AI News AI资讯 11h ago Updated 1h ago 更新于 1小时前 49

In the Token Battle, Huawei Cloud Chose the Third Path | Frontline Token大战中,华为云选择了第三条路|最前线

When Zhou Yuefeng stated on stage, "We don't care about the total number of Tokens," some in the audience might have dismissed it as mere rhetoric. After all, over the past year, the entire industry has been fiercely competing over token call volumes and pricing. When Volcengine ignited the price war with a rate of 0.0008 yuan per thousand tokens—followed by Alibaba, Baidu, and Tencent joining the fray—any cloud provider claiming "we don't care" would sound like a detached remark on the battlefi 周跃峰在台上说出“我们不在乎Token总量”这句话时,台下估计有人觉得他在说场面话。毕竟,过去一年整个行业都在为了Token的调用量和价格杀红了眼。当火山引擎用0.0008元/千Token的价格点燃战火,阿里、百度、腾讯相继入场时,任何一家云厂商说“我不在乎”,都像是战场上的风凉话。但华为云这次,似乎真的想把这场仗,拉到另一条赛道上。

70
Hot 热度
70
Quality 质量
70
Impact 影响力

Analysis 深度分析

When Zhou Yuefeng declared "We don't care about the total number of Tokens" on stage, some in the audience might have seen it as empty talk. After all, the entire industry has been locked in a cutthroat battle over token call volumes and pricing over the past year. With Volcengine igniting the price war at 0.0008 yuan per thousand tokens—followed by Alibaba, Baidu, and Tencent entering the fray—any cloud provider claiming "we don't care" would sound like a detached remark on the battlefield. Yet Huawei Cloud seems genuinely intent on pulling this competition onto an entirely different track.

This is not merely a shift in rhetoric. While other vendors highlight "trillions of tokens" and "ultra-low unit prices" as core metrics, Huawei Cloud emphasizes "health" and "productivity enhancement"—essentially acknowledging an uncomfortable reality: amid the frenzied token production pipeline, much of it is fueled by "emotional value" bubbles. The example of casually asking an AI question on a phone precisely punctures the industry's facade of prosperity painted by call data. Tokens generated from casual chat and those produced in factory scheduling or hospital pathology diagnostics hold vastly different values. The former's boom may just be an AI reincarnation of mobile internet traffic anxiety, while even slight growth in the latter signifies real-world change.

This move is risky but smart. With computing power supply—especially high-end GPUs—continuously constrained, fighting head-on over the scale of token production pipelines would be like engaging in a war of attrition on the opponent's turf. Huawei Cloud's proposed "Agentic Infra" paradigm and its AICS Lingqu AI computing cluster, dubbed a "token factory," aim to "build the factory on our own foundation." Behind parameters like 100,000-card clusters and 10-millisecond latency lies a clear subtext: leveraging domestic computing power such as Ascend chips to establish a sustainable token production line unbound by supply bottlenecks. While competitors are still comparing whose factory consumes less energy, Huawei Cloud is saying: "First, I must ensure my factory can run reliably."

This represents a strategic "flanking maneuver." Price wars target market share in existing segments, but Huawei Cloud aims to define the rules for emerging markets. By collaborating with model developers like Zhipu, DeepSeek, and Kimi on "Hundred Models, Thousand Forms," and open-sourcing the OpenJiuwen agent platform, it is more akin to building an "Android Alliance for the AI era." Zhou Yuefeng’s repeated emphasis on being "the most open cloud" is not about selling its own models but becoming the "silicon-based fertile ground" for deploying all models. This path mirrors Huawei's traditional approach in telecommunications: not necessarily reigning as the flashiest terminal but becoming an indispensable player at the level of underlying infrastructure and protocols.

Of course, strategic elegance requires tangible results to back it up. When clients are still troubled by monthly cloud bills, the argument of "productivity value" demands solid case studies to be convincing. Huawei Cloud’s introduction of a medical pathology foundation model is a strong example—enabling county-level hospitals to achieve tier-three diagnostic capabilities addresses a clear societal pain point with undeniable value. The CloudRobo platform for embodied intelligence similarly tackles the common challenge of small and medium enterprises "affording computing power." These scenarios cannot be measured by "call volumes," but they serve as the most robust footnotes for Huawei Cloud’s "third path."

The sharpest point of this launch lies in its near-direct critique of a current AI competition trap: we may be using the old internet’s "traffic mindset" to operate an entirely new intelligent productivity system. Stacking token volumes resembles early internet competitions over website page views—seemingly vibrant but far from generating real economic value. Huawei Cloud’s pivot is a correction to the industry’s collective unconscious, even though this "deviation" is currently the hotspot chased by capital and traffic.

Avoiding price wars means Huawei Cloud will face pressure in acquiring new public cloud customers in the short term. However, Huawei Cloud appears willing to bet on its identity as a "long-termist." It wagers that when the industry’s frenzy subsides, the value of a cloud provider will ultimately be measured not by how many cheap tokens it produces, but by how many efficiency screws its token factory has tightened for the real economy, and how many critical pieces it has contributed to the domestic technology stack. The war over tokens may have only just entered deeper waters.

周跃峰在台上说出“我们不在乎Token总量”这句话时,台下估计有人觉得他在说场面话。毕竟,过去一年整个行业都在为了Token的调用量和价格杀红了眼。当火山引擎用0.0008元/千Token的价格点燃战火,阿里、百度、腾讯相继入场时,任何一家云厂商说“我不在乎”,都像是战场上的风凉话。但华为云这次,似乎真的想把这场仗,拉到另一条赛道上。

这不仅仅是换一套说辞。当其他厂商将“万亿Token”、“超低单价”作为核心战报时,华为云把“健康度”和“生产力提升”挂在嘴边,本质上是承认了一个尴尬的现实:狂飙的Token流水线里,大量是“情绪价值”的泡沫。那个“闲来无事在手机上问AI一个问题”的例子,精准地刺破了行业用调用数据粉饰太平的窗户纸。AI生成的闲聊Token,和在工厂产线排程、医院病理诊断中产生的Token,价值天差地别。前者的繁荣可能只是移动互联网流量焦虑的AI翻版,而后者的哪怕一点增长,都意味着真实世界的改变。

这步棋走得险,但也走得聪明。在算力供应,特别是高端GPU供应持续受限的背景下,硬拼Token流水线的规模,无异于在别人的地盘上进行消耗战。华为云提出的“Agentic Infra”范式,以及那个号称“Token工厂”的AICS灵衢智算集群,核心诉求是“把工厂建在自己的地基上”。10万卡级集群、10毫秒时延,这些参数背后真正的潜台词是:用昇腾等国产算力,支撑起一条不被卡脖子的、可持续的Token生产线。当对手们还在比拼谁的工厂能耗更低时,华为云在说:我首先要确保我的工厂能稳定开下去。

这是一种战略上的“绕后”。价格战打的是存量市场的份额,而华为云想定义的是增量市场的规则。它联合了智谱、DeepSeek、Kimi等一众模型厂商搞“百模千态”,开源openJiuwen智能体平台,更像是在搭建一个“AI时代的安卓联盟”。周跃峰反复强调的“最开放的云”,目标不是卖自家模型,而是成为所有模型落地的“硅基黑土地”。这条路径很像华为在通信领域的传统打法:不一定在最光鲜的终端上称王,但要在底层的基础设施和协议上成为不可或缺的角色。

当然,战略的优雅需要现实的业绩来支撑。当客户还在为每月云账单头疼时,“生产力价值”这个理由需要极其扎实的案例来说服。华为云拿出的医疗病理大模型是个好例子——让县域医院获得三甲诊断能力,这是切中社会痛点的、无可辩驳的价值。具身智能的CloudRobo平台也是,解决了中小公司“用不起算力”的共性难题。这些场景不是用“调用量”能衡量的,但它们能成为华为云“第三条路”最硬的注脚。

这场发布会最辛辣的地方,在于它近乎直接地指出了当前AI竞赛的一个陷阱:我们可能正在用旧互联网的“流量思维”来运营一个全新的智能生产力体系。堆砌Token数量,就像早期互联网比拼网站PV一样,看起来热闹,但离创造真正的经济价值还很远。华为云的转向,是一次对行业集体无意识的纠偏,虽然这个“偏”正是当下资本和流量追捧的热点。

不打价格战,意味着短期内在公有云的拉新上会承受压力。但华为云似乎更愿意把赌注押在“长期主义者”的身份上。它赌的是,当行业的狂欢潮水退去,最终衡量云厂商价值的,将不是你生产了多少廉价Token,而是你的Token工厂为实体经济拧紧了多少效率的螺丝,为国产技术栈补上了多少关键的拼图。这场关于Token的战争,可能才刚刚进入深水区。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

大模型 大模型 推理 推理 芯片 芯片
Share: 分享到: