Nvidia Corrects GB300 NVL72 DeepSeek V3 Benchmark Claim
Synopsis
Nvidia issued a public correction on 17 June 2026 stating its GB300 NVL72 system delivers 1.6x performance on DeepSeek V3 pretraining at 512 GPU scale, citing internal benchmark references 6.0-0022 and 6.0-0101 to substantiate the revised figure.
Key Takeaways
Nvidia corrected a prior claim, stating the GB300 NVL72 delivers 1.6x performance on DeepSeek V3 pretraining .
The benchmark was measured at 512 GPU scale , a rack-level configuration relevant to hyperscale AI training.
Nvidia cited two internal result references — 6.0-0022 and 6.0-0101 — to support the corrected figure.
DeepSeek V3 , a mixture-of-experts model from Chinese AI lab DeepSeek , has become a standard throughput benchmark for AI hardware vendors.
The correction reflects the high scrutiny applied to Nvidia performance claims by AI labs and cloud operators making large infrastructure investments.
Chip giant Nvidia issued a public correction on Wednesday, 17 June 2026, clarifying that its GB300 NVL72 system delivers 1.6x performance on DeepSeek V3 pretraining at 512 GPU scale, citing internal result references 6.0-0022 and 6.0-0101.
Context
The post, framed as a reply correction, states verbatim: 'Correction: GB300 NVL72 delivers 1.6x performance on DeepSeekV3 pretraining at 512 GPU scale.' The two result identifiers — 6.0-0022 and 6.0-0101 — appear to be internal benchmark reference codes Nvidia has cited for transparency, allowing technical audiences to trace the specific test configuration.
DeepSeek V3 is a mixture-of-experts large language model released by Chinese AI laboratory DeepSeek in late 2024. Its open-source availability has made it a widely used benchmark target for AI hardware vendors seeking to demonstrate real-world training throughput.
Policy Backdrop
Nvidia introduced the Blackwell GPU architecture and the GB200 NVL72 rack-scale system in 2024 as its flagship platform for large-scale AI pretraining. The GB300 NVL72 represents a successor iteration of the same 72-GPU liquid-cooled rack form factor, now based on updated Grace Blackwell superchips.
Nvidia routinely publishes comparative performance data for its rack-scale systems against frontier open-source models, a practice that serves both technical marketing and the broader geopolitical narrative around US-China AI infrastructure competition. Issuing a formal correction underscores the company's stated commitment to benchmark accuracy, particularly when figures are scrutinised by hyperscale operators and AI laboratories making multi-billion-dollar infrastructure decisions.
Stakeholders and Impact
AI laboratories and hyperscale cloud operators are the primary audiences for this correction. A 1.6x pretraining speedup at 512 GPU scale is a commercially significant claim: it directly affects total-cost-of-ownership calculations for organisations planning large DeepSeek V3 training runs or fine-tuning workloads.
For Indian cloud and AI infrastructure buyers — including government-backed initiatives under the IndiaAI Mission and private hyperscalers expanding GPU capacity — Nvidia's benchmark accuracy matters because procurement decisions for next-generation rack systems are often anchored to published performance-per-watt and throughput figures.
What's Next
The benchmark community will likely seek independent replication of the 1.6x figure using the cited result references. Attention will also turn to how competing platforms — including AMD Instinct accelerators and custom silicon from major cloud providers — perform on DeepSeek V3 pretraining at equivalent scale.
Follow-on disclosures comparing the GB300 NVL72 against rival systems on additional mixture-of-experts models are expected as the AI hardware race intensifies heading into the second half of 2026.
Point of View
Credibility is a competitive asset. Geopolitically, the choice of DeepSeek V3 as the benchmark model is not incidental; it places US chip infrastructure in direct, measurable contest with Chinese AI model efficiency. For Indian policymakers and cloud buyers navigating the IndiaAI Mission's GPU buildout, the accuracy of these figures carries real fiscal weight.
NationPress
1 Aug 2026
Frequently Asked Questions
What did Nvidia correct about the GB300 NVL72?
Nvidia corrected its performance claim, stating the GB300 NVL72 delivers 1.6x performance on DeepSeek V3 pretraining at 512 GPU scale, and cited internal result references 6.0-0022 and 6.0-0101 to back the revised figure.
What is the GB300 NVL72?
The GB300 NVL72 is Nvidia's rack-scale AI training system featuring 72 liquid-cooled Grace Blackwell superchip GPUs, designed for large-scale model pretraining workloads.
What is DeepSeek V3 and why is it used as a benchmark?
DeepSeek V3 is an open-source mixture-of-experts large language model released by Chinese AI laboratory DeepSeek in late 2024. Its open availability and large scale make it a common benchmark for AI hardware vendors measuring training throughput.
Why does Nvidia publish benchmarks on Chinese AI models?
Nvidia uses frontier open-source models like DeepSeek V3 as benchmarks because they reflect real-world training workloads. Publishing results on widely used models also positions Nvidia's hardware within the broader US-China AI infrastructure competition.
How does the GB300 NVL72 benchmark affect Indian AI buyers?
Indian cloud operators and government-backed AI initiatives use published GPU performance figures to make procurement decisions. A corrected 1.6x throughput claim on a standard model directly influences cost and capacity planning for large training runs.
