Pinduoduo’s domestic discount mechanic survived ten years of attempted Western replication because its viral copy was untranslatable, not because the unit economics were unique. 36kr’s 2026 cross-border survey flags AI translation tools as an efficiency lever for Chinese platforms entering new markets, and a Sichuan University benchmark published in May confirms that current LLMs can define Chinese internet slang from social-media context, including the family of viral imperatives that built Pinduoduo. The most consequential of those imperatives is “砍一刀” (kǎn yī dāo, literally “cut one knife”), the e-commerce verb that proved hardest to carry intact into English copy. If the slang now travels intact, the next platform launching into Jakarta, São Paulo, or Warsaw does not need a year of localizer headcount before it can ask a stranger for a discount in their own dialect.

The moat was a phrase, not a feature. “砍一刀” means “help me slash one” — a request that turns a checkout into a friend-graph activation, and Pinduoduo built it into the only e-commerce verb mainland internet users recognize at a glance. Temu launched into the US in 2022 with the slogan “Team up, Price down,” converting the original imperative into a noun pair a US compliance team could clear. The slogan is a different speech act from the source. The risk was never purely regulatory; it was that no localizer could phrase the appeal without flattening it, and the appeal was the appeal.

Recent benchmark work confirms the inflection. The CHEER buzzword paper released in May 2025 tests LLM performance on Chinese internet buzzwords drawn from social media. Current best methods outperform baselines, though the authors flag underdeveloped inferential abilities as a persisting challenge alongside over-reliance on prior exposure. That improvement matters: a model that gains ground through social-context inference, even without closing the gap, gives vendors a starting capability that Temu’s 2022 launch team lacked when they cut the phrase. Vendors no longer wait for slang to enter Mandarin dictionaries before they can attempt a rendering in another language; they can read the WeChat thread that incubated it.

The translation problem was always asymmetric: Pinduoduo built one phrase that hooked the world’s largest social graph, while ten years of US growth engineers could not write a single English imperative that hooked Americans the same way.

The next entrant will not face the same trade. Temu’s first US growth team had to choose between a slogan that translated correctly and a slogan that converted; the slogan that converted required a localized imperative no English copywriter could draft and no agency would clear. A platform shipping today can write the imperative directly from a model that has read “砍一刀” in tens of thousands of contexts and can produce its functional analogue in Bahasa, Portuguese, or Polish. AI-generated multilingual copy is already in production across Chinese cross-border merchants. What the new benchmarks signal is that the upstream viral copy now sits within range of the same systems. The translation tax that bought Temu’s English-speaking competitors three years of lead-time has compressed toward zero.

The strongest objection is that translation was never the constraint; the social graph was. By this account, “砍一刀” worked in China not because the language had a verb for it but because WeChat collapses payment, commerce, and friend-tier into one thread, and no LLM produces a WhatsApp-Messenger equivalent of that graph. Give the objection its best form and the LLM thesis still holds: 36kr’s own reporting captures US Temu users circulating referral links on WhatsApp and Reddit using the 砍一刀 mechanic in its functional form — a Chinese-American user in the piece recognizes it immediately as the same format from WeChat. The mechanic crossed cultures even when the copy stayed watered down, which means the residual gap is verbal, not structural. What the model removes is the last asymmetry between the team that designs the funnel and the team writing its copy in the local imperative.

The price of the translation moat closing is paid by the localizer, not the platform. A category that paid agencies for years of glossary work, in-market focus groups, and bilingual culture hires can now route those line items into a prompt and a model card. If the CHEER curve holds, fashion retailers building cross-border arms, including Shein’s white-label tier and every brand that staffed a Mandarin team to read viral copy, will watch that budget consolidate into procurement. The next platform launching from Shenzhen will not need to learn how Americans ask each other for favors. It will let the model write the ask, and the verb that built Pinduoduo will arrive in English without anyone noticing it was translated.