Senior Product Manager – AI Inference Performance

NVIDIA•Us, CA

55d

About The Position

NVIDIA is looking for a highly technical Product Manager to own the products that help customers extract the best possible performance from AI models and applications running on NVIDIA hardware. Every inference deployment — from a single-GPU workstation to a multi-thousand-GPU data center — lives or dies on latency, efficiency, and cost per token. Your job is to make NVIDIA the obvious place to run inference by turning deep optimization techniques into products that a broad range of customers can actually adopt. The work spans the entire inference stack — optimization techniques, the frameworks that deliver them; and the benchmarking and operational tooling customers rely on to trust the results. You will translate what our best-performing internal deployments do into capabilities that ship, are detailed, and work for everyone else. The Product Management organization at NVIDIA is a small, high-leverage team driving the company’s Deep Learning and Generative AI strategy. We need a self-starter who can operate with minimal direction, form a point of view from data and customer conversations, and drive it to a shipped result.

Requirements

12+ years in product management at a technology company, or comparable time as a founder, engineering lead, or technical product owner.
Depth in AI inference optimization: KV caching and reuse, quantization, speculative decoding, disaggregated serving. Know how each one moves accuracy, latency, and cost.
Familiarity with the inference and orchestration frameworks customers use: TensorRT-LLM, vLLM, SGLang, NVIDIA Dynamo, and the surrounding serving and scheduling ecosystem.
Proven track record of working independently — you can take an ambiguous problem space, define the strategy, and drive it to a shipped outcome without waiting to be told what to do next.
Operational experience running a live product: release management, quality and regression rigor, customer issues, and support processes.
Skill at translating low-level capability into business value — lower TCO, faster response, better GPU utilization — for engineers and executives alike.
BS, MS, or PhD in Computer Science, Computer Engineering, or another relevant area of study (or equivalent experience).

Nice To Haves

Engineering experience with LLM inference performance: profiling, kernel-level or serving-level optimization, or building a serving stack!
Open-source contributions or product leadership in vLLM, SGLang, TensorRT-LLM, Triton Inference Server, or Dynamo.
Production experience at scale counts too: capacity planning, autoscaling, SLA management, or stateful multi-turn applications.
A habit of reading the relevant research and translating it into roadmap decisions — you have intuition for where model architectures and serving techniques are heading next!

Responsibilities

Own the inference performance roadmap. Set direction across the stack: how models are represented, how memory and state are managed, how requests are scheduled and served, and how tokens get generated. The techniques change fast. Judge which ones matter, then decide what we build, what we adopt, and what we retire.
Build platforms, not one-offs. Deliver capabilities that generalize across model families, deployment topologies, and customer sizes. Build for easy adoption, sane defaults, and extensibility.
Define the performance strategy for agentic applications, where long-running sessions, tool-call stalls, and unpredictable output lengths break the assumptions built into single-turn serving. Drive capabilities around cross-turn cache reuse, request prioritization, and efficient handling of idle time in agent loops.
Define how our optimizations land across TensorRT-LLM, vLLM, SGLang, and NVIDIA Dynamo. Partner with open-source communities and internal engineering teams so customers get great performance on NVIDIA hardware.
Own how performance is measured, published, and reproduced. Define the benchmark methodology, the metrics that matter (TTFT, ITL, throughput per GPU, cost per million tokens), and the guardrails that keep our numbers credible.
Run the product day to day. Own release readiness, quality bars, regression tracking, customer blocking issues, and the feedback loop from production deployments back into the roadmap.