全球互联网：不要在对手犯错时打断他

Global Internet: Never interrupt your opponent while he ismaking a mistake A short, fictitious story designed to teach a moral lesson.To paraphrase benchmarkoperators Artificial Analysis, the launch and subsequent un-launch of Anthropic’s ClaudeFable 5 (ostensibly by order from the US Department of Commerce, on national securitygrounds) was the first time that the AI frontier moved backwards. Our base case is thedecision will be reversed before too long, and Anthropic retakes its position at the top ofvarious charts. But the weekend’s news is worth thinking about beyond benchmark scores. Robin Zhu+852 2123 2659robin.zhu@bernsteinsg.com Mark Shmulik+1 917 344 8508mark.shmulik@bernsteinsg.com To err is human; to really foul things up requires a computer.We assume the reasonsfor the about-face are legitimate, but can’t help but wonder how Anthropic now feels aboutcalling Mythos “too dangerous to be released to the public”, or about its blog post arguingfor a coordinated slowdown in global AI research. The developer furor around ClaudeFable 5 throttling performance on tasks related to LLM research was partly doused whenAnthropic removed the throttle over the weekend, but the irony of Andrej Karpathy (AIpioneer, recent Anthropic hire) and Amanda Askell (head of Anthropic personality alignment)becoming technically barred from using Claude Fable 5 wasn’t lost on the community. Charles Gou+852 2123 2618charles.gou@bernsteinsg.com Min-Joo Kang+852 2123 2644minjoo.kang@bernsteinsg.com Hyrum Caesar+81 3 6777 6979hyrum.caesar@bernsteinsg.com A strong argument for sovereign AI.Anyone who uses AI models regularly will be familiarwith issues like server congestion and rate limits. But the notion that a frontier AI modelcan simply be switched off on a whim from across the pond - no matter the reason - feelsmore like something out of a Bond movie. As agentic AI capabilities improve, these tools willbecome embedded in both existing services and new workflows. As these services becomeincreasingly mission-critical, it would surprise us if large enterprise and sovereign operatorsoutside the US didn’t start asking uncomfortable questions about the dependability ofcloud-deployed AI models where access is beyond their immediate control. AI never sleeps.It probably wasn’t an accident that Kimi and Z.ai, two AI labs we considerfrontier in China, announced new models over the weekend. K2.7 Code promised betterreasoning efficiency and reasoning capability upgrades. Z.ai’s GLM-5.2 meanwhilepositions itself against Claude Opus 4.8; both online developer feedback and our ownlocally-hosted benchmark tests returned encouraging initial results. Real counter-arguments remain, but on the margin these releases support the idea that China’s leadinglabs can continue to keep pace with global peers. The idea that Chinese open-weightsmodels might now occupy the rhetorical high ground too is a novel development. AI in an era of geostrategic competition.What happens next to Claude Fable 5?Anthropic called the episode a misunderstanding, and there’s a decent chance the USadministration lives up to the Street’s favourite four-letter acronym… again. That said,an interesting precedent was set that’s probably not bullish for the US SOTA labs’ globalaspirations in use cases where dependability is important… or stock market narrativescontingent on their upcoming IPOs and continued success. When Chinese regulators shookup their Internet sector a few years ago, valuation multiples investors were willing to acceptcompressed dramatically. INVESTMENT IMPLICATIONS We think administration’s decision to restrict frontier model access by nationality serve as a reminder of the potential for(geo)political narratives to intersect with AI development, especially as frontier AI capabilities become ever more powerful. BERNSTEIN TICKER TABLE I. REQUIRED DISCLOSURES References to "Bernstein" or the “Firm” in these disclosures relate to the following entities: Bernstein Institutional Services LLC(April 1, 2024 onwards), Sanford C. Bernstein & Co., LLC (pre April 1, 2024), Bernstein Autonomous LLP, BSG France S.A. (April 1,2024 onwards), Sanford C. Bernstein (Hong Kong) Limited盛博香港有限公司,Sanford C. Bernstein (Canada) Limited, SanfordC. Bernstein (India) Private Limited (SEBI registration no. INH000006378), Sanford C. Bernstein (Singapore) Private Limited,Sanford C. Bernstein Japan KK(サンフォード・C・バーンスタイン株式会社)and analysts employed by Société GénéraleAfrica Technologies & Services to produce Bernstein research under a Global Services Agreement in place between Bernsteinand Société Générale. Bernstein is part of a joint venture between Société Générale (SG) and AllianceBernstein, L.P. (AB). Unless specifically notedotherwise, for purposes of these disclosures, references to Bernstein’s “affiliates” relate to both SG and AB and their respectiveaffiliates. VALUATION METHODOLOGY Tencent Holdings Ltd We value Tencent at HK$780 per share, on a 20x FY+1 (i.e. Quarters 5-8