0804｜时政 中国AI攻势为美国竞争对手创造了 死亡区

Technology 技术

中国的人工智能突击令美国竞争对手模型制造商陷入“死亡区”

By Bloomberg News 彭博新闻

August 4, 2026 at 12:38 PM GMT+8

A flurry of model launches from China’s AI sector is rapidly narrowing the gap with Silicon Valley and creating what’s been described as a death zone for anyone without frontier-pushing technology or market-breaking pricing.

中国人工智能领域一连串的模型发布正在迅速缩小与硅谷的差距，并且正在形成被形容为“死亡区”的局面，对那些没有前沿技术或颠覆性定价的企业极为不利。

Alibaba Group Holding Ltd. became the latest in a parade of Chinese debuts hitting the top of global benchmarks with its Qwen3.8-Max this week, its most advanced model to date that appeared to match or exceed Anthropic PBC’s flagship Fable 5. Two weeks earlier, Moonshot AI’s Kimi K3 showed

performance comparable to the priciest US options built on a much humbler budget, sparking fresh questions over the effectiveness of US chip sanctions intended to slow China’s tech ascent.

本周，阿里巴巴集团控股有限公司以其迄今为止最先进的模型 Qwen3.8-Max 成为一系列中国首秀中最新登顶全球基准测试的产品，该模型似乎可与或超越 Anthropic PBC 的旗舰产品 Fable 5。两周前，Moonshot AI 的 Kimi K3 在整体预算远逊的情况下展现了可与美国最昂贵方案相媲美的性能，这引发了人们对旨在放慢中国技术崛起的美国芯片制裁有效性的新的质疑。

Elsewhere, ByteDance Ltd. has pushed aside all rivals in video generation with its Seedance, version 2.5 of which has just come out. DeepSeek, the original Chinese disruptor of US hegemony in artificial intelligence, returned this summer with V4 Flash, a breakthrough model in terms of pricing. What was once an isolated shockwave is now a cascade of unexpected depth – delivering not just low-cost alternatives, but high-end capabilities in reasoning, coding and complex tasks that give OpenAI and its US peers real competition.

在其他方面，字节跳动有限公司凭借其视频生成产品 Seedance 将所有竞争对手甩在身后，其 2.5 版本刚刚发布。DeepSeek——最早挑战美国在人工智能领域霸主地位的中国颠覆者——今年夏天以 V4 Flash 回归，该模型在定价方面实现了突破性进展。曾经的孤立冲击现在演变为一连串出乎意料的深度冲击——不仅提供低成本替代方案，还在推理、编码和复杂任务方面展现出高端能力，给予 OpenAI 及其美国同行真正的竞争压力。

“The most important change since January 2025 is that China’s progress no longer looks like a single-company breakthrough,” said Poe Zhao, a Beijing-based tech analyst and founder of the Hello China Tech newsletter. “The first DeepSeek moment looked exceptional. The recent releases suggest China now has a repeatable system for producing models close to the global frontier.”

“自 2025 年 1 月以来最重要的变化是，中国的进展不再看起来像是单一公司的突破，”总部位于北京的科技分析师、Hello China Tech 简报创始人赵泊（Poe Zhao）说道。“第一次的 DeepSeek 时刻看起来是例外。最近的发布表明，中国现在有了一个可重复的系统，能够生成接近全球前沿的模型。”

Taken together with Z.ai’s GLM-5.2 in June, the highest-ranked open-source model globally at the time, this blitz of five models in eight weeks shows Chinese developers approaching or in some cases surpassing the US pioneers that are commonly held up as frontier AI leaders. The core value proposition isn’t just low price – it’s radical efficiency.

加上 Z.ai 在六月推出的 GLM-5.2（当时全球排名最高的开源模型），这八周内推出的五款模型显示中国开发者在接近甚至在某些情况下超越那些通常被视为前沿 AI 领导者的美国先行者。其核心价值主张不仅仅是低价——而是激进的效率。

In benchmark testing by independent evaluator Artificial Analysis, executing a complex real-world workload costs just $0.03 with DeepSeek V4 Flash, compared to $3.15 with Claude Fable 5. When China asks for cents where American companies charge dollars, the contest between the two nations for AI customers – especially ones in the rest of the

world – starts to look materially different.

在独立评估机构 Artificial Analysis 的基准测试中，执行一项复杂的真实工作负载，使用 DeepSeek V4 Flash 的成本仅为 $0.03，而使用 Claude Fable 5 则为 $3.15。当中国的价格以美分计而美国公司以美元计费时，两国在争夺 AI 客户（尤其是世界其他地区客户）方面的竞争开始显得实质性不同。

China’s AI Vanguard 中国的 AI 先锋

China’s AI Vanguard

Five domestic developers have rolled out frontier-class technology this year

Source: Bloomberg

Bloomberg

Source: Bloomberg 来源：Bloomberg

“A lot of countries are looking at the competition between US and China, and they don’t want to take a side right now because the competition’s just starting,” said George Chen, partner and Digital Practice chair at The Asia Group. Both President Donald Trump and his counterpart Xi Jinping have made AI leadership a national priority. "When Xi visits the White House in September, because of Kimi, because of Alibaba and the model they just released last night, Xi will definitely feel like he has, in Trump’s language, more cards

to play.”

“很多国家都在关注美中之间的竞争，但他们现在不想选边站，因为竞争才刚刚开始，”亚洲集团合伙人兼数字业务主管陈乔治说。美国总统唐纳德·特朗普和中国国家主席习近平都已将人工智能领导地位列为国家优先事项。“当习近平在九月访问白宫时，因为 Kimi、因为阿里巴巴以及他们昨晚刚发布的那个模型，习近平肯定会感觉到，用特朗普的话说，他手上有了更多可用的牌。”

The rapid advances are deepening US-China tech friction, prompting Washington to threaten sanctions over intellectual property concerns after Moonshot’s Kimi breakthrough. Trump has signaled that his administration is weighing China’s threat against the need for additional safety controls on products like Anthropic’s Fable or OpenAI’s GPT series. “We have to be careful in both ways,” he told reporters in the Oval Office last Wednesday. “We don’t want to restrict them where all of a sudden, we come in second to China.”

快速进展正在加剧美中科技摩擦，在 Moonshot 的 Kimi 突破后，华盛顿因知识产权问题威胁要实施制裁。特朗普已表示，他的政府正在权衡中国构成的威胁与对 Anthropic 的 Fable 或 OpenAI 的 GPT 系列等产品实施额外安全控制的必要性。“我们必须在两方面都小心，”他上周三在椭圆形办公室对记者说。“我们不想限制它们，以至于突然之间，我们在中国之后排名第二。”

To win over developers, AI contenders must now either beat DeepSeek on price or surpass it on pure capability. On a widely shared benchmark chart from Artificial Analysis, there’s now emerged a so-called DeepSeek death zone.

Charge more for the same product or be less capable for the same price and you might as well not try. Clearing the boundaries of that danger zone has become essential for long-term viability. Mid-market rivals are now under pressure to either slash prices to match the Hangzhou-based startup or spend heavily to build smarter models. The bigger GLM-5.2, Kimi K3 and Qwen3.8-Max models all sit above it in a higher performance bracket.

要赢得开发者青睐，AI 竞争者现在必须在价格上击败 DeepSeek，或在纯粹能力上超越它。在 Artificial Analysis 广泛传播的一个基准图表上，现在出现了所谓的 DeepSeek 死亡区。为相同产品收取更高费用，或以相同价格提供更低能力，那等于不必尝试。冲出该危险区的边界已成为长期存续的必要条件。中端市场的竞争对手现在面临压力，要么削价以匹配这家总部位于杭州的初创公司，要么大举投入以打造更智能的模型。更大的 GLM-5.2、Kimi K3 和 Qwen3.8-Max 模型都位于更高性能档次，处于其上方。

The recent 3 model releases set up new frontier to open source models!

近期这三款模型的发布为开源模型开启了新边界！

The gap is getting closed very quickly and it only happened within a month or two.

差距正在迅速缩小，而且只是在一两个月内发生的。

How would this figure look like by the end of this year?

到今年年底，这一形势会是什么样子？

Impressive and keep up with your great work!
@Kimi_Moonshot… pic.twitter.com/uCVmqba4Wn

令人印象深刻，继续保持你们的出色工作！
@Kimi_Moonshot… pic.twitter.com/uCVmqba4Wn
– Tiezhen WANG (@Xianbao_QIAN) July 31, 2026
– Tiezhen WANG (@Xianbao_QIAN) 2026 年 7 月 31 日

“Where I think DeepSeek V4 Flash has really changed the economics is agentic workloads,” said Dermot McGrath, founder of Shanghai-based startup consultancy ZenGen Labs.

“我认为 DeepSeek V4 Flash 真正改变经济效益的地方是具代理性的工作负载，”总部位于上海的初创咨询公司 ZenGen Labs 创始人 Dermot McGrath 说道。

When researching hundreds of companies, McGrath said he still uses Anthropic’s Claude Code as an architect to draft an action plan and summary brief. He then hands off execution to DeepSeek within the same environment, which deploys specialized AI agents to run queries and finish the task. The cost of AI work has inflated over recent months and efficiency has risen to be an even greater priority. That opens opportunity for services to augment, not necessarily

replace, the most advanced US products.

在研究数百家公司时，McGrath 表示他仍然使用 Anthropic 的 Claude Code 作为架构师来起草行动计划和摘要简报。然后他在同一环境中将执行交给 DeepSeek，后者部署专门的 AI 代理来运行查询并完成任务。近几个月来，AI 工作的成本有所上升，效率已经成为更优先的考虑。这为增强而非必然替代最先进美国产品的服务创造了机会。

“A few months ago I wouldn’t have done that. The Chinese models weren’t as reliable at tool calling or long-running agent workflows,” McGrath said. Alibaba, addressing his exact pain point, made improvements in the so-called long-horizon execution of its new Qwen release. Previous models “were slower, more brittle and sometimes didn’t fully understand what they were supposed to do.”

“几个月前我不会那样做。中国的模型在工具调用或长时运行的代理工作流方面还不够可靠，”McGrath 说。阿里巴巴针对他确切的痛点，在其新发布的 Qwen 版本中对所谓的长时执行进行了改进。以前的模型“更慢、更脆弱，有时无法完全理解它们应该做什么。”

WATCH: Micah Hill-Smith, CEO of AI model benchmarking platform Artificial Analysis, says his firm’s analysis shows that leading Chinese AI models continue to lag their US counterparts by three to nine months, even after the release

of breakthrough models. Bloomberg: The China Show.”

观看：Micah Hill-Smith，AI 模型基准测试平台 Artificial Analysis 的首席执行官，说他的公司分析显示，即使在突破性模型发布之后，领先的中国 AI 模型仍比其美国同行落后 3 到 9 个月。Bloomberg: The China Show.

Beijing-based Z.ai also made a priority of enhancing its long-horizon performance with the latest release. Chinese companies have spent the past year iterating at a blistering speed, with internal competition helping accelerate timelines and sharpen focus. DeepSeek disrupted the Chinese contest as much as it did the global picture.

总部位于北京的 Z.ai 在最新版本中也优先提升了其长远表现。过去一年里，中国公司以惊人的速度不断迭代，内部竞争推动了时间表的加速并使重心更加明确。DeepSeek 对中国赛场的冲击与对全球格局的影响同样显著。

Moonshot and Alibaba made the riskier bet of building very large models, each with more than 2 trillion parameters, an indication of their sophistication. They then moderated their energy and computing costs by only activating segments of the model at a time. OpenAI and Anthropic, by contrast, treat their parameter counts as confidential and do not disclose specific figures.

Moonshot 和阿里巴巴押注更高风险的路线，构建了非常大的模型，每个模型参数均超过 2 万亿，这显示出它们的复杂性。随后它们通过一次只激活模型的部分段落来降低能耗和计算成本。相比之下，OpenAI 和 Anthropic 将其参数数量视为机密，不披露具体数字。

Parameters are the synapses of an AI brain that help systems store, process and respond with the help of more information.

参数是人工智能“大脑”的突触，帮助系统借助更多信息进行存储、处理和响应。

The two premier US AI labs are both planning initial public offerings with aspirations for market valuations of at least a trillion dollars, a staggering mark that hinges on high-margin business models. Yet China’s cheap and open-source competition now threatens their pricing power.

这两家美国顶级人工智能实验室都计划进行首次公开募股，目标市值至少达 1 万亿美元，这是一个惊人的标志，取决于高利润率的商业模式。然而，中国廉价且开源的竞争现在威胁到了它们的定价权。

“If there weren’t these Chinese open-source models, OpenAI and Anthropic would be laughing all the way to the bank,” said Kai-Fu Lee, an AI pioneer whose startup 01.ai offers open models to global clients. “Now there’s an alternative, and it’s cheaper.”

“如果没有这些中国的开源模型，OpenAI 和 Anthropic 早就笑着进银行了，”AI 先驱、其初创公司 01.ai 向全球客户提供开源模型的李开复说。“现在有了替代品，而且更便宜。”

Alibaba’s New Flagship AI Model Comes With Attractive Pricing

阿里巴巴的新旗舰 AI 模型定价具有吸引力

Alibaba’s New Flagship AI Model Comes With Attractive Pricing

The latest from Alibaba ranks high on benchmarks and low on cost

■ Input ■ Output

Source: Company websites

Bloomberg

Source: Company websites

来源：公司网站

To be sure, Chinese companies are themselves sacrificing the prospect for significant profit in the near term in an all-out rush for users, developers and tech leadership recognition. “China’s AI providers have become trapped in a brutal price war, prioritizing market share over profitability,” Firms Too Willing to Sacifice Margin for Market Share Bloomberg Intelligence senior analyst Rob Lea.

确实，中国公司本身正在为了争夺用户、开发者和技术领导地位而全面冲刺，短期内放弃可观利润的前景。“中国的 AI 提供商已陷入一场残酷的价格战，优先考虑市场份额而非盈利能力，”Bloomberg Intelligence 高级分析师 Rob Lea 表示。

Beyond core language models, China’s AI momentum is most visible in the way it’s taken over the realm of video

generation. That market has been left wide open after OpenAI shelved its Sora tool, citing the high costs of operating a resource-intensive service. Kuaishou Technology’s Kling AI video tool, stepping into the void, drew backing from Alibaba and Tencent Holdings Ltd. recently in a $2.8 billion financing round. A significant part of China’s strength comes from the willingness of companies and investors to disregard concern about immediate returns.

除了核心语言模型外，中国在视频生成领域的势头最为明显。OpenAI 因运营资源密集型服务成本高昂而搁置 Sora 工具后，该市场被彻底放开。快手科技的 Kling AI 视频工具填补了这一空白，最近在一轮 28 亿美元的融资中获得了阿里巴巴和腾讯控股有限公司的支持。中国实力的重要来源之一是公司和投资者愿意无视对短期回报的顾虑。

No company has exemplified that approach more than ByteDance, according to BI’s Lea. It has poured money into making its Seedance the global leader on quality for AI video creation and stoked the price wars. How aggressive? “A 99% discount to prevailing market rates,” Lea said.

根据 BI 的 Lea，没有哪家公司比字节跳动更能体现这种做法。该公司投入巨资使其 Seedance 在 AI 视频创作质量方面成为全球领先者，并激发了价格战。力度有多大？“比当前市场价格打 99% 的折扣，”Lea 说道。

– Zheping Huang, Nectar Gan and Saritha Rai

– Zheping Huang、Nectar Gan 和 Saritha Rai