Layer 2 · head to head
Tenstorrent Blackhole vs NVIDIA RTX PRO 6000 Blackwell
The only fully open stack against the one everything is written for.
| Tenstorrent Blackhole | NVIDIA RTX PRO 6000 Blackwell | |
|---|---|---|
| Can you get it | You can purchase this outright. | You can purchase this outright. |
| Memory | GDDR6 rather than HBM | 96 GB GDDR7 |
| Bandwidth | Below HBM parts by design | 1.8 TB/s |
| Compute | RISC-V based Tensix cores | Blackwell-generation FP4/FP8 |
| Power | Lower than HBM accelerators | 600 W |
| Interconnect | Ethernet-native | PCIe |
| Software | TT-Metalium, TT-Buda — fully open | CUDA |
| Workload | both | both |
Tenstorrent Blackhole
For: The only credible open-stack option. You can buy a card, and the entire software stack is open source, which no other vendor here offers.
The catch: GDDR6 instead of HBM caps bandwidth well below the frontier parts. This is a developer and sovereignty choice, not a frontier-training choice.
Economics: The sovereignty play at this layer. If open stack and purchasability matter more than peak throughput, nothing else on this list qualifies.
NVIDIA RTX PRO 6000 Blackwell
For: Local and on-premise AI. The honest way into this layer for anyone who is not a hyperscaler — you can actually buy one.
The catch: GDDR7, not HBM, so bandwidth is a fraction of a datacentre part. Fine for local inference and fine-tuning; not a training cluster.
Economics: 96 GB in a workstation runs models that used to need a rented multi-GPU node. For sustained local work it repays quickly against GPU-hour rental.
Before either — can you power it?
Lower than HBM accelerators against 600 W. In 2026 that comparison usually matters more than the FLOPS one: the US interconnection queue exceeds 2,600 GW with waits approaching five years, and roughly 80% of projects withdraw before energising.