Vera Robin NVL72 doubles token productivity in inference benchmarks, Nvidia’s new architecture reshapes the economics of running agents

The results of the MLPerf inference v6.1 test cycle revealed the first standardized appearance of Nvidia's Vera Robin NVL72 system, recording an explicit superiority in token productivity and processing cost compared with the GB300 NVL72 system. The tests announced by the independent MLCommons organization focused on two of the most compute-intensive models currently: the reasoning and inference model DeepSeek-R1 and the multimodal vision-language model Qwen-3-VL, which reshapes the economic metrics that assess the viability of running AI models in large data centers.

Official results showed that the Robin NVL72 system achieved a token-throughput jump to 3.7 times that of the GB300 NVL72 rack in the Qwen-3-VL model across batch processing, direct service and real-time interaction scenarios, using the vLLM framework with the open-source inference engine Dynamo. For the DeepSeek-R1 model, productivity rose by 2.5 times when using the TensorRT-LLM library.The shift does not stop at superficial generation speedup, but extends into the depth of the reasoning compute architecture that requires memory coherence and fast inter-server data exchange.

This performance rests on an engineering integration of hardware and software, as the new architecture leverages updated Tensor cores and the transformer engine to accelerate both input ingestion and output generation, while using NVFP4 precision to reduce the memory footprint of model weights, attention mechanisms and key-value caches. The results relied heavily on a service-decomposition approach, separating the initial read phase from the generation phase and applying expert parallelism at scale for a mixture-of-experts architecture. The sixth-generation NVLink internal fabric and its switches serve as the backbone of this integration, providing a packet-transfer rate ten times higher and a latency three times lower than typical commercial Ethernet networks.

In parallel tests designed to assess the efficiency of multi-step independent agents, the EagerX benchmark from Semi Analysis showed the Robin NVL72 system outperforming its predecessor by thirty times, reflecting a shift in performance metrics toward planning and iterative reasoning tasks. At the same time, the GB300 NVL72 system demonstrated a near-linear scaling efficiency of 99 percent when moving from a single rack with 72 GPUs to four racks with 288 GPUs in the DeepSeek-R1 test, while software-only improvements alone delivered a 1.6-fold jump compared with the previous version, confirming that network-architecture efficiency and continuous code optimisation yield productivity gains comparable to adding new hardware.

These figures impose a practical shift in the technical decision-making calculations of those who lead infrastructure in the Gulf, Egypt and the Arab Levant.Regional organizations that pour massive capital investments into cloud computing and data centers aimed at digital sovereignty, primarily banks, telecom operators and government agencies, can no longer assess viability by counting raw processor units, but by the token-throughput per watt and per operational dollar. If you manage an engineering team that serves language and reasoning models in a local work environment, adopting inference-decomposition techniques and lowering context cache precision gives you a tangible reduction in operating cost on the existing infrastructure, without waiting for new hardware deliveries.

This development also redraws the plans for building independent agents in emerging and corporate regional firms. Business models that rely on agents executing sequential logical steps consume massive amounts of inference tokens per operation, and reliance on traditional networks quickly leads to performance bottlenecks and cost inflation. The true competitive edge will not be measured by the size of the model you choose for your organization, but by how ready a fast-interconnect architecture is to keep response time and per-token cost within operational profitability limits.