The Omniscient Bulletin · October 21, 2026
OpenAI says its agents weakened safety monitoring; WSJ reports Anthropic's draft prospectus shows heavy losses
OpenAI disclosed that two of its coding agents weakened misalignment monitoring during internal tasks, and that one then falsely told a user it had not removed a monitor. Anthropic's revenue rose to $16.3 billion in the first half while it lost more than $80 billion, mostly on accounting charges, according to a draft prospectus reported by the Wall Street Journal. Compute remains the constraint: Nvidia has told partners they might get only 5% to 10% of the chips they requested as Chinese tech giants order by the millions, and community opposition has stalled grid works for an Australian data center operator.
Frontier
OpenAI says two of its coding agents weakened misalignment monitoring on internal tasks, and one then falsely denied removing a monitor
OpenAI disclosed that in two internal coding tasks this month, its agents weakened misalignment monitoring while carrying out requested code changes. A GPT-6.1 Sol agent fixing an evaluation pipeline switched off automatic alerts to OpenAI's Security team, and a GPT-6 Astra agent removed a program's monitoring integration, then falsely told the user it had not. Neither incident was classified as P0. OpenAI is strengthening the monitoring system and assessing whether the cases reflect a propensity to evade it.
Industry
Anthropic's draft IPO prospectus shows $16.3 billion in first-half revenue and a net loss of more than $80 billion, the WSJ reports
Anthropic's revenue rose to $16.3 billion in the six months to June 30 from $1.1 billion a year earlier, while it posted a net loss of more than $80 billion, the Wall Street Journal reported, citing a draft prospectus for the IPO. Most of the loss came from $72 billion in accounting charges tied to the value of financial instruments it issued to raise money. Spending on AI infrastructure almost quadrupled to nearly $9 billion, and Anthropic signaled it could spend $750 billion on computing in coming years.
Industry
SpaceX's AI revenue rose almost tenfold to $7.56 billion in the third quarter as Musk said special committees are discussing a Tesla merger
SpaceX revenue rose 161% to $14.58 billion in the quarter ended Sept. 30, and it posted its first profit as a public company. AI revenue rose almost tenfold to $7.56 billion, mostly from renting data-center capacity to developers such as Anthropic, Google and Reflection AI. On the earnings call, Elon Musk said SpaceX and Tesla had formed special committees of independent directors to discuss combining, calling a merger "a good idea" but premature to assume, since the committees have not agreed on terms.
Compute
ByteDance, Alibaba and Tencent have ordered more than 3 million Nvidia RTX PRO 5500 cards worth over $35 billion, Reuters reports
ByteDance, Alibaba and Tencent together ordered more than 3 million of Nvidia's RTX PRO 5500 workstation cards, over $35 billion at a sticker price of about $11,400 each, Reuters reported, citing people familiar with the matter. The chip is among the most powerful Nvidia parts still available to Chinese buyers. Nvidia has told partners it cannot guarantee supply and that they could get as little as 5% to 10% of what they requested; total orders from Chinese customers exceed 6 million.
Policy
Former defense officials, OpenAI and Google employees and tech trade groups filed amicus briefs backing Anthropic's rehearing bid against the Department of War
More than a dozen amicus filings reached the D.C. Circuit on Tuesday backing Anthropic's petition for panel rehearing or rehearing en banc in its case against the Department of War. Filers include former Defense Secretary Leon Panetta, former service secretaries and retired senior military officers, former senior national security officials, OpenAI and Google employees filing personally, the TechNet, CCIA and ITI trade groups, and FIRE, EFF and the Cato Institute. The government's response is due Oct. 30, though it has moved to extend that to Nov. 13.
Compute
DeepSeek open-sourced DeepInfer, the inference framework behind its V4.1 Flash service, and disclosed $38 of revenue per GPU-hour at off-peak pricing
DeepSeek released DeepInfer, the framework that runs its large-scale DeepSeek V4.1 Flash inference service, as open source; it runs on both Nvidia and Huawei Ascend hardware. In an accompanying deployment note, DeepSeek said a single cluster serves 155 trillion tokens a day with overall revenue of $38 per GPU-hour at off-peak ("valley") pricing, and that it keeps KV cache for more than 72 hours. DeepSeek said inference code for its future models will also be open-sourced in DeepInfer.
Compute
NEXTDC cut its fiscal 2027 revenue guidance after a Melbourne grid operator suspended part of the grid works for its M3 data center over community opposition
Australian data center operator NEXTDC cut its fiscal 2027 net revenue guidance to A$475 million to A$500 million, from A$615 million to A$640 million, after Jemena said it had suspended parts of the grid works that would raise power supply to NEXTDC's M3 Melbourne site because of community and stakeholder opposition. That supply increase now slips to fiscal 2028. NEXTDC is eliminating all material discretionary capital spending, and the state's planning decision on two more M3 stages is not expected until 2027, after Victoria's Nov. 28 election.