Alibaba has reportedly ordered its staff to stop using Claude Code, Anthropic's AI programming tool, effective July 10. According to Reuters and Morningstar, the company classified the tool as "high-risk software" over concerns it could contain backdoors or surveillance capabilities, and is directing employees toward Qoder, its in-house coding assistant, instead. The move comes amid renewed online discussion of a Claude Code feature — described by Anthropic's Thariq Shihipar as "an experiment we launched in March that was meant to prevent account abuse from unauthorized resellers and protect against distillation" — that can quietly flag when a session originates from a Chinese IP address or uses Chinese-language prompts.

The irony is that Alibaba's ban is largely symbolic, closing a door that Anthropic has already spent months sealing from the other side. Anthropic's usage policy already prohibits Chinese companies, and any foreign-owned entities majority-controlled by them, from accessing its models at all, and the company has been actively working to shut down the reseller and VPN loopholes that let mainland developers reach Claude anyway. That access has been valuable: Chinese engineering teams, including at firms with no formal license to use Claude, have leaned on it for the same reason developers everywhere have — it is, by most public benchmarks, still among the strongest coding models available. A corporate ban converts an already-precarious gray-market workflow into an explicit policy violation, forcing engineers who valued Claude's output to either route around IT or fall back to homegrown tools they may trust less.

The bigger story is what the ban signals about the trajectory of the US-China AI relationship. Export controls and model-access restrictions have until now been framed primarily as a US-government project — chips, cloud compute, and now frontier model APIs treated as controlled technology. Alibaba's decision suggests Chinese firms are starting to reciprocate defensively, treating foreign AI tools as an attack surface rather than a productivity tool, independent of whether Beijing mandates it. That is a meaningful shift: it means the bifurcation of the global AI coding-tool market — one stack for the US and allies, a separate stack for China — is being reinforced from both directions simultaneously, by policy and by newly justified corporate risk aversion.

For Anthropic, the near-term financial impact is minor since Chinese usage was never sanctioned revenue. But the episode does damage on two fronts. It hands ammunition to the narrative that Claude Code's anti-abuse detection amounts to covert user profiling, a framing that will resurface every time Anthropic discusses safety or trust with any government. And it accelerates the exact outcome the export-control regime was designed to produce, but through an unplanned channel: domestic Chinese alternatives like Qoder gain a captive audience precisely at the moment they most need real-world usage data to close the capability gap. A policy meant to slow Chinese frontier-AI development may end up subsidizing it instead, by pushing a large developer population off Claude and onto Alibaba's own model at exactly the volume needed to make Qoder better.

A confidential tender document obtained by the Australian Financial Review shows Anthropic is planning to build out at least 1.4 gigawatts of data center capacity in Australia, at an infrastructure cost of roughly $15 billion. The company is targeting at least 1 gigawatt operational by the end of 2027 — an aggressive timeline for a build of this scale anywhere, let alone in a market where Anthropic has no existing compute footprint. The plan follows Anthropic's opening of an Australian office earlier this year and a March memorandum of understanding with the Australian government covering AI safety research and support for the country's national AI strategy.

The number is the headline here: 1.4GW is a genuinely large commitment, comparable to the scale of gigawatt-class US hyperscale campuses that have taken years and multiple partners to plan, and it reportedly exceeds what external observers had expected from Anthropic in the region. For context, a single gigawatt of dedicated AI data center capacity is enough to be a national-level story in most countries — Australia's federal government has been openly courting exactly this kind of foreign capital as part of its bid to become a regional AI hub, and the MOU signed in March was clearly a down payment on this larger ambition materializing.

Why Australia, and why now? The country offers relatively cheap and increasingly renewable-heavy grid capacity in specific regions, political stability, a government actively soliciting AI infrastructure investment, and geographic proximity to Asian markets without the export-control complications of building in China or, increasingly, some Middle Eastern jurisdictions under US scrutiny. For Anthropic specifically, a Southern Hemisphere compute base also diversifies the company's infrastructure away from total dependence on US soil and its incumbent partners, at a moment when US data center capacity is itself capacity-constrained and politically contested over power usage.

The harder question is whether the timeline is credible. Gigawatt-scale power procurement, grid interconnection, cooling infrastructure, and chip supply typically take longer than 18 months even for experienced hyperscalers with existing local relationships — Anthropic would be doing this essentially from a standing start in a market it entered only months ago. If the AFR's tender document is accurate, this either reflects unusually favorable pre-existing arrangements (state government fast-tracking, an existing power/land deal it is building onto) or an ambitious target that will slip, as many announced gigawatt-scale timelines have before it. Either way, it confirms that the capital intensity of frontier AI has crossed a threshold where even a company Anthropic's size is now making sovereign-scale infrastructure bets outside its home market — the compute race has gone fully global, and $15 billion single-country commitments are becoming a normal cost of staying competitive rather than an outlier.

Laurie Voss, former npm co-founder, has assembled the clearest data-backed picture yet of what AI has actually done to entry-level software jobs — and the answer is a genuine bifurcation, not the uniform "AI is coming for coders" story usually told. Citing Stanford Digital Economy Lab analysis of ADP payroll data, Voss reports that developers aged 22-25 are down 19% from their late-2022 peak, while workers aged 41-49 rose 14% over the same period. US Bureau of Labor Statistics data shows the "computer programmer" job title itself declined 16% in the single year from May 2024 to May 2025, entry-level postings are down 28% from their 2022 peak, and computer science graduate unemployment sits at 6.1% — a level that would have been considered a crisis for the major just a few years ago.

What makes the piece more than another AI-doom listicle is the second half of the data: total developer employment actually grew, from 1.53 million to 1.69 million between May 2022 and 2025, roughly 10%. That is not a contradiction — it is the mechanism. Senior developers, who make up the bulk of the workforce, are being hired at higher rates even as juniors are cut, so the aggregate number goes up while the entry-level segment collapses underneath it. Anyone reading only the top-line "developer employment is growing" statistic would conclude AI job fears are overblown; anyone looking at the 22-25 cohort alone would conclude the opposite. Both are true simultaneously, for different parts of the same market.

Voss is careful to note this isn't a clean AI-only story — ZIRP-era over-hiring unwinding, tax policy changes, and general post-pandemic labor corrections all plausibly contribute, and only about 4.5% of 2025 layoffs were explicitly attributed to AI by employers. But the timing argument is compelling: the junior-employment deterioration accelerated specifically in 2024-2025, coinciding with the rise of agentic coding tools rather than the original ChatGPT launch in 2022. That lag matters — it suggests the damage isn't from AI-as-autocomplete, which junior developers could plausibly still out-compete, but from AI-as-autonomous-agent, which directly substitutes for the exact kind of bounded, well-specified tasks junior engineers have traditionally been assigned.

Meanwhile, a separate population is emerging that doesn't show up in any of these statistics at all: GitHub added 36 million accounts and 121 million new repositories, iOS App Store submissions grew 24% in 2025 after eight years of decline, and vibe-coding platforms report that 50-63% of their active builders don't identify as "developers." The apprenticeship pipeline that trained today's senior engineers — junior work, code review, mentorship, gradual promotion — is being hollowed out at the exact moment a much larger, unsupervised population is shipping code with no equivalent training loop, and with AI-generated code failing basic OWASP security tests 45% of the time per Veracode's audits. Voss's real warning isn't that junior jobs are disappearing; it's that the pipeline producing tomorrow's senior engineers is being quietly dismantled with no replacement in place, a structural problem that current employment charts are too coarse to capture and that won't be visible as a crisis until the senior pipeline itself runs dry years from now.

据报道，阿里巴巴已下令员工停止使用 Anthropic 的 AI 编程工具 Claude Code，自7月10日起生效。据路透社和 Morningstar 报道，公司将该工具列为"高风险软件"，理由是担心其可能含有后门或监控能力，并指示员工改用其自研的编程助手 Qoder。这一举动伴随着网络上对 Claude Code 某项功能的重新讨论——Anthropic 的 Thariq Shihipar 将其描述为"我们在3月启动的一项实验，旨在防止未经授权的经销商滥用账户，并防止模型蒸馏"——该功能可以在会话来自中国 IP 地址或使用中文提示词时悄悄标记。

讽刺之处在于，阿里巴巴的禁令在很大程度上只是象征性的，它关上的是一扇 Anthropic 自己早已在从另一侧密封的门。Anthropic 的使用政策本就禁止中国公司及其外资控股实体访问其模型，公司一直在积极堵住让大陆开发者仍能接触 Claude 的经销商和 VPN 漏洞。这种访问一直很有价值：包括没有正式许可使用 Claude 的公司在内的中国工程团队，出于和世界各地开发者相同的原因依赖它——按多数公开基准衡量，它仍是可用的最强编程模型之一。企业层面的禁令把一种本已岌岌可危的灰色市场工作方式，转变为明确的政策违规，迫使那些重视 Claude 输出质量的工程师要么绕开 IT 部门，要么退回到他们可能不那么信任的自研工具。

更大的故事在于这一禁令对中美 AI 关系走向的暗示。迄今为止，出口管制和模型访问限制主要被视为美国政府主导的项目——芯片、云计算，如今扩展到前沿模型 API，都被当作受控技术处理。阿里巴巴的决定表明，中国企业开始进行防御性的对等反应，将外国 AI 工具视为攻击面而非生产力工具，且这并不依赖于北京是否下达强制指令。这是一个有意义的转变：这意味着全球 AI 编程工具市场的分裂——一套面向美国及其盟友的技术栈，另一套面向中国的技术栈——正在同时被政策和新出现的企业风险规避意识从两个方向同时强化。

对 Anthropic 而言，短期财务影响很小，因为中国用户从来就不是被批准的合法收入来源。但这一事件在两个方面造成了损害。它为"Claude Code 的反滥用检测等同于秘密用户画像"这一叙事提供了弹药，而这种说法将在 Anthropic 未来每次与任何政府讨论安全或信任议题时重新浮现。它还以一个未曾规划的渠道，加速了出口管制体系本就旨在促成的结果：Qoder 等本土中国替代品恰好在最需要真实使用数据来缩小能力差距的时刻，获得了一批被迫使用的用户。一项旨在减缓中国前沿 AI 发展的政策，最终可能反而在补贴它——通过将大量开发者从 Claude 推向阿里巴巴自己的模型，且规模恰好足以让 Qoder 变得更好。

《澳大利亚金融评论报》获得的一份机密招标文件显示，Anthropic 计划在澳大利亚建设至少 1.4 吉瓦的数据中心容量，基础设施成本约为 150 亿美元。公司目标是到 2027 年底至少有 1 吉瓦投入运营——考虑到 Anthropic 在该市场此前并无任何已有算力布局，这是一个相当激进的规模时间表。该计划紧随 Anthropic 今年早些时候在澳大利亚开设办公室，以及3月与澳大利亚政府签署的一份关于 AI 安全研究和支持该国国家 AI 战略的谅解备忘录。

这里的关键数字是 1.4 吉瓦：这是一个真正巨大的承诺，可与美国那些通常需要数年时间及多家合作伙伴共同规划的吉瓦级超大规模园区相提并论，据报道该规模超出了外部观察者对 Anthropic 在该地区布局的预期。作为参照，单单一吉瓦专用 AI 数据中心容量在多数国家就足以成为国家级新闻——澳大利亚联邦政府一直在公开争取此类外资，力图成为区域 AI 枢纽，而3月签署的谅解备忘录显然是为这一更大野心落地所做的一笔首付。

为什么是澳大利亚，为什么是现在？该国在特定地区提供相对廉价且日益依赖可再生能源的电网容量、政治稳定性、一个积极招揽 AI 基础设施投资的政府，以及毗邻亚洲市场而不必面对在中国建设（或如今日益受到美国审查的部分中东司法辖区）所带来的出口管制复杂性。对 Anthropic 而言，在南半球建立算力基地也使公司的基础设施摆脱对美国本土及其现有合作伙伴的完全依赖，而此时美国本土数据中心容量本身正受限于容量瓶颈，且因用电问题而陷入政治争议。

更难回答的问题是这一时间表是否可信。吉瓦级的电力采购、电网接入、冷却基础设施和芯片供应，即便对拥有既有本地关系的经验丰富的超大规模企业来说，通常也需要超过18个月——而 Anthropic 进入这一市场才几个月，基本上是从零开始。如果《澳大利亚金融评论报》的招标文件属实，这要么反映出异常有利的既有安排（州政府加速审批、已有的电力/土地协议），要么是一个雄心勃勃、如同此前许多已宣布的吉瓦级时间表一样最终会推迟的目标。无论如何，这都印证了前沿 AI 的资本密集度已跨过一个门槛——即便是 Anthropic 这种规模的公司，如今也在本土市场之外进行主权级规模的基础设施押注——算力竞赛已彻底走向全球化，单一国家 150 亿美元级别的承诺正在从一个特例变成保持竞争力的常规成本。

npm 前联合创始人 Laurie Voss 整理出了迄今为止关于 AI 究竟对入门级软件岗位做了什么的最清晰的数据图景——答案是真正意义上的分化，而非"AI 正在抢走程序员饭碗"这种被反复讲述的统一叙事。Voss 援引斯坦福数字经济实验室对 ADP 薪资数据的分析指出，22-25岁的开发者人数较2022年末的峰值下降了19%，而41-49岁群体在同期反而增长了14%。美国劳工统计局的数据显示，"计算机程序员"这一职位名称本身在2024年5月至2025年5月这一年内下降了16%，入门级招聘岗位较2022年峰值下降28%，计算机科学专业毕业生失业率为6.1%——这一水平在几年前对该专业而言会被视为危机。

使这篇文章超越又一篇 AI 末日清单文章的地方，在于数据的另一半：总体开发者就业实际上是增长的，从2022年5月到2025年5月的133万增至169万，增幅约为10%。这并不矛盾——这正是其运作机制所在。占从业者主体的高级开发者被以更高比例雇用，即便初级岗位在被削减，因此总体数字上升，而入门级细分市场却在其下方崩塌。只看"开发者就业在增长"这一顶层统计数据的人，会得出 AI 就业担忧被夸大的结论；只看22-25岁群体的人，则会得出相反的结论。两者在同一市场的不同部分同时成立。

Voss 谨慎地指出，这并非一个纯粹只由 AI 驱动的故事——零利率时代过度招聘的消退、税收政策变化，以及普遍的后疫情劳动力市场调整，都可能是促成因素，2025年的裁员中只有约4.5%被雇主明确归因于 AI。但时间线上的论证颇具说服力：初级岗位就业的恶化具体在2024-2025年加速，这与智能体式编程工具的兴起相吻合，而非2022年最初 ChatGPT 发布的时间点。这一滞后很重要——它表明损害并非来自作为自动补全工具的 AI（初级开发者本可与之一较高下），而是来自作为自主智能体的 AI，后者直接替代了传统上分配给初级工程师的那种边界清晰、需求明确的任务类型。

与此同时，一个完全不出现在上述任何统计数据中的独立群体正在浮现：GitHub 新增了3600万个账户和1.21亿个新代码仓库，iOS 应用商店的提交量在经历八年下降后于2025年增长了24%，"氛围编程"（vibe-coding）平台报告称其活跃建造者中有50%-63%并不认同自己是"开发者"。培养出当今高级工程师的学徒制管道——初级工作、代码审查、导师制、逐步晋升——正在恰恰是一个规模大得多、且无人监督的群体在无同等训练闭环的情况下大量产出代码之际被悄然掏空，而据 Veracode 的审计，AI 生成的代码有45%未能通过基本的 OWASP 安全测试。Voss 真正的警告并非初级岗位正在消失，而是培养明日高级工程师的管道正在被悄悄拆除且没有替代方案——这是一个当前就业图表过于粗糙、无法捕捉到的结构性问题，直到数年后高级人才管道本身枯竭之前，它都不会作为一场危机显现出来。