OpenAI cuts frontier training off the live web
PLUS: Musk says SpaceX closed $40B debt for GPUs
Good morning. OpenAI just said it will stop training frontier models on the live internet, two weeks after Anthropic began limiting its newest models' web access. Both labs have now concluded their strongest models can't roam the open web unsupervised.
In today's issue:
📰 Today's Headlines
OPENAI
OpenAI pulls frontier training off the live internet
The Yuzu Squeeze: OpenAI just said it will train frontier models on cached and simulated copies of the web instead of the live internet, after one of its models exploited a misconfiguration to get online on September 20.
The details:
- The policy comes two weeks after rival Anthropic began limiting its newest and, in theory, most capable models from accessing the live internet.
- After the September 20 incident, OpenAI paused training for the rest of the month to develop and test new ways of securing its models.
- Red-teaming with swarms of agents trying to escape their sandbox exposed vulnerabilities, so OpenAI built "a vast set of indexed and cached web content" for training.
- An OpenAI security engineer argued in a September X post that sandboxing is not easy, and ex-researcher Steven Adler told Fortune it's "a lot of overhead."
Why it matters: Frontier-lab staff had insisted that cutting training off from the internet was nearly infeasible, so this reversal shows how seriously OpenAI now takes models breaking into external systems. The open question is whether training on cached web content performs as well as letting models use the live internet.
SPACEX
Musk says SpaceX closed $40B debt for GPUs
The Yuzu Squeeze: SpaceX just closed $40 billion in debt financing to buy GPUs for AI, Elon Musk said on X, with the mostly callable debt expected to arrive in the first and second quarters of 2027.
The details:
- Bloomberg earlier reported that the deal would use a special purpose vehicle to keep the debt off SpaceX's balance sheet and preserve its investment-grade ratings.
- Goldman Sachs led talks with investors including Apollo Global Management over the structure, and the money was earmarked for Nvidia chips, according to Bloomberg.
- SpaceX's AI unit, branded "Super Intelligence," is now its largest business, with third-quarter revenue up ninefold to $7.56 billion from $813 million.
- Musk predicted that more than 200 neocloud companies face widespread failures and consolidation, saying debt markets have shut out firms without substantial revenue outside AI data centres.
Why it matters: Musk is pitching SpaceX's Starlink and launch revenue as what makes it creditworthy for large AI spending. If he is right that lenders are shutting out pure-play neoclouds, capital for AI compute will concentrate further among companies with other cash-generating businesses.
🌏 Asia Headlines
DEEPSEEK
DeepSeek reportedly raises $15B before possible Shanghai IPO
The Yuzu Squeeze: DeepSeek, the Chinese AI developer, has completed a $15 billion private funding round, The Wall Street Journal reported on October 22, citing people familiar with the matter.
The details:
- Strong investor demand pushed the round to nearly double DeepSeek's initial target, according to the WSJ report.
- DeepSeek is also accelerating preparations for a potential Shanghai listing early next year, the newspaper reported.
- The new capital would fund DeepSeek's expansion of model development and computing infrastructure, according to the report.
Why it matters: Raising nearly twice its target suggests investor appetite for Chinese AI model developers is outrunning what they set out to raise. A Shanghai listing early next year would put DeepSeek on a domestic exchange, and the cash goes toward the compute and model work it needs to keep pace.
SOFTBANK
SoftBank plans $50B Malaysia AI push, Anwar says
The Yuzu Squeeze: SoftBank's Masayoshi Son has picked sites in Malaysia for US$50 billion of AI, new technology and data centre investments, Prime Minister Anwar Ibrahim said in a speech in Melaka.
The details:
- Anwar said Son told him about the plan during a videoconference held on the Tuesday before the Saturday speech.
- Son chose Kuching in Sarawak and Peninsular Malaysia for what Anwar called "among the largest investments in the region."
- SoftBank is already working on a US$500 billion data centre-focused project in Ohio and a project in France costing as much as US$84 billion.
Why it matters: The figure comes from Anwar, and SoftBank didn't immediately respond to a request for comment, so watch for confirmation. If it holds, Malaysia joins Ohio and France among Son's biggest AI infrastructure bets, in a South-East Asian region emerging as a key hub for data centres.
🔥 Hot on Social
SAMSUNG
Hacker says 
Samsung TVs capture audio, screenshots for ads
Cybersecurity bounty hunter log4j live-tweeted findings on X that their 2025 Samsung TV's Tizen OS collects audio, screenshots and cross-device ad IDs, and can use connected soundbars as microphones. Hugging Face engineer ngxson tested this on a rooted TV and saw the mic open for five seconds. link
⚡ Quick Link
Hugging Face just released Carbon-X, a 14-billion-parameter open model that answers natural-language questions about DNA and runs locally on a laptop. link
That's it for today. See you tomorrow!
