Ampere

And the reason is, at least as of right now, whenever you bring a data center on, it fills up immediately.

And by the way, the four generations ago NVIDIA chip, which is called the Ampere, it was the one before the Hopper.

The current one is Blackwell.

They're about to go with Vero Rubin.

While NVIDIA owned figures for eight-bit floating point on its Hopper architecture claim two to two times higher token generation than sixteen-bit precision on the same chip.

On the hardware generation side, the Ampere to Hopper transition improved peak sixteen-bit tensor throughput by roughly three point two times and memory bandwidth by about sixty-seven percent, and NVIDIA owned published Blackwell Ultra figures drawn from third-party benchmarking at sites claim up to fifty times higher throughput per megawatt and up to thirty-five times lower cost per token than Hopper at low latency, with more than ten times more tokens produced per watt.

These last figures are vendor published rather than independently reproduced at the time of this report and should be read as the upper bound of a real but as yet unverified improvement, not as a demonstrated audited number on the order of the OpenAI pricing history or the vLLM throughput benchmark.

Key advantages and limitations, the inference cost decline central advantage for a deep tech founder is that it decouples product ambition from balance sheet size.