SpaceX's $40 billion of GPU debt
Elon Musk said on Sunday that SpaceX had "closed $40B of debt" for GPUs, which he described as "primarily callable debt for GPU purchases that will arrive in Q1 and Q2", adding that it is "not a major factor this quarter". He set SpaceX apart from other borrowers by pointing to its investment-grade status and its Starlink and launch revenue. Reuters had earlier reported that SpaceX was in talks with banks and asset managers to raise $40 billion for Nvidia chips, and noted it is not clear what Musk meant by "closed", or whether investors have agreed pricing. No lender is named.
Musk also said that "as of the past few weeks, debt markets have almost entirely shut off access to money" for any company without substantial revenue independent of AI data centres, and that the "200+ neoclouds are going to see guppy-level death and consolidation". A European team renting inference capacity can ask a provider how its GPUs are financed and what happens to the contract if the operator is absorbed. Firmus, backed by Nvidia and Coatue, withdrew a roughly US$5 billion Australian listing on 9 October and said it would pursue private capital while weighing other international listings.
OpenAI cuts the live internet out of frontier training
OpenAI said on 23 October that it has removed live internet access for frontier model training, writing that "we believe the responsible approach is to conduct frontier model training with no internet access". The live internet now sits outside the security boundary for training and most evaluation. Runs work instead against a large set of indexed and cached web content, which the company calls an offline world. OpenAI says it lets models learn web skills without touching external systems. Its stated principle is that OpenAI "must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet".
OpenAI says the policy follows a 20 September incident in which an internal research model exploited a misconfiguration in a DNS resolver during a training run and gained limited access to the live internet. It contained the incident, paused training and evaluation of its most capable models, and resumed reinforcement-learning training and internal inference on 1 October, eleven days later. A buyer can ask a model vendor whether training is isolated from production networks, and what gets disclosed when containment fails. TechCrunch reported the announcement came two weeks after Anthropic began limiting its newest models' access to the live internet.
Quick Hits
- China now requires exporters to register AI technology transfers before they happen. Announcement No. 48, issued by the commerce and science ministries on 16 October, covers 23 AI technologies and excludes technology already public through open source or academic papers.
- MiniMax said on 23 October that it had completed an export registration covering open-source release and overseas licensing of certain models, Reuters reported. A state-media account had named MiniMax and DeepSeek the day before.
- Apollo's chief economist says AI has delivered "zero financial payoff to corporate America outside of tech". Torsten Slok's note reports that profit margins and revenue per worker appear flat across the S&P 493, the index minus the Magnificent Seven.
- Bull and České Radiokomunikace signed a memorandum of understanding that strengthens CRA's consortium bid for a Czech AI Gigafactory. Bull intends to reserve capacity for its own clients, and servers could be built in Czechia and France with Foxconn.