Vera Rubin reshapes inference economics as the processor race turns to token output per megawatt

The compass for evaluating advanced computing infrastructure has changed fundamentally, as raw device power is no longer measured by peak theoretical compute performance but by the rate of tokens completed per megawatt of power consumed. During the AI Infrastructure Summit in Santa Clara, whose attendance doubled to exceed eight thousand participants compared with about 3,500 last year, Ian Buck, vice president of high-performance computing at Nvidia, unveiled the outlines of the next generation of data centers aimed at running autonomous agents on the Vera Rubin and DSX platforms.

The workloads of intelligent agents impose an operational pattern that differs from traditional conversational queries, as a single programming session consumes hundreds of thousands of input tokens due to the sequence of reasoning steps, tool calls, and sub-agent generation, which is equivalent to fifteen times the size of a simple query.Results from the AgentX board of Semi Analysis showed that the Vera Rubin NVL72 system achieved a jump of up to thirty times the token-per-megawatt productivity compared with the GB300 NVL72 system on the DeepSec V4 Pro model, while reducing the cost per million tokens by roughly forty-five times.The system is completed by integrating Grok 3 LPX processors to provide low-latency deterministic inference, recording a speed of 2,529 tokens per second per user within a context extending to one hundred thousand tokens on the Quine 3.8 model, which has twenty-seven billion parameters.

On the level of industry-standard benchmarks, the Vera Rubin NVL72 system recorded, in its first appearance in the MLPerf Inference v6.1 tests, a productivity advantage of 3.7 times compared with the GB300 system, while four combined racks containing 288 GPUs achieved a scaling efficiency of 99 percent. This hardware advancement is accompanied by an accelerated software stack, as optimizations through the Dynamo software, Nemo libraries, and TensorRT engines raised performance by a factor of 1.6 relative to the previous test version without hardware changes.

Performance tests extended to the Vera central processor, where startups documented advanced operational results in agent workloads and databases. Perplexity recorded a 1.9-fold acceleration in launching secure isolation environments for its platform, while the processor achieved the highest recorded speed at ClickHouse in analytical data tests, Deep Infra logged a reduction of roughly 2.2 times in the time to process coordination steps, and Red Panda measurements showed a 5.5-fold cut in response latency with a 73 percent productivity increase compared with conventional processors. To ensure continuous operation when linking hundreds of thousands of chips, the system supported the sixth generation of NVLink with a multi-layer protection architecture that detects and corrects physical faults without disrupting applications.

On the power-management side, the DSX platform, with Max LBS software, introduced mechanisms to dynamically redistribute power among racks according to the differing needs of training and inference workloads. Lambda demonstrated the operation of 19 compute nodes within the electrical budget allocated for only sixteen nodes, raising token productivity by 24 percent and improving performance per watt by 23 percent. In parallel, Emerald AI, in collaboration with the Silicon Valley Power facility, showcased a real-time response system for flexible loads via DSX Flex software, where data centers automatically reduce consumption based on grid signals without harming critical computing tasks.

This shift sets a mandatory benchmark for infrastructure developers in the Gulf, Egypt and the Arab Levant, where large-scale expansion plans clash with electrical feed limits and station efficiency requirements.When investment evaluation shifts from chip ownership to token-per-megawatt efficiency, the ability to run agent workloads at a cost forty-five times lower becomes the decisive factor in the economic viability of regional data centers. This trajectory also requires local engineers and technology officers to recalibrate their inference architectures to exploit flexible load management instead of reserving surplus, unused electrical capacity.

The computing race is entering an industrial-maturity phase governed by precise energy equations and the costs of producing complex tokens. Building AI factories is no longer merely stacking processors, but an integrated system that starts with electrical grid stability and extends to the most refined inference algorithms in autonomous-agent environments.