Blog

Notes from the machine room.

Guides, hardware deep dives and product news from the people who rack, run and bill Nexpod's 8-GPU servers.

6 posts5 topicslatest Oct 24, 2026

/

Choosing between B200 and B300 for 70B inference

Both fit a 70B model on one GPU. What differs is the KV cache left over, and that sets how many users each GPU can serve.

When the meter starts, when it stops, and what a 47-minute job on a B300 server actually costs.

Where the seconds go between a request and the first token, and four changes that remove most of them.

LatestOct 24, 2026Product

One binary for every Nexpod product, JSON output on every command, and a watch mode that follows a deploy until it is ready.

From one 8-GPU server to four: create the cluster, test the fabric, and launch torchrun across 32 GPUs.

How region pinning works, what stays in the region you choose, and one setting to check before you upload anything.

Monthly

One email a month: new guides, hardware notes and what changed in nexctl. No tracking pixels.

Browse

Try it

Pick a server, pick a region, and connect over SSH. Stop it whenever you are done; the meter stops with it.

On-demand ratesEUR · excl. VAT