What does TensorWave actually sell?

TensorWave sells AMD Instinct GPU-hours and reserved clusters — MI300X, MI325X and MI355X — positioned around memory capacity per dollar for large-model training and high-throughput inference.

Who should choose TensorWave?

TensorWave fits inference platforms and memory-bound workloads whose teams can validate on ROCm. Production customers Fireworks AI and Luma AI demonstrate the AMD software stack working at scale. Teams with hard CUDA dependencies should look elsewhere.

How big is TensorWave in 2026?

As of mid-2026 TensorWave has raised about $493 million including a $350 million Series B co-led by AMD Ventures at a $1.55 billion valuation, operates 8,192 MI325X GPUs, and has secured more than 2 gigawatts of long-term capacity.

TensorWave: Sells GPU-hours on the contrarian silicon — The Compute Ladder

The two-sided view

What TensorWave sells — and what it buys

Atomic unit sold

AMD Instinct GPU-hours and reserved clusters (MI300X / MI325X / MI355X). The real product is memory-per-dollar.

What they buy

AMD accelerators at volume ($350M earmarked largely for MI355X), 2GW+ of long-term data-center capacity (incl. 20MW with TecFusions in PA/AZ), and networking via a Credo partnership.

Verified 4 July 2026

Scale proof

$350M

Series B at a $1.55B valuation (June 2026), co-led by Magnetar and AMD Ventures — roughly $493M raised since 2023

8,192

MI325X GPUs online — one of the largest all-AMD training clusters in North America, with MI355X deployments underway

2GW+

of long-term data-center capacity secured

named inference customers — Fireworks AI and Luma AI — production evidence that answers the ROCm software-lift objection

The strategy

The bet

That memory capacity (192GB+ HBM per GPU) plus an open ecosystem wins the inference era — and that being AMD’s flagship cloud is worth more than being NVIDIA’s five-hundredth partner.

Deal-making

Natural counterparty

Inference platforms and memory-bound workloads — long-context LLMs, video models — that have or will build ROCm competence. A multi-tenant deployment layer is the de-risking piece most buyers still need.

Buyer fit

Choose TensorWave when

Your workload is memory-bound and cost-per-token sensitive, and you can invest in ROCm validation or partner with someone who has.

Look elsewhere when

Your stack has hard CUDA dependencies you can’t touch.

Private ledger

Five fields we track but don’t publish

Everything above is public and verifiable. The fields below change weekly and are verified directly with each provider — they live in a private supply-demand ledger, not on this page.

Available capacity, next 90 daysGPU type · quantity · location · online date

Effective price floorreserved · per-GPU-hour, at volume

Idle-inventory posturecurrent utilization pressure

Preferred deal structureattach terms · co-sell terms

Capacity-partnerships ownerthe person who signs

Ask about TensorWave’s current posture → Answered case-by-case, with the provider’s consent.