NVIDIA ships Groq 3 LPX and starts quoting inference in tokens per user, not per rack

Take the one number NVIDIA published and convert it into a unit of work. At 3,400 output tokens per second ^[3], a 10,000-token agent step lands in about 2.9 seconds ^[13]. The 4x framing ^[17] implies the unnamed nearest platform runs that step at roughly 850 tokens per second, or about 11.8 seconds ^[14]. Over a 200-step coding loop, that is under ten minutes against roughly 39 minutes ^[15].

Which is a real difference, and it is not the difference the same release advertises when it says coding tasks go from hours to minutes ^[18]. A 4x reduction applied to a two-hour task yields thirty minutes ^[16]. Tens of minutes is not minutes, so the two claims are measured against two different baselines, and neither baseline is named ^[5].

The specification change underneath is worth more than the benchmark. NVIDIA is defining interactivity as the token rate seen by an individual user, which sets how fast an agent closes each step of its loop ^[9]. Capacity has generally been bought as aggregate tokens per second per rack, or simply as accelerator count. A per-request floor at a stated context length is a different line item, and for a 100,000-token agent it is the line item that decides whether the thing is usable. The gap in the disclosure is concurrency: neither release says how many simultaneous requests were in flight when the 3,400 figure was recorded ^[5]. A box that holds that rate for one request and sags at fifty is a different machine from one that does not, and the published material cannot tell you which you are getting.

Then there is the awkward part for anyone hoping to specify this at all. Nebius CTO Danila Shtan says the acceleration arrives through the same API developers already use, with no migration to a new stack ^[7]. That is good for adoption and bad for procurement leverage: if LPX is invisible behind an existing endpoint, the buyer cannot order it, only ask for a latency number and let the provider choose the silicon. The two NVIDIA releases do not even agree on where Nebius is. One says Nebius is the first AI cloud to adopt the part ^[4]; the other says Nebius plans to bring it to its Token Factory inference platform ^[6]. Full production for the chip is not the same as a price and a date for the tokens.

The rest of the announcement reads as the same unbundling applied to the rack. NVIDIA says Vera Rubin now spans seven chips and five purpose-built racks ^[10]. CoreWeave is putting Spectrum-X Multiplane into production to link Vera Rubin racks across parallel switches ^[11], and SpaceXAI is taking Vera CPUs for the CPU-bound parts of agentic work, including orchestration, tool use and code execution ^[12]. The unit being sold is a rack matched to a phase of the agent loop. Capacity planning that still counts GPUs is measuring the wrong thing.