CoreWeave (CRWV.US) launched twice a day: the physical AI workload template and zero trust security framework simultaneously landed at the Berlin GTC, and the pre-market stock price rebounded more than 2%

The Zhitong Finance App learned that on Wednesday, AI cloud service provider Coreweave (CRWV.US) released two products at the Nvidia (NVDA.US) Berlin GTC event — Physical AI Launchables for robotics and autonomous driving, and the Zero Trust AI Framework for the intelligent AI era. After the news was announced, as of press time, CoreWeave's pre-market share price had risen by more than 2%, taking a break after experiencing a deep retracement of more than 25% last week.

The two announcements probably point to a strategic intention at the same time: CoreWeave is trying to extend itself from a simple “computing power supplier” to a physical AI engineering layer and security control layer, and occupy a position closer to customer workflows in the Nvidia ecosystem.

Physical AI Launchables: compressing the “pain before running” into one command

Physical AI Launchables is not positioned as a new computing power product or model tool, but rather a set of workload templates built and tested by CoreWeave engineers. Robotics, autonomous driving, and simulation teams can launch tasks with a single command without having to rebuild the environment, configure storage, and wait for the scheduler to be ready before each run.

Richard Ahlfeld, senior vice president of physical AI at CoreWeave, said, “Robotics and autonomous driving teams are working on some of the most demanding engineering jobs in this industry, and they should spend their time on it. Physical AI Launchables provided a tested starting point for the tools they were already using.”

The first batch of templates covered five tools: NVIDIA Isaac Sim, CARLA, MuJoco, NVIDIA AlpaSim, and NVIDIA Cosmos 3, running through the three key aspects of data generation, evaluation and scoring, and strategy training in the physical AI cycle. In addition, Launchables also integrates Cosmos Data Factory's intelligence skills for generative data enhancement, automatic labeling, data organization, and retrieval of videos, images, and text; each template also comes with optimization skills for programming agents such as Claude Code and Codex, so that the agent can not only “start an operation once”, but also read the operation output and determine whether the results are useful.

Notably, each tool also comes with a blank template, so teams don't have to be limited to CoreWeave's pre-packaged workloads and can point to their own scenarios and data on a tested base. Nvidia's director of product management, Nakul Arora, said that Nvidia's reference agent skills for physical AI are designed to run on any cloud, and CoreWeave has built its own version optimized for the AI cloud, making these workflows “run natively on its infrastructure.”

Zero Trust AI Framework: Six-layer protection that “excludes CoreWeave from the chain of trust”

The Zero Trust AI Framework, which was released on the same day, confronts changes in the security paradigm brought about by intelligent AI. When AI agents retrieve context, call tools, and trigger actions between multiple systems, sensitive data no longer remains within the boundaries of a single application, but rather traverses between reminders, embedded vectors, model weights, operation trajectories, and outputs. Under the traditional cloud model, customers have to extend their trust to cloud service providers — service providers manage encryption keys, have privileged access to servers, and even operate their own models that may learn from customer interactions.

Chen Goldberg, executive vice president of product and engineering at CoreWeave, said, “The security of AI workloads cannot depend on the provider's commitment to its own actions. The Zero Trust AI Framework uses architectures instead of guarantees, keeping critical control in the hands of the customer.”

According to the introduction, the framework consists of the following six layers of protection:

Zero Key Access: Through remote key encryption, customers continue to use their existing key infrastructure — whether it's a software key manager, hardware security module, or KMS deployed on other clouds — key control is not transferred;

Zero Node Access: Supports access management so that service provider access becomes the customer's explicit decision, and the customer authorizes access according to a defined scope;

Zero Trust Infrastructure: Establishing verifiable trust at the infrastructure layer based on a hardware root of trust;

Zero In-Use Access: The plan is to expand Nvidia's third-generation confidential computing to CPUs and GPUs to cover data usage. Relevant capabilities will be implemented in the fourth quarter with HGX Rubin NVL8;

Zero Model Data: CoreWeave clearly does not use customer data to train its own models;

Zero Trust Network Encryption: Secure inter-node communication with MACSec backbone encryption.

The overall design goal is summarized by CoreWeave as “excluding itself from the customer's trust chain” — security no longer depends on service provider promises, but is fulfilled by keys, access authorizations, and encryption architectures owned by the customer.

Additionally, CoreWeave said it plans to use Nvidia's NVLink interconnect technology, including the Nvidia HGX Rubin NVL8 for its confidential computing aspects, and the HGX Rubin NVL8 will be launched in the fourth quarter.

Stock price background: rebound after deep retracement, differences still exist

The two announcements occurred at a node where CoreWeave's stock price fluctuated sharply. Over the past week and a half, the stock has fallen by more than 20%. Triggers include market concerns about its debt burden of about 50 billion US dollars, credit default swap spreads that once hit difficult levels, and Tesla CEO Musk's “yikes” statement on social platforms amplifying market sentiment.

However, the Wall Street analysts' camp was not unanimously bearish: Bank of America reiterated the “buy” rating and target price of $140 on the day before the press conference, believing that CoreWeave's core advantage is to “launch active power capacity on a large scale,” and that its five-year major customer contract structure provides high revenue visibility.

In contrast, HSBC maintains a “holdings reduction” rating and a target price of $39, on the grounds that the company expects an average return on invested capital of about 8.9% from 2026 to 2030, which is lower than the cost of borrowing. “The more investment, the greater the value damage.” According to LSEG data, out of 45 analysts covering CoreWeave, 32 still give “buy” or “strong buy” ratings, and HSBC's views are in the minority.

In this context, the simultaneous release of Physical AI Launchables and the Zero Trust AI Framework provided CoreWeave with new material that went beyond the “debt and computing power” narrative to a certain extent: the former points to competition for entry into physical AI workflows, while the latter responds to the most sensitive security control issues of enterprise customers in large-scale intelligent deployment. Whether the two products can be converted into verifiable revenue and customer retention will likely become the core of future testing of the market.