For two years the open web has faced an impossible bargain: to appear in Google's search results, a publisher had to let Google's crawler feed the same pages into its AI models. Refusing the training use meant vanishing from search — commercial suicide for anyone who lives on traffic. On September 15 Cloudflare severed that link, rolling out a "Disallow AI Training" directive that lets a site stay indexed while opting out of model training.

The mechanism is deceptively simple: a robots.txt signal that "Accountable" crawlers agree to honor. Cloudflare grants that label only to operators — so far Apple, Google and Microsoft — that meet four tests: a working opt-out, transparency into how pages are used, URL-level visibility, and a binding promise that opting out of training will not quietly demote a site in search rankings. Training-only crawlers from Amazon, Anthropic, Meta and OpenAI get no such middle path; they are simply blocked. Roughly 17% of Cloudflare sites already refuse training outright.

The significance is less technical than political. Cloudflare sits in front of a large slice of the internet, which gives it the leverage to turn a voluntary preference into an enforceable norm — and to decide who counts as "accountable." That is real power over the terms of the AI economy, and it lands squarely on the industry's rawest nerve: the value of training data now that models have exhausted the easy corpus. By decoupling discovery from ingestion, Cloudflare hands publishers something they have lacked since ChatGPT arrived — a credible way to say no without disappearing.

The unresolved question is enforcement. A robots.txt flag is a request, not a fence; it works only for crawlers that choose to identify themselves and behave. The labs left outside the "Accountable" club have every incentive to route around it, and Cloudflare's real deterrent is its ability to block traffic at the network edge, not the directive itself. Still, the framing matters. For the first time a neutral infrastructure layer has drawn a line between reading the web to serve users and consuming it to build a competitor — and dared the largest AI companies to respect it.

On the same day its founder was crowned Asia's richest man, ByteDance quietly reminded the market that TikTok is no longer the whole story. Anew Labs, the company's AI drug-discovery unit, closed a $290 million Series A at a $1.5 billion valuation, led by HSG, IDG Capital and Hillhouse, with ByteDance keeping a 56% majority stake. The spinoff turns an internal research group into a standalone company — and a rare Chinese pure-play in AI-designed therapeutics.

What Anew brings is not a pipeline of luck but a stack of models. Its Protenix protein-structure predictor and PXDesign generative-design system underpin four drug candidates, including an AI-designed IL-17 inhibitor presented at an immunology conference in April. Operations span Shanghai, Singapore and San Jose with only about 50 core staff, and Volcano Engine — ByteDance's cloud arm — continues to supply the compute. That last detail is the strategic tell: the deal keeps Anew tethered to ByteDance's infrastructure even as it is set free to raise outside money.

The logic for spinning out is candid. AI drug discovery, ByteDance concedes, "follows a different industry logic and management approach" than a business built on engagement and ads — clinical trials run in years, not quarters, and are governed by regulators, not recommendation algorithms. Housing that inside a consumer-tech giant strangles it; setting it loose lets specialist investors price the risk and specialist managers run it on biotech time.

The wider signal is about where China's AI ambitions are migrating. As foundation-model competition commoditizes and the domestic chatbot market saturates, the frontier of value is shifting toward AI-for-science — protein design, materials, molecules — where a good model can produce an asset a regulator will approve and a patient will pay for. ByteDance is not alone; the round drew Hillhouse, Primavera, Boyu and a Shanghai state industry fund, a coalition of commercial and quasi-state capital betting that the next Chinese tech champions may be grown in a wet lab. The risk is equally clear: models can propose molecules, but only trials validate them, and biology has humbled better-funded believers. A $1.5 billion valuation buys optionality, not a cure.

iFlytek's new Spark-Audio-1.0-Preview would be a solid speech model on its merits: a 0.65-billion-parameter encoder paired with a 30-billion-parameter mixture-of-experts decoder, covering 99 languages and 202 dialects, moving from transcription and translation into speaker identification, emotion analysis, environmental-sound detection and audio question-answering. iFlytek frames the leap as going from "hearing clearly" to "understanding thoroughly." But the specification that matters most is not in the model card's capability list — it is the line stating the system was trained entirely on domestically developed computing power.

That claim, if it holds, is the real headline. For every Chinese lab, the binding constraint has not been talent or data but access to Nvidia's accelerators under a widening US export regime. A frontier-grade audio model trained end to end on domestic silicon is a proof point that the sanctioned path is becoming viable — that the gap between wanting sovereign compute and shipping a product on it is closing. iFlytek, already on the US entity list, has more reason than most to make the point loudly, and audio is a shrewd place to make it: the workloads are demanding enough to matter yet less punishing than training a trillion-parameter language model, making it an ideal showcase for maturing homegrown hardware.

The 202-dialect coverage is its own strategic statement. Dialect and low-resource-language speech is exactly the kind of localized, data-rich problem where a domestic incumbent with two decades of Chinese-language voice data holds an edge no foreign model can easily match — and where the customer is often the state, from public services to education to security. This is AI built for a market Western labs neither prioritize nor fully understand.

The caveats are real. "Domestic compute" spans a wide quality range, and training efficiency, cost and the true performance ceiling remain unverified against Western baselines. A preview is not a benchmark. But the direction is unmistakable: while Washington debates whether export controls still bite, Chinese labs are increasingly answering the question by shipping. Spark-Audio is less a product launch than a status report on decoupling — and the report reads: progressing.

过去两年，开放网络一直被迫接受一桩不公平的交易：想出现在谷歌搜索结果里，出版商就必须允许谷歌的爬虫把同样的网页喂进它的 AI 模型。拒绝训练用途就意味着从搜索中消失——对任何靠流量为生的人来说无异于商业自杀。9 月 15 日，Cloudflare 斩断了这条捆绑，推出"禁止 AI 训练"指令，让网站既能保持被搜索收录，又能拒绝被用于模型训练。

机制看似简单：在 robots.txt 中发出一个信号，由"守信"爬虫承诺遵守。Cloudflare 只把这一标签授予满足四项条件的运营方——目前是苹果、谷歌与微软：提供可用的退出机制、公开网页被如何使用、支持 URL 级别的可见性，以及作出具约束力的承诺——退出训练不会悄悄拉低网站的搜索排名。而亚马逊、Anthropic、Meta 与 OpenAI 这类纯训练爬虫则没有这条中间道路，直接被拦。目前约 17% 的 Cloudflare 站点已彻底拒绝训练。

其意义与其说是技术性的，不如说是政治性的。Cloudflare 位于互联网相当大一部分流量的入口，这让它有能力把一项自愿偏好变成可执行的规范，也让它得以裁定谁算"守信"。这是对 AI 经济游戏规则的实实在在的权力，而且正戳在行业最敏感的神经上：在模型已耗尽易得语料之后，训练数据的价值。通过把"被发现"与"被吞噬"解耦，Cloudflare 交给出版商一样自 ChatGPT 出现以来他们一直缺失的东西——一种既能说不、又不至于消失的可行办法。

悬而未决的是执行力。robots.txt 标记是请求，而非围墙；它只对愿意表明身份并守规矩的爬虫有效。被挡在"守信"俱乐部之外的实验室有一切动机绕道而行，而 Cloudflare 真正的威慑来自它在网络边缘拦截流量的能力，而非指令本身。但话语框架很重要：这是第一次，一个中立的基础设施层在"为服务用户而读取网络"与"为打造竞争对手而吞噬网络"之间画出了一条线——并敢于要求最大的那几家 AI 公司予以尊重。

就在创始人加冕亚洲首富的同一天，字节跳动悄然提醒市场：TikTok 已不再是故事的全部。其 AI 药物研发部门 Anew Labs 完成 2.9 亿美元 A 轮融资，估值 15 亿美元，由 HSG、IDG 资本与高瓴领投，字节跳动保留 56% 的多数股权。这次分拆把一个内部研究团队变成了独立公司，也造就了中国罕见的 AI 制药"纯玩家"。

Anew 拿出的不是碰运气的管线，而是一套模型底座。其 Protenix 蛋白质结构预测器与 PXDesign 生成式设计系统支撑起四款候选药物，其中一款 AI 设计的 IL-17 抑制剂已于 4 月在一场免疫学会议上亮相。业务横跨上海、新加坡与圣何塞，核心团队仅约 50 人，算力则继续由字节跳动云业务火山引擎供给。最后这个细节正是战略要害：交易让 Anew 在被放出去自行融资的同时，仍紧紧拴在字节跳动的基础设施上。

分拆的逻辑很坦率。字节跳动承认，AI 药物研发"遵循与主业不同的行业逻辑与管理方式"——临床试验以年计而非以季度计，受监管机构约束而非推荐算法左右。把它塞在一家消费科技巨头体内只会扼杀它；放它出去，专业投资人才能给风险定价，专业管理者才能按生物科技的节奏来经营。

更宽的信号，关乎中国 AI 雄心正在向何处迁移。当基础模型竞争走向同质化、国内聊天机器人市场趋于饱和，价值前沿正转向"AI for science"——蛋白质设计、材料、分子——在那里，一个好模型能产出监管机构会批、病人会付费的资产。字节跳动并不孤单：本轮吸引了高瓴、春华、博裕以及一支上海国资产业基金，一个商业与准国有资本的联盟正押注下一批中国科技冠军或许诞生在实验台上。风险同样清晰：模型能提出分子，但只有试验能验证分子，而生物学曾让资金更雄厚的信徒折戟。15 亿美元买来的是可能性，而非解药。

单看本身，科大讯飞新发布的 Spark-Audio-1.0-Preview 已是一款扎实的语音模型：0.65B 参数编码器搭配 30B 参数的 MoE 解码器，覆盖 99 种语言、202 种方言，能力从转写、翻译延伸到说话人识别、情绪分析、环境声检测与音频问答。讯飞将这一跃迁概括为从"听清"走向"听懂"。但真正最重要的规格并不在能力清单里——而是那句"全程使用国产算力训练"。

若属实，这才是真正的头条。对每一家中国实验室而言，约束从来不是人才或数据，而是在不断收紧的美国出口管制下对英伟达加速卡的获取。一款前沿级别的音频模型从头到尾在国产芯片上训练，是一个证明点：被制裁下的路线正变得可行——"想要自主算力"与"拿它把产品做出来"之间的差距正在收窄。已被列入美国实体清单的讯飞，比多数人更有理由把这一点讲响；而选择音频来讲，也很精明：其算力负荷足够可观、有分量，却又不像训练万亿参数语言模型那般严苛，恰好是成熟中的国产硬件的理想橱窗。

202 种方言的覆盖本身就是一份战略声明。方言与低资源语种语音，正是那种本地化、数据密集的问题——一家握有二十年中文语音数据的本土巨头在此拥有外国模型难以匹敌的优势，而客户往往就是国家：从公共服务到教育再到安防。这是为一个西方实验室既不优先、也未必真正理解的市场量身打造的 AI。

该有的保留意见也真实存在。"国产算力"覆盖的质量跨度很大，训练效率、成本以及相对西方基线的真实性能上限都尚未经验证。预览版不是跑分。但方向确凿无疑：当华盛顿还在争论出口管制是否依然有效时，中国实验室正越来越多地用"发货"来作答。Spark-Audio 与其说是一次产品发布，不如说是一份关于"脱钩"的进度报告——而报告写着：进行中。