Early buy — delivery March 2027
VR200 NVL72
The VR200 NVL72 is Rubin-generation capacity on early buy: 288 GB of HBM4 per GPU at 22 TB/s in a 72-GPU liquid-cooled rack, with delivery in March 2027.
Specification
- GPU memory
- 288 GB HBM4 · 22 TB/s
- Intra-node
- NVLink 6 · 3.6 TB/s per GPU
- Inter-node
- ConnectX-9 · 1.6T XDR
- Node config
- NVL72 rack · liquid-cooled
- Best for
- Rubin-generation reserve capacity
NVL72 rack, liquid-cooled, on the same dedicated per-tenant InfiniBand between racks as the rest of the fleet — 1:1, no shared spine.
Why this machine
Early buy is a supply instrument, not a pre-order button. Rubin-generation racks are allocated before they are built, so reserving now is how you hold a place in a commissioning schedule rather than joining a queue in 2027. HBM4 at 22 TB/s per GPU is the headline; the allocation is the actual product.
The terms are the same ones that apply to a cluster we hand over next month. You get a delivery date in writing, the exact burn-in report format your rack must pass, and hot spares on the same fabric — because a reservation that does not say what "working" means is not a reservation.
What it is for
Rubin-generation reserve capacity
Holding an allocation in a commissioning schedule rather than bidding for capacity after it exists.
Multi-year training roadmaps
Programs whose 2027 compute budget has to be committed now, with a delivery date and acceptance criteria attached.
Next-generation frontier work
288 GB of HBM4 per GPU at 22 TB/s in a 72-GPU liquid-cooled rack.
VR200 NVL72 questions
When does VR200 deliver?
March 2027. Early buy is open now, and the reservation carries the delivery date and the acceptance-test format your rack must pass.
What does VR200 early buy cost?
Quoted per rack. Send GPU count, interconnect and timeline and a firm quote comes back within 48 hours.
What are the VR200 specs?
288 GB of HBM4 per GPU at 22 TB/s, in a 72-GPU NVL72 rack with direct-to-chip liquid cooling.
What if the schedule slips?
The delivery date is in the contract, alongside the goodput and node-replacement SLAs. We publish yield live at status.tensorbay.com for the same reason — a factory that will not show its numbers is asking you to guess.