DeepSeek founder Liang Wenfeng is among 130+ authors on a new paper detailing DSec, the elastic compute sandbox platform behind the company's agent RL training. DSec serves about 3 million sandboxes per day, with peak concurrency over 380,000 and creation speeds above 5,000 per second. One production unit runs roughly 160 CPU nodes, 30,000 cores, and 250TB of memory, hosting petabyte-scale images. A single training job can pull up 32,000 sandboxes at once. All RL training and evaluation from V3.2 through V4.1 ran on DSec. The paper is dated September 19. Our read: DeepSeek is showing the infrastructure moat behind its agent push, not just model weights. This is scale most labs won't replicate. h/t @WallStreetCN

DeepSeek Nears a CFO Hire as Its IPO Clock Starts DeepSeek's CFO search has narrowed to Yan Wentao, a partner at Hillhouse Venture Capital, who is already working through his exit, per Phoenix Technology, cited by ITHome on September 14. Sources say Liang Wenfeng also spoke with principals from 5Y Capital, Sequoia China, and Longzhu Capital about the seat. Several industry people read those conversations as less about the job than about access to a potential investment. Liang reportedly wants the hire to be born in the 1990s. Yan was born in 1991 and studied at Fudan. The consensus read is that this is a routine pre-IPO hire. The composition of the candidate pool says something narrower: DeepSeek is recruiting from the venture side, not from investment banks or listed-company finance departments. Yan's track record includes ByteDance, Zhipu, MiniMax, Xiaohongshu, Webull, and J&T. That is a portfolio of growth-stage bets, not a career of quarterly reporting and disclosure controls. To us, the CFO seat is being filled for the capital-markets phase, not the operating one. DeepSeek closed its first external round in June at roughly RMB 51 billion (~$7B), about $7.2 billion. Reuters reported on September 9 that CITIC Securities is preparing a STAR Market listing, with the process to start within the year. Fundraising size and target valuation are still undetermined. A venture partner's network is useful for the shareholder list and the next round. It is less obviously the skill set a newly public company needs. That gap is the one to watch once the listing actually starts. Yan's hire is not confirmed, and DeepSeek has not commented.

Moore Threads says its MTT S5000 cards and MUSA stack hit Day-0 support for Zhipu's open-source GLM-5.3-Flash (320B-A18B), the first native multimodal model in the GLM-5 series. The model scores 57 on the AA index, matching Claude Opus 4.8. Our read: this is about the domestic stack closing the loop. Zhipu ships a frontier-adjacent model, and a Chinese GPU vendor keeps pace on day one. $688795.SS