TensorWave, founded in December 2023 by Darrick Horton (CEO), Jeff Tatarchuk, and Piotr Tomasik, sells training and inference compute built on AMD Instinct accelerators rather than NVIDIA, positioning itself as an AMD-first neocloud. It is based in Las Vegas and has raised real money for a company this young: a $43M SAFE round in October 2024, then a $100M Series A in May 2025 co-led by Magnetar and AMD Ventures, with participation from Nexus Venture Partners, Maverick, and Prosperity7. The company has said it is deploying more than 8,000 AMD Instinct MI325X GPUs for a dedicated training cluster and has described building one of the largest liquid-cooled AMD GPU deployments. For candidates, the takeaway is plain: this is a capital-intensive infrastructure company whose product is GPU compute, so most engineering roles touch hardware, networking, or the systems that keep large clusters running.
Process
TensorWave publishes little about its loop, and there is almost no Blind or Glassdoor detail on it yet, so read this as the likely shape for a company at this stage rather than a fixed script. Expect a recruiter screen first (background, why AMD infrastructure, whether a small fast team fits you), then a technical round. The coding at startups this size tends to be practical: write clean working code against a real systems problem, often in Rust, Go, or Python, and reason about complexity and failure out loud rather than grinding pure algorithm trivia. A system design or infrastructure round usually maps to the actual product, meaning GPU cluster design, scheduling, networking, and keeping a training job alive across thousands of accelerators. A hiring-manager or founder conversation closes it out, weighted toward ownership and behavioral signal about how you operate when nothing is documented and something is on fire. Plan for roughly two to three weeks end to end, faster when the team moves quickly on a candidate it likes.
What they actually ask
- GPU fleet and cluster design on AMD: how you would stand up, schedule, and operate thousands of Instinct GPUs on bare-metal Kubernetes, and what fails first at that scale.
- ROCm versus CUDA: where AMD’s software stack differs from NVIDIA’s, what porting or tuning a model for ROCm involves, and how you debug a kernel or driver issue.
- Networking for training: RDMA, RoCE or InfiniBand, collective operations, and why fabric design and topology decide whether a distributed run is fast or stalls.
- Inference serving: standing up low-latency model serving on Instinct hardware, batching, memory limits, and holding throughput steady under load.
- Linux and distributed systems: core troubleshooting, kernel and driver behavior, containers, and infrastructure-as-code with Terraform, Helm, or Ansible.
- Backend and platform, plus on-call: services that schedule jobs, meter usage, and expose an API over the fleet, and the reality of paging when physical hardware misbehaves.
Levels and comp (2026)
TensorWave is a well-funded Series A startup, so read comp as base plus equity, with a bonus less central than at big tech. Public comp data is thin: levels.fyi shows only a handful of TensorWave submissions, with a reported software engineer figure around $180K total and a company-wide median near $138K, both from small samples, so do not treat either as a band. For a funded AI-infrastructure startup, senior and staff systems or infrastructure hires generally land in a rough cash range of about $170K to $260K depending heavily on level and location, but the company does not publish bands, so verify rather than trust one number. Equity is the real variable: grants come as stock options in a private company on a standard four-year vest with a one-year cliff, and the paper value can be large but carries no liquidity until an exit or a secondary sale. What matters is the unit count, the current strike or 409A price, and the preferred-share price behind any headline figure. Ask for base, unit count, current 409A, and vesting terms, then check levels.fyi and recent Blind threads for current TensorWave and comparable Series A infrastructure bands so you can value the offer yourself.
Prep priorities
Put most of your time into systems and infrastructure, not algorithm puzzles. Be ready to talk through running GPUs at scale: bare-metal Kubernetes, scheduling, networking and RDMA, storage for datasets and checkpoints, and observability with tools like Prometheus, Grafana, and Loki. Know the AMD angle, since it is the whole thesis: ROCm, the Instinct line, and how training and inference behave on it against CUDA. Public postings list Go, Rust, C, Zig, and JavaScript alongside NixOS, Terraform, and Ansible, so pick the language closest to the role and be able to write and reason in it under a clock. Have two or three project stories where you owned a system through real trade-offs and on-call pain, because a small hardware-heavy team screens hard for people who stay calm when a cluster degrades. If your background is NVIDIA-only, read up on ROCm and be upfront about the gap while showing you can carry systems skills across.
Frequently Asked Questions
Is the TensorWave interview LeetCode-heavy? There is little public signal, but for a company whose product is GPU infrastructure, expect practical coding and systems or infrastructure design over obscure algorithm tricks. The strongest lever is reasoning about clusters, networking, and Linux at scale, not memorizing puzzles.
Do I need AMD or ROCm experience to get hired? It helps but is not always required. Strong Linux, Kubernetes, and networking skills carry most infrastructure roles, and many good candidates come from NVIDIA or general infra backgrounds. For GPU-specific and ML-systems roles, hands-on ROCm and Instinct experience counts for more, so show you can pick it up fast if you lack it.
How long does the process take? Plan for roughly two to three weeks from first call to decision, with the usual caveat that a small startup can move faster when it likes a candidate. Keep your scheduling open and follow up quickly.
Is the job in Las Vegas or remote? TensorWave is headquartered in Las Vegas, and some roles, especially hands-on data-center and hardware work, expect presence there, while other engineering roles are listed as remote. Confirm location, time-zone overlap, and any on-site expectation with the recruiter early.
Similar company guides
Prepping for TensorWave? Put it to work:
