Marvell与Google扩大AI芯片合作并授予最高122亿美元认股权，消息已提前扩散并被股价重定价

The companies I talked to at the event were eager to tinker with AI models for robots and accommodate their limitations. **Foxglove**, the data startup that hosted the conference, announced new software that lets customers use text prompts to search through troves of video and other physical AI data collected from sensors on their robots, drawing on one of Nvidia’s Cosmos world models to understand which video clips are most similar. That’s a relatively humble use of Cosmos, given Nvidia’s ambitions that world models—AI models designed to produce physically-realistic predictions—will someday power robots directly or generate simulated worlds for robots to learn in. But it stands to make Cosmos genuinely useful right now.

Robotics companies are also falling back on humans when AI makes a mistake. **Avea Robotics**, a startup providing software for remote controlling robots, showed off how an operator in Mexico City could take over a robot when it made a mistake and guide it to successfully plug in an ethernet cable, similar to the way self-driving vehicle companies like **Waymo** hand over control to remote human operators if they need to.

The idea of a “ChatGPT moment” for robotics also calls to mind humanoid robots in particular, whose five-fingered hands could allow them to handle a wide variety of work. But the dream for robotics foundation models is that they will power a variety of robots, from robot dogs and drones, all the way to highly specialized machinery.

That might feel far off right now, but if anything, the vibes in robotics are even more positive than they were for language models in the GPT-2 era.

It’s worth remembering that developers who used GPT-2, three years before ChatGPT, were over the moon about using the model for primitive chatbots and autocompleting lines of code.

While scaling up from GPT-2 looked promising, it was not yet known whether language models could grow up into the sophisticated models we have today. Now that language models have blazed the trail, and self-driving cars have extended AI into the physical world, roboticists expect progress in AI that can power a variety of robots to perform a variety of tasks will follow.

“We are on an inflection point where we are seeing signs of generalization, and we can see models getting better and better. And very soon we’ll have a ChatGPT moment,” said **Jay Chooi**, co-founder and CEO of **Robocurve**, a startup that evaluates the capabilities of physical AI models.

Sure, there are questions about the path to that moment, such as what kind of training data and what kind of AI model will make it possible. But roboticists seem confident that progress is assured.

One sign of that progress arrived Wednesday, when startup **Generalist** announced its latest AI model for powering robots. Generalist found that the model could sometimes control a pair of robot arms to complete tasks such as unzipping a purse and pulling out dollar bills after seeing an example of a person or robot doing the same task.

A robot model that can learn from a real-world example would be more akin to **GPT-3**, which OpenAI released in 2020. To achieve general-purpose robots, “there’s a very long road ahead,” said Generalist CEO **Pete Florence**, but the ability to pick up new skills from examples in a prompt has been thought of as “the most vivid, important kind of waypoint along the journey.”

Whether Generalist’s new model ends up bringing robotics into the GPT-3 era remains to be seen. For now, the model will only be available to a handful of companies that Generalist is working with.

*Here’s what else is going on…*

## Overheard

**OpenAI** is preparing to[roll out a system](https://www.theinformation.com/briefings/openai-launch-security-analysis-system-better-privacy-protections) to analyze user interactions with its models for cybersecurity and other safety concerns without storing customer data, in a move that positions the company to potentially lure customers frustrated over how Anthropic has handled similar issues. OpenAI plans to launch the new system, dubbed Private Safety Processing, next month, executives said on Wednesday.

President **Donald Trump** [said U.S. regulators](https://www.theinformation.com/briefings/trump-says-regulators-working-bring-hyperliquid-u-s) are working to bring Hyperliquid, a popular offshore venue for perpetual futures, to the U.S. The price of Hyperliquid token jumped 22% after the remark.If Hyperliquid enters the U.S., it will bring a competitor to Coinbase and Kalshi, which recently got approval to offer perpetual futures, which are contracts that allow traders to bet on the price of crypto or other assets with leverage without an expiry date.

## People on the Move

Tim Luke**,** a vice chairman in **Barclays global technology**investment banking group, [is joining](https://www.theinformation.com/briefings/veteran-barclays-semiconductor-banker-tim-luke-join-morgan-stanleyhttps://www.theinformation.com/briefings/veteran-barclays-semiconductor-banker-tim-luke-join-morgan-stanley)**Morgan Stanley**as a senior investment banker, according to people familiar with his move.

## Big Number

Shares of Chinese humanoid robot maker **Unitree Robotics** [soared on the debut on the Shanghai Stock Exchange](https://www.theinformation.com/briefings/chinese-humanoid-maker-unitree-shares-soar-shanghai-debut) on Wednesday, after raising 6.1 billion yuan ($905 million) in an initial public offering. The shares surged more than seven times its IPO price before closing 460% up, valuing the company at slightly more than $60 billion, underscoring Chinese investors’ euphoria in humanoids.

27.3% of the companies dissolved by **SimpleClosure**, a startup that helps businesses shut down and sell off their assets, were in the SaaS sector from January to June 2026, compared to just 14.4% in AI. Companies that shut down typically recover between $10,000 and $100,000 by liquidating assets to tech buyers, research labs, and AI companies.

## Deals and Debuts

*See The Information’s [Generative AI Database](https://www.theinformation.com/projects/generative-ai) for an exclusive list of private companies and their investors.*

**Marvell Technology** granted **Google** [rights to acquire a stake of up to $12.2 billion](https://www.theinformation.com/briefings/marvell-gives-google-right-buy-12-2-billion-stock-part-chip-deal) as part of an expanded partnership between the two companies.

**Stripe** [announced Wednesday that it agreed to acquire](https://www.theinformation.com/briefings/stripe-confirms-acquiring-ai-marketplace-startup-openrouter)**OpenRouter**, a marketplace that aggregates AI models, seeking to position itself at the center of payments to AI firms. Stripe is paying $7.5 billion for OpenRouter, of which $1.5 billion is paid to the startup’s founders and $6 billion to investors, the New York Times reported.

**ALSO**, a company that offers small electric vehicles — consumer electric bikes and commercial delivery quads — and is now extending them to drive themselves, raised $150 million in a Series D funding round led by **Prysm Capital**.

**Veeda AI**, a startup founded by Sanja Fidler, who ran Nvidia's Toronto AI lab for eight years, and two of her Nvidia colleagues, raised more than $90 million in a seed funding round from investors including **Khosla Ventures** and**Radical Ventures.**

**Network Bio**, a company that trains AI models on tissue, blood, molecular and clinical data pooled from hospital sample banks, raised $50 million in a funding round from investors including **Section 32**, **Thiel Bio** and **Founders Fund**.

**Flip**, a company that makes a phone app for workers who don't sit at a desk or have a company email address, raised $25 million in a funding round from investors including **Notion Capital**, **HV Capital** and **L-Bank**.

**Queen One**, a company that sells retailers a single system for knowing who their customer is and what to sell them, raised $25 million in a funding round from investors including **Mercury Fund**, **Full In**, **Connecticut Innovations**, **CP Overture**, **Charge Ventures** and **Inspired Capital.**

**Prevalent AI**, a company that uses AI to pull hundreds of scattered corporate data sources into one continuously updated map of what a company owns, how it all connects and where the gaps are, raised $22 million in a growth funding round from **Integrity Growth Partners**.

**Boom**, a company that makes software for landlords who rent out houses and manufactured homes, raised $15 million in a Series A funding round led by **S3 Ventures**.

**Thunder Compute**, which aims to address the GPU capacity shortage in data centers by eliminating unused resources, announced a $13 million Series A round led by **Matrix Partners**.

**Medly AI**, a company that offers students an AI tutor for exam preparation, raised $8 million in a seed funding round led by **Felix Capital**.

**Hypercubic**, a company that uses AI to move banks, insurers and government agencies off mainframes, raised $5.3 million in a seed funding round led by **CIV**.

**Guidelight**, which creates technical standards for AI safety, released a report card comparing the practices of different AI companies for controlling potentially rogue AIs.

## What We’re Reading

#### [Humanoid robots don’t deserve their superhuman valuations](https://www.ft.com/content/ba2a65de-7b9a-4938-bde6-91d41722ea08?syn-25a6b1a6=1)

#### [The Hidden Debt That Apple Owes to the CIA](https://www.wsj.com/tech/steve-jobs-apple-next-cia-161b65f9?mod=tech_lead_story)

#### [A County Got Rich From Data Centers. Some Question ‘At What Cost?’](https://www.nytimes.com/section/technology)

### Upcoming Events

#### [Wednesday, September 23 — AI Agenda Live SF 2026](https://www.theinformation.com/events/ai-agenda-live-2026)

Join us to explore the real-world shifts, new architectures, and enterprise strategies defining the next era of AI. Secure your spot today to hear directly from top minds at Google DeepMind, OpenAI, Atlassian, and more.

[More details](https://www.theinformation.com/events/ai-agenda-live-2026)

#### [Tuesday, October 27 — The Information’s 2026 WTF Summit](https://www.theinformation.com/events/wtf-summit-2026)

Tickets on sale now: The Information returns to Napa Valley October 27–28 to convene senior women across tech, media, and finance. Join us for two days of intimate, candid conversations with the leaders navigating today’s global shifts.

[More details](https://www.theinformation.com/events/wtf-summit-2026)

*Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [stephanie@theinformation.com](mailto:stephanie@theinformation.com).*

*If you think someone else might enjoy this newsletter, please pass it forward or they can sign up [here](https://www.theinformation.com/newsletters/ai-agenda).*