Combining Groq with Vera Rubin under Dynamo scheduling improves overall performance by 35 times and reaches inference performance NVLink 72 could not previously reach.

Judged as of 2026-03-17. The claim has two parts: (a) Groq technology combined with Vera Rubin under Dynamo scheduling gives a 35x improvement in overall performance, on the metric NVIDIA states (e.g. throughput, tokens per second per user, or per-watt/per-MW output), against the baseline it states; and (b) this reaches an inference performance level an NVLink 72 system alone could not reach. NVIDIA technical materials published by that date (keynote slides, technical blog or whitepaper with stated methodology and baseline) count as evidence of what was measured. Independent benchmarks or analyses that contradict them take precedence. True: published results show about 35x (roughly 32x–38x) on the stated metric and baseline, plus an operating point beyond what NVL72 alone achieves. Mostly true: both parts hold but the multiple differs modestly (e.g. 30x or 40x), or it holds only for a narrower set of conditions than 'overall' implies, with that caveat stated in the source. Mixed: one part is supported and the other is not. Mostly false: the 35x figure comes from a single cherry-picked point (e.g. one latency regime) and is presented as overall, or the gain is real but much smaller than 35x. False: no such combined system or result existed as of the date, or evidence shows the gain is far below the claim (e.g. under 10x) or that NVL72 alone reaches the same performance.

Set before the research began.

1. E1 Inside NVIDIA Groq 3 LPX: The Low-Latency Inference Accelerator for the NVIDIA Vera Rubin Platform | NVIDIA Technical Blog NVIDIA (developer.nvidia.com technical blog) · 16 March 2026 Own record

   “the combination of Vera Rubin NVL72 and LPX delivers up to 35x higher TPS per megawatt at 400 TPS per user compared with the NVIDIA GB200 NVL72”

2. E2 Inside NVIDIA Groq 3 LPX: The Low-Latency Inference Accelerator for the NVIDIA Vera Rubin Platform | NVIDIA Technical Blog NVIDIA (developer.nvidia.com technical blog) · 16 March 2026 Own record

   “In Figure 10, that premium tier is represented by a 2-trillion-parameter MoE model with a 400K input context window operating at roughly 400 TPS per user and beyond.”

3. E3 NVIDIA Spent $20 Billion Because GPUs Alone Can't Win the Inference Era | Barrack AI Barrack AI · 17 March 2026

   “These figures are NVIDIA projections. No independent MLPerf results exist for Groq 3 LPU. They should be evaluated accordingly.”

4. E4 Nvidia's 35x inference number lost its denominator on the way to the headline Barista Labs · 16 March 2026

   “Not 35x faster. Not 35x more tokens per second per chip. 35x more inference throughput per megawatt.”

NVIDIA's own technical blog (E1) gives the 35x figure as 'up to 35x higher TPS per megawatt at 400 TPS per user' against a GB200 NVL72 baseline. That is a peak result at one operating point and per-megawatt metric, not an overall performance improvement. Independent analyses (E4, E3) say the figure is a per-megawatt metric, not a general speed-up, and that it is an NVIDIA projection with no independent MLPerf results. E2 describes a premium tier at roughly 400 TPS per user, but nothing quoted shows that an NVL72 system alone could not reach that level.