compute resources. What are the key steps between receiving hardware and making it available for customers, and where do organizations most commonly encounter bottlenecks?
Our engineering team at Hydra is investing a lot of time into automating the tools that help us deploy clusters at speed and perform all the quality checks, such as burn-in tests and NCCL testing. We’re developing tools that help us visualize the data center virtually before anything has been delivered or installed. These products then feed a very detailed guide for the deployment team on what and how to deploy and connect. Once the cluster is powered on, our software automates provisioning and large-scale validation.
We cannot afford to turn over a faulty or low-performing cluster to the customer. Any issues the customer experiences after the deployment are going to result in downtime, which will be lost revenue for us and compute owners, which is why we spend a lot of focus on pre-deployment checks and QA.
With demand for AI compute continuing to accelerate, many companies are looking to build infrastructure as quickly as possible. How do you balance speed of deployment with reliability, testing, and long-term operational stability?
Slow is smooth. Smooth is fast. We try not to rush the deployments. When things are rushed and teams forced towards unreasonable deadlines, mistakes and delays will happen, guaranteed. Planning thoroughly and validating every stage of the deployment are the best investments of our time. Setting realistic expectations gives our teams the time they need to deploy, test, and validate the cluster properly, without unnecessary pressure to cut corners. In the end, the successful deployment is the one that doesn’t have to be re-done later.
Looking ahead, what do you think the industry still misunderstands about large-scale AI infrastructure, and what operational capabilities will become increasingly important as next-generation GPU clusters continue to grow in size and complexity?
I think where the industry still has a long way to go is the standardization of the data center space. We get to work with dozens of facility operators around the world, and the way they design, run and maintain their facilities varies dramatically.
And methods keep changing as new generations of the GPUs are released, and require new technologies such as direct liquid cooling support. More standardization should happen on the cooling and power distribution fronts. It is generally easier to standardize when building a new data center from scratch, but there is also a lot of already existing space that needs retrofitting.
However, retrofitting existing data centers or other industrial facilities is sometimes even more complex and time consuming, and the total budget for retrofitting a facility can sometimes approach the cost of new construction. That’s why the industry still needs more modular, standardized, and scalable approaches to power distribution and cooling. The easier it becomes to upgrade existing facilities, or deploy in simpler environments, the faster we can bring new AI capacity online around the world.
