Key Takeaways
- NVIDIA used GTC Berlin to release more open models, an OpenShell agent-boundary architecture, and DSX infrastructure products, while naming European enterprises and EBU as partners.
- SpaceX's Terafab is in talks with Korean chip suppliers, with a 10-year target of 1 million wafers per month; Musk also addressed a possible Tesla merger and SpaceX's Q3 results.
- OpenAI acquired Orchestra Bio and reported GPT-Rosalind drug-discovery results plus a PHP vulnerability found through its Patch the Planet program.
- Anthropic added physicist Christopher Stubbs to its Long-Term Benefit Trust, and Sam Altman welcomed Aleksander back to lead safety and security at OpenAI.
- Hugging Face highlighted Bria's Fibo 2.0 image-editing model, the Flora robotics data engine, and Sentence Transformers v6.2.0.
1. NVIDIA's Berlin Releases and European AI Push
- NVIDIA released additional open models at GTC Berlin for European teams to build AI on their own data and controlled infrastructure, covering specialized agents, robot training, protein design, and quantum computing. It also introduced OpenShell, a reference architecture combining compute for running agents at scale with runtime access controls; CEO Jensen Huang said it lets enterprises set agent boundaries and stop behavior that crosses them. A separate Agent Knowledge Foundation addition to Agent Toolkit is intended to help agents find relevant information and support decisions, with OpenShell governing what they can access and execute. 1 2 3 4
- NVIDIA said L'Oréal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestlé, and HSBC are among European companies working with it on AI, and announced a collaboration with the European Broadcasting Union on sovereign AI standards intended to preserve data control and editorial independence. It also released DSX, which European infrastructure operators can adopt to cut power use and cooling water while helping manage electricity demand, and DSX POD, a containerized Vera Rubin AI factory that NVIDIA says shortens on-site deployment from six months to 14 days. 1 2 3 4 5
- European industrial users including Agile Robots, Audi, Dassault Systèmes, Schaeffler, and Siemens are using CUDA-X, Omniverse libraries, and Cosmos and Nemotron open models to speed design testing, robot simulation, and factory operations, according to NVIDIA. Huang said in Berlin that all companies will become AI companies and that Europe will have a thriving AI economy. 1 2 3
2. SpaceX's Terafab, Financials, and Tesla Merger Talk
- SpaceX's Terafab is in talks with nearly all Korean chip suppliers, from basic tubing parts to process materials and test components, according to Elon Musk. He cited Hanmi Semiconductor discussing TC bonders, HPSP agreeing to supply one high-pressure hydrogen annealing tool to an R&D line, and MiCo's Komico signing a three-way NDA with Tesla and SpaceX about a month earlier. The stated target is 1 million wafers per month within 10 years, with initial investment of $16.8 billion, more than 3,000 jobs, and first-phase capacity of 100,000 wafers per month expected before 2030; logic chips would involve Intel and ongoing TSMC talks, with memory built separately. 1
- Musk said Terafab's output value would put a single building among the top 20 national GDPs and argued that bringing large-scale compute online quickly would be easier with double the time, but reality does not allow it; he called "power shortage" an oversimplification. SpaceX reported Q3 2026 revenue of $14.6 billion, up 161% year over year, and net profit of $4.4 billion, with AI at 52% of revenue; its first revenue-generating Starship mission deployed 26 Starlink V3 satellites with downlink capacity about 10 times a single Falcon 9 Starlink launch, and nominal compute reached 1.7 GW as of September 30. 1 2 3 4
- Musk called a SpaceX-Tesla merger "a good idea" and said the companies have formed independent board special committees at shareholder request, with him not participating due to a conflict of interest. He said claims that a merger is a done deal are exaggerated, the committees have not agreed on terms, and there is no guarantee; the Tesla side would require approval from 50% of all shares. He also said SpaceX reaching $1 trillion in revenue in 2028 is "no longer impossible" and relayed that SpaceX expects at least 13 GW of compute by the end of 2027, mainly using Nvidia Vera Rubin, while Grok Bot will open to enterprises within weeks; Huang said Nvidia has listed SpaceX as a cloud service provider customer and engineers are co-designing space servers. 1 2 3 4
3. OpenAI's Science and Security Moves
- OpenAI acquired Orchestra Bio, a platform for coordinating complex drug-development workflows, and said it will combine it with its Rosalind life-science research models and tools to speed therapies to patients. Greg Brockman described the acquisition as an important application for accelerating drug discovery. 1 2
- OpenAI reported that GPT-Rosalind produced a new drug-discovery lead: in mouse brain data, disruption of the Perry syndrome-related gene DCTN1 increased expression of cholesterol-synthesis genes, and the model hypothesized that cholesterol may not be reaching where neurons need it, offering a testable mechanism and potential therapeutic direction for neurodegenerative disease. With PerturbAI and Scripps Research, thousands of GPT-Rosalind agents worked in parallel to link gene-expression results from millions of mouse brain cells to scientific literature in 32 minutes, and the collaboration is expanding to identify new drug targets. 1 2
- Through its Patch the Planet program, OpenAI said its models helped researchers find a severe security vulnerability in PHP, which it says underpins about 70% of websites, and it worked with maintainers, researchers, and security vendors for two months on a fix requiring major changes to the PHP engine. OpenAI called it one of the most consequential recent vulnerabilities and said it shows AI can find flaws with internet-scale impact, describing a "Defender's Window" in which frontier AI found the flaw before attackers; it released a Codex skill users can point at repositories to check whether their sites are affected. 1 2 3 4
4. Safety Leadership, Research Tools, and Evaluation Practice
- Anthropic said experimental physicist and former Harvard Faculty of Arts and Sciences dean Christopher Stubbs joined its Long-Term Benefit Trust; it noted he has advised the U.S. government on national security and arms control. Separately, Sam Altman welcomed Aleksander back to lead safety and security work at OpenAI, citing the scale of the challenge around superintelligence; the quoted post said Aleksander had thought he was leaving OpenAI but treated it more like a sabbatical, and that safety and security must be a core priority as AI enters a new phase, with him working alongside Merett. 1 2
- Hugging Face highlighted Bria AI's Fibo 2.0, which it described as the top-ranked open-weight model for image editing, with Diffusers support, JSON-native control, improved text rendering, and Turbo and Dev versions. It also highlighted the open robotics data engine Flora, which it said reaches Astra-level performance at 125 times lower cost and has been used internally to label 200,000 hours of video per day on 500 GPUs, and the Sentence Transformers v6.2.0 release with mergekit model merging, faster loading, and cross-GPU evaluator sharing. 1 2 3
- Practitioners emphasized evaluation limits: Hamel Husain shared research testing decision models including Jev, d1-ultra, clef, pplx-decider, and openai decisions on failure detection, calibration, and threshold trade-offs, and said his product's arbitrary 0.8 threshold works for now but should be tuned, prioritizing recall because false positives cost about a minute while misses can lose parent trust. He also shared warnings that evaluations which pass 100% of the time lose information, that public benchmarks should not be trusted blindly, and that human domain experts are needed to review agent outputs for subjective work. 1 2 3 4 5 6 7
