OpenAI aims to close a $30 billion fundraising before the end of next week at up to $1.25 trillion pre-money, with Coatue and UAE investors leading. The valuation is a step down from the $1.4 trillion OpenAI was earlier reported to be seeking. OpenAI is separately negotiating a deal with the UAE to secure more compute capacity. Terms are still being finalized and could change.
Why it matters
The price of staying in the frontier race just got a public number, and even OpenAI had to shave about $150 billion off its asking price to get there. Watch what this does to every other lab's fundraising: if OpenAI takes a haircut at $1.25T, Anthropic's $2T IPO talk and every neocloud's leverage math get repriced against it. The UAE pairing money with compute is the new pattern to watch, because the round is really about buying electricity and chips, not equity.
DeepSeek closed its second funding round at 106 billion yuan ($15.9 billion), more than double its 50 billion target and roughly double its first-ever round from June. The raise values the company above 600 billion yuan (about $90 billion), and it is preparing to list on the Shanghai Stock Exchange. The money funds a two-track compute strategy: buying more of Nvidia's Blackwell chips while shifting training toward Huawei chips.
Why it matters
Days after Washington ordered DeepSeek off federal systems, Chinese capital poured in more than twice what the company asked for. The decoupling is now complete on both sides of the Pacific: America is walling off investment while China funds its champion at home and toward a Shanghai listing. For investors, the read is that export controls raised DeepSeek's price of compute, and its backers simply paid it.
Anthropic's prospectus is expected within days ahead of a listing next month, but some investors who met the company last week left unconvinced its growth justifies the nearly $2 trillion valuation bankers have floated privately. Inside the firm, 91% of 1,300 surveyed employees believe AI could deliver a Nobel-level breakthrough by 2030 (versus 39% of Americans), and AI agents inside the company now work six times the hours of its human staff. JPMorgan's finance chief said AI's benefits 'aren't observable in our current quarterly earnings' and will be offset by token costs; Peter Thiel called AI 'the furthest thing from a monopoly imaginable.'
Why it matters
This is the first real test of whether public markets will pay private-market AI prices. If the IPO lands below $2T, it sets the ceiling every other AI company is valued against, including OpenAI's $1.25T round. The deeper signal is Thiel's: when Google, Meta, Grok, and open-source models are all 'not that far behind,' the moat investors are buying is execution speed, not a monopoly. Founders should read the 91%-versus-39% gap as a warning: the people building AI and the people buying it no longer live in the same reality.
Anthropic opened a research preview of Claude Duet, a standalone app where Claude works as a proactive agent: it sends a morning brief of what needs attention across connected apps, keeps working after the user steps away on its own virtual machine, and asks before it sends messages, pays, signs, or deletes anything. Built on Claude Opus 5.5, it plugs into 6,000+ Claude Marketplace connectors (Gmail, Slack, QuickBooks, Stripe, Shopify) and can spend money through Stripe's Link digital wallet or Shopify's Shop Pay, paying with virtual cards. It is waitlisted, US-only, for Pro and Max subscribers to start.
Why it matters
The agent race just moved from 'chat that does tasks' to 'a coworker that spends money while you sleep.' The important design choice is the approval line: Duet shows exactly what it will do before paying or sending, which is the difference between an agent businesses trust and one they sue. Watch this as the template every enterprise agent will copy, because the first agent that reliably handles invoices, payroll drafts, and supply orders becomes the operating system for small business.
The US Treasury proposed expanding outbound investment restrictions so Americans can no longer invest in any AI systems, or in key semiconductor and microelectronics technology like high-bandwidth memory chips, in China, Russia, North Korea, Venezuela, and Cuba. The current rule only blocks AI investments above a compute threshold; the proposal also adds hypersonic systems and biotech, and is open for public comment until December 7.
Why it matters
This closes the loopholes venture funds were using: small AI bets, robotics AI, and application-layer deals all fall inside the fence now. Any fund with China exposure in AI, chips, or biotech needs to review its portfolio before December 7, not after. Paired with the DeepSeek raise, the two economies are now building separate AI capital stacks on purpose.
OpenAI made its Decisions API generally available. Built on GPT-6 Luna, it answers bounded questions (a probability, a choice from your options, or a rubric score) up to 10x faster than calling Luna through the Responses API, at $0.10 per million input tokens with no charge for output tokens. It plugs natively into GPT-Live so voice agents can decide mid-conversation whether to clarify, route, or escalate. Early customers include Box, which saw about 5x faster routing on insurance certificates at the same accuracy, plus Datadog, Ramp, Sierra, and Speak.
Why it matters
This is the industrialization of the small stuff. Most of what agents do all day is not brilliant reasoning, it is thousands of tiny yes-or-no calls, and OpenAI just made those ten times faster and priced them like a utility. The input-only pricing is the tell: OpenAI wants to own the routing layer of every agentic workflow, and every fraction of margin on those micro-decisions now flows through its pipes. If you build agents, the cost floor of your product just dropped; if you compete with OpenAI on infrastructure, your price umbrella just got smaller.
Anthropic donated Agent Skills, the open standard for packaging what an agent knows how to do (a SKILL.md file plus scripts), to the Agentic AI Foundation under the Linux Foundation, which it co-founded with Block and OpenAI last year after donating MCP. More than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code, have integrated Skills; 1 million have been published and nearly 280 million installed. Engineers from Anthropic, Google, OpenAI, and others will maintain the spec as individuals, with versioned releases and security guidance for writing and installing skills on the roadmap.
Why it matters
Anthropic just did for agent skills what it did for MCP: gave away the standard so it can shape the ecosystem. When 280 million installs run through one format, whoever writes the spec decides how every agent on earth gets its know-how. The security-guidance part matters most, because a skill is executable instructions an agent follows blindly, and that is a malware channel waiting to happen. Founders building agent tooling should build on this format rather than around it.
On its earnings call, Blackstone argued the forces straining the AI buildout favor incumbents: CEO Steve Schwarzman said power constraints and public opposition to new data centers make existing facilities more valuable, while president Jon Gray said the firm could demand higher returns on projects without signed leases. Its infrastructure bets are outperforming on 'extraordinary' data-center leasing. But the strain is real: CoreWeave shares fell another 10% Thursday, Firmus pulled its IPO, Blackstone has lent billions to CoreWeave and backed Firmus, and its $5 billion Google TPU cloud venture may need up to $25 billion with site delays. AI chief Jas Khaira admitted: 'We have not mapped out who's going to buy all the debt... we don't know where all of it will come from.' Separately, SoftBank's SB Energy is discussing delaying its IPO to December or early next year; it is building an OpenAI data center in Ohio.
Why it matters
The smartest money in the room just said the quiet part out loud: nobody knows who funds the rest of this buildout. Blackstone's pitch is that scarcity is an asset if you already own the scarce thing, which is true right up until the debt behind it reprices. Watch CoreWeave's bonds, not its stock: if yields keep climbing, the neocloud financing model breaks before any data center does. The SB Energy delay is the canary for the IPO market's willingness to fund AI power.
SpaceX's AI division launched the SpaceX Super Intelligence Partner Network, its official channel for companies building on Grok: Transformation Partners (consultancies like Bain, Booz Allen, and EPAM) that deploy Grok inside enterprises, and Marketplace Partners that build plugins and integrations. More than 130 companies joined the early access, with engineers trained at over 30 organizations through a Forward Deployed Engineer accreditation program. Launch partners also include Databricks, NVIDIA, Snowflake, MongoDB, Twilio, Sentry, Supabase, Zscaler, and NTT DATA.
Why it matters
Grok is going enterprise through the oldest playbook in software: systems integrators. The 130-company roster says SpaceXSI knows the model alone won't win; the winners will be whoever shows up inside a bank or a factory with people who can install it. This is also a direct shot at OpenAI's and Anthropic's enterprise channels, which lean more on self-serve developers. Watch which consultancies staff up fastest on Grok, because in AI services, bodies are the moat.
