DD-滴滴

Kimi’s latest model, K3, has climbed to the top of several key benchmarks this month.
The 2.8-trillion-parameter open-weight system from Moonshot AI now matches or beats leading closed models from OpenAI and Anthropic on coding and long-context tasks.
That breakthrough has drawn fresh attention to China’s generative AI race.
Moonshot, founded in 2023 by former Meta and Google researcher Yang Zhilin, built its name on the Kimi chatbot. The company moved fast: it released the trillion-parameter K2 series in 2025, then pushed K3 in July 2026 with a million-token context window and native multimodal support. Full weights are due out by late July. Early users and independent tests show strong results in agent workflows and complex coding.
Revenue has kept pace. Annual recurring revenue crossed $200 million in April, with API sales making up more than 70 percent. Growth came from both consumer subscriptions and enterprise deals. Investors noticed. After a $4.3 billion valuation at the end of 2025, Moonshot raised roughly $2 billion in May at over $20 billion post-money, led by Meituan’s venture arm. It has now pulled in nearly $4 billion in six months and is reportedly targeting a $30 billion valuation in the next round. Backers include Alibaba, Tencent, and China Mobile.
The company is still private but has floated the idea of a Hong Kong listing. Its story shows how quickly private capital and open-weight models can scale in China when compute constraints push teams toward efficiency.
Two Listed Peers in Hong Kong
Two close rivals have already gone public on the Hong Kong exchange, giving investors direct exposure.
MiniMax focuses on multimodal generation across text, image, video, and audio. Its latest open model, M3, handles a million-token context. The company listed in January 2026 and raised about $619 million. Shares surged on debut but later pulled back sharply. In July it announced plans to raise another $2 billion through new shares and convertible bonds to fund infrastructure and model work. Customer numbers have grown quickly on both the enterprise and consumer sides.
Zhipu AI , known for its GLM models, listed around the same time. GLM-5.2, released and open-sourced earlier this year, ranks among the strongest open models globally, especially in coding and long-horizon tasks. Shares climbed fast after listing, briefly pushing market value above HK$1 trillion before moderating. In July the company raised roughly $4 billion in a share placement and is preparing a possible Shanghai STAR listing. Founder Tang Jie has signaled a long-term push toward advanced agents and AGI-level capabilities rather than quick monetization.
Both stocks have been volatile, typical for early-stage AI names. Zhipu has held up better on model momentum; MiniMax has faced more pressure from dilution and lock-up expirations.
Established Player and Broader Exposure
SenseTime offers a more mature route into the same theme. The company built its business on computer vision and has expanded into generative models through the SenseNova line. Revenue rose 33 percent last year, with generative AI contributing strongly. Analysts generally rate the stock a buy, seeing room for further gains as multimodal tools move into commercial use.
Hong Kong’s broader AI theme also includes exposure through Tencent, Alibaba, and chip names such as SMIC and Biren. Many of these groups have invested in or compete directly with the startups above.
What Wall Street Banks Are Saying
Major banks have started coverage and see clear differentiation among the players.
Goldman Sachs initiated on Zhipu with a target around HK$1,880 and picked Zhipu, DeepSeek, and ByteDance as its top Chinese model names. The bank is constructive on the full China AI value chain — power, chips, infrastructure, and models — noting that global funds remain underweight despite China’s roughly $4 trillion in related market value.
Bank of America started both Zhipu and MiniMax with Buy ratings (targets HK$1,250 and HK$500). It describes a “two-speed” global market: the U.S. leads on raw frontier performance while Chinese models win on cost and efficiency for high-volume work. BofA also flags big long-term spending on Chinese AI data centers.
JPMorgan has been the most bullish on Zhipu, raising its target multiple times to HK$2,000 and higher while keeping an Overweight rating. The bank argues the sector is moving into a “winner-takes-more” phase where the strongest open models convert wide distribution into better monetization. It has been more cautious on MiniMax.
Morgan Stanley has also lifted Zhipu targets. Other houses have issued Buys on MiniMax at times. Common threads across the reports: open weights are accelerating adoption, Chinese teams are optimizing around domestic chips, and enterprise demand is strong even as compute remains tight.
Investment Angles
Direct plays sit in the two pure-play listings. Zhipu currently carries the strongest bank support tied to its enterprise focus and open-model upside. MiniMax offers multimodal leverage but comes with more near-term share supply pressure. SenseTime gives a steadier, vision-heavy alternative with improving margins.
Indirect routes run through Tencent and Alibaba, both investors in several of these companies and developers of their own competing models. Infrastructure names add another layer for those wanting exposure to the compute build-out.
The bullish case rests on cost advantages, open-source momentum, and practical enterprise use cases. Chinese models are gaining traction on platforms outside China and in cost-sensitive markets. Continued model leadership from names like Kimi and GLM could drive further revenue growth.
Risks are straightforward:
Fierce competition inside China, repeated capital raises that dilute shareholders, high valuations, and ongoing limits on advanced chips. Execution on turning open models into sustained profits remains unproven at scale.

羅素2000 $XIWM 指數分析
截至2026年7月中旬
羅素2000指數（Russell 2000）由約2,000家美國小型上市公司組成，是觀察美國小型股表現最常使用的指標之一。市場上最具代表性的相關ETF是 iShares Russell 2000 ETF，代號 IWM。
近期表現
2026年上半年，羅素2000表現相當強勢，累計上漲約19%至22%，創下1991年以來最佳上半年表現。
同期標普500指數約上漲9%，羅素2000明顯跑贏大型股，兩者的上半年報酬差距也是2003年以來最大。
指數在6月底一度來到約3,033至3,046點的歷史高位，之後於7月初出現回檔，目前大致在2,970至3,000點附近震盪。IWM則約在293至300美元區間波動。
這一波上漲的原因
首先是估值修復。
2026年初，美國小型股相對大型股的估值折價，一度接近近30年來最大的水準。當市場資金開始從估值偏高的大型科技股轉向其他板塊，小型股自然成為輪動的主要受益者。
其次，小型公司的營收來源更集中在美國本土。羅素2000成分公司的收入約有77%來自美國，因此相較跨國大型企業，受到美元升值、關稅及海外地緣政治因素的影響較小。
盈利預期上修也是重要推力。市場原先預估羅素2000成分股2026年每股盈餘將增長約23%，之後陸續上調至38%左右。部分增長來自AI基礎設施相關的小型企業，包括晶片供應鏈、電子設備、電力及資料中心周邊公司。
此外，AI行情也開始由大型科技公司向中小型供應鏈擴散。市場不再只追逐少數幾家大型企業，而是逐漸尋找AI資本支出背後的設備、零組件及基礎設施受益者。
美國經濟仍具韌性，加上財政支出及製造業復甦預期，也改善了市場對小型企業營運環境的看法。
不過，6月指數重組後，羅素2000中AI基礎設施相關公司的權重，由約15%降至7%。這代表前期推動指數上漲的重要力量有所減弱，後續能否延續漲勢，仍要看其他產業的盈利能否接上。
投資銀行與機構觀點
高盛：漲勢可能開始降溫
高盛在2026年7月的報告中，對小型股後續表現轉為較謹慎。
策略師Ben Snider認為，羅素2000經過上半年的快速上漲後，估值已不再像年初那麼便宜，預估未來12個月的報酬可能只剩低個位數。
高盛主要擔心幾個問題。
第一，市場已經提前反映不少盈利復甦與估值修復的預期。
第二，美國經濟增長目前接近長期趨勢，未必足以支持小型股持續大幅上漲。
第三，指數重組後，AI相關公司的比重下降，題材帶來的推動效果可能減弱。
另外，羅素2000仍有約四分之一的成分公司處於虧損狀態，未盈利公司的市值占比約23%。一旦盈利不如預期，股價容易出現較大修正。
小型公司也比大型企業更容易受到利率影響。約30%的小型企業債務採用浮動利率，標普500企業則約為7%。假如聯準會維持高利率更久，甚至重新轉向緊縮，小型公司的融資成本會更快上升。
美國銀行：盈利與製造業仍有支撐
美國銀行策略師Jill Hall的看法相對正面。
她認為，羅素2000的領先表現不只是估值反彈，也反映小型企業盈利改善及美國製造業動能回升。
美國銀行同時看好中型股，原因是中型企業通常兼具一定的盈利能力與成長空間，對利率的敏感度又低於小型企業，風險報酬可能更平衡。
其他機構
LPL Financial仍維持羅素2000成分股2026年盈利增長約38%的預估。
Fidelity、T. Rowe Price及Royce等長期研究小型股的機構，整體仍偏向正面。他們認為，小型股相對大型股的估值仍不算昂貴，未來也可能受惠於盈利復甦、併購活動增加及資金重新配置。
從更長的五年至十年來看，小型股仍有機會結束過去長時間落後大型科技股的局面，重新進入相對有利的市場週期。
值得注意的是，高盛及摩根士丹利等機構在2025年底至2026年初時，對小型股的態度也較為樂觀。當時主要看好經濟增長、利率環境改善及市場上漲範圍擴大。隨著羅素2000已經累積可觀漲幅，部分機構才開始調低後續預期。
主要風險
羅素2000目前最大的問題，是成分公司品質差異很大。
部分企業盈利能力穩定，也有不少公司仍在虧損。當市場情緒良好時，投資人通常願意提高風險承受度；但只要經濟或盈利預期轉弱，未盈利企業往往會率先受到拋售。
利率也是不能忽略的風險。小型公司較依賴銀行貸款及浮動利率債務，缺乏大型企業低成本發債的能力，因此高利率環境對小型股的壓力通常更大。
另外，上半年漲幅已經相當可觀，短期出現獲利了結、估值消化或區間整理都很正常。假如美國經濟放緩、通膨重新升溫，或聯準會態度轉鷹，小型股可能比大型股承受更大的下跌壓力。
後續展望
短期來看，羅素2000在快速上漲後，繼續單邊走高的難度已經增加。市場可能需要一段時間整理，等待企業財報及盈利預期確認。
中長期則仍有幾個值得觀察的條件：
美國經濟能否維持溫和增長、企業盈利能否兌現、利率是否逐步回落，以及資金是否繼續由大型科技股流向其他產業。

其實從昨天下午，很多老師大儒們看到溢價就很興奮的叫人套利
然後就是如圖的
從1300點一路被 嘎
為何會這樣呢
不是套利嗎 不是會收斂嗎
先說結論：這個套利在機制上被“單向轉換”限制，直接買入首爾正股、轉換成 ADR 再賣出的路徑並不可行。這也是 SK 海力士 ADR 溢價能夠長期存在的重要原因。
SK 海力士今年 7 月在納斯達克發行保薦 ADR，代碼為 SKHY，發行價 149 美元。上市首日，ADR 相較首爾正股溢價約 16%；本週一溢價一度擴大至約 37%，到 7 月 14 日已收窄至約 19%。因此，市場上看到的 30% 溢價更接近短期高點，實際價差變化很快。
為什麼不能通過“買正股、換 ADR、賣出”鎖定利潤？
關鍵在於這套 ADR 轉換機制並不對稱。
投資者可以將 ADR 轉換為首爾上市的普通股，但普通股能否重新轉換成 ADR，需要獲得公司或存托機構同意。海力士在相關申報文件中也明確提示，普通股可能無法重新轉換為 ADR。
這意味著，投資者即使買入價格較低的首爾正股，也不一定能將其轉換成可在美國市場出售的 ADR。經典 ADR 套利所依賴的雙向轉換閉環，因此無法成立。
另一種思路是進行價差收斂交易，也就是買入首爾正股，同時做空 ADR，押注兩者價差縮小。理論上可行，但並不是無風險套利。
首先，溢價可能長期維持。台積電 ADR 就是典型案例。自 2024 年以來，其 ADR 溢價平均約為 19%，並長期處於較高區間。美國市場對交易便利性、流動性和可及性的需求，本身就可能形成持續溢價。
其次，新上市 ADR 的融券成本可能很高，可借數量也可能有限。一旦市場繼續追高，空頭可能面臨追加保證金、被迫回補甚至逼空風險。此前溢價快速擴大至 37%，已經說明這類價差可能在短期內進一步失衡。
此外，還需要考慮韓元與美元之間的匯率風險、韓美市場交易時間不重疊帶來的隔夜跳空、韓國賣空規定、交易稅以及其他摩擦成本。
真正能夠壓縮溢價的力量，可能不是散戶套利，而是公司增加 ADR 供給。例如海力士未來增發 ADR，或通過其他資本運作提高美國市場流通量，都可能降低稀缺性溢價。市場討論中的股票面額分割，也可能改善交易便利性，但未必能從根本上解決 ADR 供給不足的問題。
因此，押注 SK 海力士 ADR 溢價收斂，本質上是在判斷公司何時擴大 ADR 供給，以及美國市場的需求溢價能否下降，而不是一個可以穩定鎖定利潤的無風險套利機會。
以上內容僅用於梳理交易機制與市場風險，不構成投資建議。此類跨市場價差交易涉及融券、匯率和流動性風險，尤其是空頭一側，潛在虧損可能超過最初預期。
參考資料
The Korea Times：SK hynix ADRs to follow Seoul-listed shares as premium narrows
Financial News：같은 SK하이닉스인데 美에서 16% 더 비싸…‘역김치 프리미엄’ 얼마나 갈까
Insight：액면분할만으론 못 좁힌다…SK하이닉스 ADR 23.7% 프리미엄의 구조