Nebius $NBIS Achieves Top NVIDIA AI Speeds, Promising 30% Cost Savings

Nebius Group’s latest benchmark triumph underscores a pivotal edge in the race for efficient AI infrastructure, where raw speed on NVIDIA hardware can dictate market positioning for cloud providers vying to host demanding large language models. Independent tests from Artificial Analysis highlight Nebius delivering unmatched performance for Llama 70B at 245 tokens per second and Qwen3-32B at 212 tokens per second on public NVIDIA GPU setups, a feat that signals not just technical prowess but potential revenue acceleration in an industry hungry for optimised inference capabilities.

Benchmark Metrics and Competitive Edge

These speeds, clocked on standard NVIDIA configurations, outpace rivals in public leaderboards, offering a tangible metric for developers and enterprises selecting platforms for AI deployment. For models like Llama 70B, which demand high-throughput processing for real-time applications such as chatbots or content generation, Nebius’s results imply reduced latency that could shave operational costs by up to 30% compared to slower alternatives, based on comparative analyses from similar NVIDIA-based runs. The Qwen3-32B model, known for its efficiency in multilingual tasks, benefits similarly, with Nebius’s optimisations—likely involving custom speculators and fine-tuned kernel integrations—pushing boundaries that others struggle to match without proprietary hardware tweaks.

This leadership emerges amid a broader push for inference-as-a-service platforms, where providers like Nebius leverage NVIDIA’s Hopper and Ampere architectures to deliver what Artificial Analysis deems the “top Inference-as-a-Service” offering. Historical benchmarks from MLPerf, for instance, show NVIDIA GPUs scaling effectively for training workloads, but Nebius extends this to inference, achieving near-linear gains when clustering resources. Investors eyeing this should note how such metrics correlate with user adoption; platforms posting superior speeds often see 20-40% quarterly upticks in hosted workloads, as evidenced by trailing data from competing clouds in 2024 filings.

Implications for AI Model Hosting Economics

The economic ripple of these speeds is profound, particularly as AI inference costs dominate budgets for scaling models beyond training phases. Nebius’s performance allows for 50% lower pricing structures while rivalling outputs from premium models like those from OpenAI, according to Artificial Analysis rankings published in late 2024. This pricing power stems from efficient GPU utilisation—Nebius reports custom adaptations that train speculators tailored to specific fine-tunes, ensuring even bespoke models run at peak velocity without the lock-in risks of custom silicon.

Comparatively, earlier 2024 benchmarks from NVIDIA’s own tests on Llama variants showed H200 GPUs boosting throughput by factors of 2-3x with optimisations like XQA kernels, yet Nebius amplifies this on public boards, suggesting proprietary software layers that enhance NVIDIA’s baseline. For Qwen3-32B, which features a 262K context window ideal for complex logic and coding tasks, these speeds translate to faster iteration cycles, potentially accelerating enterprise AI integrations by weeks. Analyst sentiment from Seeking Alpha, as of mid-2025, labels Nebius a “rare early-stage shot at a GPU superpower,” with forward EPS projections improving to -1.39 for the current year, reflecting optimism around such technical wins driving revenue growth despite trailing losses of -1.65 per share.

Market Reaction and Valuation Context

Recent market activity has been volatile, reflecting broader tech sector pressures alongside company-specific movements. The benchmark news, however, could provide a firm catalyst for a rebound, especially when considering the stock’s significant climb over the past year. Historical patterns show that AI infrastructure firms often gain 10-15% following positive independent validations.

NVIDIA’s ecosystem benefits indirectly, with its market cap holding steady above $4.2 trillion despite the dip, underscoring the symbiotic relationship where optimised providers like Nebius amplify demand for Hopper GPUs. Sentiment from verified accounts on platforms like Wccftech praises NVIDIA’s open-source efforts, such as the Llama 3.1 Nemotron 70B, which outperforms peers in benchmarks, but Nebius’s hosting speeds ensure these models deploy efficiently in production environments. Model-based forecasts from firms like Morningstar suggest Nebius could achieve 25% year-over-year revenue growth if these speeds attract more fine-tune workloads, building on 2024’s expansion into global data centres.

Strategic Positioning in a Crowded Field

Nebius’s edge in these benchmarks positions it against heavyweights like AWS and Google Cloud, where NVIDIA GPU availability often bottlenecks performance. The Artificial Analysis data, current as of 2025, reveals Nebius not only leading in speed but also in cost-efficiency for tasks like molecular dynamics simulations, where H200 GPUs hit 555 ns/day at $15.26 per 100 ns—13% cheaper than AWS equivalents. For Llama 70B and Qwen3-32B, this translates to scalable inference that supports zero-retention hosting, appealing to privacy-focused enterprises.

Expanding on this, Nebius’s ability to customise speculators for user fine-tunes—without mandating CUDA overhauls—lowers barriers for developers migrating from slower platforms. This flexibility could erode market share from competitors, with analyst models from StockTitan projecting Nebius capturing an additional 5-7% of the inference market by 2026, driven by these public board dominances. In contrast, NVIDIA’s own benchmarks, such as those for Llama 3.1 405B trained in under 125 minutes on 1,024 Hopper GPUs, highlight the hardware’s potential, but Nebius operationalises it for inference, creating a moat in high-performance AI services.

Investor Considerations Amid Volatility

For investors, these speeds represent more than bragging rights; they signal Nebius’s maturation from a Yandex spin-off into a focused AI enabler, with a price-to-book ratio of 3.92 suggesting undervaluation relative to growth peers. The 52-week range from $14.09 to $58.16, with the current price 8.38% above the 50-day average of $47.98, indicates resilience despite broader market headwinds. Sentiment from Fierce Electronics notes NVIDIA’s dominance in MLPerf for generative AI, but Nebius’s application to public inference boards extends this narrative, potentially boosting partnerships and contract wins.

Looking ahead, with earnings slated for 7 August 2025, any guidance affirming these benchmark advantages could lift sentiment, countering the current year’s projected EPS of -1.39. Dark wit aside, in a sector where speed kills competition, Nebius’s results might just be the accelerator needed to outrun profitability doubts.

References

Artificial Analysis. (2025, August). AI Inference Benchmark Data. [Obtained via social media post].

Dai, X. (n.d.). GPU-Benchmarks-on-LLM-Inference. GitHub. Retrieved August 5, 2025, from https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

Fierce Electronics. (2023, September 12). Nvidia, Intel tout marks in MLPerf benchmarks running Llama 2 70B. Retrieved August 5, 2025, from https://www.fierceelectronics.com/ai/nvidia-intel-tout-marks-mlperf-benchmarks-running-llama-2-70b

Javilopen. (2024, July 16). [Post on Qwen2 72B performance]. X. https://x.com/javilopen/status/1846591717211795695

NVIDIA Developer Blog. (2024, July 23). Boost Llama 3.1 70B Inference Throughput 3x with NVIDIA TensorRT-LLM and Speculative Decoding. Retrieved August 5, 2025, from https://developer.nvidia.com/blog/boost-llama-3-3-70b-inference-throughput-3x-with-nvidia-tensorrt-llm-speculative-decoding/

NVIDIA NGC. (n.d.). Llama 3.1 70B on DGX H100 (1xH100). Retrieved August 5, 2025, from https://catalog.ngc.nvidia.com/orgs/nvidia/teams/dgxc-benchmarking/resources/llama31-70b-dgxc-benchmarking-a

Reddit. (2024, April 27). Quick speed benchmark for Llama 3 70b on 1x 3090. r/LocalLLaMA. Retrieved August 5, 2025, from https://www.reddit.com/r/LocalLLaMA/comments/1caofxm/quick_speed_benchmark_for_llama_3_70b_on_1x_3090/

Reach_vb. (2024, July 16). [Post on Qwen 2 72B performance]. X. https://x.com/reach_vb/status/1846484958342168953

Schmid, P. (2024, January 31). [Post on Llama 70B inference speeds]. X. https://x.com/_philschmid/status/1752707016681042239

Seeking Alpha. (2025, July 30). Nebius Group: A Rare Early-Stage Shot At A GPU Superpower In The Making. Retrieved August 5, 2025, from https://seekingalpha.com/article/4803419-nebius-group-rare-early-stage-shot-gpu-superpower-making

StockSavvyShay. (2025, August 4). [Post on NBIS stock performance]. X. https://x.com/StockSavvyShay/status/1930577580408934793

StockTitan. (2024, July 10). Nebius AI Studio, a High-Performing Inference-as-a-Service Platform. Retrieved August 5, 2025, from https://www.stocktitan.net/news/NBIS/nebius-ai-studio-a-high-performing-inference-as-a-service-platform-cbkaqs9c93f4.html

Subhrajit. (2024, July 29). NVIDIA’s Llama 3.1 Nemotron-70B: A New Benchmark in AI Performance. Medium. Retrieved August 5, 2025, from https://medium.com/@subhraj07/nvidias-llama-3-1-nemotron-70b-a-new-benchmark-in-ai-performance-11fd10c60d80

Wall St. Engine. (2025, August 4). [Post on NBIS stock performance]. X. https://x.com/wallstengine/status/1930576175820071156

Wccftech. (2024, July 29). NVIDIA Releases Open-Source Llama 3.1 Nemotron-70B Instruct LLM, Surpassing OpenAI’s GPT-4o & Others In Key Benchmarks. Retrieved August 5, 2025, from https://wccftech.com/nvidia-open-source-llama-3-1-nemotron-70b-instruct-llm-surpassing-openai-gpt-4o/