About The Serving Desk

The Serving Desk covers the layer nobody demos: the models, the serving stack and the GPU economics underneath every AI product. Cost per token is treated here as a first-class metric, because for anyone running this in production it is the one that decides whether the product survives.

Our readers are the engineers who have watched a latency graph at three in the morning. We assume you know what a KV cache is and we do not pad the piece explaining it.

What we cover

• Models — what actually changed in a release, read against the release notes
• Serving — throughput, latency and the configuration that decides both
• Costs — what a workload really costs per token, per request and per month
• Infrastructure Startups — the companies selling into this stack, and what they are really offering

How we work

We test claims where we can and read the release notes where we cannot, and we are explicit about which of those two a given number came from. Benchmarks carry the conditions they were measured under, because a throughput figure without its batch size is decoration. Estimates are labelled as estimates. Every piece passes human editorial review.