DeepSeek Seeks 50 Billion Yuan Round and Bets on Huawei Chips; OpenAI Forms Math Advisory Group
DeepSeek is finalizing a 50 billion yuan funding round and betting on Huawei AI chips, according to The Information. OpenAI has formed an independent math advisory group after its new model solved more than 100 open problems, while Amazon Bedrock added Kimi K3 with a one-million-token context.
Two people familiar with the matter said DeepSeek expects to receive a batch of new Huawei training chips in the fourth quarter of this year or the first quarter of next year. The company still uses Nvidia chips, and Huawei chip supply is constrained by shortages of advanced memory and other components, the report said. DeepSeek is also training a 2-trillion-parameter model and planning an 8-trillion-parameter model, according to the report.
OpenAI announced the formation of an independent advisory group on mathematics and AI, called AGMAI, based at the Institute for Advanced Study in Princeton. The group has nine leading mathematicians from Cambridge, Harvard, Stanford and other institutions, aiming to bridge academia and technical teams. OpenAI previously disclosed that an internal model trained since Aug. 28 had solved more than 100 open mathematics problems in addition to the Navier-Stokes Millennium Prize problem, exceeding internal expectations and prompting public concern in the mathematics community about negative externalities of AI problem-solving. The group will operate independently and unpaid, may publish its views publicly, and will assess the significance of results, academic norms and tool use, but will not interfere with OpenAI’s internal research and development schedule.
Amazon AWS said on Sept. 18 that Moonshot AI’s Kimi K3 is available on Amazon Bedrock. Developers can call the model through Bedrock Runtime’s OpenAI-compatible Responses and Chat Completions interfaces, as well as the Invoke and Converse APIs. Kimi K3 supports native visual input and a one-million-token context, and is the first open-weight model on Bedrock to support explicit prompt caching. The model is available through US geographic regions and a global cross-region inference configuration. AWS said inference data will not be shared with the model provider or used to train the underlying model, with zero data retention and zero operator access enabled by default.
ZCode said it would open-source its project under the Apache 2.0 license, with source code published in Zhipu’s official GitHub repository. It said ZCode will not store or use user code for training. The announcement also responded to an earlier data security controversy: after checks by CAICT, zcode-prod object storage no longer contains relevant user data; NSFOCUS confirmed that the bucket and related objects had been deleted and that current versions have no path that could trigger local repository snapshots or file exfiltration. ZCode v3.14.0 removed the Repo Wiki function and the entire chain for uploading local repository snapshots.
Oura filed an updated prospectus with the US Securities and Exchange Commission on Sept. 21, planning to sell 50 million shares in an IPO at $40 to $44 each, to raise up to $2.2 billion. At the top of the range, the company would be valued at about $14.1 billion. The company will issue 13.5 million new shares, while existing shareholders will sell the remaining 36.5 million; Oura will not receive proceeds from that portion. Yingli has expressed interest in buying up to $100 million of shares, and a Dragoneer-affiliated fund intends to buy up to $300 million. Oura applied to list on Nasdaq under the ticker OURA. The prospectus shows the company sold 3.6 million Oura Rings in the 12 months ended June 30, and revenue for the nine months ended June 30 was $1.2145 billion, up 74% year over year, with net profit of $60.8 million.