ByteDance kept 56% of its AI drug unit after HSG's $290M bet
Morning on three fronts: a Chinese super-app prices its science arm separately from the ads business, a laptop finally gets enough memory and watts to argue about 300-billion-parameter models, and China starts asking what school is for when translation is free.
ByteDance has closed a $290 million round for Anew Labs, the Shanghai-based AI drug discovery unit it spun out of the group, at a $1.5 billion valuation, according to Reuters, which spoke to two people with knowledge of the matter. The round — Anew Labs' first from outside the parent — was led by HSG (the firm formerly known as Sequoia China), IDG Capital and Hillhouse Investment, with 5Y Capital co-leading; Gaorong Ventures, Primavera Venture Partners, Boyu Capital, strategic investor SBP Group and the state-backed Shanghai Future Industries Fund also took part. ByteDance keeps 56% after the money lands, and one source said the spin-off exists because AI drug discovery "follows a different industry logic and management approach" from the core business. That 56% is the detail that matters: this is not an exit, it is a ring-fence. ByteDance gets a separately priced science asset — with a state-linked fund on the cap table — without giving up control, and the round clears the unicorn mark on the strength of a single deal rather than a product line. Worth holding next to our own reporting on the sector's evidence problem — More molecules, same failures: the crack in AI drug design. A $1.5 billion valuation prices optimism, not clinical outcomes.
HP announced the ZBook Ultra G3a 16, a 16-inch mobile workstation with 192GB of soldered unified memory that it says runs language models up to 300 billion parameters locally. The configuration pairs the AMD Ryzen AI Max+ PRO 495 — the top Strix Halo PRO part, 16 cores, 32 threads — with LPDDR5X-8533 that can hand up to 160GB to the integrated Radeon 8065S GPU, plus a 55 TOPS NPU. The headline spec is not the memory. It is the cooling: HP's new thermal design sustains up to 100W where the previous ZBook held 55W, and StorageReview's preview notes the honest weakness of the last generation was generation speed — large models crawled at single-digit tokens per second inside that 55W envelope. Capacity decides which model loads; watts decide whether it is usable. HP's other bet is software: Perplexity Computer ships on the machine with a connector into Autodesk Revit so a local model can work on project files without uploading them, alongside HP-only task Skills and a local-AI ROI calculator. Availability in October, pricing undisclosed.
A viral post from Tang, a Chinese math teacher whose students have repeatedly scored full marks on state exams, has reopened the argument over whether English should stay a core school subject now that AI translation is good enough. English is taught from third grade, carries full weight on the graduation exam and gates entry to master's programs; proposals circulating include cutting its weighting to 120 points from 150. Tang's case is economic — children start English before they have mastered Chinese, for a subject "completely irrelevant to the majority of people in their future professional lives." The strongest rebuttal came from Hu Xijin, the former Global Times editor-in-chief, who argues English as a core subject is what gives rural children access to the world, and that most scientific publishing still runs through it. The New York Times led with the debate; DW reported the details. The telling part is that neither side is arguing about translation quality. Model capability just moved the question to who bears the cost of a curriculum that no longer pays for itself.
What to watch: whether any provincial education bureau moves the English weighting before the next exam cycle.
If a model can read the paper, hold the meeting and translate the contract — is learning a language still a skill, or just a credential? Tell us in the comments.
