The February jobs report was a genuine shock. The US economy shed 120,000 non-farm jobs — economists had expected a gain of 180,000. The unemployment rate ticked up to 4.2%, its highest since the pandemic recovery plateau. Government layoffs, driven by Elon Musk's Department of Government Efficiency (DOGE) cuts, accounted for roughly 60,000 of those losses.
At the same time, oil surged past $85 a barrel on fears of supply disruption through the Strait of Hormuz. Iran-linked militia activity and new US sanctions threats have put roughly 20% of the world's oil transit at risk. Brent crude is at levels not seen since 2023.
The Dow fell 890 points (2.1%) on Friday. The S&P 500 dropped 1.7%, and the Nasdaq shed 2.3% as tech stocks — already under pressure from tariff uncertainty — led the decline. The VIX (fear index) spiked above 25.
Bond yields moved sharply: the 10-year fell to 4.05% as traders priced in higher probability of a June rate cut. Gold hit $2,950 — a new record.
Jobs weakness alone would be manageable — the Fed would simply cut sooner. Oil spikes alone would be manageable — temporary supply shocks pass. But both together create a stagflation signal: weakening demand plus cost-push inflation. That's the scenario central banks hate most because rate cuts fuel inflation while rate holds deepen the slowdown.
USD exposure: The dollar weakened 0.6% against a basket. If this continues, it's a tailwind for CNY-denominated assets but raises import costs for oil-dependent economies like China. Watch the Fed's March 18-19 meeting for guidance.
HK equities: HSI actually held up better than US markets on Friday (closed before the worst of the sell-off). Monday's open will be the tell — expect some contagion but potentially less than feared. Beijing's stimulus signals at the annual parliament sessions could provide a floor.
Portfolio action: Don't panic-sell into a fear spike. The oil component is likely transient (Hormuz crises historically resolve within weeks). The jobs data is the more structural concern — wait for next month's revision before drawing conclusions.
2月就业报告是真正的冲击。美国经济减少了12万非农就业——经济学家预期增加18万。失业率升至4.2%，为疫后复苏以来最高。政府裁员——受马斯克政府效率部(DOGE)推动——占约6万个岗位流失。
与此同时，油价突破每桶85美元，霍尔木兹海峡供应中断恐慌升级。伊朗相关民兵活动和美国新制裁威胁使全球约20%石油运输面临风险。布伦特原油达到2023年以来最高水平。
道指周五暴跌890点(2.1%)。标普500下跌1.7%，纳斯达克跌2.3%，科技股——已受关税不确定性压制——领跌。VIX恐慌指数飙升至25以上。
债券收益率剧烈波动：10年期降至4.05%，交易员定价6月降息概率上升。黄金创纪录达2,950美元。
单独的就业疲软可以应对——美联储提前降息即可。单独的油价飙升也可以应对——暂时供应冲击会过去。但两者叠加形成滞胀信号：需求疲软加成本推动型通胀。这是央行最恐惧的场景——降息助长通胀，维持利率加深衰退。
美元敞口：美元对一篮子货币贬值0.6%。若持续，利好人民币计价资产，但推高中国等石油依赖经济体的进口成本。关注美联储3月18-19日会议。
港股：恒指周五表现实际好于美股（在最大跌幅前收盘）。周一开盘是关键——预计有传导但可能弱于预期。两会期间北京的刺激信号可能提供支撑。
组合操作：不要在恐慌中抛售。石油部分很可能是暂时的（霍尔木兹危机历史上通常在数周内解决）。就业数据是更具结构性的担忧——等下月修正数据再下结论。
Global venture capital hit $189 billion in February 2026, according to Crunchbase — a single-month record. But the distribution tells the real story: OpenAI ($40B), Anthropic ($25B), and Waymo ($15B) accounted for more than 40% of all funding. Strip those three out and the remaining landscape looks surprisingly normal.
AI infrastructure dominates. Foundation model companies, chip designers, and data centre operators are absorbing the lion's share. Application-layer startups — the ones building on top of these models — are raising too, but at more modest valuations and with more scrutiny on unit economics.
Family offices are increasingly direct investors: 41 direct AI investments tracked in February alone, up from single digits a year ago. They're bypassing traditional VC funds entirely, attracted by the potential returns and frustrated by fund timelines.
This level of concentration is historically unusual and echoes the late-1990s pattern where a handful of companies absorbed most capital before a broader correction. It doesn't mean a crash is imminent — the underlying technology is far more mature than 1999's internet — but it does mean the bar for new entrants is rising fast.
If evaluating Qrio's fundraising: The window is open but narrowing. What gets funded today: vertical specificity (not "AI for everything"), defensible data moats, and clear monetisation paths within 18 months. "We'll figure out revenue later" pitches are getting rejected even in this climate.
As an investor: Public market AI exposure (MSFT, NVDA, GOOGL) is the safer way to ride this wave. Direct startup investment carries extreme concentration risk — you're essentially betting on one team at one valuation in a market that could correct 30-50% within 18 months.
据Crunchbase，2026年2月全球风投达到1890亿美元，创单月纪录。但分布才是重点：OpenAI(400亿)、Anthropic(250亿)、Waymo(150亿)占全部融资的40%以上。去掉这三家，其余市场其实相当正常。
AI基础设施占主导。基础模型公司、芯片设计商和数据中心运营商吸收了大部分资金。应用层创业公司也在融资，但估值更温和，对单位经济效益审查更严。
家族办公室越来越多直投：2月追踪到41笔AI直投，一年前还是个位数。它们绕过传统VC基金，被潜在回报吸引，对基金周期不满。
这种集中度历史上不常见，类似90年代末模式——少数公司吸收大部分资本，随后出现广泛调整。不意味崩盘迫在眉睫——底层技术比1999年的互联网成熟得多——但新入场者门槛在快速提高。
如果评估Qrio融资：窗口开放但在缩小。今天能拿到钱的：垂直细分（非"万能AI"）、可防御的数据护城河、18个月内清晰的变现路径。"以后再想收入"的方案即便在这个市场也被拒绝。
作为投资者：公开市场AI敞口（MSFT、NVDA、GOOGL）是更安全的方式。直投创业公司有极端集中风险——本质上是在一个可能18个月内修正30-50%的市场中，以一个估值押注一个团队。
Twenty-four US states, led by California and New York, filed a joint lawsuit challenging President Trump's authority to impose blanket 10% tariffs on all imports. The legal argument centres on the Trade Act of 1974: the states argue it was designed for specific national security threats, not routine trade deficits.
The suit was filed in the Court of International Trade in New York. Legal experts give it a reasonable chance of at least a preliminary injunction — the court has previously narrowed executive trade authority.
Treasury Secretary Scott Bessent has been quietly signalling that the tariffs are a negotiating tool, not permanent policy. In a CNBC interview this week, he suggested rates could "revert within five months" if trading partners make concessions. But markets aren't waiting for diplomatic timelines.
The tariffs have already triggered retaliatory measures from Canada (25% on US goods), the EU (targeted tech tariffs), and China (expanded rare earth export controls). A trade war spiral is underway regardless of the lawsuit's outcome.
If the court grants an injunction, expect a sharp relief rally in export-dependent stocks — particularly Asian manufacturers, European luxury, and US retailers who've been pricing in higher input costs. HSI could see a 3-5% bounce.
If the court declines, the tariffs continue and the next catalyst becomes the US-China summit expected in late April. Beijing has signalled willingness to negotiate but won't be seen as conceding under pressure.
HK equities are essentially a tariff trade right now. The HSI's 2026 performance correlates almost perfectly with tariff headlines. This lawsuit adds a new variable — legal resolution could come faster than diplomatic resolution.
Don't trade on the lawsuit outcome. Binary legal bets are coin flips. Instead, position for the broader theme: tariff uncertainty is the new normal for at least 6 months. Favour companies with pricing power and domestic China revenue over export-dependent plays.
以加州和纽约为首的24个州联合提起诉讼，挑战特朗普总统对所有进口商品征收10%统一关税的权力。法律论点集中在1974年贸易法：各州认为该法是为特定国家安全威胁设计的，而非常规贸易逆差。
诉讼在纽约国际贸易法院提起。法律专家认为至少获得初步禁令的可能性合理——该法院此前曾缩小过行政贸易权力。
财长贝森特一直在悄悄暗示关税是谈判工具而非永久政策。他在本周CNBC采访中暗示，如果贸易伙伴作出让步，税率可能"五个月内回调"。但市场不等待外交时间表。
关税已引发加拿大（对美商品征25%）、欧盟（定向科技关税）和中国（扩大稀土出口管制）的报复措施。无论诉讼结果如何，贸易战螺旋已经开始。
如果法院批准禁令，预期出口依赖型股票大幅反弹——尤其是亚洲制造商、欧洲奢侈品和已定价更高投入成本的美国零售商。恒指可能反弹3-5%。
如果法院拒绝，关税继续，下一个催化剂是预计4月下旬的中美峰会。北京表示愿意谈判，但不会在压力下被视为让步。
港股现在本质上是关税交易。恒指2026年表现与关税头条几乎完全相关。这场诉讼增加了新变量——法律解决可能比外交解决更快。
不要押注诉讼结果。二元法律赌注是掷硬币。应关注更大主题：关税不确定性至少是未来6个月的新常态。偏好有定价权和国内中国收入的公司，而非出口依赖型标的。
OpenAI quietly removed direct purchase capabilities from ChatGPT's shopping features. The product recommendations remain — you can still ask "what's the best noise-cancelling headphone under $300" and get curated results — but the checkout step now redirects to the merchant's own site or app.
This is a strategic retreat. OpenAI initially positioned ChatGPT as a full-funnel commerce tool: discover, compare, buy — all within the conversation. Now it's discovery-only.
Three reasons. First, payment compliance is expensive — PCI-DSS, regional payment regulations, fraud liability. Second, merchants resisted giving up the customer relationship at checkout. Third, the margin wasn't there: taking a cut of transactions requires scale that ChatGPT's shopping feature hadn't achieved.
There's now a clear white space in the market: the middleware layer between AI discovery and merchant checkout. Think of it as "Stripe for AI commerce" — a platform that handles the handoff when an AI agent says "buy this" and a merchant needs to fulfil it.
This isn't theoretical. Google Shopping, Amazon, and Shopify are all building AI discovery features. None of them have solved the cross-platform checkout problem — the moment a user discovers a product through an AI but wants to buy from a different merchant.
As a startup evaluator: If you see a pitch for "AI commerce infrastructure" — the plumbing between discovery and purchase — it's worth serious attention. The market is real, OpenAI just validated the problem by failing to solve it in-house.
For Microsoft/SharePoint: The same pattern applies to enterprise procurement. Copilot will increasingly surface purchasing recommendations (office supplies, software licenses, vendor services). The checkout layer for B2B AI-assisted purchasing is equally unsolved.
OpenAI悄悄从ChatGPT购物功能中移除了直接购买能力。产品推荐仍在——你仍可以问"300美元以下最好的降噪耳机"并获得策划结果——但结账步骤现在跳转到商家自己的网站或应用。
这是战略性撤退。OpenAI最初将ChatGPT定位为全链路商业工具：发现、比较、购买——全在对话内完成。现在只做发现。
三个原因。第一，支付合规成本高——PCI-DSS、区域支付法规、欺诈责任。第二，商家抵制在结账环节放弃客户关系。第三，利润率不够：抽取交易分成需要ChatGPT购物功能尚未达到的规模。
市场出现明确空白：AI发现与商家结账之间的中间件层。可以理解为"AI商业的Stripe"——一个处理AI代理说"买这个"到商家完成履约之间交接的平台。
这不是理论。Google Shopping、Amazon和Shopify都在建AI发现功能。没有人解决了跨平台结账问题——用户通过AI发现产品但想从另一个商家购买。
作为创业评估者：如果看到"AI商业基础设施"的项目——发现与购买之间的管道——值得认真关注。市场是真实的，OpenAI刚刚通过无法内部解决来验证了这个问题。
对微软/SharePoint：同样的模式适用于企业采购。Copilot将越来越多地推荐购买（办公用品、软件许可、供应商服务）。B2B AI辅助采购的结账层同样未解决。
Microsoft updated Copilot in Edge and Windows to open web links directly inside the Copilot panel rather than launching a new browser tab. When you ask Copilot a question and it cites a source, clicking that source now loads an embedded reader view within the assistant itself.
This is subtle but significant. Copilot is no longer a sidebar that sends you elsewhere — it's becoming a self-contained reading surface.
This follows the same trajectory as social media platforms absorbing news: Facebook Instant Articles, Twitter's reader mode, Apple News. The platform that controls the reading surface controls the attention. Microsoft is applying this to productivity.
The next logical step — already in preview — is Copilot summarising SharePoint pages without users ever visiting SharePoint. They'll get the answer, never see the page, and the page designer's work becomes invisible.
If users consume content through Copilot, the design of the page matters less than the quality of its metadata. Copilot extracts structured information — titles, descriptions, tagged fields, table data. It ignores visual hierarchy, custom layouts, and brand design.
This means SharePoint strategy shifts from "design beautiful pages" to "structure data so Copilot extracts it correctly." It's the SEO-to-AI optimisation transition that happened to Google Search, now arriving in the enterprise.
Immediate: Audit your team's SharePoint templates. Are they optimised for human reading or machine extraction? Most are the former. Start adding structured metadata — especially to knowledge base articles, policy documents, and FAQ pages.
Strategic: This is a massive opportunity to reframe your role. The person who understands how to design information architecture for AI consumption is the most valuable person in the enterprise knowledge stack right now. Position accordingly.
微软更新了Edge和Windows中的Copilot，在Copilot面板内直接打开网页链接而非打开新浏览器标签。当你问Copilot问题并引用来源时，点击来源会在助手内加载嵌入式阅读视图。
这很微妙但意义重大。Copilot不再是把你送到别处的侧边栏——它正成为自包含的阅读界面。
这与社交媒体平台吸收新闻的轨迹相同：Facebook即时文章、Twitter阅读模式、Apple News。控制阅读界面的平台控制注意力。微软正将此应用于生产力。
下一个逻辑步骤——已在预览中——是Copilot在用户从未访问SharePoint的情况下总结SharePoint页面。他们得到答案，从未看到页面，页面设计师的工作变得不可见。
如果用户通过Copilot消费内容，页面设计不如元数据质量重要。Copilot提取结构化信息——标题、描述、标签字段、表格数据。它忽略视觉层次、自定义布局和品牌设计。
这意味着SharePoint策略从"设计漂亮页面"转向"结构化数据使Copilot正确提取"。这是SEO到AI优化的转变——Google搜索已经历的，现在到达企业。
立即：审计团队的SharePoint模板。它们是为人类阅读还是机器提取优化的？大多数是前者。开始添加结构化元数据——尤其是知识库文章、政策文档和FAQ页面。
战略：这是重新定位角色的巨大机会。理解如何为AI消费设计信息架构的人是当前企业知识栈中最有价值的人。据此定位。
At HIMSS 2026 (healthcare IT conference), Microsoft unveiled Dragon Copilot's evolution from a clinical documentation tool to a full agent platform. Third-party developers can now build specialised AI agents — for radiology, pharmacy, patient scheduling — that plug into Dragon's infrastructure.
Previously, Dragon was a dictation tool that happened to have AI. Now it's a platform that happens to start with dictation.
Microsoft is running the same playbook across verticals: documentation → copilot → platform → agent marketplace. Dragon in healthcare. GitHub Copilot in development. Dynamics Copilot in business operations. The next logical domain is knowledge management — your territory.
Expect SharePoint Copilot to evolve from "search and summarise" to a platform where third-party agents handle specific knowledge tasks: compliance checking, content governance, automated publishing workflows.
Start thinking in agents, not features. Your design work should anticipate a world where SharePoint hosts multiple AI agents — each handling a slice of the knowledge workflow. The design challenge becomes orchestration: how do users understand what each agent does, trust its outputs, and manage conflicts between agents?
This is a career-defining design problem. Multi-agent UX is almost completely uncharted territory. The person who defines the patterns here shapes the next decade of enterprise software.
在HIMSS 2026（医疗IT大会），微软展示了Dragon Copilot从临床文档工具到完整Agent平台的进化。第三方开发者现在可以构建专业AI代理——放射学、药房、患者排班——接入Dragon基础设施。
此前Dragon是一个碰巧有AI的听写工具。现在是一个碰巧从听写开始的平台。
微软在垂直领域运行同一套playbook：文档→copilot→平台→Agent市场。医疗用Dragon。开发用GitHub Copilot。业务运营用Dynamics Copilot。下一个逻辑领域是知识管理——你的领地。
预期SharePoint Copilot将从"搜索和总结"进化为第三方Agent处理特定知识任务的平台：合规检查、内容治理、自动发布流程。
开始以Agent思维思考，而非功能。你的设计工作应预见SharePoint托管多个AI代理的世界——每个处理知识工作流的一个切面。设计挑战变成编排：用户如何理解每个Agent的功能、信任其输出、管理Agent之间的冲突？
这是定义职业的设计问题。多Agent UX几乎完全是未开垦领域。在这里定义模式的人将塑造企业软件的下一个十年。
Microsoft expanded its AI content safety filters across Copilot products, blocking a wider range of terms and topics. The changes affect Copilot in Edge, Windows, Microsoft 365, and Teams. Several enterprise customers reported that legitimate business queries were being flagged — legal firms researching case law, medical teams discussing treatments, HR departments working on policy documents.
Microsoft is caught between two pressures. Regulators and media want tighter content controls — every AI "hallucination" or inappropriate response makes headlines. But enterprise customers want flexibility — they need AI that can discuss sensitive topics relevant to their industry without getting blocked.
The current approach is blanket filtering with no tenant-level customisation. A pharmaceutical company can't adjust filters to allow drug discussions. A law firm can't enable legal terminology that triggers safety flags. It's a one-size-fits-all approach applied to radically different contexts.
This is a significant UX design challenge that your team should own. The solution isn't just "make filters configurable" — it's designing a system where:
• IT admins can set organisation-wide policies without needing to understand AI safety
• Compliance teams can audit what's being filtered and why
• End users understand when content is filtered (vs. when Copilot genuinely doesn't know)
• The system adapts to context (legal department gets different defaults than marketing)
Flag this internally. Content policy controls will be one of the top enterprise Copilot feature requests in 2026. If your team can propose the design framework before it becomes a crisis, you position as the owner of this problem space.
微软扩大了Copilot产品线的AI内容安全过滤，屏蔽了更多用词和话题。变更影响Edge、Windows、Microsoft 365和Teams中的Copilot。多个企业客户报告合法业务查询被标记——律所研究判例法、医疗团队讨论治疗方案、HR部门制定政策文件。
微软夹在两种压力之间。监管者和媒体要求更严控制——每次AI"幻觉"或不当回应都成头条。但企业客户要求灵活性——他们需要AI能讨论与行业相关的敏感话题而不被屏蔽。
当前做法是无租户级自定义的统一过滤。制药公司不能调整过滤以允许药物讨论。律所不能启用触发安全标记的法律术语。用一刀切方法应对截然不同的语境。
这是你团队应该主导的重大UX设计挑战。解决方案不仅是"让过滤可配置"——而是设计一个系统：
• IT管理员可设置组织级策略而无需理解AI安全
• 合规团队可审计过滤内容及原因
• 终端用户理解何时内容被过滤（vs. Copilot真的不知道）
• 系统根据上下文适应（法务部门默认与营销部门不同）
在内部提出这个问题。内容策略控制将成为2026年企业Copilot排名前列的功能需求。如果你的团队能在危机到来前提出设计框架，就能成为这个问题领域的负责人。
Five Chinese companies released new AI models in the first week of March: Baidu (ERNIE 5.0), Alibaba (Qwen 2.5), ByteDance (Doubao Pro), Zhipu AI (GLM-5), and MiniMax (abab7). Each claims significant improvements over their previous versions in reasoning, code generation, and Chinese language understanding.
UBS published a note ranking Alibaba's Qwen 2.5 above DeepSeek for enterprise applications, citing better structured output and API reliability. DeepSeek remains the consumer favourite for conversational use.
Unlike the US where OpenAI and Anthropic dominate, China's model landscape is genuinely competitive. No single player has more than 20% market share. This is partly regulatory (Beijing prefers distributed capability) and partly commercial (each tech giant has different distribution channels — Alibaba through cloud, ByteDance through consumer apps, Baidu through search).
The practical result: Chinese developers have real choice and leverage. They can switch models based on cost, performance, and feature needs. This is healthy for the ecosystem and keeps prices low.
If Qrio is building for the Chinese market, these models — not GPT-4 or Claude — are the integration targets. Specifically:
• Qwen 2.5 is the strongest for structured educational content — it handles curriculum-aligned generation well
• Doubao Pro has the best voice/audio capabilities — relevant if Qrio adds voice interaction for kids
• MiniMax abab7 has the most creative/playful text generation — fits Qrio's creative learning positioning
Strategic recommendation: Design Qrio's AI layer to be model-agnostic from day one. The Chinese model landscape will keep shifting every quarter. Lock-in to one provider is a strategic mistake. Build an abstraction layer that lets you swap models per feature.
五家中国公司在3月第一周发布新AI模型：百度(ERNIE 5.0)、阿里(通义千问2.5)、字节(豆包Pro)、智谱AI(GLM-5)、MiniMax(abab7)。各家都声称在推理、代码生成和中文理解方面显著优于前版。
瑞银发布报告，在企业应用方面将阿里通义千问2.5排在DeepSeek之上，理由是更好的结构化输出和API稳定性。DeepSeek在对话式消费领域仍是首选。
不同于美国OpenAI和Anthropic主导，中国模型市场真正有竞争。没有单一玩家占超过20%份额。部分因为监管（北京倾向分布式能力），部分因为商业（各科技巨头有不同分发渠道——阿里通过云、字节通过消费应用、百度通过搜索）。
实际结果：中国开发者有真正的选择权和议价力。可根据成本、性能和功能需求切换模型。这对生态健康且保持低价。
如果Qrio面向中国市场，这些模型——而非GPT-4或Claude——才是集成目标。具体来说：
• 通义千问2.5在结构化教育内容方面最强——课程对齐的内容生成表现好
• 豆包Pro有最好的语音/音频能力——如果Qrio为儿童添加语音交互则相关
• MiniMax abab7有最具创意/趣味的文本生成——契合Qrio创意学习定位
战略建议：从第一天起将Qrio的AI层设计为模型无关。中国模型格局每个季度都在变。锁定单一供应商是战略失误。建一个抽象层，让你可以按功能切换模型。
China's annual parliamentary sessions (全国人大 and 政协) kicked off in Beijing with a clear technology focus. Premier Li Qiang's government work report outlined AI, quantum computing, space technology, and advanced robotics as the "new productive forces" that will define China's next economic chapter.
Key numbers: GDP growth target of 5% (same as last year, widely seen as aspirational), defence spending up 7.2%, and a new ¥100 billion ($14B) AI development fund for provincial governments to deploy.
The emphasis on "new productive forces" (新质生产力) is code for pivoting away from real estate and construction as growth engines. Beijing is signalling that the old model of infrastructure-driven growth is over, and the future is technology-led.
For consumer spending: the report acknowledged "challenges in domestic demand" — political-speak for the fact that Chinese consumers are saving, not spending. No major consumer stimulus was announced, disappointing markets that expected consumption vouchers or tax cuts.
AI policy tailwind: AI in education is explicitly in the "encouraged" category. This means no regulatory crackdown risk for Qrio — unlike gaming (heavily restricted) or private tutoring (effectively banned under "double reduction" 双减 policies).
But cautious spending limits the market: If Chinese parents are saving more and spending less, pricing strategy matters enormously for Qrio. Premium pricing will face resistance. Consider freemium models or partnerships with schools (where government budgets, not parent wallets, pay).
For HK equities: The lack of consumer stimulus disappointed bulls. The HSI needs either tariff resolution or domestic stimulus to break upward — it got neither from the parliament sessions.
中国两会在北京开幕，科技成为明确焦点。李强总理的政府工作报告将AI、量子计算、航天科技和先进机器人列为定义中国下一个经济篇章的"新质生产力"。
关键数据：GDP增长目标5%（与去年相同，被广泛视为理想化目标）、国防开支增长7.2%、新设1000亿元AI发展基金供各省部署。
"新质生产力"的强调实际是从房地产和建筑业作为增长引擎转型的信号。北京表明基础设施驱动增长的旧模式结束，未来是技术引领。
消费方面：报告承认"内需面临挑战"——政治话语意味着中国消费者在储蓄而非消费。没有宣布重大消费刺激，令期待消费券或减税的市场失望。
AI政策利好：教育AI明确属于"鼓励"类别。意味着Qrio无监管打压风险——不同于游戏（严格限制）或私教辅导（双减政策下实际被禁）。
但谨慎消费限制市场：如果中国父母多存少花，定价策略对Qrio极其重要。高端定价会遇到阻力。考虑免费增值模式或与学校合作（政府预算而非家长钱包买单）。
对港股：缺少消费刺激令多头失望。恒指需要关税解决或国内刺激才能向上突破——两会两样都没给。
AI startups are reaching $1 billion valuations faster than any category in venture history. The median time from founding to unicorn status for 2025-2026 AI companies is under 24 months — compared to 7 years for the 2010s cohort of cloud/SaaS companies.
Several examples from this week: Cognition (Devin AI) hit $2B at 18 months old. Harvey AI (legal) reached $1.5B in under 2 years. Glean (enterprise search) doubled its valuation in 6 months to $4.6B.
Three factors. First, revenue ramps are genuinely fast — AI products can charge per-use from day one, unlike freemium SaaS that took years to monetise. Second, the market is enormous — every company wants AI and most don't have in-house capability. Third, FOMO among investors — missing the next OpenAI is a career-ending mistake for a VC, so they bid aggressively.
This velocity historically precedes a correction. Not a crash — a correction. The 2021 fintech boom saw similar speed, followed by a 40-60% valuation reset in 2022-23. The companies that survived (Stripe, Plaid) are now worth more than ever. The ones that were pure hype (most crypto projects) are gone.
The AI cohort will follow the same pattern: real companies with real revenue will emerge stronger from a correction. Hype-driven companies with no unit economics will struggle to raise follow-on rounds.
If considering leaving Microsoft: The fundraising climate is favourable right now. But "right now" might mean 6-12 months, not indefinitely. If Qrio needs external funding, the optimal window is closing gradually — every quarter the bar rises.
If staying at Microsoft: Understand that many of these unicorns are building features that compete with Microsoft's product suite. Copilot's competitive moat isn't the AI — it's the distribution through 400 million Office users. That distribution advantage gets tested if startups build superior, specialised alternatives.
AI创业公司达到10亿美元估值的速度超过风投历史上任何类别。2025-2026年AI公司从成立到独角兽的中位时间是不到24个月——相比2010年代云/SaaS公司的7年。
本周案例：Cognition(Devin AI)成立18个月估值达20亿美元。Harvey AI(法律)不到2年达15亿。Glean(企业搜索)6个月估值翻倍至46亿。
三个因素。第一，收入增长确实很快——AI产品从第一天就可以按使用收费，不像免费增值SaaS需要数年变现。第二，市场巨大——每家公司都想要AI，大多没有内部能力。第三，投资者FOMO——错过下一个OpenAI对VC是致命失误，所以激进出价。
这种速度历史上通常预示调整。不是崩盘——是调整。2021年金融科技繁荣有类似速度，随后在2022-23年估值回调40-60%。幸存的公司（Stripe、Plaid）现在价值更高。纯炒作的（大多数加密项目）已消失。
AI群体将遵循同样模式：有真实收入的真实公司将从调整中变得更强。无单位经济效益的炒作公司将难以获得后续融资。
如果考虑离开微软：当前融资环境有利。但"当前"可能意味着6-12个月，而非永远。如果Qrio需要外部资金，最优窗口在逐渐关闭——每个季度门槛都在提高。
如果留在微软：理解许多独角兽正在构建与微软产品套件竞争的功能。Copilot的竞争护城河不是AI——是通过4亿Office用户的分发。如果创业公司构建更好的专业化替代品，这个分发优势将受到考验。
Intuit launched a construction-specific AI platform this week — not a generic chatbot with a construction template, but a purpose-built system that understands project timelines, material pricing, permit workflows, and subcontractor management. It took 2 years and $200M to build.
The results so far: construction firms using it report 40% faster project estimation and 25% fewer change orders. Generic AI tools (ChatGPT, Gemini) applied to the same tasks achieved roughly 10% improvement — four times worse.
Horizontal AI is commoditising fast. GPT-4 and Claude are essentially interchangeable for most general tasks. The premium — and the defensible margin — goes to vertical tools that encode domain expertise.
What makes vertical AI defensible: domain-specific training data (construction documents, medical records, legal briefs), workflow integration (plugging into existing industry tools), and regulatory compliance (meeting industry-specific requirements that generic tools can't).
Qrio's bet on creative learning for young children is a vertical play in a market where everyone else is building horizontal. Generic AI tutors (Khanmigo, Duolingo AI) handle convergent tasks well — math problems, language drills, test prep. But they're terrible at divergent thinking — imagination, creative storytelling, open-ended exploration.
That's not a limitation that improves with better models. It's a product design problem. How do you design an AI experience that encourages kids to explore rather than converge? That's Qrio's moat — and it requires design leadership, not just better AI.
At Microsoft: This pattern validates the move from "Copilot does everything" to "Copilots specialised per domain." Your role designing knowledge management Copilot is exactly this vertical play within Microsoft.
For Qrio: Don't compete on model quality — compete on experience quality. The AI layer should be commodity (swappable models). The design layer — the interactions, the creative scaffolding, the child psychology — that's where value accrues.
Intuit本周推出了建筑业专用AI平台——不是套了建筑模板的通用聊天机器人，而是理解项目时间线、材料定价、许可流程和分包管理的专用系统。花了2年2亿美元打造。
目前结果：使用它的建筑公司报告估算速度提升40%，变更单减少25%。通用AI工具（ChatGPT、Gemini）应用于相同任务只提升约10%——差四倍。
通用AI正快速商品化。GPT-4和Claude在大多数通用任务上基本可互换。溢价——和可防御的利润率——属于编码了领域专业知识的垂直工具。
垂直AI可防御的原因：领域专用训练数据（建筑文件、医疗记录、法律摘要）、工作流集成（接入现有行业工具）、监管合规（满足通用工具无法满足的行业特定要求）。
Qrio在儿童创意学习上的押注是在所有人都做通用的市场中的垂直打法。通用AI辅导（Khanmigo、Duolingo AI）处理收敛任务很好——数学题、语言练习、考试准备。但在发散思维方面很差——想象力、创意故事、开放式探索。
这不是更好模型能改善的局限。这是产品设计问题。如何设计一个鼓励孩子探索而非收敛的AI体验？这是Qrio的护城河——需要设计领导力，不仅是更好的AI。
在微软：这个模式验证了从"Copilot做一切"到"按领域专业化Copilot"的转变。你设计知识管理Copilot的角色正是微软内部的垂直打法。
对Qrio：不要在模型质量上竞争——在体验质量上竞争。AI层应该是商品化的（可切换模型）。设计层——交互、创意脚手架、儿童心理学——才是价值积累之处。
JP Morgan's Asia equity strategists published their Q2 outlook with an unusually blunt headline: "Tariffs are far more important for Asian exports than most investors appreciate." Their analysis shows that a sustained 10% US tariff on Asian goods would reduce regional export growth by 3-4 percentage points — enough to tip several economies into export recession.
JP Morgan models three scenarios:
• Resolution (40% probability): Tariffs reduced to 2-3% through negotiation. Asia ex-Japan equities rally 15-20% from current levels. HSI target: 24,000.
• Status quo (35% probability): 10% tariffs remain but no escalation. Markets range-bound. HSI stays 18,000-20,000.
• Escalation (25% probability): Retaliatory spiral pushes effective tariffs to 20%+. Asia equities drop 10-15%. HSI could test 15,000.
Key dates to watch:
• March 18-19: Fed meeting — rate decision signals how seriously they take the jobs/oil double hit
• Late April: Expected US-China summit — the main tariff negotiation event
• June 1: Canada/EU retaliatory tariff deadline — escalation risk if not resolved
Your HK portfolio is a tariff trade whether you want it to be or not. The correlation between tariff headlines and HSI movements is running at 0.85 — higher than any fundamental factor including earnings or GDP.
Practical action: Don't rebalance until after the April summit. If you must act, favour companies with domestic China revenue (less tariff-sensitive) over exporters. In HK, that means tech platforms (Tencent, Alibaba) over manufacturers (Li & Fung, Techtronic).
The contrarian opportunity: If tariffs resolve, the snapback rally in beaten-down export stocks will be violent and fast — 20-30% in weeks. Consider building a small watchlist of high-quality exporters at distressed valuations, ready to buy on a resolution headline.
摩根大通亚洲股票策略师发布Q2展望，标题异常直白："关税对亚洲出口的重要性远超多数投资者的认知。"分析显示，美国对亚洲商品持续征收10%关税将使区域出口增长降低3-4个百分点——足以让多个经济体进入出口衰退。
摩根大通模拟三种场景：
• 解决(40%概率)：通过谈判将关税降至2-3%。亚洲（日本除外）股票从当前水平反弹15-20%。恒指目标：24,000。
• 维持现状(35%概率)：10%关税维持但不升级。市场区间震荡。恒指在18,000-20,000之间。
• 升级(25%概率)：报复性螺旋将有效关税推至20%以上。亚洲股票下跌10-15%。恒指可能测试15,000。
关注日期：
• 3月18-19日：美联储会议——利率决定反映对就业/石油双重冲击的重视程度
• 4月下旬：预期中美峰会——主要关税谈判事件
• 6月1日：加拿大/欧盟报复性关税截止日——未解决则升级风险
你的港股组合就是关税交易，无论你是否愿意。关税头条与恒指走势的相关性达0.85——高于任何基本面因素包括盈利或GDP。
实际操作：4月峰会前不要调仓。如必须行动，偏好国内中国收入（关税敏感度低）的公司而非出口商。在港股，意味着科技平台（腾讯、阿里）而非制造商（利丰、创科）。
逆向机会：如果关税解决，被打压的出口股反弹将猛烈且迅速——数周内20-30%。考虑建一个困境估值下高质量出口商的小型观察名单，准备在解决消息出来时买入。
Two major announcements landed this week. Apple confirmed its "Visual Intelligence" features for smart glasses — a lightweight AR platform focused on real-time object recognition, navigation overlays, and contextual information display. Meanwhile, OpenAI is working with a hardware partner (widely reported to be Jony Ive's team) on AI-native glasses that would bring ChatGPT-level intelligence to a wearable form factor.
These join Meta's Ray-Ban AI glasses (already shipping) and Google's rumoured Project Iris revival. Four of the five biggest tech companies are now actively building AR glasses.
Three technologies converged: small enough AI chips (Qualcomm's Snapdragon AR2 Gen 2), good enough displays (MicroLED waveguides with daylight visibility), and AI models small enough to run locally (sub-7B parameter models running at 30 tokens/sec on-device). None of these existed at consumer-viable cost 18 months ago.
If AR glasses go mainstream — and "mainstream" might mean 2028-2030 — information consumption shifts from pull to push. Today, you navigate to a SharePoint page to find information. In an AR world, relevant information appears in your visual field based on context — where you are, what you're looking at, who you're meeting with.
This is a fundamental redesign of knowledge management. Instead of organising information into pages, sites, and folders, you'd organise it by context triggers: location, activity, people, time. The entire information architecture paradigm changes.
Don't design for glasses today — the platform is 3-4 years from mainstream adoption. But start thinking about context-triggered information delivery. If your SharePoint knowledge base could surface the right document at the right moment without the user searching — that's the AR mental model, achievable today through Copilot.
For your career: Designers who understand spatial computing and contextual information design will be extremely scarce and valuable by 2028. Start building the conceptual framework now, even if the hardware isn't ready. When it arrives, you'll be years ahead.
本周两个重要公告。苹果确认了"视觉智能"智能眼镜功能——轻量级AR平台，专注实时物体识别、导航覆盖和上下文信息显示。同时OpenAI正与一个硬件伙伴（广泛报道为Jony Ive团队）合作开发AI原生眼镜，将ChatGPT级智能带入可穿戴形态。
加上Meta的Ray-Ban AI眼镜（已发售）和Google传闻中的Project Iris复活。五大科技公司中的四家正在积极构建AR眼镜。
三项技术汇聚：足够小的AI芯片（高通骁龙AR2 Gen 2）、足够好的显示（日光可见的MicroLED波导）、足够小的本地AI模型（低于70亿参数的模型在设备端以30 tokens/秒运行）。18个月前这些都不存在于消费者可承受的成本。
如果AR眼镜成为主流——"主流"可能意味着2028-2030——信息消费从拉取转为推送。今天你导航到SharePoint页面找信息。在AR世界，相关信息根据上下文出现在视野中——你在哪里、看什么、与谁会面。
这是知识管理的根本重新设计。不再按页面、站点和文件夹组织信息，而是按上下文触发器组织：位置、活动、人物、时间。整个信息架构范式改变。
今天不要为眼镜设计——该平台距主流采用还有3-4年。但开始思考上下文触发的信息传递。如果你的SharePoint知识库能在正确时刻推送正确文档而无需用户搜索——这就是AR思维模型，今天通过Copilot可实现。
对你的职业：理解空间计算和上下文信息设计的设计师到2028年将极度稀缺和有价值。现在开始构建概念框架，即使硬件还没准备好。当它到来时，你将领先数年。
Moganshan (莫干山) is entering its best month. Bamboo forests are lush, wildflowers blooming along the mountain trails, and temperatures are comfortable for hiking (12-18°C).
Several new boutique stays have opened along the western trail — getting strong reviews on Xiaohongshu for design-forward interiors and valley views. Book weekend slots now; by April Golden Week it'll be packed.
Getting there: Drive or take the high-speed rail to Deqing (德清), then 30min taxi. Friday evening departure recommended to maximise Saturday morning hiking.
莫干山正在进入最佳月份。竹林茂密，山间步道野花盛开，温度适合徒步（12-18°C）。
西侧步道旁新开了几家精品民宿——在小红书上因设计感内饰和山谷景观获得好评。现在订周末，四月黄金周就满了。
交通：自驾或高铁到德清，再打车30分钟。建议周五晚出发，最大化周六早晨徒步时间。
teamLab's permanent Borderless museum has opened at West Bund, replacing the Tokyo original that closed. The Shanghai venue is the largest teamLab installation globally.
Best experienced on weekday afternoons (significantly less crowded). Allow 2-3 hours. The infinity mirror rooms have separate queues — head there first.
Tickets: ¥299 weekday / ¥399 weekend. Book on their WeChat mini-program at least a week ahead for weekend slots.
teamLab永久无界美术馆在西岸开幕，取代已关闭的东京原馆。上海场馆是全球最大的teamLab装置。
最佳体验时间是工作日下午（人少很多）。预留2-3小时。无限镜屋有单独排队——先去那里。
门票：工作日¥299/周末¥399。周末票至少提前一周在微信小程序预订。
Bird — one of Shanghai's best cocktail bars — has opened a second location in Jing'an. The original on Yongkang Lu has been a staple for years; the new spot keeps the same intimate, no-frills energy.
The new menu leans mezcal-forward with several agave-based cocktails that aren't available at the original. Walk-ins only, no reservations. Best on weeknights.
Location: Near Jing'an Temple station. Look for the small bird sign — it's easy to miss.
Bird——上海最好的鸡尾酒吧之一——在静安开了第二家店。永康路原店多年来一直是标杆；新店保持同样亲密、不做作的氛围。
新酒单偏梅斯卡尔风格，有几款原店没有的龙舌兰基酒。仅接受walk-in，不预约。工作日晚最佳。
位置：静安寺站附近。找小鸟标志——容易错过。
Shanghai Fashion Week's Spring/Summer showcase is running through March 14. The main runway shows require invites, but several designer pop-ups around Xintiandi are open to everyone.
Worth checking out even if fashion isn't your primary interest — the pop-ups often feature interesting product design, spatial design, and brand experience work that's relevant from a design leadership perspective.
Best day: Saturday afternoon. The Xintiandi area comes alive with installations and street-style crowds.
上海时装周春夏展持续到3月14日。主秀场需邀请，但新天地周边多个设计师快闪对公众开放。
即使时尚不是主要兴趣也值得一看——快闪活动常有有趣的产品设计、空间设计和品牌体验，从设计领导力角度很有参考价值。
最佳时间：周六下午。新天地区域装置艺术和街拍人群让整个区域活跃起来。
A growing number of Chinese parents are independently using AI chatbots to tutor their children — not through schools or official programs, but on their own initiative. DeepSeek, ByteDance's Doubao, and Google's Gemini are the most popular choices, with parents sharing prompting strategies on Xiaohongshu and WeChat groups.
The use cases are overwhelmingly convergent: solving math problems, checking homework answers, explaining textbook concepts, generating practice tests. Parents treat AI as a private tutor replacement — filling the gap left by the 2021 "double reduction" crackdown on after-school tutoring.
Every AI education tool on the market is optimising for the same thing: better test scores. This is understandable — it's measurable, parents want it, and AI is genuinely good at it. But it creates a market where every product looks the same.
The result: AI tutoring is already commoditising. DeepSeek is free. Doubao is free. Kimi is free. When the product is free and interchangeable, there's no business model — only a race to the bottom on compute costs.
Qrio's focus on creative learning — divergent thinking, imagination, open-ended play — is the opposite of what every other player is building. This is either a brilliant differentiation or a market-reading error. The evidence suggests differentiation:
• Chinese parents increasingly worry about "exam robots" — kids who score well but can't think creatively. This concern is rising on Xiaohongshu and parenting forums.
• International school enrollment in Shanghai and Beijing is growing 15% annually — parents paying premium for creativity-focused education.
• The government's own education reform rhetoric emphasises "comprehensive quality" (综合素质) over pure academics — even if the gaokao system hasn't changed.
Position Qrio as a complement, not competitor, to AI tutoring. The messaging should be: "Your child has an AI tutor for homework. Now they need an AI creative companion for everything else." Don't fight the convergent AI trend — ride alongside it.
越来越多中国家长自发使用AI聊天机器人辅导孩子——不是通过学校或官方项目，而是自主行动。DeepSeek、字节豆包和Google Gemini是最受欢迎的选择，家长在小红书和微信群分享提示词策略。
使用场景主要是收敛型：解数学题、检查作业答案、解释教材概念、生成练习试卷。家长将AI当作私教替代品——填补2021年"双减"打压课后辅导留下的空白。
市场上每个AI教育工具都在优化同一件事：更好的考试成绩。这可以理解——可衡量、家长想要、AI确实擅长。但创造了每个产品看起来一样的市场。
结果：AI辅导已经在商品化。DeepSeek免费。豆包免费。Kimi免费。当产品免费且可互换，没有商业模式——只有算力成本竞底赛。
Qrio专注创意学习——发散思维、想象力、开放式游戏——与其他所有玩家构建的方向相反。这要么是出色的差异化，要么是市场误判。证据支持差异化：
• 中国家长越来越担忧"做题机器"——成绩好但不会创造性思考的孩子。这种担忧在小红书和育儿论坛上升。
• 上海和北京的国际学校入学率年增长15%——家长为创意教育支付溢价。
• 政府自身的教育改革论述强调"综合素质"而非纯学术——即使高考制度未变。
将Qrio定位为AI辅导的补充而非竞争者。信息传递应该是："你的孩子有AI辅导做作业了。现在他们需要一个AI创意伙伴做其他一切。"不要对抗收敛AI趋势——与之并行。
China's annual parliament sessions explicitly included AI integration across education as a priority under the "new productive forces" umbrella. The Ministry of Education's supplementary briefing went further: they outlined plans for AI teaching assistants in 50,000 primary and secondary schools by 2027, starting with pilot programs in Beijing, Shanghai, Guangdong, and Zhejiang.
Crucially, creative and STEAM education tools are in the "encouraged" category — the same classification that helped robotics education companies thrive over the past five years.
The good news: Government support means no regulatory crackdown risk. Unlike private tutoring (双减) or gaming (version number restrictions), AI creative education is aligned with stated policy goals. This is a rare green light in China's education market.
The bad news: Policy support attracts big players. ByteDance is reportedly building an AI education product for its Doubao platform. Tencent has hired a 200-person team for "AI-native education." Alibaba's DingTalk is integrating Qwen into classroom tools. When the government signals "encouraged," the giants mobilise.
The window for a startup to establish position before the giants arrive is 12-18 months. After that, ByteDance and Tencent will have products in market with distribution advantages that no startup can match (hundreds of millions of existing users on their platforms).
What a startup can do that giants can't: move fast, take design risks, build deeply for a niche. ByteDance will build for the mass market (test prep). Tencent will build for the social market (collaborative learning). The creative niche — divergent thinking, imagination — is too small for giants to prioritise but large enough for a focused startup to own.
Speed to market matters more than product perfection. Launch an MVP that's 70% right and iterate, rather than spending 18 months building something perfect only to find ByteDance has occupied the market.
Government partnerships: The 50,000-school pilot program is a potential distribution channel. If Qrio can get into the pilot cohort, it gains instant credibility and scale that would take years to build independently. Explore this path aggressively.
两会明确将AI融入教育列为"新质生产力"框架下的优先事项。教育部补充简报更进一步：计划到2027年在5万所中小学部署AI教学助手，从北京、上海、广东、浙江的试点开始。
关键是，创意和STEAM教育工具属于"鼓励"类别——与过去五年帮助机器人教育公司蓬勃发展的分类相同。
好消息：政府支持意味着无监管打压风险。不同于私教辅导（双减）或游戏（版号限制），AI创意教育与既定政策目标一致。这在中国教育市场是罕见的绿灯。
坏消息：政策支持吸引大玩家。字节据报为豆包平台构建AI教育产品。腾讯已招聘200人团队做"AI原生教育"。阿里钉钉正在将通义千问集成到课堂工具。当政府发出"鼓励"信号，巨头就会动员。
创业公司在巨头到来前建立位置的窗口是12-18个月。之后字节和腾讯将有产品上市，带着创业公司无法匹敌的分发优势（平台上数亿现有用户）。
创业公司能做而巨头不能的：快速行动、承担设计风险、为细分市场深度构建。字节会为大众市场构建（应试）。腾讯会为社交市场构建（协作学习）。创意细分——发散思维、想象力——对巨头太小不值得优先，但对聚焦的创业公司足够大可以占领。
上市速度比产品完美更重要。推出70%正确的MVP然后迭代，而不是花18个月构建完美产品却发现字节已占领市场。
政府合作：5万所学校试点计划是潜在分发渠道。如果Qrio能进入试点，将获得独立构建需要数年的即时信誉和规模。积极探索这条路。