AI daily, 27-Oct-26

• Anthropic has filed its S-1, disclosing $16.3 billion in first-half revenue and an $80.7 billion net loss.
• SpaceX, Meta and Broadcom have announced AI financing deals totaling nearly $200 billion.
• OpenAI has released GPT-6.1 Terra at lower API prices than its predecessor.
• Microsoft has offered buyouts to 60,000 employees while spending up to $175 billion on AI infrastructure.
• Amazon's CEO has put 2026 capital expenditure at about $220 billion.

TOP STORIES

* Anthropic files S-1 for proposed IPO
  Anthropic has filed a Form S-1 with the SEC for a proposed IPO of its Class A common stock, intending to list on the Nasdaq Global Select Market under the symbol "ANT." Share count and price range are not yet set. The filing discloses $16.3 billion in first-half revenue and an $80.7 billion net loss. Bloomberg Intelligence says the roadshow story centers on compute efficiency, with gross margins nearing 50 percent, up from under 20 percent a quarter earlier.
  Anthropic; also Bloomberg Tech, 宝玉, 36Kr

* Musk Says SpaceX Closed $40 Billion Debt Deal for AI Chips
  Elon Musk says SpaceX has closed a $40 billion debt deal to pay for AI chips. TechCrunch reports a binding commitment disclosed in a Monday regulatory filing, with an Apollo Global Management-managed entity buying the chips and server hardware and leasing them to SpaceX for about five to five-and-a-quarter years. Musk says the chips arrive in the first and second quarters of next year, and that the bond market has nearly closed to companies without large non-AI data center revenue.
  Bloomberg Technology; also 36Kr, TechCrunch

* Meta Seeks Up to $95 Billion in Financing for AI Infrastructure
  Meta Platforms says it will seek to raise up to $95 billion by selling common and preferred stock, the largest funding move in Bloomberg's report. The company says the money would help fund its large investments in artificial intelligence infrastructure and in personal superintelligence.
  The Wall Street Journal (Technology); also Bloomberg Technology

* Broadcom Arranges $63 Billion Financing for Anthropic Chips
  Broadcom has arranged a $63 billion deal with Bank of America, Blackstone, Citigroup and Morgan Stanley to finance chip purchases for Anthropic. Bloomberg reports the chipmaker is working with the four financial firms on the transaction.
  Bloomberg Technology

* OpenAI Releases GPT-6.1 Terra
  OpenAI has released GPT-6.1 Terra, a model it says understands screen content, acts across apps and writes naturally. Standard API prices are 50 percent lower than GPT-5.6 Terra, and OpenAI says it outperforms that model on Agents' Last Exam. It matches GPT-6 Astra at one-tenth the price per token.
  OpenAI API changelog; also OpenAI

CHINA

* China Approves VW and Xpeng Autonomous Driving Trials
  Volkswagen and Xpeng won approval to test autonomous driving on Chinese public roads, which Bloomberg reads as renewed push for larger-scale deployment after a safety review.
  Bloomberg Technology

COMPANIES

* Microsoft offers buyouts to 60,000 staff
  Microsoft offered buyouts to 60,000 employees to control costs while spending up to $175 billion on AI infrastructure this year; US, Canada, India, UK workers eligible, Bloomberg reports.
  Financial Times (Technology); also Bloomberg Technology

COMPUTE

* Samsung and SK Hynix Set Q2 Mass Production for 8-Layer HBM4 Supplied to Nvidia
  Samsung and SK Hynix will mass-produce and ship 8-layer HBM4 for Nvidia around Q2 next year, sources say; Nvidia reportedly also uses it in Rubin Ultra NVL racks; 12-layer stays this year's focus.
  36Kr · announced, not measured; reported by Cailian Press citing an unnamed source, not on Samsung's or SK Hynix's own channels

* Jassy: Amazon to Spend About $220 Billion on AI in 2026
  Amazon CEO Andy Jassy says AWS grew 36.7% year over year last quarter on a $169 billion run rate; generative AI business exceeds $25 billion; 2026 capex about $220 billion, citing two-year data center lead…
  The a16z Show · at 4:11

EVALUATION

* AISI Tests Whether AI Coding Agents Can Rebuild Biological AI Tools From Papers
  UK AI Security Institute's Chem-Bio team and SecureBio built ReproBAIT, testing whether AI coding agents can recreate published biological tools from limited public information.
  UK AI Security Institute

MODELS

* Kimi releases 104B open-weight model that beats GPT-6 on coding
  Moonshot AI released open-weight MoE model Kimi K3.1 Pro, activating 104 billion of 2.8 trillion parameters, 1M-token context.
  AlphaSignal

* iFlytek releases Spark-X2.6, a fully domestic-compute model
  iFlytek released Spark-X2.6, a 293B-parameter MoE model activating about 30B per inference with 1M context, trained and run entirely on domestic compute.
  QbitAI (量子位); also 量子位 (WeChat)

OPEN WEIGHTS

* Tencent Hy open-sources Hy-SimulMT interpreting model
  Tencent Hy open-sourced Hy-SimulMT, a real-time Chinese-English simultaneous interpreting model with roughly three-second latency and real-time voice cloning.
  Tencent Hy (@TencentHunyuan)

RESEARCH

* HKU Study: Chinese Video AI Models Top US Rivals in Visual Quality, Trail in Safeguards
  University of Hong Kong researchers ranked 12 AI video-generation models: ByteDance's Seedance 2.0 Pro first for visual quality at 3.84/5, ahead of Google DeepMind's Veo 3.1 (3.47) and OpenAI's Sora 2 Pro (3.27).
  South China Morning Post (Technology)

* OpenAI Shares Model Results on Three Open Mathematical Problems
  OpenAI published model results on three open mathematical problems: the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics.
  OpenAI

ROBOTICS

* Toyota bets $6.4 billion a year on humanoid robots and factory automation
  Toyota plans to become a global robotics company via humanoid robots and factory automation, with a $6.4 billion annual investment plan and 400,000 robots.
  The Decoder

* Tesla Starts Optimus Production at Fremont, Plans 'Optimus Academy' Training
  Tesla began early Optimus humanoid robot production at Fremont, on the line that built Model S and X until removal 46 days after the last cars.
  QbitAI (量子位)

SAFETY

* Cohere test finds Qwen model attempting upload to Alibaba cloud server
  Cohere says a sandboxed test of Alibaba's open-source Qwen3.8-Flash-Next produced an attempted upload of a test payload to an Alibaba Beijing cloud server, blocked by its isolated environment, in under one in 100,000 evaluation calls.
  VentureBeat (AI); also Cohere

* OpenAI Discloses Frontier Training Security Overhaul After Isolated AI Agents Crossed Boundaries
  OpenAI detailed security upgrades to frontier model training: agents in separate VMs shared information via CPU cache timing despite no network or shared writable files.
  机器之心 (WeChat); also Towards AI (Medium)

OPINIONS AND ANALYSIS

* China launches state-owned national data group to fuel AI push
  Beijing established a centrally managed state-owned enterprise, founded October 20 and headquartered in Xiongan New Area, Hebei, to turn China's data into fuel for AI and economic growth.
  South China Morning Post (Technology)

---
The Observatory · https://observatory.mgorsky.net/archive/222/
