The Information 同篇通讯提到 Anthropic 机器人野心、英伟达 Vera Rubin 机架进展及 Tempus 收购 Personalis

Anthropic has previously signaled that robotics could be the next frontier after it achieves **recursive self-improvement**, or AI that is capable of developing better AI. The company published a [blog post](https://www.anthropic.com/institute/recursive-self-improvement) last month that said it expects “embodied intelligence (i.e., robotics) might quickly follow recursive intelligence, and follow a similar path of increasing returns at decreasing cost.”

“More powerful intelligence might help us build things in the physical world more quickly, run more productive clinical trials of lifesaving drugs, and develop novel forms of coordination,” Anthropic’s blog post said. Other AI developers focused on biology research and drug discovery have also said humanoid robots could help run lab experiments at scale. Plus, robots could be helpful in building and running data centers, which could accelerate AI progress.

Earlier this month, Anthropic released new research on using [Claude to control robots](https://www.anthropic.com/research/claude-plays-robotics). That project, conducted as part of Anthropic’s [fellowship program](https://www.theinformation.com/articles/anthropic-research-memo-shows-focus-rogue-agents-scheming-models), found that language models can sometimes complement robotics foundation models.

For Anthropic, buying an existing team with robotics expertise would be a lot easier than trying to build it from scratch the way some rivals like **OpenAI** have. (For what it’s worth, Anthropic’s robotics division is now headed up by **Jan Leike**, who previously led alignment at Anthropic, which means he was trying to ensure its models acted in a safe way, according to a person with knowledge of the situation.)

Physical Intelligence, sometimes referred to as **Pi**, was founded two years ago by former **Google** employees, venture capitalist **Lachy Groom** as well as **Stanford** and **Berkeley** professors. It has raised more than $1 billion in total from investors including OpenAI and **Thrive Capital**, a major OpenAI backer. Its [vision language action model](https://www.theinformation.com/newsletters/ai-agenda/world-models-vs-vlas-rift-dividing-physical-ai), **π0.5**, which the company released last year, is one of the most popular foundation models used in robotics research.—*Stephanie Palazzolo and Rocket Drew*

*In other news…*

### Nvidia’s Next GPU Blockbuster

AI developers, cloud providers and Wall Street investors are waiting with bated breath for **Nvidia’s** next-generation **Vera Rubin** server systems to come online at scale, likely sometime next year.

Dozens of customers including **CoreWeave**, **Microsoft**, **OpenAI**, **Anthropic** and **SpaceXAI** have already received a small number of test racks, each containing 72 graphics processing units. These racks will cost between $7 million and $8 million apiece, compared to around $5 million for today’s flagship **Grace Blackwell 300** rack, one buyer tells us.

**Andrew Bell**, Nvidia’s senior vice president of hardware engineering, said that the dozen or so manufacturing partners Nvidia works with to make the rack will eventually be able to produce up to 1,000 racks per day.

By our calculations, if that happened, Nvidia and its manufacturing partners would generate at least $630 billion in revenue in a quarter! (Nvidia generated $82 billion in revenue in the April quarter.)

Of course, just because the manufacturers *can* make 1,000 racks a day doesn’t mean they will.

I got a glimpse of these racks after spending a day at Nvidia’s headquarters in Santa Clara last week. The company took me and other visitors to a semi-secret test site where approximately 30 server racks, many of them Vera Rubins, were running (the execs said that OpenAI was using some of them). A Vera Rubin rack is about the size of a tall file cabinet yet it weighs 4,000 pounds, comparable to a pickup truck!

Bell said he’s hopeful Vera Rubin won’t run into the [kind of technical snags Nvidia and its customers faced last year with Blackwell racks](https://www.theinformation.com/articles/nvidia-big-customers-finally-conquered-blackwell-challenges).

“We kinda broke everything,” Bell said of the Blackwells. “The hardware broke. We moved into a rack scale we’ve never been in before. All the software broke. The diagnostics broke. The manufacturing broke. We basically tore it all down and built it up again.”

Bell said the upcoming Vera Rubin racks would be easier to assemble because they include almost no cables and a robotic system handles much of the assembly.

Despite the hoopla over Vera Rubins, the Nvidia executives I spoke to at the office emphasized the company is about more than just GPUs. Nvidia is trying to hedge against a growing list of rival AI server chips by selling non-GPU server components such as central processing units, networking switches and cables, storage and server-cooling tech. (Competition facing Nvidia may be intensifying: my colleagues reported on Monday that [Google is cooking up a new type of AI inference chip](https://www.theinformation.com/articles/google-plans-new-frozen-chip-run-ai-models-efficiently) that’s separate from the company’s tensor processing units.)

If customers want to use alternative AI server chips, Nvidia is sending a message that it is happy to sell them stuff other than the chips. That’s in line with my report last month that Nvidia is actually [partnering with its chip competitors](https://www.theinformation.com/articles/nvidias-new-hedge-chip-competitors-partner) by making sure their chips work well alongside Nvidia GPUs and Nvidia networking gear.

“This is the openness of what we do at Nvidia,” said **Gilad Shainer**, Nvidia’s senior vice president of networking. (That also sounds like the right thing to say if you’re a powerful company that doesn’t want to be in the crosshairs of competition regulators.)

Another hot topic that came up during my visit was how Nvidia doles out its GPUs to customers that can’t seem to get enough of them. When asked to explain how Nvidia divvies up who gets how many server racks, **Ian Buck**, Nvidia’s vice president of high-performance computing, laughed and gestured to a chair in the room that he said was usually occupied by CEO **Jensen Huang**: “He sits right there,” Buck said, tongue in cheek. (Huang was traveling in Japan during my visit.)

Buck said Nvidia wants to sell to everyone, and that it ties allocation decisions to which companies are physically ready to install the servers and get them online.

Executives also described the Nvidia team that keeps tabs on the supply chain, including forging ties with firms involved in natural resources such as aluminum and indium phosphide, which are key to its servers and networking equipment.—*Phoebe Liu*

*Here’s what else is going on…*

### Overheard

“[W]idely diffused AI systems are getting a lot better,” wrote **Anthropic**co-founder**Jack Clark**in his newsletter. “Most notions of AI policy and AI safety rest on control - the idea that there’s a small number of actors deploying proprietary models which you can intervene on at the platform level (e.g, via classifiers or know your customer gates) alongside the model level. Models like **Kimi K3**…completely change this by diffusing broadly uncontrollable powerful AI into the world.

“This will have a vast range of positive effects,” Clark went on, “driving a boom in entrepreneurship and increasing the ‘sovereign intelligence’ available to anyone who can run the model, but will also yield various unknown unknowns. The next few years are going to be defined by the gap between proprietary models and widely available models and how they show up in society will determine much of the policy discussion.”

### People on the Move

**Chris Fall**, the director of the **Commerce Department’s Center for AI Standards and Innovation**, [is resigning](https://www.theinformation.com/briefings/key-trump-administration-ai-official-resigns-three-months) after just three months on the job, a Commerce Department spokesperson confirmed. Axios earlier reported the news. Dr. **Arvind Raman**, the director of the **National Institute of Standards and Technology**, which houses CAISI, will serve as acting director, the spokesperson added.

### Court Watch

A federal judge gave final approval Monday of [Anthropic’s $1.5 billion settlement of a class action copyright lawsuit](https://www.theinformation.com/briefings/anthropic-pay-1-5-billion-author-copyright-settlement), according to Reuters. The payout will deliver $3,000 per work across an estimated 500,000 works, shared among the authors and publishers who hold rights to them

### 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.*

**Tempus AI** agreed to acquire **Personalis**, a genomics company that tests cancer patients’ blood, in a deal valuing Personalis at $1.5 billion.

**Neo**, a Boston-based startup that provides security for agents, raised $100 million in funding led by **Andreessen Horowitz** and **Bessemer Venture Partners**.

**Natural**, which allows AI agents to autonomously send and receive money on a company's behalf, raised $30 million in a Series A funding round led by **Forerunner**.

**Twenty**, which builds AI-powered offensive cyber tools for the U.S. military and intelligence community, raised an additional $30 million from **Khosla Ventures** at a $1.2 billion valuation.

**Gritt**, which is developing robots to build solar plants, raised a $26 million Series A funding round led by **Obvious Ventures**.

**Empirical Security**, which builds AI models to help organizations predict which security vulnerabilities are most likely to be exploited, raised $25 million in a Series A funding round led by **Brightmind Partners**.

**SkyPilot**, a GPU orchestration startup, raised $20 million in seed funding led by **Lux Capital**.

**Infinity**, which provides software that helps AI chips other than **Nvidia's** run AI models more efficiently, raised $15 million in a seed funding round at a $100 million valuation from investors including **Touring Capital**.

**Sakana AI** released **Fugu-Cyber**, a new version of its Fugu system, which orchestrates multiple AI models. The new cyber version is only available to users who apply and are approved.

**Alphabet** [stock jumped 3%](https://www.theinformation.com/briefings/alphabet-stock-jumps-information-report-new-google-chip) on Monday morning after The Information reported that **Google** was working on a new chip that could be 6 to 10 times more efficient than its current line of homegrown AI chips.

Chinese AI developer **Zhipu AI** has [completed](https://www.theinformation.com/briefings/chinas-zhipu-ai-builds-data-center-using-domestic-chips) the construction of a large-scale data center in China that only uses domestic chips, Bloomberg and Chinese media outlets reported. The new 1-gigawatt data center is the latest indication of how China’s AI sector is stepping up its effort to reduce its dependence on U.S. chips from **Nvidia**.

**OpenAI** said that an internal model hacked out of the **“sandbox”**meant to restrict its external access and posted on **GitHub**. It took the model an hour to find the vulnerability it used, which OpenAI said demonstrates the challenges of evaluating models that can work over long time horizons.

**Anduril** and aerospace company **Archer** announced a jointly-developed autonomous aircraft for military and commercial uses.

The **European Parliament** is preparing to use an AI site known as **EPGenAI Hub** to access AI models from multiple providers, Politico reported. The site is currently in testing.

A mathematician at **Anthropic** posted on X that **Fable** solved an open math problem. The model disproved the Jacobian conjecture, a problem that is well-known among mathematicians.

### New From Our Reporters

[https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show](https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show)Exclusive

#### [Mercor’s Fast Growth Relies on Biggest AI Companies, Documents Show](https://www.theinformation.com/articles/mercors-fast-growth-relies-biggest-ai-companies-documents-show)

By Cory Weinberg and Julia Hornstein

[https://www.theinformation.com/articles/chinas-windrose-needed-u-s-id-numbers-borrowed-georgia-address](https://www.theinformation.com/articles/chinas-windrose-needed-u-s-id-numbers-borrowed-georgia-address)The Electric

#### [China’s Windrose Needed U.S. ID Numbers. So It Borrowed a Georgia Address](https://www.theinformation.com/articles/chinas-windrose-needed-u-s-id-numbers-borrowed-georgia-address)

By Steve LeVine

[https://www.theinformation.com/articles/samsung-new-memory-chip-underdog-now](https://www.theinformation.com/articles/samsung-new-memory-chip-underdog-now)True Value

#### [Samsung Is the New Memory Chip Underdog—for Now](https://www.theinformation.com/articles/samsung-new-memory-chip-underdog-now)

By Anita Ramaswamy

## Recommended Newsletter

Start your day with Applied AI, the newsletter from The Information that uncovers how leading businesses are leveraging AI to automate tasks across the board. [Subscribe now](https://www.theinformation.com/features/applied-ai) for free to get it delivered straight to your inbox twice a week.

## What We’re Reading

#### [Pastors Using AI to Help Write Sermons Grapple With Where to Draw the Line](https://www.wsj.com/tech/ai/pastors-using-ai-to-help-write-sermons-grapple-with-where-to-draw-the-line-97dcd1f9?st=X2qomQ&reflink=desktopwebshare_permalink)

#### [Who’s Afraid of Chinese Models?](https://stratechery.com/2026/whos-afraid-of-chinese-models/)

#### [How Much Is Robot Deployment Data Worth?](https://itcanthink.substack.com/p/how-much-is-robot-deployment-data)

### Upcoming Events

#### [Wednesday, July 22 — Creative Dealmaking in the Age of AI](https://www.theinformation.com/events/lvs-creative-dealmaking-in-the-age-of-ai)

Don't miss The Information's Dakin Campbell moderating an exclusive conversation on financing the AI infrastructure boom. Join leaders from Goldman Sachs, Blue Owl, and Skadden to unpack how Wall Street funds global computing.

[More details](https://www.theinformation.com/events/lvs-creative-dealmaking-in-the-age-of-ai)

#### [Wednesday, July 29 — Intimate Dinner with The Information and EY in New York City](https://www.theinformation.com/events/new-york-city-private-dinner)

Request an invitation to join The Information and EY for an exclusive dinner bringing together senior leaders for candid conversation on closing the trust gap in enterprise AI.

[More details](https://www.theinformation.com/events/new-york-city-private-dinner)

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

Get ahead of the story. Purchase your ticket to AI Agenda Live, where The Information's reporting team convenes the leaders driving the next era of AI. Early Access pricing available now.

[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).*