Chip giant Nvidia on Tuesday, 14 July 2026 claimed its GB300 NVL72 rack-scale system delivers up to 25 times higher performance per watt than its own previous-generation Hopper architecture when running the DeepSeek V4 Pro inference workload.

Context

The post, shared from Nvidia's official corporate account, states: 'On DeepSeek V4 Pro, NVIDIA GB300 NVL72 systems deliver up to 25× higher performance per watt than NVIDIA Hopper.' The claim positions the GB300 NVL72 — a liquid-cooled, rack-scale platform built around the Blackwell Ultra generation — as a decisive leap over the H100-based Hopper systems that currently power the majority of the world's large AI deployments.

The Hopper architecture was unveiled in 2022 and introduced the H100 GPU alongside NVLink-based rack-scale configurations. It became the de-facto standard for AI training and inference at hyperscale data centres globally, including those operated by major cloud providers serving Indian enterprises.

Policy Backdrop

Nvidia has followed a cadence of releasing new GPU architectures roughly every two to three years — from Ampere (2020) to Hopper (2022) — with each generation targeting large uplifts in performance per watt to keep pace with the exponential growth in AI model sizes. Data-centre power consumption has become a binding constraint for governments and operators worldwide, making efficiency gains as commercially significant as raw throughput.

In India, the government's push to build sovereign AI compute capacity under initiatives such as the IndiaAI Mission has placed GPU procurement and power efficiency at the centre of policy discussions. A 25-fold improvement in performance per watt, if independently validated, would materially change the economics of building and operating AI infrastructure in power-constrained markets.

Stakeholders and Impact

AI hyperscalers and data-centre operators are the primary audience for this benchmark claim. For them, performance per watt directly translates into total cost of ownership: fewer racks, lower cooling bills, and reduced power-purchase agreements for an equivalent amount of AI compute.

Indian cloud providers and public-sector AI projects evaluating next-generation hardware will likely scrutinise this figure closely. The DeepSeek V4 Pro model — a large language model that has attracted significant attention for its efficiency profile — serves as the benchmark workload here, lending the comparison particular relevance to inference-heavy deployments that dominate commercial AI usage today.

What's Next

Independent benchmark results from third-party research groups and hyperscaler customers will be critical to validating Nvidia's claim. The company's own successive architecture comparisons have historically been borne out by external testing, though the precise multiplier can vary by workload and configuration.

Analysts and data-centre planners will watch for detailed technical disclosures on GB300 NVL72 power envelopes, rack density, and availability timelines. For the broader AI industry, a confirmed 25× efficiency gain over Hopper would accelerate the retirement of existing H100 clusters and reshape capital expenditure plans for 2026 and beyond.