What happened

Per Mike Wheatley reporting in SiliconANGLE and cited on Techmeme, Nvidia says its dedicated artificial intelligence inference accelerator Groq 3 LPX has entered full production, Nebius has signed on as the first customer, and SpaceX will deploy Vera CPUs.

CNBC coverage states Nvidia said Groq racks will be online this year following a reported twenty-billion-dollar purchase tied to Groq LPUs, and The Register reports Nvidia claimed Groq 3 LPX racks delivered 3,400 tokens per second in an Artificial Analysis benchmark running Gemma 4 31B with a 100,000-token input sequence.

Evidence

• Nvidia says Groq 3 LPX entered full production with Nebius as first customer

  Techmeme · attributed

  Nvidia says its inference accelerator Groq 3 LPX has entered full production and Nebius has signed on as the first customer, and SpaceX will deploy Vera CPUs

• Nvidia claimed 3,400 tokens per second on a 100,000-token Gemma 4 31B benchmark

  Techmeme · attributed

  Nvidia says its Groq 3 LPX racks delivered 3,400 tokens per second in an Artificial Analysis benchmark running Gemma 4 31B with a 100,000-token input sequence

• Nvidia said Groq racks will be online this year after a reported $20 billion purchase

  CNBC AI · attributed

  Nvidia says Groq racks will be online this year following $20 billion purchase

• Groq rebuilt as an inference cloud after licensing LPU technology to NVIDIA

  MarkTechPost · attributed

  Groq rebuilt itself as an inference cloud after licensing its LPU technology to NVIDIA

• Nvidia denied a China-specific LPU product roadmap

  Tom's Hardware AI · attributed

  Nvidia denies report it will ship Groq-based LPUs to China by year-end — says there is 'no China-specific LPU product in our roadmap'

• The Groq 3 LPX label may conflate Nvidia reporting with Groq Inc. inference chips

  Techmeme · attributed

  Nvidia's naming of its inference accelerator 'Groq 3 LPX' conflates it with the independent Groq startup, and the claim warrants scrutiny because Groq-branded inference chips already exist from a different company.

Why it matters

If Groq LPU racks reach commercial scale, operators gain a specialized inference option for long-context serving beyond GPU clusters, but product identity and workload-specific benchmarks must be validated before infrastructure commitments.

Limits and uncertainties

The Groq 3 LPX production label may be a reporting error because Groq Inc. already sells inference chips under the Groq brand.

SiliconANGLE and Techmeme excerpts are truncated and lack verified performance, capacity, or deployment specifics.

The 3,400 tokens per second figure reflects one model, one context length, and one benchmark profile rather than mixed production traffic.

CNBC and Tom's Hardware excerpts in the packet are brief and do not fully document deal terms or the scope of Nvidia's denial.

Practical implications

Treat Nebius as an early signal for LPU rack availability but confirm product naming and specs through Nvidia official channels before engineering plans.

Request benchmarks that match your actual request patterns, batch sizes, and context lengths rather than relying on a single Artificial Analysis datapoint.

Factor Groq's pivot to an inference cloud and Nebius SEC reporting into neocloud vendor comparisons alongside headline GPU pricing.

What to watch

Nvidia official statements clarifying Groq LPU rack naming, production status, and Vera CPU deployment timelines.

Nebius public disclosures or service announcements confirming Groq LPU rack capacity and customer availability.

Independent benchmark releases covering short-context and mixed-traffic inference beyond the Gemma 4 31B 100,000-token test.