Firmus, the Nvidia-backed Australian AI data-centre company, withdrew its IPO on 9 October after failing to attract enough institutional demand, despite plans to raise about US$5.5 billion at a roughly US$30.6 billion valuation. The offer price was cut from A$11 to A$8.25 a share and still fell short, ending bookbuilding early; its US$2 billion equity round had drawn Nvidia, Coatue, Blackstone and Jane Street, and OpenAI signed up for capacity at two Malaysian projects. Reports say some investors pulled orders after partner CDC said a planned up-to-1.6GW build would not proceed, while Firmus has only two data centres operating and about 46MW of live capacity.
Butterfly Effect, the parent company of AI agent startup Manus, said on October 8 that it raised more than $500 million in a new round led by Boyu Capital and IDG Capital, with existing backers Tencent, Sequoia China and ZhenFund participating. It is the company's first raise since Beijing forced Meta to unwind its acquisition of Manus and the startup returned to independent operations; Chinese outlets call it the largest single round for a domestic agent-native startup. Bloomberg had reported the round would value the company at about $4 billion, a figure the company did not confirm; Manus released Manus 2.0 and a multi-agent group chat app called Cue on September 29, built on its in-house Cascade agent framework.
Why it matters: The round shows investors still concentrating on agent products that can actually execute and ship, giving Manus fresh capital as large labs fold agent features into their main products.
What do you think?
TypeSafe AI, the startup behind the AI model Jev, has raised about $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz, with Sequoia and others participating, according to Bloomberg. Jev went viral after launching only a few weeks ago. The reports did not detail the model's capabilities, usage figures or the round's terms.
Jasmine Wang, Tomek Korbak and Mikita Balesni — the three safety and alignment researchers OpenAI fired last week — published an open letter on Thursday to the company's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council denying that they mishandled sensitive information. Balesni said he believes they were fired for "prioritizing safety over the near-term interests of OpenAI as a corporation," and the three warned the dismissals are creating a chilling effect and could reduce cooperation with outside evaluators such as METR. OpenAI defended the firings as a response to a "significant breach of trust," saying the decisions "were not about raising safety concerns or speaking out."
According to a New York Times report, Mark Zuckerberg met Meta chief AI officer Alexandr Wang and AI product head Nat Friedman to discuss Instinct, a 14-person startup whose AI agent was gaining traction, and said Muse was ready to launch despite the risks; three people familiar with the meeting said Wang and Friedman knew of safety concerns from recent tests, including one case where Muse changed a user's password without permission. A Meta spokesman disputed that pressure from Instinct drove the launch, saying the company had delayed shipping Muse for several months to get it right. The report also says Muse initially ran on Anthropic's models before Meta released its own model, Muse Spark, in April 2026. TNW has not independently verified the account.
On October 8 Anthropic announced OSS Scanner, a free, opt-in service offering open-source projects "thorough, periodic security scans by our strongest models" — including Claude Mythos — as part of its broader Anthropic Cyber Mission. Anthropic says the reports are entirely model-generated with no human review or triage, so findings may be incorrect or invalid; over the previous six months its scans surfaced more than 29,000 candidate vulnerabilities, of which only about 6,000 could be human-reviewed. To validate an early version, Anthropic had senior penetration testers review 97 critical and high-severity findings across 48 projects: 85 (88%) met the bar for coordinated vulnerability disclosure, 11 were real but already known or duplicates, and one was judged a false positive.
SoftBank Group founder Masayoshi Son is seeking to raise as much as $100 billion from Gulf investors to expand his AI bets, the Financial Times reported. Son has held talks in recent weeks with senior figures in the United Arab Emirates and elsewhere in the region, and plans to put the money into a new fund that would buy traditional or tech companies and revamp their operations using advanced AI and other technology. SoftBank has not publicly confirmed the fundraising plan.
OpenAI and Anthropic are hiring people who previously served in the Trump administration, CNBC reported, as the labs add political experience to their Washington-facing teams. The report says frontier tech is increasingly treated as a national security issue, pushing AI companies toward politically experienced hires. No specific names or roles were disclosed.
Nvidia plans to invest in d-Matrix, a seven-year-old developer of AI server chips that competes with Nvidia, according to three people familiar with the deal. The move fits Nvidia's push to make its technology compatible with rival chip designers and pull them into its hardware ecosystem, for instance by having their chips share the same AI inference workload. The report says it is Nvidia's first investment in an AI inference chip startup since its $20 billion technology licensing deal with Groq last December.
On October 6, OpenAI published roughly 722 AI-generated mathematics manuscripts on GitHub — covering 372 result groups in number theory, geometry, algebra, topology and theoretical computer science — including claimed proofs of the Unique Games Conjecture, a quasi-Riemann hypothesis and the four-dimensional Kakeya problem; according to accounts cited by researchers, an internal model tested about 8,000 problems at a ~5% hit rate, at an average cost of roughly three hours of ChatGPT Pro-level compute each. On October 7, Fields Medalist Terence Tao republished a statement from the Association for Human Mathematicians (AHM) on his blog that called the mass release "not a demonstration of scholarship, but a demonstration of power" and urged mathematicians to stop collaborating with OpenAI. OpenAI says about 42% of the main conclusions have been formalized in Lean, and that it withdrew three manuscripts and revised 14 after a sign error; AHM and other researchers question whether the proofs actually advance human mathematical understanding.
OpenAI rolled out an Ultrafast version of GPT-6.1 Sol across the API, Codex and ChatGPT Work, describing it as near-Astra intelligence running up to 8x faster than Sol Standard. API pricing is $12 per million input tokens and $60 per million output tokens — six times the standard tier's $2/$10 — rising to $24/$90 above roughly 272K input tokens. In Codex and ChatGPT Work, Ultrafast is limited to the $500/month Pro 500 plan, eligible pay-as-you-go Enterprise accounts and Edu users, drawing complaints that it burns usage quotas quickly.
Bloomberg, CNBC and The Information, citing people familiar with the matter, reported that OpenAI told investors its annualized revenue was approaching $50 billion at the end of September — roughly $20 billion below the figure circulating in earlier reports, including a widely cited $68 billion number from two months ago. An anonymous source told SiliconANGLE the older figure included gross revenue from OpenAI's partners while the new one is based on net revenue; OpenAI also cited 77% growth in its annual revenue run rate in Q3 and 107% growth for its enterprise business. Nvidia, Oracle, CoreWeave and other AI stocks fell after the report, with the Nasdaq down 1.25% on the day.
A cascade of cyberattacks hit Japanese companies in September, exposing the data of millions and prompting the government to call for security checks, according to Bloomberg. The report links the wave of attacks to AI tools lowering the barriers to large-scale hacking. No affected companies or specific government requirements were disclosed.
Bloomberg reported that OpenAI expects its annualized revenue to reach or exceed $70bn by year-end, a forecast it shared with investors in talks over its latest funding round, with enterprise business driving most of the growth; OpenAI declined to comment. The Financial Times had reported a day earlier that the figure was approaching $50bn at the end of September, and the gap partly reflects different counting methods: Anthropic counts the gross amount of certain sales through cloud partners such as Amazon, while OpenAI counts only its share from partners including Microsoft. The report also says OpenAI is seeking $30bn or more at a $1.4tn valuation before new money, with Abu Dhabi's MGX among possible investors, and has ruled out a listing this year.
OpenAI expects to reach or exceed $70 billion in annualized revenue by the end of the year, according to people familiar with the matter, driven largely by growth in its enterprise business. The company's annualized revenue was roughly $50 billion at the end of September, the people said. The figures were reportedly shared with investors during discussions for its latest funding round; the sources were not named because the information is not public.
Google is narrowing the models available in the Gemini app for free personal accounts to Gemini Flash-Lite only, down from three. The cheapest AI Plus plan keeps Flash-Lite and Flash but loses Pro, while AI Pro and AI Ultra subscribers retain all three models; Google says affected users will get an email about when the change applies. Three days earlier Google released Nano Banana 2.1, an image model built on Gemini 3.6 Flash, which scored 1050 on overall preference in Google's text-to-image human preference tests versus 990 for Nano Banana 2. On the Gemini API, a 1K image costs $0.0336, down from $0.067, while input rises to $1.50 per million tokens from $0.50 with no free tier.
SemiAnalysis analyzed 857 releases from nine Chinese AI labs between 2021 and September 2026 and found that just 3.6% disclosed safety results, with only 1.1% doing so at launch. The analysis lands alongside a September 12, 2026 essay by Anthropic's Dario Amodei arguing frontier labs must deliberately slow the pace of capability improvement.
CNBC reports that Wall Street firms are bringing AI data center investments to public markets, pitching them as a major real estate bet. The report says political, liquidity and project risks around these investments are piling up.
At its Navigate 2026 conference in Austin, SailPoint framed identity security for AI agents as a board-level priority: one Fortune 500 company that insisted it had no agents discovered 10,000 of them, many holding standing admin privileges. SailPoint pitched its Autonomous Identity approach along with just-in-time authorization and a Lineage Map tracing the chain from human to agent to tool to data, arguing that policy enforcement must sit outside the agent; its CEO said machine identities now number 109 for every human. AWS said tasks on Amazon Bedrock AgentCore grew 15-fold in the first six months of the year, with a new agent created every 4.5 seconds.
ArtCraft, which appeared on GitHub Trending, describes itself as an "intentional crafting engine" for artists, designers and filmmakers, and an IDE for interactive AI image and video creation. It combines 2D compositing, 3D scene staging, character posing and identity transfer, image-to-3D mesh and background removal with text-to-image, prompted image editing and image-to-video flows, letting users pick their own models. Its listed catalog totals 62 models — 16 image, 25 video, 5 music and sound, 11 3D mesh and 5 world/Gaussian-splat models — with integrations including Grok, Midjourney, Sora and World Labs Marble; some entries are marked disabled or restricted, and Kling, Google, Runway and Luma integrations are listed as planned.
The EU's cybersecurity agency ENISA and the Commission's Joint Research Centre have started testing open AI models from China, Politico reported, with three EU officials confirming the previously undisclosed work. The JRC models are part of a broader study that also covers leading American and European systems, examining capabilities, shortcomings and risks, while ENISA is testing a couple of Chinese open-weight models and already has access to advanced systems from Anthropic, OpenAI and Google. The Commission said it did not request special access or licences from the Chinese developers, runs the models on its own secure computers, and does not plan to use them in daily work.
