WHAT SHIPPED
● NVIDIA Groq 3 LPX — an interactive inference accelerator, now in full production. Announced 24 August at Hot Chips.
● Up to 256 accelerators per rack, extending Vera Rubin NVL72.
● The claim: 4x faster responsiveness for agents than the nearest alternative platform.
● The timing: OpenAI published Jalapeno benchmarks against the GB300 the next day, and did not test against Vera Rubin.
What it is
Groq 3 LPX is a purpose-built inference accelerator, not a general-purpose GPU. It extends the Vera Rubin NVL72 platform rather than replacing anything, and NVIDIA has it in full production as of 24 August. Nebius is the first AI cloud to adopt it.
The architecture splits a job that used to run on one class of chip:
| Component |
Handles |
| Rubin GPUs |
Large-scale context ingestion and processing |
| Groq 3 LPX |
Latency-critical decode — the token-by-token generation |
In a demonstration, a Vera Rubin NVL72 with Groq 3 LPX produced 3,400 tokens per second on Artificial Analysis running Gemma 4 31B at a 100,000-token context — which NVIDIA says is the fastest recorded on that model.
Why split context from decode
AGENTS CREATE TWO DIFFERENT PROBLEMS
An agent reasoning through a task generates enormous volumes of tokens across hundreds or thousands of steps. Processing the context is one workload. Generating tokens fast enough that the agent can iterate without stalling is a completely different one.
Optimising a single chip for both means compromising on each. Splitting them is the design answer, and it is why a dedicated decode accelerator exists at all.
NVIDIA's framing is that faster generation gives an agent more time to inspect files, write and test code, call tools, verify results and iterate — while the person waiting still gets a responsive experience. It claims agentic tasks like coding move from hours to minutes.
The timing next to OpenAI
Both companies presented at Hot Chips 2026 within a day of each other.
| Date |
Announcement |
| 24 August |
NVIDIA: Groq 3 LPX in full production, extending Vera Rubin. Vera CPU also in mass production |
| 25 August |
OpenAI: first Jalapeno benchmarks, claiming 1.5 to 1.9x work per watt over GB200 and GB300 |
| 26 August |
NVIDIA quarterly earnings |
Jalapeno was benchmarked against GB200 and GB300 — the previous generation. It was not compared against Vera Rubin, and specifically not against a Vera Rubin configuration with LPX handling exactly the decode workload Jalapeno targets.
That is not a gotcha. Comparing against shipping hardware you can buy is reasonable, and LPX entered production the day before. But anyone reading "beats NVIDIA" should know which NVIDIA the comparison used.
The claim to be careful with
NVIDIA says LPX delivers 4x faster responsiveness than the nearest alternative platform. That phrasing does a lot of work. It does not name the alternative, does not specify the workload, and comes from NVIDIA's own testing.
It is the same category of claim as OpenAI's 1.5 to 1.9x, measured by the party that benefits. Both are worth knowing and neither is independent benchmarking.
What it means if you are buying inference
| If you are... |
The useful read |
| Renting inference from a cloud |
Nebius first, others likely to follow. Ask specifically whether LPX is in the configuration |
| Running latency-sensitive agents |
This is aimed squarely at you. Decode speed is what makes an agent feel responsive |
| Doing batch processing |
Less relevant. Throughput matters more than interactivity and you are already served |
| Just using an API |
Nothing changes today. Cheaper serving reaches rate cards eventually, not immediately |
FAQ
What is NVIDIA Groq 3 LPX?
A dedicated interactive inference accelerator extending the Vera Rubin platform, in full production as of 24 August 2026. Up to 256 accelerators per rack, with Rubin GPUs handling context processing and LPX handling latency-critical decode.
How fast is it?
NVIDIA claims 4x faster responsiveness than the nearest alternative platform, and demonstrated 3,400 tokens per second on Gemma 4 31B at 100,000-token context, which it describes as the fastest recorded on that model.
Who is using it?
Nebius is the first AI cloud to adopt it. CoreWeave has separately deployed Spectrum-X Multiplane connecting Vera Rubin racks.
Does this respond to OpenAI Jalapeno?
Not directly — LPX was announced a day before the Jalapeno benchmarks. But it targets the same decode workload, and Jalapeno was not compared against a Vera Rubin configuration including it.
Can I buy one?
Not individually. It ships inside Vera Rubin NVL72 rack configurations through cloud providers.
Should I trust the 4x figure?
Treat it as a vendor claim. It does not name the alternative platform or specify the workload, and it comes from NVIDIA's own testing — the same caveat that applies to OpenAI's Jalapeno numbers.