Anthropic Files for IPO, Warns Future AI Models Could Soon Self-Improve
Anthropic's public filing reveals financial details and a self-improvement warning, as Google and Moonshot release new models in a rapidly evolving AI software market.
The story
Anthropic made headlines this week with its amended S-1 filing for an initial public offering on Nasdaq, under the ticker ANT. The filing, made public on Monday, November 3, revealed the company's financial performance, including $16.3 billion in first-half 2026 revenue and an operating profit of $974 million in the second quarter alone.
Crucially, the prospectus also contained a warning that upcoming models could "imminently" pass the company's internal safety threshold for AI-driven self-improvement. This disclosure highlights the accelerating pace of AI development and the complex challenges labs face in governing advanced systems.
Meanwhile, Google DeepMind released Gemini 4 Carbon on Wednesday, a new model designed for cybersecurity, achieving top benchmark scores. Moonshot AI also released the full open weights for its Kimi K3.1 Pro model by November 4, further diversifying the accessible frontier models.
Who moved
Anthropic
What Changed: Filed an amended S-1 for its IPO, revealing financial performance and a warning about potential AI self-improvement capabilities in future models.
Consequence: This move provides unprecedented transparency into the economics of a major AI lab and signals increasing scrutiny on advanced AI safety.
Google DeepMind
What Changed: Released its new Gemini 4 Carbon model, optimized for cybersecurity tasks and achieving high benchmark scores.
Consequence: The release targets a specialized, high-stakes domain, indicating a focus on practical, secure AI applications for defenders.
Moonshot AI
What Changed: Made the full open weights for its Kimi K3.1 Pro model publicly available.
Consequence: This expands the landscape of powerful open-weight models, allowing broader access for developers and researchers to build upon.
OpenAI
What Changed: An internal model reportedly disproved the 27-year-old Exponential Time Hypothesis in complexity theory.
Consequence: This research breakthrough demonstrates advanced reasoning capabilities in AI, potentially impacting theoretical computer science.
Hg (with Anthropic)
What Changed: Invested in a new AI transformation company based in Europe to advance AI adoption and product innovation.
Consequence: This initiative aims to accelerate the integration of AI, particularly Claude, into European businesses and foster local AI development.
New models
Gemini 4 Carbon
Lab: Google DeepMind
What: Achieved top scores on the Artificial Analysis Intelligence Index and excels at vulnerability finding tests.
Use: Initially rolled out to cybersecurity defenders and trusted partners for defense applications and vulnerability analysis.
Kimi K3.1 Pro
Lab: Moonshot AI
What: A 3-trillion parameter class model with a 1-million token context window, now with full open weights.
Use: Available for developers and researchers to deploy, fine-tune, and customize for various applications on their own infrastructure.
Market signals
Anthropic's IPO filing revealed $16.3 billion in first-half 2026 revenue, with $8.9 billion in computing and infrastructure costs, and a $974 million operating profit in Q2 2026.
Implication: The figures show significant revenue growth in the AI sector, alongside the substantial capital expenditures required for model development and inference, with a recent shift towards profitability.
Microsoft, Alphabet, Amazon, Meta, and Oracle collectively recorded $556.9 billion in capital expenditures against $542.9 billion in operating cash flow in the first nine months of 2026.
Implication: This indicates that major tech companies' AI infrastructure spending is exceeding their operating cash generation, resulting in a combined $14 billion free cash flow deficit for the period.
Safe Superintelligence secured a $5 billion investment from Nvidia at a $32 billion valuation, despite having no consumer product, API, or revenue stream.
Implication: This highlights continued substantial investment in foundational AI research, even for ventures without immediate commercialization plans, driven by long-term AGI ambitions.
Builders saying
Sam Altman, OpenAI CEO. Altman stated that having DRAM sit idle in personal computers is an inefficient use of limited global memory resources, advocating for maximum utilization in cloud data centers.
What we'll be watching
- Amazon Research Awards proposal deadline on November 11, 2026.
- Web Summit 2026 in Lisbon, Portugal, from November 9–12, 2026.
- Microsoft Ignite 2026 from November 17–20, 2026.
- DeepSeek model release, forecast near November 23, 2026.
- AWS re:Invent 2026 from November 30 – December 4, 2026.
- The AI Summit New York 2026 from December 9–10, 2026.
Reporting + analyst voices: grounded via Google Search at publish time.