LLamjinLab
AI NewsWords 1421Read time≈ 4 min

Groq Will Deploy NVIDIA Groq 3 LPX and Vera Rubin NVL72 With Dell

Groq plans to add NVIDIA Groq 3 LPX and Vera Rubin NVL72 capacity to GroqCloud through a deployment partnership with Dell.

AILLMAgentTool Use

1. Groq Commits to NVIDIA’s New Inference Platform

Groq plans to deploy NVIDIA Groq 3 LPX alongside NVIDIA Vera Rubin NVL72 in GroqCloud, giving the independent inference provider a path to offer NVIDIA’s newest low-latency architecture through its existing cloud platform. Dell Technologies will supply deployment infrastructure and support the systems’ rollout at scale.

The announcement came on August 24, 2026, the same day NVIDIA said Groq 3 LPX had entered full production. Groq described itself as “among the first adopters,” not the first: NVIDIA identified Nebius as the first AI cloud scheduled to bring the platform into production, with Groq following as another early provider.

That distinction matters because neither company said Groq’s LPX capacity is already available to customers. Groq did not disclose an activation date, deployment region, rack count, supported model list, API configuration, reservation process, or token pricing. Its announcement consistently describes customer access in the future tense.

Once the capacity is online, Groq intends to expose it through infrastructure already used for production inference. The company says GroqCloud serves more than six million developers and thousands of AI-native companies, processes trillions of tokens each week, and operates data centers across North America, Europe, the Middle East, and Asia-Pacific. Those operating figures are company disclosures rather than independently audited measurements.

Dell’s stated role is to help turn the NVIDIA rack architecture into deployable infrastructure using its compute, networking, integration, and supply-chain capabilities. The parties did not identify the specific Dell systems involved or publish the commercial terms of the arrangement.

2. What Groq 3 LPX Adds to Vera Rubin

NVIDIA Groq 3 LPX is a rack-scale inference accelerator built around 256 Groq 3 LPU chips. Unlike a standalone replacement for Vera Rubin NVL72, it is designed to work beside the GPU rack and specialize in latency-sensitive token generation.

A complete LPX rack contains 32 liquid-cooled 1U compute trays, each carrying eight LPUs. NVIDIA lists 315 FP8 petaflops of inference compute, 128 GB of aggregate on-chip SRAM, 40 petabytes per second of SRAM bandwidth, and 640 terabytes per second of scale-up bandwidth per rack. Each LPU provides 500 MB of SRAM, 150 TB/s of SRAM bandwidth, and 2.5 TB/s of scale-up bandwidth.

The use of SRAM is central to the design. SRAM offers much less capacity than the HBM attached to modern data-center GPUs, but it provides extremely high bandwidth and predictable access times. Groq’s compiler-scheduled architecture also plans computation, memory movement, and chip-to-chip communication in advance, aiming to reduce the latency variation that becomes visible during interactive generation.

Vera Rubin NVL72 supplies the more general and memory-intensive side of the system. NVIDIA describes three ways the racks can divide inference work.

In a conventional prefill-and-decode split, Vera Rubin processes the prompt and creates the key-value cache, then transfers that cache to LPX for token generation. In an attention-and-feed-forward split, Rubin GPUs retain the cache and run attention while LPX executes feed-forward or mixture-of-experts layers. LPX can also run a smaller draft model for speculative decoding while the larger model on Vera Rubin verifies proposed tokens.

These configurations are intended to assign large-context processing and flexible, high-throughput workloads to GPUs while directing predictable, latency-sensitive generation to LPUs. That division is especially relevant to coding agents and other systems that repeatedly inspect context, generate a response, call a tool, observe the result, and begin another inference turn.

3. The Published Performance Evidence

The most concrete performance result is a test conducted by Artificial Analysis on an NVIDIA-operated Groq 3 LPX system. Using the 31-billion-parameter Gemma 4 model with 100,000 input tokens, the system produced a median 3,431 output tokens per second.

NVIDIA rounded that result to 3,400 tokens per second in its announcement and described it as approximately four times the 870-token-per-second result of the fastest public endpoint in the comparison. At a shorter 10,000-token context, the LPX system recorded a median 3,382 output tokens per second, versus 1,402 for the fastest public endpoint included in that chart.

Artificial Analysis and NVIDIA also reported that the tested outputs did not lose model precision or quality under the benchmark configuration. However, the LPX measurements came from a system running in NVIDIA’s own data centers rather than a generally available GroqCloud endpoint. They therefore establish the potential speed of the hardware configuration, not the price, capacity, congestion behavior, or end-to-end latency customers will receive from Groq.

NVIDIA separately ran the open-source SPEED-Bench coding workload on the same model. It reported a median generation rate of 4,767 output tokens per second and a P80 result of 5,520 tokens per second. That coding result was vendor-run, unlike the Artificial Analysis measurement, and should be treated accordingly.

Output-token speed also measures only the generation phase after a system begins returning tokens. An application’s total response time still includes prompt processing, queueing, network delay, tool execution, model reasoning, and any transfers between the Rubin and LPX racks. Faster decode can shorten agent loops substantially when generation is the bottleneck, but it does not make every step of an agentic task proportionally faster.

4. Why the Deployment Changes GroqCloud

The planned rollout deepens a relationship that began with Groq and NVIDIA’s non-exclusive inference-technology licensing agreement in December 2025. Under that agreement, Groq founder Jonathan Ross, then-president Sunny Madra, and other team members joined NVIDIA to advance the licensed technology. Groq remained an independent company, and GroqCloud continued operating.

NVIDIA subsequently incorporated the technology into the NVIDIA-branded Groq 3 LPU and LPX platform. Groq joined the NVIDIA Cloud Partner program on August 12, 2026, certifying it to design and operate NVIDIA accelerated-computing infrastructure according to NVIDIA reference architectures and operational standards.

The new announcement moves that relationship from technology licensing and certification toward an identified production deployment. Groq, which built its cloud identity around its own LPU infrastructure, now plans to operate NVIDIA’s commercial implementation of licensed Groq technology together with Vera Rubin GPUs.

For developers, the practical change is the prospect of accessing this heterogeneous hardware through Groq’s existing cloud rather than buying and operating complete rack-scale systems. Groq previously said customers would receive future NVIDIA capacity without changing application code, but it has not yet documented whether LPX will use existing model identifiers and API endpoints or require a distinct service tier.

The decision also broadens GroqCloud beyond a single hardware lineage. It allows Groq to combine its experience running LPU inference with NVIDIA’s GPU, networking, orchestration, and rack-scale ecosystem, while Dell handles part of the physical deployment path.

What remains unconfirmed is commercially important. Groq has not announced when developers can send production requests to LPX, which models will be supported, whether access will be generally available or reserved, or how its price and service-level guarantees will compare with existing GroqCloud offerings. NVIDIA’s declaration that LPX is in full production means the platform has reached manufacturing scale; it does not mean Groq’s service is already live.

Frequently Asked Questions

Is Groq the first cloud provider to deploy NVIDIA Groq 3 LPX?

No. NVIDIA named Nebius as the first AI cloud bringing the platform to production. Groq says it will be among the earliest adopters.

When will Groq 3 LPX become available on GroqCloud?

Groq has not announced a launch date. It also has not disclosed initial regions, capacity, supported models, or access requirements.

What does Dell Technologies provide?

Dell is helping Groq deploy the LPX and Vera Rubin NVL72 infrastructure at scale, including integrated compute, networking, and supply-chain support. The exact system configuration was not published.

Is Groq 3 LPX a replacement for Vera Rubin NVL72?

No. LPX is designed to operate alongside Vera Rubin NVL72, accelerating latency-sensitive token generation while Rubin GPUs handle complementary inference work.

Does the 3,431-token-per-second benchmark represent GroqCloud performance?

No. Artificial Analysis measured an NVIDIA-operated LPX system. Groq has not published production endpoint performance or pricing for its planned deployment.

Sources

Share

Share this article