Single-tenant by allocation
Hosted workloads are assigned dedicated capacity rather than sharing a pool. Your throughput does not change because of what someone else is running.
Infrastructure
We own the metal.
That changes the economics.
Consultancies that rent compute pass the meter on to you. We built a platform capable of running the largest open-weight models available, which means experiments stay affordable and your workload never shares a machine.
Platform
The machine
A single-node system built specifically for large mixture-of-experts models, where system memory capacity and bandwidth matter as much as GPU memory.
Processor
AMD EPYC 9755 — 128 cores / 256 threads, Zen 5
System memory
768 GB DDR5-5600, 12-channel — measured 408 GB/s copy bandwidth
Accelerators
2 × NVIDIA RTX PRO 6000 Blackwell — 192 GB combined GPU memory
Serving storage
12.8 TB enterprise NVMe — 5.5 GB/s write, 4.4 GB/s read
Archive
37 TB redundant array for model and dataset storage
Connectivity
Dual-WAN — fibre primary with independent 5G failover on a separate carrier
Facility
Hosted with conditioned power, dedicated cooling, and redundant connectivity
Measured performance
Numbers we ran ourselves
Single-stream decode throughput, measured on this hardware. These are not vendor figures and not estimates — they are what the machine does with these models loaded.
Why the model sizes matter. A 554 GB model does not fit in any single machine's GPU memory — not ours, and not a rack of consumer cards. Running it at usable speed requires holding the model in system memory and moving the right parts to the GPUs at the right time. That is an engineering problem, and solving it is the reason these numbers exist.
Measured concurrency
What it does under load
Single-stream figures describe one user. This is the number that matters for a team or an API: total tokens per second with many people asking at once. Ramped in-house, identical prompts, zero errors at every level.
Latency holds while throughput scales. At 48 simultaneous requests, 95th-percentile time to first token was 0.03 seconds — indistinguishable from a single user. Aggregate throughput rises almost linearly to that point and then turns over, which is why capacity is sold below it rather than at it. Benchmarks use uniform prompts; real traffic is burstier, and quoting the peak would be quoting a number you would not reliably get.
Operating principles
How the platform is run
No prompt or output retention
Inference traffic is not logged to disk, retained, or used for any downstream purpose. Where you need an audit trail, it is built to your specification and stays under your control.
Commercially licensed weights only
We deploy only models whose licences permit commercial use, and we confirm the licence terms for your specific use case before anything goes into production.
Need a model benchmarked on your workload?
Send us the task. We will measure it on this hardware and show you the methodology alongside the result.