Nvidia Says Its Groq 3 LPX Rack Is in Full Production as Vera Rubin Ramps; What the Hardware Is and Which Figures Are Nvidia’s Own

Nvidia said on 24 August 2026 that its Groq 3 LPX inference rack is “now in full production”. In prepared remarks for its third-quarter fiscal 2027 results, released on 17 November, chief financial officer Colette Kress said the Vera Rubin platform generated “nearly $20 billion in revenue” in its first quarter of production shipments. This post sets out what the hardware is according to Nvidia’s own technical pages, which figures are Nvidia’s claims rather than independent measurements, and where the pages differ.
What the Groq 3 LPX rack is, according to Nvidia
Nvidia’s developer blog post “Inside NVIDIA Groq 3 LPX” (16 March 2026, by Kyle Aubrey and Farshad Ghodsian) describes a rack of 256 interconnected Groq 3 LPU accelerators, labelled LP30 chips in its table, arranged in 16 liquid-cooled 2U trays of 16 chips each. It lists 500 MB of on-chip SRAM per chip (SRAM is fast memory built into the processor itself), 128 GB of SRAM per rack, 150 TB/s of on-chip memory bandwidth per chip and 40 PB/s per rack, 640 TB/s of scale-up bandwidth per rack, and 96 chip-to-chip links of 112 Gbps per chip. Its table gives 315 PFLOPS of FP8 compute per rack.
On the division of labour, the same post says Vera Rubin GPUs handle prefill and the attention step over the accumulated context (the KV cache), while the LPX rack runs the feed-forward layers of sparse mixture-of-experts models and other pointwise operations. Nvidia’s Dynamo software orchestrates the two, and intermediate activations pass between them for each token. A second developer blog post dated 24 August (by Seth Weidman, Kirthi Devleker and Andrew Ling) lists three configurations, including one in which LPX performs “the entire decode step” (decode is the stage that produces output tokens one after another) and one in which a small draft model runs on LPX for speculative decoding.
Neither post states a process node or foundry. Tom’s Hardware (19 March 2026) reported 512 MB per chip and a Samsung 4nm process; Nvidia’s pages say about 500 MB.
The performance claims and what each is measured against
- 35x and 10x (Nvidia’s claims). The developer blog says LPX paired with Vera Rubin delivers “up to 35x higher inference throughput per megawatt” and “up to 10x more revenue opportunity for trillion-parameter models”, compared with the GB200 NVL72 at 400 tokens per second per user. The workload in its chart is described as a 2-trillion-parameter mixture-of-experts model with a 400K-token input context; the model is not named. Nvidia’s 16 March press release gives the same two figures without a baseline or the 400 tokens-per-second condition. Neither page labels the figures as measurements, and the 24 August developer blog calls its own scale-up chart “a projection” for a GPT-OSS model scaled to 2 trillion parameters. Chart text on the March post gives estimated annual revenue per gigawatt of $30 billion for Blackwell, $150 billion for Rubin and $300 billion for Rubin with LPX.
- 3,400 tokens per second (measured by Artificial Analysis). Nvidia’s 24 August press release says LPX delivered “a record 3,400 output tokens per second” in Artificial Analysis benchmarking of Gemma 4 31B at a 100,000-token context, and “4x faster responsiveness” than “the nearest alternative platform”. The developer blog gives the median as 3,431 tokens per second at 100K context and 3,382 at 10K context, and says the run used an LPX system Nvidia had set up in its own data centers. Chart text on that page lists 870 tokens per second as the fastest public endpoint at 100K context (Artificial Analysis, 20 August 2026) and 1,402 at 10K context; 3,431 divided by 870 is about 3.9 (Lens arithmetic). Nvidia’s own run of the SPEED-Bench coding benchmark gave a median of 4,767 output tokens per second.
Availability and the Groq transaction
Nvidia’s 24 August release, datelined Hot Chips, says Nebius is “the first AI cloud to adopt NVIDIA Groq 3 LPX” and also that Nebius “plans to bring” it to its Token Factory service. It adds: “Following Nebius, purpose-built AI inference cloud Groq plans to be among the platform’s earliest adopters.” In March Nvidia had said LPX would be available “in the second half of this year”. The release footer says “Groq and LPU are used under license from Groq, Inc.” and does not describe the transaction. Tom’s Hardware calls the 24 December 2025 arrangement a “$20 billion licensing and talent deal”; Jon Peddie Research calls it an “asset acquisition”. Neither page says what Groq the company now consists of.
The Vera Rubin ramp, as Nvidia’s CFO described it
Nvidia’s results release calls Vera Rubin “the fastest product ramp in the company’s history” but does not give a revenue figure for it. The figures below come from Kress’s prepared remarks as published in the Motley Fool’s transcript of the call.
- The Rubin platform generated “nearly $20 billion in revenue” in its first quarter of production shipments, which began in August.
- Hyperscalers reached production in about a month, half the time Grace Blackwell needed. Racks are running at CoreWeave, Google Cloud, Azure, OCI and Nebius.
- Kress cited 10 times the tokens per watt of GB300 for Vera Rubin, from MLPerf Endpoints.
- B300 rental rates are up more than 60% since launch, and Hopper contracts were renewed at up to twice the original price.
- Networking revenue was above $21 billion, and co-packaged optics (optical links built into the switch chip) are in volume production.
- Rubin Ultra is due next year and Feynman in 2028. Purchase commitments stand at $390 billion.
Groq and LPX are not mentioned in those remarks. Lens covered the headline results in a separate post. Related coverage: the SpaceX Vera Rubin financing, the Nvidia and Oracle superclusters and a Cerebras CS-4 explainer, since Cerebras also builds around on-chip memory.
Where reports differ
- SRAM per chip: about 500 MB (Nvidia developer blog) against 512 MB (Tom’s Hardware).
- Rack SRAM bandwidth: 40 PB/s in Nvidia’s rack table, against 38.4 PB/s from 256 chips at 150 TB/s each (Lens arithmetic). The blog does not reconcile the two.
- Rack compute: 315 PFLOPS FP8 in the table, against 307.2 from 16 trays at 19.2 PFLOPS each (Lens arithmetic).
- Baseline for 35x: GB200 NVL72 in the blog text, “Grace Blackwell NVL72” in chart text, and none in the March release. The 35x and 10x are “up to”; the 4x is unqualified.
- Speed figure: 3,400 (press release) against 3,431 (developer blog).
- Rubin CPX: Tom’s Hardware said it “appears to exit the roadmap”. Nvidia has not said so on the pages Lens read, and the HTML title of Nvidia’s 24 August blog post mentions “LPX and CPX” while its headline and body do not.
- The Groq deal: “licensing and talent deal” (Tom’s Hardware) against “asset acquisition” (Jon Peddie Research); Nvidia’s release says only “under license”.
Sources and notes
Lens read Nvidia’s two developer blog posts, its 24 August blog post and press release, its 16 March press release “NVIDIA Vera Rubin Opens Agentic AI Frontier”, its third-quarter fiscal 2027 results release, the Motley Fool transcript of the call (prepared remarks only), Tom’s Hardware (19 March) and Jon Peddie Research (17 March). Chart values on Nvidia’s pages were readable only as image text. Lens did not open Nvidia’s LPX product page, two PDF decks linked from the 24 August blog post, or Artificial Analysis’s methodology page, and has not independently verified any speed, efficiency or revenue multiplier in this post; those are Nvidia’s figures or Nvidia’s account of a third-party benchmark.
Image: File photo of an Nvidia HGX B200 NVL8 baseboard, a Blackwell-generation data-center GPU board, by Pokiiri (2025), CC BY-SA 4.0, via Wikimedia Commons. It is a stand-in: it does not show Vera Rubin or Groq 3 LPX hardware, and the file page does not say where it was taken.

