Product Manager · Infrastructure Platforms & Economics
I work on turning infrastructure hardware into products customers can actually buy: which compute, storage and networking choices we expose, how those configurations stay buildable as hardware changes, and what they cost to run.
I am now applying the same thinking to AI infrastructure, especially the economics of training, inference and retrieval.
Turning hardware choices into a smaller set of configurations that customers can buy and teams can actually support.
Understanding what training, inference and retrieval need from compute, storage, networking, power and cost.
Looking beyond purchase price to what the full system costs to run over time.
Getting product, engineering, supply, factory, configuration and finance aligned so a change actually ships.
A few of the decisions I spend most of my time on.
The catalog started as hundreds of fixed server designs, each one effectively its own predefined build.
I work on the product model behind a configurable private cloud platform: customers choose across processor, memory, storage, accelerators and networking, and those choices still resolve into governed, buildable designs. My part covers the catalog model, the configuration rules and the work across teams that makes a design orderable.
Read the caseProcessors were being chosen on the quoted price difference, months before anyone worked out what the core count would do to software licensing.
Compared 27 processor designs across 15 licensing scenarios and found that a $369K hardware difference could drive $1.2M in software cost over five years. The lesson was simple: the cheaper server is not always the cheaper system.
Read the caseEvery launch was being negotiated from scratch across product, engineering, supply chain, the factory, quoting and finance.
I led new product introduction across nine ProLiant server platforms against one shared commercial structure, so product, engineering, supply chain, the factory, quoting and finance were ready for the same server at the same time.
Read the caseHardware choices rarely stay isolated. A processor can change software cost, a drive can change the chassis, and a GPU can change power and networking.
Explore the StackCPU, memory, drives and accelerators, and how one choice moves the others.
DomainStart with how the application accesses its data, then choose the platform.
QuestionWhat it costs to own, rather than what it costs to buy.
QuestionTraining, inference and retrieval land on the same four domains.
Two things I am working on in the open, outside the day job.
AI infrastructure economics
An infrastructure cost model for AI inference. I am testing how accelerator choice, utilization, memory and serving assumptions change the real cost of running a workload.
In developmentInfrastructure configuration agent
An experiment in making configuration easier to explain and check. Ask a configuration question in normal language, and have the answer checked against the same hard rules that decide whether a build is possible.
In developmentI started in electrical engineering, building hardware with Pravega Racing at VIT, then spent three years at ZS working on commercial analytics. At Texas A&M I added an MBA and MS in Analytics before moving into product management at HPE. Outside work, I am a trustee at Adarsh Women’s Polytechnic in India.
“… becoming a trusted leader on critical projects and earning respect from clients and team members.”
“Where Kushagra shines isn’t just his strategic mindset or technical skills; it’s his heart for the team.”
“I’ve taught a lot of great students, but not many students who were as engaged as he was in the life of our program.”
