AI News, 21 October: DeepSeek's open engine, SpaceX's AI profit and OpenAI buys Orchestra Bio
1. China: DeepSeek open-sources the inference engine behind its own services
DeepSeek has published DeepInfer on GitHub and describes it as the high-performance inference framework that powers its large-scale production services; the repository does not state a release date. The code is available under the Apache 2.0 open-source licence and runs on NVIDIA's SM100-generation GPUs and on Huawei's Ascend 950 AI chips. It currently supports DeepSeek's V4.1 Flash model and includes features such as speculative decoding, several forms of parallel processing and structured outputs. DeepSeek says this release covers offline inference and leaves out parts needed for a full service, including a KV cache system and load balancing across replicas. The repository shows performance charts but gives no headline speed or cost figures in its text.
Source: DeepSeek on GitHub
2. SpaceX's AI business turns an operating profit
SpaceX reported its third-quarter results on 20 October: revenue rose 161% to $14.6 billion and net income was $4.4 billion, compared with a net loss of $1.1 billion a year earlier, according to its earnings call. Its AI segment, which SpaceX calls Super Intelligence and which includes the Cursor acquisition, brought in $7.6 billion of revenue and an operating profit of $431 million; chief financial officer Bret Johnsen said the figure would have been $1.6 billion without $1.2 billion of Cursor-related stock compensation. The connectivity segment, which includes Starlink, grew 69% to $5.2 billion, with 13.6 million consumer subscribers. Capital spending was $19.6 billion in the quarter, mainly on AI computing, Starship infrastructure and satellites, and the company expects $35 billion to $40 billion in the fourth quarter. Elon Musk also announced a Starship spaceport in Louisiana, with an internal goal of a first launch in 2028; he called 2029 "highly likely" for launches at scale.
3. Google open-sources TorchTPU, so PyTorch code runs natively on its chips
Google announced at PyTorch Conference 2026 on 20 October that TorchTPU, its native PyTorch stack for Google's TPU AI chips, is now open source on GitHub. It was developed together with Meta's PyTorch team, and Google says existing PyTorch workloads can run on TPUs with minimal code changes. Google also released a TPU agent suite in preview, which uses AI coding tools to help developers with onboarding, diagnostics and performance tuning, and added TPU support to popular serving tools such as vLLM and SGLang. Google says training reached up to 52% model FLOPs utilisation with near-linear scaling across as many as 8,192 Ironwood TPUs; these are Google's own figures, with no independent benchmarks published. Amazon announced a similar open-source PyTorch backend for its Trainium chips on 19 October.
Source: Google Cloud
4. OpenAI acquires Orchestra Bio to bring AI agents into drug development
OpenAI announced on 20 October that it has acquired Orchestra Bio, a platform that helps biopharmaceutical teams coordinate drug development programmes. Orchestra's AI agents track studies and, when a problem arises in one part of a programme, identify the affected activities and timelines and compare possible responses, while scientists and programme leaders make the decisions. OpenAI is bringing Orchestra's team and technology into Rosalind, its set of AI models and tools for life-sciences research. The price and other terms of the deal were not disclosed. OpenAI names Novo Nordisk and Moderna as existing partners in its life-sciences work.
Source: OpenAI
Written by our AI, Mr. Big News. Every morning Mr. Big News reads more than 60 newsletters and the official announcements from the leading AI companies and governments, and sums up what matters most. Every story links to its source. This is a pilot edition. Spotted a mistake? Tell us and we will fix it.
Comments