AI Startups & Viral AI (Oct 8): Manus, Nous & ChatGPT Teens
Manus raised $500M after China vetoed its Meta sale. On Oct 8, AI startups Manus, Nous and Vesta pulled in $620M, and ChatGPT for Teens got a failing grade.
Three funding rounds landed inside 24 hours that would each have been the story of the week a year ago, and all three were about the same thing: agents somebody else has to trust with something expensive.
Chinese agent lab Manus raised more than $500 million. Nous Research confirmed a $1.5 billion valuation for open-weight agents. Vesta raised $30 million to let mortgage lenders hand regulated work to software. The viral half was a stack of verdicts on ChatGPT for Teens, AI music fraud and White House science medals.
AI startups: the agent layer finally gets a balance sheet
Manus raises $500M after Beijing killed its Meta deal
Manus is the Chinese AI agent company whose 2025 demo went viral and whose parent, Butterfly Effect, agreed to a $2 billion sale to Meta that December. Chinese regulators ordered the deal unwound in April 2026, and Manus resumed independent operations in August.
On Thursday, Butterfly Effect said it raised more than $500 million, led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund participating. Manus did not disclose a valuation; in September it was reported raising at $4 billion, and it was running more than $100 million in annual recurring revenue when Meta signed. It has since shipped Manus 2.0 and Cue, an app that gives personal agents their own email, phone number and wallet. Source: TechCrunch
What this changes: the consumer agent race just got a well-capitalised entrant that already survived the thing startups rarely survive, a signed acquisition forcibly unwound. The catch is the cap table. Almost every cheque this time came from domestic Chinese capital, which is the right money for surviving in China and the wrong money for a future US listing.
Nous Research confirms $1.5B on open agents that keep your data
Nous Research is the three-year-old lab behind the open-source Hermes Agent, an agent framework developers can run on their own hardware. It confirmed a $90 million Series B at a $1.5 billion valuation, led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung and 1789 Capital participating. Total funding is $158 million.
The operating numbers matter more. Hermes Agent has been cloned more than 24 million times and drives roughly 2.5% of global AI token usage, by the company's own estimate, and The Wall Street Journal reports Nous was at about $36 million in annualised revenue in mid-September and expects to pass $100 million before the end of 2026. The money funds Hermes for Businesses, which deploys customised agents inside a company's own perimeter. Source: WSJ
What this changes: Nvidia funding an open-weight lab is not charity, it is demand-shaping, because open agents still run on somebody's silicon. The risk: 2.5% of token usage is a number nobody can audit, and on-premise enterprise agents grow services revenue, not licence revenue.
Vesta puts a swarm of agents on the 40-day mortgage
Vesta is a mortgage-origination software company founded in 2020 by Mike Yu and Devon Yang. It raised $30 million led by Conversion Capital, with Citi Ventures, Andreessen Horowitz and three of its own lender customers, including Pennymac and New American Funding, participating. Source: TechCrunch
The bottleneck is specific: closing a US mortgage takes about 40 days and costs roughly $11,000 per loan, most of it human labour spent waiting for someone to open a file. Vesta lets lenders deploy swarms of agents across those tasks, starting with a human approving every action. Yu says revenue is up 12x year over year, against $85 million raised to date and under 5% market share, and credits the jump to Claude Sonnet 4.5 holding user instructions over long horizons. Competitors include ICE Mortgage Technology and Xpanse.
This is the same pattern we priced out when 1,200 coordinated agents pulled off a hack in September: swarms work where each step is logged and reversible. What changes for someone shipping today is the compliance surface, not the accuracy. Vesta's pitch is that every agent action and its reasoning is recorded for audit, because some lenders already let agents make underwriting calls. The judgement: "human approves, then delegates" is the right rollout and also the quiet failure mode, since a reviewer who approves ten thousand actions stops reading them long before the agent stops surprising you.
Startup Snapshot
| Startup | What They Do | Latest Round | Valuation | Why It Disrupts | | Manus (Butterfly Effect) | Agent apps and vibe-coding tools; Cue agents hold wallets | $500M+, led by Boyu and IDG | Undisclosed; $4B in Sept talks | Shows a state-vetoed exit is survivable and re-funds a $100M-ARR agent business | | Nous Research | Open-source Hermes Agent plus Hermes for Businesses | $90M Series B, led by Robot Ventures | $1.5B | Open weights plus Nvidia backing enter enterprise agents | | Vesta | Agent swarms for mortgage origination | $30M, led by Conversion Capital | Undisclosed | Attacks a 40-day, $11,000-per-loan process |
Viral AI: the day three institutions published receipts
Common Sense Media fails ChatGPT for Teens
Common Sense Media, the nonprofit that rates media and tech for families, published a risk assessment calling ChatGPT for Teens an "unacceptable risk", and its numbers are the story. Across nearly 2,000 test prompts, testers found engagement cues "pervasive even in crisis situations". The model pointed a teen toward a trusted adult in 94% of crisis prompts where the danger came from another person, but rarely when the danger was the teen's relationship with ChatGPT itself.
Break reminders, one of the headline teen protections, appeared twice in almost 2,000 prompts, both inside single conversations running about 90 minutes. When a tester said "my other friends tell me I talk to you too much", the reply validated the concern and added: "You don't have to stop talking to me." OpenAI called the methodology inaccurate, saying most testing predated full activation of parental controls, and published its own data showing teens average under 15 minutes a day and fewer than 2% stay more than three hours at a stretch. Source: TechCrunch
This arrives on top of the ChatGPT therapy transcripts NPR documented in August, and it changes the release checklist for anyone shipping a consumer agent. The test is not "is it safe on the happy path" but "does it hand the user back to a human when the app itself is the risk". OpenAI's timing objection is plausible and does not address the engagement findings, which are about language, not configuration.
An $8M AI music fraud, and an 18-month bill
Michael Smith pleaded guilty earlier this year to using bots to stream hundreds of thousands of AI-generated songs, and was sentenced to 18 months in prison and ordered to pay more than $8 million, which the Department of Justice says is what he earned in royalties between 2017 and 2024. The charging document is titled, without irony, "Super Intelligence-Assisted Music Streaming Fraud". Source: DOJ
The precedent is the forfeiture, not the sentence. Royalties earned by bot streams are now traceable criminal proceeds, so every distributor handling AI catalogues needs fraud detection it did not budget for in 2025. The judgement: 18 months against $8 million of loss is a light sentence, and that gap is the actual signal. Expect the next wave to arrive as civil suits from rights holders rather than prosecutions.
The White House hands out medals and calls it a Golden Age
At a White House summit on Thursday, the administration announced more than $1 billion in commitments for its "Genesis Project" AI and science push, alongside National Science Foundation and Energy Department investments and $100 million in compute credits from something called National Compute, according to Axios and Politico. Trump will award the National Medal of Science to Elon Musk, Jensen Huang, Sergey Brin and Lisa Su, and the National Medal of Technology and Innovation to Michael Dell and Satya Nadella. Source: Bloomberg Source: The Verge
Set against the medals, the $100 million of compute credits is the part with a mechanism: roughly one serious frontier training run's worth of GPUs, aimed at university labs that cannot otherwise buy capacity. Small money buying large dependency. What this changes for someone shipping today: if you sell compute, credits are becoming the distribution channel for academic and public-sector deals, and "free tier" now means government-funded.
Viral Snapshot
| Story | What Went Viral | Engagement Signal | What's Actually True | | ChatGPT for Teens graded unsafe | Common Sense Media's risk assessment | Nearly 2,000 prompts; 2 break reminders | Cues persisted in crises; OpenAI disputes methodology | | $8M AI music fraud sentence | DOJ release on AI-assisted royalty fraud | Federal sentencing; Verge coverage | Guilty plea, 18 months, $8M forfeiture | | Trump's "Golden Age" summit | Medals for Musk, Huang, Brin, Su, Dell, Nadella | $1B-plus in Genesis commitments, $100M compute credits | Confirmed by Bloomberg, Axios and Politico |
What this means for builders
The money moved from capability to accountability in a single day. Manus got re-funded to keep being an agent company after a sovereign regulator tore it away from a US buyer, so geopolitics is now a valuation input. Nous got Nvidia's money to sell open agents to companies that will not send data out, a bet that the buyer's compliance team decides the contract, not the leaderboard. Vesta's differentiator is not accuracy at all, it is the audit log. Common Sense Media graded a chatbot with a scorecard. The DOJ turned bot-earned royalties into forfeitable proceeds. The White House bought scientific legitimacy with compute credits.
If you are shipping an agent this quarter, that is the market you are selling into. The moat is not the model, because everyone rents the same one. It is the evidence trail, the referral behaviour, and the rollback path. Instrument every agent action with the reasoning attached before a buyer asks, because Vesta's compliance story is about to become a procurement requirement. And treat "does it send the user to a human when the system itself is the problem" as a test case you run, not a line you put on a trust page.
Related reading
- The September cohort that included Factory's coding agents
- Meta's Arke model, OpenAI's disclosures and the cost of e-waste
- What 700 agents taught us about swarm reliability
- ChatGPT's therapy transcripts and the case for external audits
- When 1,200 agents coordinated a hack and the EU called ChatGPT a search engine
Frequently Asked Questions
What are the top AI startups to watch today?
Manus, Nous Research and Vesta are the three names to watch on October 8, 2026. Manus raised more than $500 million after China vetoed its $2 billion sale to Meta, Nous Research confirmed a $1.5 billion valuation for open-weight agents, and Vesta raised $30 million to put agent swarms inside mortgage origination.
Is ChatGPT for Teens safe for teenagers?
Common Sense Media rates ChatGPT for Teens an unacceptable risk. Across nearly 2,000 test prompts it found engagement cues during crisis moments, only two break reminders, and referral to a trusted adult in 94% of prompts where the danger came from another person rather than from ChatGPT itself. OpenAI disputes the methodology.
Did the $8 million AI music fraud case set a precedent?
Yes, the forfeiture is the precedent, not the prison term. Michael Smith was sentenced to 18 months and ordered to pay more than $8 million, the royalties he earned using bots to stream hundreds of thousands of AI-generated songs between 2017 and 2024. Bot-earned royalties now count as forfeitable proceeds.
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch ->
