---
格式版本: 2
标题: "Mixture of Experts Powers the Most Intelligent Frontier Models | NVIDIA Blog"
原文链接: "https://blogs.nvidia.com/blog/mixture-of-experts-frontier-models/"
发布日期: "2026-06-30"
发布时间校准状态: "found"
发布时间来源: "llm:strict_original_body"
发布时间证据: "time class=related-news-date nvidia-article-date datetime=2026-06-30T08:00:57-07:00: Jun 30, 2026"
发布时间校准原因: "该日期来自HTML中的time标签，且class包含nvidia-article-date，位于标题附近，符合文章发布时间的特征。"
发布时间校准置信度: "1"
发布时间候选数量: 34
发布时间严格候选数量: 8
发布时间原页读取状态: "原页面来自已抓取 HTML"
发布时间未找到原因: "候选日期无效或 LLM 未确认"
发布时间校准时间: "2026-07-20T11:07:03+08:00"
发现时间: "2026-07-20T09:24:42+08:00"
入库时间: "2026-07-20T03:10:15.557Z"
来源平台: "NVIDIA Blog 搜索"
搜索渠道: "source_template"
搜索词: "https://blogs.nvidia.com/?s=NVL72"
匹配关键词:
  - "NVL72"
  - "GPU"
  - "Nvlink"
  - "Vera Rubin"
相关厂家:
  - "NVIDIA"
  - "Microsoft"
  - "Google"
  - "Oracle"
  - "OpenAI"
相关专家:
  []
内容类型: "网页"
抓取工具: "AgentKey Scrape"
清洗工具: "AgentKey Markdown + LLM 正文裁剪"
原始附件:
  []
AI优质: "是"
AI打分: 88
AI分档: "高置信优质"
AI质检状态: "通过"
AI打分理由: "NVIDIA官方博文，深度解析GB200 NVL72机柜级系统架构（72 GPU NVLink互联、30TB共享内存、130TB/s带宽）及MoE模型扩展瓶颈，含性能/成本数据及多家云厂商部署信息，技术细节与商业信号明确。"
AI质检模型: "qwen3.6-plus"
AI质检时间: "2026-07-20T11:10:15+08:00"
AI主题相关性: 18
AI来源权威性: 15
AI新颖性: 18
AI技术细节: 17
AI商业部署信号: 12
AI完整性: 8
图片摘要:
  - "★ ./assets/img-ac07e26e.png | chart | 展示前沿模型向MoE架构演进趋势，标注Kimi-K2、DeepSeek-R1等主流模型的时间线与智能/规模关系。"
  - "★ ./assets/img-27b94716.png | diagram | 对比Hopper H200与Blackwell NVL72架构，展示NVL72通过All-to-All NVLink Switch实现72 GPU单rack互…"
  - "★ ./assets/img-936903d0.jpg | chart | DeepSeek-R1模型在GB200上实现10x能效提升，展示Tokens/MW与用户交互TPS的性能曲线对比。"
  - "★ ./assets/img-52d7c7f9.jpg | chart | 展示GB200相比H200在DeepSeek R1推理中实现1/10 Token成本，对比每百万Token成本与端到端延迟。"
  - "★ ./assets/img-4e2f94c5.jpg | chart | Kimi-K2-Thinking模型在GB200上实现10x能效提升，展示Tokens/MW与用户交互TPS的性能曲线对比。"
  - "★ ./assets/img-4149dcbc.jpg | chart | Mistral Large 3模型在GB200上实现10x能效提升，展示Tokens/MW与用户交互TPS的性能曲线对比。"
  - "✗ ./assets/img-83d2f103.jpg | photo | 与正文主题无关，疑似通用配图或推荐阅读图片"
  - "✗ ./assets/img-1f6575f1.png | diagram | 与正文主题无关，涉及Vera CPU，正文未提及"
  - "✓ ./assets/img-3511966c.jpg | other | "
  - "✓ ./assets/img-e47e99dc.png | other | "
采集批次: "2026年7月20日9点23分34秒"
采集批次ID: "20260720-092334-062"
去重键: "https://blogs.nvidia.com/blog/mixture-of-experts-frontier-models"
---

- *The top 10 most intelligent open-source models all use a mixture-of-experts architecture.*
- *Kimi K2 Thinking, DeepSeek-R1, Mistral Large 3 and others run 10x faster to enable one-tenth the cost per token on NVIDIA GB200 NVL72.*

A look under the hood of virtually any frontier model today will reveal a [mixture-of-experts](https://www.nvidia.com/en-us/glossary/mixture-of-experts/) (MoE) model architecture that mimics the efficiency of the human brain.

Just as the brain activates specific regions based on the task, MoE models divide work among specialized “experts,” activating only the relevant ones for every [AI token](https://blogs.nvidia.com/blog/ai-tokens-explained/). This results in faster, more efficient token generation without a proportional increase in compute.

The industry has already recognized this advantage. On the independent [Artificial Analysis (AA) leaderboard](https://artificialanalysis.ai/models/open-source), the top 10 most intelligent open-source models use an MoE architecture, including DeepSeek AI’s DeepSeek-R1, Moonshot AI’s Kimi K2 Thinking, OpenAI’s gpt-oss-120B and Mistral AI’s Mistral Large 3.

However, scaling MoE models in production while delivering high performance and low cost per token is notoriously difficult. The extreme codesign of [NVIDIA GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/) systems combines hardware and software optimizations for maximum performance and efficiency, making it practical and straightforward to scale MoE models.

The Kimi K2 Thinking MoE model — ranked as the most intelligent open-source model on the AA leaderboard — sees a 10x performance leap on the NVIDIA GB200 NVL72 rack-scale system compared with NVIDIA HGX H200. This 10x performance leap, also seen on other MoE models such as [DeepSeek-R1](https://blogs.nvidia.com/blog/blackwell-inferencemax-benchmark-results/) and Mistral Large 3, enables one-tenth the cost per token and underscores why NVIDIA’s full-stack inference platform is the key to unlocking their full potential.

## What Is MoE, and Why Has It Become the Standard for Frontier Models?

Until recently, the industry standard for building smarter AI was simply building bigger, dense models that use all of their model parameters — often hundreds of billions for today’s most capable models — to generate every token. While powerful, this approach requires immense computing power and energy, making it challenging to scale.

Much like the human brain relies on specific regions to handle different cognitive tasks — whether processing language, recognizing objects or solving a math problem — MoE models comprise several specialized “experts.” For any given token, only the most relevant ones are activated by a router. This design means that even though the overall model may contain hundreds of billions of parameters, generating a token involves using only a small subset — often just tens of billions.

Like the human brain uses specific regions for different tasks, mixture-of-experts models use a router to select only the most relevant experts to generate every token.

By selectively engaging only the experts that matter most, MoE models achieve higher intelligence and adaptability without a matching rise in computational cost. This makes them the foundation for efficient AI systems optimized for performance per dollar and per watt — generating significantly more intelligence for every unit of energy and capital invested.

Given these advantages, it is no surprise that MoE has rapidly become the architecture of choice for frontier models, adopted by over 60% of open-source AI model releases this year. Since early 2023, it’s enabled a nearly 70x increase in model intelligence — pushing the limits of AI capability.

![图片](./assets/img-ac07e26e.png)

Since early 2025, nearly all leading frontier models use MoE designs.

“Our pioneering work with OSS mixture-of-experts architecture, starting with Mixtral 8x7B two years ago, ensures advanced intelligence is both accessible and sustainable for a broad range of applications,” said Guillaume Lample, cofounder and chief scientist at Mistral AI. “Mistral Large 3’s MoE architecture enables us to scale AI systems to greater performance and efficiency while dramatically lowering energy and compute demands.”

## Overcoming MoE Scaling Bottlenecks With Extreme Codesign

Frontier MoE models are simply too large and complex to be deployed on a single GPU. To run them, experts must be distributed across multiple GPUs, a technique called expert parallelism. Even on powerful platforms such as the NVIDIA H200, deploying MoE models involves bottlenecks such as:

- **Memory limitations**: For each token, GPUs must dynamically load the selected experts’ parameters from high-bandwidth memory, causing frequent heavy pressure on memory bandwidth.
- **Latency**: Experts must execute a near-instantaneous all-to-all communication pattern to exchange information and form a final, complete answer. However, on H200, spreading experts across more than eight GPUs requires them to communicate over higher-latency scale-out networking, limiting the benefits of expert parallelism.

The solution: extreme codesign.

NVIDIA GB200 NVL72 is a rack-scale system with 72 NVIDIA Blackwell GPUs working together as if they were one, delivering 1.4 exaflops of AI performance and 30TB of fast shared memory. The 72 GPUs are connected using NVLink Switch into a single, massive NVLink interconnect fabric, which allows every GPU to communicate with each other with 130 TB/s of NVLink connectivity.

MoE models can tap into this design to scale expert parallelism far beyond previous limits — distributing the experts across a much larger set of up to 72 GPUs.

This architectural approach directly resolves MoE scaling bottlenecks by:

- **Reducing the number of experts per GPU**: Distributing experts across up to 72 GPUs reduces the number of experts per GPU, minimizing parameter-loading pressure on each GPU’s high-bandwidth memory. Fewer experts per GPU also frees up memory space, allowing each GPU to serve more concurrent users and support longer input lengths.
- **Accelerating expert communication**: Experts spread across GPUs can communicate with each other instantly using NVLink. The NVLink Switch also has the compute power needed to perform some of the calculations required to combine information from various experts, speeding up delivery of the final answer.

![图片](./assets/img-27b94716.png)

Other full-stack optimizations also play a key role in unlocking high inference performance and lower cost per token for MoE models. The [NVIDIA Dynamo](https://developer.nvidia.com/dynamo) framework orchestrates [disaggregated serving](https://www.nvidia.com/en-us/glossary/disaggregated-serving/) by assigning prefill and decode tasks to different GPUs, allowing decode to run with large expert parallelism, while prefill uses parallelism techniques better suited to its workload. The [NVFP4](https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/) format helps maintain accuracy while further boosting performance and efficiency.

Open-source inference frameworks such as NVIDIA TensorRT-LLM, SGLang and vLLM support these optimizations for MoE models. [NVIDIA TensorRT-LLM](https://developer.nvidia.com/blog/scaling-large-moe-models-with-wide-expert-parallelism-on-nvl72-rack-scale-systems/) and [SGLang](https://lmsys.org/blog/2025-09-25-gb200-part-2/) have played a significant role in advancing large-scale MoE deployment on GB200 NVL72, helping validate and mature many of the techniques used today.

To bring this performance to enterprises worldwide, the GB200 NVL72 is being deployed by major cloud service providers and [NVIDIA Cloud Partners](https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/partners/) including Amazon Web Services, Core42, CoreWeave, Crusoe, Google Cloud, Lambda, Microsoft Azure, Nebius, Nscale, Oracle Cloud Infrastructure, Together AI and others.

“At CoreWeave, our customers are leveraging our platform to put mixture-of-experts models into production as they build agentic workflows,” said Peter Salanki, cofounder and chief technology officer at CoreWeave. “By working closely with NVIDIA, we are able to deliver a tightly integrated platform that brings MoE performance, scalability and reliability together in one place. You can only do that on a cloud purpose-built for AI.”

Customers such as DeepL are using Blackwell NVL72 rack-scale design to build and deploy their next-generation AI models.

“DeepL is leveraging NVIDIA GB200 hardware to train mixture-of-experts models, advancing its model architecture to improve efficiency during training and inference, setting new benchmarks for performance in AI,” said Paul Busch, research team lead at DeepL.

## The Proof Is in the Performance Per Watt and Per Dollar

NVIDIA GB200 NVL72 efficiently scales complex MoE models and delivers a 10x leap in performance per watt.

![图片](./assets/img-936903d0.jpg)

This performance leap isn’t just a benchmark; it lowers the cost -per -token by over 10x, transforming the economics of AI at scale in power- and cost-constrained data centers.

Simply put, because GB200 NVL72 processes ten times as many tokens using the same time and power, the cost required to generate each individual token drops. GB200 NVL72’s order-of-magnitude efficiency gains far outweigh its incremental increase in costs, resulting in significantly better token economics.

As shown in the [SemiAnalysis InferenceMax results](https://developer.nvidia.com/blog/nvidia-blackwell-leads-on-new-semianalysis-inferencemax-benchmarks/) for DeepSeek-R1, this efficiency delivers more than a 10x reduction in cost per million tokens compared to NVIDIA H200 systems. This dramatic shift allows enterprises to deploy complex “thinking” MoE models into everyday products.

![图片](./assets/img-52d7c7f9.jpg)

The chart demonstrates that NVIDIA GB200 NVL72 delivers 1/10 the cost per token of NVIDIA H200 at the same end-to-end latency — the total time required to generate a full response.

This 10x performance leap extends to other frontier models as well, significantly reducing cost per token.

“With GB200 NVL72 and Together AI’s custom optimizations, we are exceeding customer expectations for large-scale inference workloads for MoE models like DeepSeek-V3,” said Vipul Ved Prakash, cofounder and CEO of Together AI. “The performance gains come from NVIDIA’s full-stack optimizations coupled with Together AI Inference breakthroughs across kernels, runtime engine and speculative decoding.”

Kimi K2 Thinking, the most intelligent open-source model, serves as another proof point, achieving 10x better generational performance when deployed on GB200 NVL72.

![图片](./assets/img-4e2f94c5.jpg)

Fireworks AI has currently deployed Kimi K2 on the NVIDIA B200 platform to achieve the [highest performance on the Artificial Analysis leaderboard](https://artificialanalysis.ai/models/kimi-k2-thinking/providers).

“NVIDIA GB200 NVL72 rack-scale design makes MoE model serving dramatically more efficient,” said Lin Qiao, cofounder and CEO of Fireworks AI. “Looking ahead, NVL72 has the potential to transform how we serve massive MoE models, delivering major performance improvements over the Hopper platform and setting a new bar for frontier model speed and efficiency.”

Mistral Large 3 also achieved a 10x performance gain on the GB200 NVL72 compared with the prior-generation H200. This generational gain translates into better user experience, lower per-token cost and higher energy efficiency for this new MoE model.

![图片](./assets/img-4149dcbc.jpg)

## Powering Intelligence at Scale

The NVIDIA GB200 NVL72 rack-scale system is designed to deliver strong performance and lower token costs beyond MoE models.

The reason becomes clear when taking a look at where AI is heading: the newest generation of multimodal AI models have specialized components for language, vision, audio and other modalities, activating only the ones relevant to the task at hand.

The [dramatic cost savings seen](https://blogs.nvidia.com/blog/inference-open-source-models-blackwell-reduce-cost-per-token/) across healthcare, gaming and customer service are driven by the efficiency of NVIDIA Blackwell.

For a large-scale MoE model, inference provider DeepInfra reduced the cost per million tokens from 20 cents on the NVIDIA Hopper platform to 10 cents on Blackwell by running large open source models on its inference platform, powered by Blackwell GPUs and NVIDIA TensorRT LLM. Using Blackwell’s native low-precision NVFP4 format reduced the cost even further to just 5 cents — a total 4x improvement in cost per token while maintaining the accuracy customers expect.

In agentic systems, different “agents” specialize in planning, perception, reasoning, tool use or search, and an orchestrator coordinates them to deliver a single outcome. In both cases, the core pattern mirrors MoE: route each part of the problem to the most relevant experts, then coordinate their outputs to produce the final outcome.

Extending this principle to production environments where multiple applications and agents serve multiple users unlocks new levels of efficiency. Instead of duplicating massive AI models for every agent or application, this approach can enable a shared pool of experts accessible to all, with each request routed to the right expert.

Mixture of experts is a powerful architecture moving the industry toward a future where massive capability, efficiency and scale coexist. The GB200 NVL72 unlocks this potential today, and NVIDIA’s roadmap with the NVIDIA Vera Rubin architecture will continue to expand the horizons of frontier models.

Learn more about how GB200 NVL72 scales complex MoE models in this [technical deep dive](https://developer.nvidia.com/blog/scaling-large-moe-models-with-wide-expert-parallelism-on-nvl72-rack-scale-systems/) and register for [NVIDIA GTC 2026 sessions](https://www.nvidia.com/gtc/sessions/scaling-ai-inference-with-nvidia/) on scaling AI inference

*This post is part of* [*Think SMART*](https://blogs.nvidia.com/blog/tag/think-smart)*, a series focused on how leading AI service providers, developers and enterprises can boost their* [*inference performance*](https://developer.nvidia.com/deep-learning-performance-training-inference/ai-inference) *and return on investment with the latest advancements from NVIDIA’s full-stack* [*inference platform*](https://www.nvidia.com/en-us/solutions/ai/inference/)*.*

![图片](./assets/img-ac07e26e.png)Since early 2025, nearly all leading frontier models use MoE designs.

![图片](./assets/img-27b94716.png)

![图片](./assets/img-936903d0.jpg)

![图片](./assets/img-52d7c7f9.jpg)The chart demonstrates that NVIDIA GB200 NVL72 delivers 1/10 the cost per token of NVIDIA H200 at the same end-to-end latency — the total time required to generate a full response.

![图片](./assets/img-4e2f94c5.jpg)

![图片](./assets/img-4149dcbc.jpg)

![NVIDIA Unlocks AI Compute at Scale, Inviting Partners to Power the AI Infrastructure Buildout](./assets/img-3511966c.jpg)

![How NVIDIA’s Inference Software Stack Powers the Lowest Token Cost](./assets/img-e47e99dc.png)