LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
NotíciasTópico · 海马云生产力系统Material4 min de leituraEditorial DisconfirmAI

Haiyun Cloud Founder Dang Jinfeng: Models Are Available to Anyone, but the Gap Lies in the Productivity System

Dang Jinfeng, CEO of Haiyun Cloud, argues that competition at the AI application layer has shifted from model capabilities to the systemic integration of computing power, toolchains, and production data. He supports this view with two mechanisms: integrated rendering and inference scheduling, and expert decision trajectories.

2026-11-03 19:13 GMT+8

The gap in large model parameters is narrowing daily, and practitioners increasingly believe that application models relying solely on API wrapping are difficult to build barriers around. Dang Jinfeng, founder and CEO of Haiyun Cloud, takes this judgment further: he believes that model routing only solves for call efficiency, not delivery capability. The true differentiator between products is a productivity system where computing power, Harness (Agents, context, toolchains), and data are deeply interlocked. In his view, AI without industry-specific Agents and expert trajectories will ultimately remain mere toys that generate text.

The specific mechanism behind this argument is RunningHub, a platform under Haiyun Cloud. Haiyun Cloud started in real-time cloud rendering in 2014, providing cloud gaming services for games such as Genshin Impact and Black Myth: Wukong, accumulating heterogeneous GPU scheduling capabilities. The key cost logic pointed out by Dang Jinfeng is that in the real production processes of audio, video, and images, AI inference and traditional graphics rendering are alternating workloads—3D scenes require rendering, Blender and Unreal runtimes rely on GPUs, and video generation and super-resolution synthesis consume massive computational resources. Traditional practices stockpile cards separately for inference and rendering based on peak loads, inevitably resulting in queuing during peaks and idle time during troughs. Haiyun Cloud's approach is to serve different workloads with the same batch of GPUs at different times, with a unified scheduling layer allocating resources in real-time based on task type, model size, and latency requirements—a concept known as integrated rendering and inference. According to Haiyun Cloud, this scheduling pushes the platform's overall computing utilization to a position 70% higher than common industry levels.

Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

A deeper barrier claim than computing scheduling is data. Dang Jinfeng believes that tools and processes can eventually be imitated; the true moat lies in compliantly obtained production site data and expert decision trajectories. The specific mechanism is as follows: when senior creators complete a production task on the platform, their choices—which model to select, which tool nodes to invoke, how many rounds a video was modified, and whether each round adjusted lighting or expressions—are fully recorded as anonymized operation trajectories. These are converted into reinforcement learning reward signals, training models to compress what was originally a cumbersome multi-step game into a single native inference. Haiyun Cloud states that approximately 60% of model calls on RunningHub are completed by intelligent dynamic routing, with Agents assigning simple tasks to small models and complex tasks to large models based on task type.

Boundaries that must be retained: All the above data are self-reported by Haiyun Cloud and published through tech media, without independent verification. The five-layer capability stack (computing layer, model layer, Harness layer, data layer, application layer) is the company's own strategic narrative framework. The article juxtaposes Claude Code, Lovable, Figma, and RunningHub to argue for Harness value, but these three belong to programming tools, website generation, and design collaboration respectively, limiting their comparability to AIGC production pipelines. If MiniMax's official reverse procurement of H3 RH Enhanced via RunningHub is true, it is indeed rare, but it is supported by only a single source.

However, the actual contribution of Dang Jinfeng's discourse lies in deconstructing the broad judgment that 'vertical AI has barriers' into three checkable engineering links: heterogeneous GPU pooling scheduling addresses the cost baseline; expert operation trajectories serve as reinforcement learning signals to address the alignment between models and production needs; and Agent dynamic routing addresses the attribution of decision-making authority for multi-model collaboration. Each of these three links has clear failure conditions—if rendering and inference workloads cannot switch efficiently on the same hardware, if the volume of collected creator operation trajectories is insufficient to form effective reward signals, or if Agent routing accuracy falls below human selection—they are not self-acting flywheels, but systematic capabilities that require continuous engineering investment to maintain. Readers evaluating any AI application company can use these three specific dimensions to replace the generic 'vertical scenario' label, but all performance figures still require independent verification before serving as the basis for investment judgments.

Fontes

  • qbitai.com量子位↗
Tópico

海马云生产力系统

海马云创始人党劲峰提出,AI应用层的竞争焦点已从模型能力转向由算力、工具链与生产数据整合而成的“生产力系统”。随着大模型参数差距收窄,单纯依赖API调用的商业模式难以形成壁垒,海马云通过其旗下平台RunningHub展示了另一种解法:利用十…

4 etapas1 peçasAtualizado 2026-11-04 00:06

Como chegamos aqui

  1. 2014-01-01海马云成立,以实时云渲染起家,积累GPU算力调度能力。qbitai.com ↗
  2. 2026-09-11智东西报道,RunningHub推出并开源H3视频生成加速方案H3 Lightning,将5秒视频生成耗时从348.8秒缩短至28.7秒,提速约12倍。news.qq.com ↗
  3. 2026-07-17世界人工智能大会(WAIC)期间,海马云展示“AI+内容”全栈解决方案,联合创始人温健透露平台每月Token消耗量达9万亿。bjnews.com.cn ↗

O que ainda está indefinido

  • 海马云提出的“渲推一体”算力调度模式能否被其他缺乏底层GPU掌控力的纯应用层公司有效复制?
  • 垂直场景后训练模型(如RH系列)反向输出给基础大模型厂商(如MiniMax采购H3 RH Enhanced)的模式是否会成为行业常态?
  • 在Agent驱动的生产流程中,如何量化“专家决策轨迹”对最终交付质量的具体贡献率?
Ler o tópico completo →
Continuar pesquisando →