AI daily, 21-Oct-26

• Anthropic’s draft prospectus discloses an $80.7 billion net loss for the first half of 2026.
• Anthropic is also negotiating a potential $45 billion computing deal with Nscale.
• Goldman Sachs warns that energy bottlenecks pose a major threat to the AI sector's financial growth.
• New open-source frameworks and powerful new models continue to launch across the market.

TOP STORIES

* Anthropic Draft Prospectus Discloses $80.7 Billion Net Loss
  Anthropic's draft prospectus, seen by The Wall Street Journal, shows an $80.7 billion net loss for the first half of 2026. The filing details rapid growth, large computing costs, and risk factors ahead of a potential public offering.
  The Wall Street Journal (Technology) · according to leaked documents, reported by The Wall Street Journal; announced, not measured

* Anthropic in Talks with Nscale for $45B Computing Deal
  Anthropic is negotiating a computing deal with Nscale, which could be worth up to $45 billion. The agreement would establish a second facility at Nscale's Monarch data center campus in West Virginia. This facility would deliver 460 megawatts of computing power utilizing Nvidia's Vera Rubin AI chips.
  Bloomberg Technology

* OpenAI Rushes Fundraising Ahead of Anthropic Debut
  OpenAI is conducting an unusually fast fundraising round before rival Anthropic's anticipated public debut in mid-November. Reuters reports that this push could value the ChatGPT maker at more than $1.4 trillion.
  Reuters (Technology)

* Gemini 4 Argon released globally
  Google announces Gemini 4 Argon, its latest frontier model, is globally available. Access is provided through the API in AI Studio, for AI Ultra subscribers in Google Antigravity, and for other tiers in the Gemini app and Search AI Mode.
  Hacker News · Google's own announcement; details differ from the announcement

* Muse Spark Open, Meta's Open Weights Model, Launches on Fireworks AI
  Fireworks AI makes Meta's open weights model, Muse Spark Open, available on its platform as of October 20, 2026. The 816B-parameter mixture-of-experts model has 42B active parameters and a 1M-token context window. Fireworks reports it ranks first on Vals AI's benchmark for Harvey's Legal Agent and outperforms Claude Opus 5.5 on Rogo's Big Finance Bench.
  Fireworks AI

COMPANIES

* Fireworks AI in early talks for funding at $30B–$40B valuation
  Fireworks AI seeks new funding at a $30B–$40B valuation, up from $17.5B.
  Business Insider (AI) · reported by Business Insider; not on Fireworks AI's own channels

* Salesforce Agrees to Buy Braze for $5.7 Billion
  Salesforce Inc. has agreed to acquire Braze Inc., an artificial intelligence-powered marketing platform, for an enterprise value of almost $5.7 billion.
  Bloomberg Technology

* Neolix raises nearly $400 million in Series E funding
  Chinese autonomous delivery company Neolix completed a nearly $400 million Series E round, co-led by Boyu Capital and China's National Artificial Intelligence Industry Investment Fund.
  TechNode; also 36Kr

* ASI Raises $325 Million, Reaching $3 Billion Valuation
  ASI raised $325 million, valuing the startup at $3 billion.
  Bloomberg Technology

COMPUTE

* Alibaba, ByteDance, Tencent order over $35 billion in Nvidia chips
  Alibaba, ByteDance, and Tencent together ordered over $35 billion worth of Nvidia chips, highlighting the entanglement of US and Chinese tech sectors.
  Semafor (Technology)

* Nvidia Backs Neoclouds With Over $170 Billion in Commitments
  Nvidia disclosed up to $164.5 billion in commitments supporting new data centers in its July quarter, including guarantees and direct investments in AI cloud providers.
  The Information (newsletters)

IN THE FIELD

* Claude AI Plans First AI-Planned Mars Rover Drive
  NASA's JPL used Anthropic's Claude on Dec 8 and 10, 2025, to plan Perseverance rover's route (400m Martian rock), marking first AI-planned drive.
  Anthropic

MODELS

* Inception launches Mercury 3 diffusion language model
  Inception released Mercury 3, outperforming GPT-6 Luna and Claude Haiku 5.5.
  The Rundown

* Mistral AI Releases Voxtral 3 Transcription and Speech Models
  Mistral AI introduced Voxtral 3, powering its voice agents API.
  Mistral AI (@MistralAI); also AI News / TestingCatalog

* MiniMax Releases M3.1 Flash, Upgraded 'Bunny' Model
  MiniMax M3.1 Flash, a lightweight model, has a 1M token context.
  MiniMax_AI (@MiniMax_AI); also 智东西 (WeChat), InfoQ (WeChat)

OPEN WEIGHTS

* DeepSeek open-sources DeepInfer inference framework
  DeepSeek open-sourced DeepInfer (Apache 2.0), which powers DeepSeek-V4.1-Flash with efficient inference.
  DeepSeek (@deepseek_ai); also The Decoder

POLICY

* US Sanctions Five Chinese Entities Tied to Google AI Trade Secret Theft
  US sanctioned five Chinese entities linked to Linwei Ding, a former Google software engineer jailed for stealing Google AI trade secrets.
  The Information

* China Condemns U.S. Ban on DeepSeek A.I. in Government Systems
  China criticized the U.S. ban on using DeepSeek's AI in government systems after the U.S. implemented usage restrictions.
  The New York Times (Technology)

OPINIONS AND ANALYSIS

* Goldman Sachs warns power bottlenecks threaten AI buildout
  Goldman Sachs warns power bottlenecks are the top risk to AI's buildout.
  Axios AM

* Anthropic draft AI bill leaks, drawing criticism
  Anthropic-pushed draft federal AI safety bill leaked to Semafor.
  Semafor (Technology) · according to leaked documents, reported by Semafor

---
The Observatory · https://observatory.mgorsky.net/archive/166/