Here is a roundup of the day's key topics from overseas AI news collected on September 12 (Japan time). Capital is concentrating even further — with investment talks around Anthropic's massive IPO and Cohere's funding round — while official case studies were published showing GPT-6 Astra being run "with fewer human checks," and on the regulatory front, a bill going as far as criminal liability for developers was introduced in the Senate. All in all, it was a day when the debate over "who trusts AI, and how much, to let it operate" moved forward on multiple fronts.
Nvidia Considers Up to $10 Billion Investment in Anthropic's IPO — Toward a $2 Trillion Valuation and One of the Largest IPOs in History
According to a Reuters report carried by Bloomberg, Nvidia is considering investing up to $10 billion in Anthropic's initial public offering (IPO). Anthropic aims to raise as much as $100 billion through the IPO, and if it goes through, its valuation is expected to reach around $2 trillion — potentially making it the largest IPO in history.
Nvidia, the largest AI chipmaker, is reportedly in talks to join as an anchor investor. The sourcing rests on anonymous insiders, however, so the outcome and terms of the negotiations could still change. How deeply Nvidia — the linchpin of GPU supply — involves itself in the capital structure of a major model developer is a key point to watch, as it could also influence the listing strategies of competitors.
Cohere Also in Final Talks to Raise $2–3 Billion at a $20 Billion Valuation
Also reported by Bloomberg, citing the Globe and Mail, enterprise AI company Cohere is in final-stage talks for a funding round of $2–3 billion. The round is expected to include investments from the Canadian government and existing investors, and would value the company at $20 billion. The deal could close as early as next week, though the timeline could still slip, according to the report.
With two large funding stories — Anthropic and Cohere — landing at the same time, it is clear that the concentration of capital into generative AI companies is accelerating further.
Perplexity and Devin Put GPT-6 Astra into Production — Human Oversight Drops Sharply
In a case study published as official OpenAI news, Perplexity is using GPT-6 Astra for end-to-end work. It has delegated drafting communications, making software changes, and monitoring production systems to Astra, and says the frequency of human-in-the-loop checks has dropped dramatically compared with when it relied on earlier models.
In another published OpenAI case, Cognition is using GPT-6 Astra to improve the self-testing capabilities of Devin, its AI software engineer. The quality of tests and working demonstrations for code Devin itself writes has improved, and the company aims to reduce the amount of code engineers must review so that more output can be shipped.
These cases show that the bar for operating agentic AI is shifting when it comes to "how much of the verification humans actually oversee." That said, the companies' public information is heavily skewed toward adopter success stories, so it is worth keeping in mind that information on failures and limitations remains scarce.
"20-Year Prison Sentences" Bill Introduced in the Senate — The Sanders Bill on Developer Accountability
According to reports including the Times of India, Senator Bernie Sanders has introduced a bill in the Senate that could send AI developers to prison for up to 20 years. Executives at major AI companies — such as Sam Altman of OpenAI and Dario Amodei of Anthropic — could reportedly be subject to punishment under the bill.
Details such as the full text of the bill and its scope remain limited based on the information available at this point. Alongside the congressional tug-of-war we covered earlier over whether safety reviews should be voluntary or mandatory, we will be watching how much traction this proposal — which reaches all the way to personal criminal liability for developers — gains going forward.
Altimeter's Gerstner Criticizes AI "Extinction" Warnings — Investors and Researchers Remain Far Apart
According to CNBC, Brad Gerstner of investment firm Altimeter has sharply criticized researchers who warn about the risk of AI-driven human extinction. Gerstner is known for stressing AI's economic benefits from an investor's standpoint, and the episode made the divide plain: he sees the warnings' tone as "exaggeration."
Researchers including Bengio have recently been arguing that "the training process itself is dangerous," and the debate over safety risks now appears to be surfacing beyond the research community as a perceptual gap with the capital side.
DeepSeek-V4.1-Flash: $0.15 per 1 Million Input Tokens — Lowering the Cost Barrier of Having AI "Read"
According to an article on Qiita, China's DeepSeek has released a new model, DeepSeek-V4.1-Flash. Its standout feature is the low cost of processing long inputs (reading in prompts), at $0.15 per 1 million tokens. The model weights are also said to be freely available.
Work that involves having AI read through long documents has always carried costs that stack up the longer the text you feed in, so this could reshape the cost structure of use cases such as RAG (retrieval-augmented generation) and long-document summarization. Detailed evaluations in Japanese are likely still to come.